xref: /freebsd/sys/kern/vfs_bio.c (revision 4c9e27bd0a5f7fda85b0c0bf750575aee300a172)
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_directio.h"
81 #include "opt_swap.h"
82 
83 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
84 
85 struct	bio_ops bioops;		/* I/O operation notification */
86 
87 struct	buf_ops buf_ops_bio = {
88 	.bop_name	=	"buf_ops_bio",
89 	.bop_write	=	bufwrite,
90 	.bop_strategy	=	bufstrategy,
91 	.bop_sync	=	bufsync,
92 	.bop_bdflush	=	bufbdflush,
93 };
94 
95 /*
96  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
97  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
98  */
99 struct buf *buf;		/* buffer header pool */
100 caddr_t unmapped_buf;
101 
102 static 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 mtx_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 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 #ifdef DIRECTIO
386 extern void ffs_rawread_setup(void);
387 #endif /* DIRECTIO */
388 
389 /*
390  *	bqlock:
391  *
392  *	Return the appropriate queue lock based on the index.
393  */
394 static inline struct mtx *
395 bqlock(int qindex)
396 {
397 
398 	if (qindex == QUEUE_DIRTY)
399 		return (struct mtx *)(&bqdirty);
400 	return (struct mtx *)(&bqclean);
401 }
402 
403 /*
404  *	bdirtywakeup:
405  *
406  *	Wakeup any bwillwrite() waiters.
407  */
408 static void
409 bdirtywakeup(void)
410 {
411 	mtx_lock(&bdirtylock);
412 	if (bdirtywait) {
413 		bdirtywait = 0;
414 		wakeup(&bdirtywait);
415 	}
416 	mtx_unlock(&bdirtylock);
417 }
418 
419 /*
420  *	bdirtysub:
421  *
422  *	Decrement the numdirtybuffers count by one and wakeup any
423  *	threads blocked in bwillwrite().
424  */
425 static void
426 bdirtysub(void)
427 {
428 
429 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
430 	    (lodirtybuffers + hidirtybuffers) / 2)
431 		bdirtywakeup();
432 }
433 
434 /*
435  *	bdirtyadd:
436  *
437  *	Increment the numdirtybuffers count by one and wakeup the buf
438  *	daemon if needed.
439  */
440 static void
441 bdirtyadd(void)
442 {
443 
444 	/*
445 	 * Only do the wakeup once as we cross the boundary.  The
446 	 * buf daemon will keep running until the condition clears.
447 	 */
448 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
449 	    (lodirtybuffers + hidirtybuffers) / 2)
450 		bd_wakeup();
451 }
452 
453 /*
454  *	bufspacewakeup:
455  *
456  *	Called when buffer space is potentially available for recovery.
457  *	getnewbuf() will block on this flag when it is unable to free
458  *	sufficient buffer space.  Buffer space becomes recoverable when
459  *	bp's get placed back in the queues.
460  */
461 
462 static __inline void
463 bufspacewakeup(void)
464 {
465 
466 	/*
467 	 * If someone is waiting for BUF space, wake them up.  Even
468 	 * though we haven't freed the kva space yet, the waiting
469 	 * process will be able to now.
470 	 */
471 	mtx_lock(&nblock);
472 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
473 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
474 		wakeup(&needsbuffer);
475 	}
476 	mtx_unlock(&nblock);
477 }
478 
479 /*
480  *	runningwakeup:
481  *
482  *	Wake up processes that are waiting on asynchronous writes to fall
483  *	below lorunningspace.
484  */
485 static void
486 runningwakeup(void)
487 {
488 
489 	mtx_lock(&rbreqlock);
490 	if (runningbufreq) {
491 		runningbufreq = 0;
492 		wakeup(&runningbufreq);
493 	}
494 	mtx_unlock(&rbreqlock);
495 }
496 
497 /*
498  *	runningbufwakeup:
499  *
500  *	Decrement the outstanding write count according.
501  */
502 void
503 runningbufwakeup(struct buf *bp)
504 {
505 	long space, bspace;
506 
507 	bspace = bp->b_runningbufspace;
508 	if (bspace == 0)
509 		return;
510 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
511 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
512 	    space, bspace));
513 	bp->b_runningbufspace = 0;
514 	/*
515 	 * Only acquire the lock and wakeup on the transition from exceeding
516 	 * the threshold to falling below it.
517 	 */
518 	if (space < lorunningspace)
519 		return;
520 	if (space - bspace > lorunningspace)
521 		return;
522 	runningwakeup();
523 }
524 
525 /*
526  *	bufcountadd:
527  *
528  *	Called when a buffer has been added to one of the free queues to
529  *	account for the buffer and to wakeup anyone waiting for free buffers.
530  *	This typically occurs when large amounts of metadata are being handled
531  *	by the buffer cache ( else buffer space runs out first, usually ).
532  */
533 static __inline void
534 bufcountadd(struct buf *bp)
535 {
536 	int old;
537 
538 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
539 	    ("buf %p already counted as free", bp));
540 	bp->b_flags |= B_INFREECNT;
541 	old = atomic_fetchadd_int(&numfreebuffers, 1);
542 	KASSERT(old >= 0 && old < nbuf,
543 	    ("numfreebuffers climbed to %d", old + 1));
544 	mtx_lock(&nblock);
545 	if (needsbuffer) {
546 		needsbuffer &= ~VFS_BIO_NEED_ANY;
547 		if (numfreebuffers >= hifreebuffers)
548 			needsbuffer &= ~VFS_BIO_NEED_FREE;
549 		wakeup(&needsbuffer);
550 	}
551 	mtx_unlock(&nblock);
552 }
553 
554 /*
555  *	bufcountsub:
556  *
557  *	Decrement the numfreebuffers count as needed.
558  */
559 static void
560 bufcountsub(struct buf *bp)
561 {
562 	int old;
563 
564 	/*
565 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
566 	 * delayed-write, the buffer was free and we must decrement
567 	 * numfreebuffers.
568 	 */
569 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
570 		KASSERT((bp->b_flags & B_INFREECNT) != 0,
571 		    ("buf %p not counted in numfreebuffers", bp));
572 		bp->b_flags &= ~B_INFREECNT;
573 		old = atomic_fetchadd_int(&numfreebuffers, -1);
574 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
575 	}
576 }
577 
578 /*
579  *	waitrunningbufspace()
580  *
581  *	runningbufspace is a measure of the amount of I/O currently
582  *	running.  This routine is used in async-write situations to
583  *	prevent creating huge backups of pending writes to a device.
584  *	Only asynchronous writes are governed by this function.
585  *
586  *	This does NOT turn an async write into a sync write.  It waits
587  *	for earlier writes to complete and generally returns before the
588  *	caller's write has reached the device.
589  */
590 void
591 waitrunningbufspace(void)
592 {
593 
594 	mtx_lock(&rbreqlock);
595 	while (runningbufspace > hirunningspace) {
596 		runningbufreq = 1;
597 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
598 	}
599 	mtx_unlock(&rbreqlock);
600 }
601 
602 
603 /*
604  *	vfs_buf_test_cache:
605  *
606  *	Called when a buffer is extended.  This function clears the B_CACHE
607  *	bit if the newly extended portion of the buffer does not contain
608  *	valid data.
609  */
610 static __inline
611 void
612 vfs_buf_test_cache(struct buf *bp,
613 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
614 		  vm_page_t m)
615 {
616 
617 	VM_OBJECT_ASSERT_LOCKED(m->object);
618 	if (bp->b_flags & B_CACHE) {
619 		int base = (foff + off) & PAGE_MASK;
620 		if (vm_page_is_valid(m, base, size) == 0)
621 			bp->b_flags &= ~B_CACHE;
622 	}
623 }
624 
625 /* Wake up the buffer daemon if necessary */
626 static __inline void
627 bd_wakeup(void)
628 {
629 
630 	mtx_lock(&bdlock);
631 	if (bd_request == 0) {
632 		bd_request = 1;
633 		wakeup(&bd_request);
634 	}
635 	mtx_unlock(&bdlock);
636 }
637 
638 /*
639  * bd_speedup - speedup the buffer cache flushing code
640  */
641 void
642 bd_speedup(void)
643 {
644 	int needwake;
645 
646 	mtx_lock(&bdlock);
647 	needwake = 0;
648 	if (bd_speedupreq == 0 || bd_request == 0)
649 		needwake = 1;
650 	bd_speedupreq = 1;
651 	bd_request = 1;
652 	if (needwake)
653 		wakeup(&bd_request);
654 	mtx_unlock(&bdlock);
655 }
656 
657 #ifdef __i386__
658 #define	TRANSIENT_DENOM	5
659 #else
660 #define	TRANSIENT_DENOM 10
661 #endif
662 
663 /*
664  * Calculating buffer cache scaling values and reserve space for buffer
665  * headers.  This is called during low level kernel initialization and
666  * may be called more then once.  We CANNOT write to the memory area
667  * being reserved at this time.
668  */
669 caddr_t
670 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
671 {
672 	int tuned_nbuf;
673 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
674 
675 	/*
676 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
677 	 * PAGE_SIZE is >= 1K)
678 	 */
679 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
680 
681 	/*
682 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
683 	 * For the first 64MB of ram nominally allocate sufficient buffers to
684 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
685 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
686 	 * the buffer cache we limit the eventual kva reservation to
687 	 * maxbcache bytes.
688 	 *
689 	 * factor represents the 1/4 x ram conversion.
690 	 */
691 	if (nbuf == 0) {
692 		int factor = 4 * BKVASIZE / 1024;
693 
694 		nbuf = 50;
695 		if (physmem_est > 4096)
696 			nbuf += min((physmem_est - 4096) / factor,
697 			    65536 / factor);
698 		if (physmem_est > 65536)
699 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
700 			    32 * 1024 * 1024 / (factor * 5));
701 
702 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
703 			nbuf = maxbcache / BKVASIZE;
704 		tuned_nbuf = 1;
705 	} else
706 		tuned_nbuf = 0;
707 
708 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
709 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
710 	if (nbuf > maxbuf) {
711 		if (!tuned_nbuf)
712 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
713 			    maxbuf);
714 		nbuf = maxbuf;
715 	}
716 
717 	/*
718 	 * Ideal allocation size for the transient bio submap if 10%
719 	 * of the maximal space buffer map.  This roughly corresponds
720 	 * to the amount of the buffer mapped for typical UFS load.
721 	 *
722 	 * Clip the buffer map to reserve space for the transient
723 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
724 	 * maximum buffer map extent on the platform.
725 	 *
726 	 * The fall-back to the maxbuf in case of maxbcache unset,
727 	 * allows to not trim the buffer KVA for the architectures
728 	 * with ample KVA space.
729 	 */
730 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
731 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
732 		buf_sz = (long)nbuf * BKVASIZE;
733 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
734 		    (TRANSIENT_DENOM - 1)) {
735 			/*
736 			 * There is more KVA than memory.  Do not
737 			 * adjust buffer map size, and assign the rest
738 			 * of maxbuf to transient map.
739 			 */
740 			biotmap_sz = maxbuf_sz - buf_sz;
741 		} else {
742 			/*
743 			 * Buffer map spans all KVA we could afford on
744 			 * this platform.  Give 10% (20% on i386) of
745 			 * the buffer map to the transient bio map.
746 			 */
747 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
748 			buf_sz -= biotmap_sz;
749 		}
750 		if (biotmap_sz / INT_MAX > MAXPHYS)
751 			bio_transient_maxcnt = INT_MAX;
752 		else
753 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
754 		/*
755 		 * Artifically limit to 1024 simultaneous in-flight I/Os
756 		 * using the transient mapping.
757 		 */
758 		if (bio_transient_maxcnt > 1024)
759 			bio_transient_maxcnt = 1024;
760 		if (tuned_nbuf)
761 			nbuf = buf_sz / BKVASIZE;
762 	}
763 
764 	/*
765 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
766 	 * We have no less then 16 and no more then 256.
767 	 */
768 	nswbuf = max(min(nbuf/4, 256), 16);
769 #ifdef NSWBUF_MIN
770 	if (nswbuf < NSWBUF_MIN)
771 		nswbuf = NSWBUF_MIN;
772 #endif
773 #ifdef DIRECTIO
774 	ffs_rawread_setup();
775 #endif
776 
777 	/*
778 	 * Reserve space for the buffer cache buffers
779 	 */
780 	swbuf = (void *)v;
781 	v = (caddr_t)(swbuf + nswbuf);
782 	buf = (void *)v;
783 	v = (caddr_t)(buf + nbuf);
784 
785 	return(v);
786 }
787 
788 /* Initialize the buffer subsystem.  Called before use of any buffers. */
789 void
790 bufinit(void)
791 {
792 	struct buf *bp;
793 	int i;
794 
795 	mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
796 	mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
797 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
798 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
799 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
800 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
801 
802 	/* next, make a null set of free lists */
803 	for (i = 0; i < BUFFER_QUEUES; i++)
804 		TAILQ_INIT(&bufqueues[i]);
805 
806 	/* finally, initialize each buffer header and stick on empty q */
807 	for (i = 0; i < nbuf; i++) {
808 		bp = &buf[i];
809 		bzero(bp, sizeof *bp);
810 		bp->b_flags = B_INVAL | B_INFREECNT;
811 		bp->b_rcred = NOCRED;
812 		bp->b_wcred = NOCRED;
813 		bp->b_qindex = QUEUE_EMPTY;
814 		bp->b_xflags = 0;
815 		LIST_INIT(&bp->b_dep);
816 		BUF_LOCKINIT(bp);
817 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
818 #ifdef INVARIANTS
819 		bq_len[QUEUE_EMPTY]++;
820 #endif
821 	}
822 
823 	/*
824 	 * maxbufspace is the absolute maximum amount of buffer space we are
825 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
826 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
827 	 * used by most other processes.  The differential is required to
828 	 * ensure that buf_daemon is able to run when other processes might
829 	 * be blocked waiting for buffer space.
830 	 *
831 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
832 	 * this may result in KVM fragmentation which is not handled optimally
833 	 * by the system.
834 	 */
835 	maxbufspace = (long)nbuf * BKVASIZE;
836 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
837 	lobufspace = hibufspace - MAXBSIZE;
838 
839 	/*
840 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
841 	 * arbitrarily and may need further tuning. It corresponds to
842 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
843 	 * which fits with many RAID controllers' tagged queuing limits.
844 	 * The lower 1 MiB limit is the historical upper limit for
845 	 * hirunningspace.
846 	 */
847 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
848 	    16 * 1024 * 1024), 1024 * 1024);
849 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
850 
851 /*
852  * Limit the amount of malloc memory since it is wired permanently into
853  * the kernel space.  Even though this is accounted for in the buffer
854  * allocation, we don't want the malloced region to grow uncontrolled.
855  * The malloc scheme improves memory utilization significantly on average
856  * (small) directories.
857  */
858 	maxbufmallocspace = hibufspace / 20;
859 
860 /*
861  * Reduce the chance of a deadlock occuring by limiting the number
862  * of delayed-write dirty buffers we allow to stack up.
863  */
864 	hidirtybuffers = nbuf / 4 + 20;
865 	dirtybufthresh = hidirtybuffers * 9 / 10;
866 	numdirtybuffers = 0;
867 /*
868  * To support extreme low-memory systems, make sure hidirtybuffers cannot
869  * eat up all available buffer space.  This occurs when our minimum cannot
870  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
871  * BKVASIZE'd buffers.
872  */
873 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
874 		hidirtybuffers >>= 1;
875 	}
876 	lodirtybuffers = hidirtybuffers / 2;
877 
878 /*
879  * Try to keep the number of free buffers in the specified range,
880  * and give special processes (e.g. like buf_daemon) access to an
881  * emergency reserve.
882  */
883 	lofreebuffers = nbuf / 18 + 5;
884 	hifreebuffers = 2 * lofreebuffers;
885 	numfreebuffers = nbuf;
886 
887 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
888 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
889 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
890 }
891 
892 #ifdef INVARIANTS
893 static inline void
894 vfs_buf_check_mapped(struct buf *bp)
895 {
896 
897 	KASSERT((bp->b_flags & B_UNMAPPED) == 0,
898 	    ("mapped buf %p %x", bp, bp->b_flags));
899 	KASSERT(bp->b_kvabase != unmapped_buf,
900 	    ("mapped buf: b_kvabase was not updated %p", bp));
901 	KASSERT(bp->b_data != unmapped_buf,
902 	    ("mapped buf: b_data was not updated %p", bp));
903 }
904 
905 static inline void
906 vfs_buf_check_unmapped(struct buf *bp)
907 {
908 
909 	KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
910 	    ("unmapped buf %p %x", bp, bp->b_flags));
911 	KASSERT(bp->b_kvabase == unmapped_buf,
912 	    ("unmapped buf: corrupted b_kvabase %p", bp));
913 	KASSERT(bp->b_data == unmapped_buf,
914 	    ("unmapped buf: corrupted b_data %p", bp));
915 }
916 
917 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
918 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
919 #else
920 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
921 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
922 #endif
923 
924 static void
925 bpmap_qenter(struct buf *bp)
926 {
927 
928 	BUF_CHECK_MAPPED(bp);
929 
930 	/*
931 	 * bp->b_data is relative to bp->b_offset, but
932 	 * bp->b_offset may be offset into the first page.
933 	 */
934 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
935 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
936 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
937 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
938 }
939 
940 /*
941  * bfreekva() - free the kva allocation for a buffer.
942  *
943  *	Since this call frees up buffer space, we call bufspacewakeup().
944  */
945 static void
946 bfreekva(struct buf *bp)
947 {
948 
949 	if (bp->b_kvasize == 0)
950 		return;
951 
952 	atomic_add_int(&buffreekvacnt, 1);
953 	atomic_subtract_long(&bufspace, bp->b_kvasize);
954 	if ((bp->b_flags & B_UNMAPPED) == 0) {
955 		BUF_CHECK_MAPPED(bp);
956 		vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
957 		    bp->b_kvasize);
958 	} else {
959 		BUF_CHECK_UNMAPPED(bp);
960 		if ((bp->b_flags & B_KVAALLOC) != 0) {
961 			vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
962 			    bp->b_kvasize);
963 		}
964 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
965 		bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
966 	}
967 	bp->b_kvasize = 0;
968 	bufspacewakeup();
969 }
970 
971 /*
972  *	binsfree:
973  *
974  *	Insert the buffer into the appropriate free list.
975  */
976 static void
977 binsfree(struct buf *bp, int qindex)
978 {
979 	struct mtx *olock, *nlock;
980 
981 	BUF_ASSERT_XLOCKED(bp);
982 
983 	olock = bqlock(bp->b_qindex);
984 	nlock = bqlock(qindex);
985 	mtx_lock(olock);
986 	/* Handle delayed bremfree() processing. */
987 	if (bp->b_flags & B_REMFREE)
988 		bremfreel(bp);
989 
990 	if (bp->b_qindex != QUEUE_NONE)
991 		panic("binsfree: free buffer onto another queue???");
992 
993 	bp->b_qindex = qindex;
994 	if (olock != nlock) {
995 		mtx_unlock(olock);
996 		mtx_lock(nlock);
997 	}
998 	if (bp->b_flags & B_AGE)
999 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1000 	else
1001 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1002 #ifdef INVARIANTS
1003 	bq_len[bp->b_qindex]++;
1004 #endif
1005 	mtx_unlock(nlock);
1006 
1007 	/*
1008 	 * Something we can maybe free or reuse.
1009 	 */
1010 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1011 		bufspacewakeup();
1012 
1013 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1014 		bufcountadd(bp);
1015 }
1016 
1017 /*
1018  *	bremfree:
1019  *
1020  *	Mark the buffer for removal from the appropriate free list.
1021  *
1022  */
1023 void
1024 bremfree(struct buf *bp)
1025 {
1026 
1027 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1028 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1029 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1030 	KASSERT(bp->b_qindex != QUEUE_NONE,
1031 	    ("bremfree: buffer %p not on a queue.", bp));
1032 	BUF_ASSERT_XLOCKED(bp);
1033 
1034 	bp->b_flags |= B_REMFREE;
1035 	bufcountsub(bp);
1036 }
1037 
1038 /*
1039  *	bremfreef:
1040  *
1041  *	Force an immediate removal from a free list.  Used only in nfs when
1042  *	it abuses the b_freelist pointer.
1043  */
1044 void
1045 bremfreef(struct buf *bp)
1046 {
1047 	struct mtx *qlock;
1048 
1049 	qlock = bqlock(bp->b_qindex);
1050 	mtx_lock(qlock);
1051 	bremfreel(bp);
1052 	mtx_unlock(qlock);
1053 }
1054 
1055 /*
1056  *	bremfreel:
1057  *
1058  *	Removes a buffer from the free list, must be called with the
1059  *	correct qlock held.
1060  */
1061 static void
1062 bremfreel(struct buf *bp)
1063 {
1064 
1065 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1066 	    bp, bp->b_vp, bp->b_flags);
1067 	KASSERT(bp->b_qindex != QUEUE_NONE,
1068 	    ("bremfreel: buffer %p not on a queue.", bp));
1069 	BUF_ASSERT_XLOCKED(bp);
1070 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1071 
1072 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1073 #ifdef INVARIANTS
1074 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1075 	    bp->b_qindex));
1076 	bq_len[bp->b_qindex]--;
1077 #endif
1078 	bp->b_qindex = QUEUE_NONE;
1079 	/*
1080 	 * If this was a delayed bremfree() we only need to remove the buffer
1081 	 * from the queue and return the stats are already done.
1082 	 */
1083 	if (bp->b_flags & B_REMFREE) {
1084 		bp->b_flags &= ~B_REMFREE;
1085 		return;
1086 	}
1087 	bufcountsub(bp);
1088 }
1089 
1090 /*
1091  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1092  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1093  * the buffer is valid and we do not have to do anything.
1094  */
1095 void
1096 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1097     int cnt, struct ucred * cred)
1098 {
1099 	struct buf *rabp;
1100 	int i;
1101 
1102 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1103 		if (inmem(vp, *rablkno))
1104 			continue;
1105 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1106 
1107 		if ((rabp->b_flags & B_CACHE) == 0) {
1108 			if (!TD_IS_IDLETHREAD(curthread))
1109 				curthread->td_ru.ru_inblock++;
1110 			rabp->b_flags |= B_ASYNC;
1111 			rabp->b_flags &= ~B_INVAL;
1112 			rabp->b_ioflags &= ~BIO_ERROR;
1113 			rabp->b_iocmd = BIO_READ;
1114 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1115 				rabp->b_rcred = crhold(cred);
1116 			vfs_busy_pages(rabp, 0);
1117 			BUF_KERNPROC(rabp);
1118 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1119 			bstrategy(rabp);
1120 		} else {
1121 			brelse(rabp);
1122 		}
1123 	}
1124 }
1125 
1126 /*
1127  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1128  *
1129  * Get a buffer with the specified data.  Look in the cache first.  We
1130  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1131  * is set, the buffer is valid and we do not have to do anything, see
1132  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1133  */
1134 int
1135 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1136     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1137 {
1138 	struct buf *bp;
1139 	int rv = 0, readwait = 0;
1140 
1141 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1142 	/*
1143 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1144 	 */
1145 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1146 	if (bp == NULL)
1147 		return (EBUSY);
1148 
1149 	/* if not found in cache, do some I/O */
1150 	if ((bp->b_flags & B_CACHE) == 0) {
1151 		if (!TD_IS_IDLETHREAD(curthread))
1152 			curthread->td_ru.ru_inblock++;
1153 		bp->b_iocmd = BIO_READ;
1154 		bp->b_flags &= ~B_INVAL;
1155 		bp->b_ioflags &= ~BIO_ERROR;
1156 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1157 			bp->b_rcred = crhold(cred);
1158 		vfs_busy_pages(bp, 0);
1159 		bp->b_iooffset = dbtob(bp->b_blkno);
1160 		bstrategy(bp);
1161 		++readwait;
1162 	}
1163 
1164 	breada(vp, rablkno, rabsize, cnt, cred);
1165 
1166 	if (readwait) {
1167 		rv = bufwait(bp);
1168 	}
1169 	return (rv);
1170 }
1171 
1172 /*
1173  * Write, release buffer on completion.  (Done by iodone
1174  * if async).  Do not bother writing anything if the buffer
1175  * is invalid.
1176  *
1177  * Note that we set B_CACHE here, indicating that buffer is
1178  * fully valid and thus cacheable.  This is true even of NFS
1179  * now so we set it generally.  This could be set either here
1180  * or in biodone() since the I/O is synchronous.  We put it
1181  * here.
1182  */
1183 int
1184 bufwrite(struct buf *bp)
1185 {
1186 	int oldflags;
1187 	struct vnode *vp;
1188 	long space;
1189 	int vp_md;
1190 
1191 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1192 	if (bp->b_flags & B_INVAL) {
1193 		brelse(bp);
1194 		return (0);
1195 	}
1196 
1197 	if (bp->b_flags & B_BARRIER)
1198 		barrierwrites++;
1199 
1200 	oldflags = bp->b_flags;
1201 
1202 	BUF_ASSERT_HELD(bp);
1203 
1204 	if (bp->b_pin_count > 0)
1205 		bunpin_wait(bp);
1206 
1207 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1208 	    ("FFS background buffer should not get here %p", bp));
1209 
1210 	vp = bp->b_vp;
1211 	if (vp)
1212 		vp_md = vp->v_vflag & VV_MD;
1213 	else
1214 		vp_md = 0;
1215 
1216 	/*
1217 	 * Mark the buffer clean.  Increment the bufobj write count
1218 	 * before bundirty() call, to prevent other thread from seeing
1219 	 * empty dirty list and zero counter for writes in progress,
1220 	 * falsely indicating that the bufobj is clean.
1221 	 */
1222 	bufobj_wref(bp->b_bufobj);
1223 	bundirty(bp);
1224 
1225 	bp->b_flags &= ~B_DONE;
1226 	bp->b_ioflags &= ~BIO_ERROR;
1227 	bp->b_flags |= B_CACHE;
1228 	bp->b_iocmd = BIO_WRITE;
1229 
1230 	vfs_busy_pages(bp, 1);
1231 
1232 	/*
1233 	 * Normal bwrites pipeline writes
1234 	 */
1235 	bp->b_runningbufspace = bp->b_bufsize;
1236 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1237 
1238 	if (!TD_IS_IDLETHREAD(curthread))
1239 		curthread->td_ru.ru_oublock++;
1240 	if (oldflags & B_ASYNC)
1241 		BUF_KERNPROC(bp);
1242 	bp->b_iooffset = dbtob(bp->b_blkno);
1243 	bstrategy(bp);
1244 
1245 	if ((oldflags & B_ASYNC) == 0) {
1246 		int rtval = bufwait(bp);
1247 		brelse(bp);
1248 		return (rtval);
1249 	} else if (space > hirunningspace) {
1250 		/*
1251 		 * don't allow the async write to saturate the I/O
1252 		 * system.  We will not deadlock here because
1253 		 * we are blocking waiting for I/O that is already in-progress
1254 		 * to complete. We do not block here if it is the update
1255 		 * or syncer daemon trying to clean up as that can lead
1256 		 * to deadlock.
1257 		 */
1258 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1259 			waitrunningbufspace();
1260 	}
1261 
1262 	return (0);
1263 }
1264 
1265 void
1266 bufbdflush(struct bufobj *bo, struct buf *bp)
1267 {
1268 	struct buf *nbp;
1269 
1270 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1271 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1272 		altbufferflushes++;
1273 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1274 		BO_LOCK(bo);
1275 		/*
1276 		 * Try to find a buffer to flush.
1277 		 */
1278 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1279 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1280 			    BUF_LOCK(nbp,
1281 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1282 				continue;
1283 			if (bp == nbp)
1284 				panic("bdwrite: found ourselves");
1285 			BO_UNLOCK(bo);
1286 			/* Don't countdeps with the bo lock held. */
1287 			if (buf_countdeps(nbp, 0)) {
1288 				BO_LOCK(bo);
1289 				BUF_UNLOCK(nbp);
1290 				continue;
1291 			}
1292 			if (nbp->b_flags & B_CLUSTEROK) {
1293 				vfs_bio_awrite(nbp);
1294 			} else {
1295 				bremfree(nbp);
1296 				bawrite(nbp);
1297 			}
1298 			dirtybufferflushes++;
1299 			break;
1300 		}
1301 		if (nbp == NULL)
1302 			BO_UNLOCK(bo);
1303 	}
1304 }
1305 
1306 /*
1307  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1308  * anything if the buffer is marked invalid.
1309  *
1310  * Note that since the buffer must be completely valid, we can safely
1311  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1312  * biodone() in order to prevent getblk from writing the buffer
1313  * out synchronously.
1314  */
1315 void
1316 bdwrite(struct buf *bp)
1317 {
1318 	struct thread *td = curthread;
1319 	struct vnode *vp;
1320 	struct bufobj *bo;
1321 
1322 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1323 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1324 	KASSERT((bp->b_flags & B_BARRIER) == 0,
1325 	    ("Barrier request in delayed write %p", bp));
1326 	BUF_ASSERT_HELD(bp);
1327 
1328 	if (bp->b_flags & B_INVAL) {
1329 		brelse(bp);
1330 		return;
1331 	}
1332 
1333 	/*
1334 	 * If we have too many dirty buffers, don't create any more.
1335 	 * If we are wildly over our limit, then force a complete
1336 	 * cleanup. Otherwise, just keep the situation from getting
1337 	 * out of control. Note that we have to avoid a recursive
1338 	 * disaster and not try to clean up after our own cleanup!
1339 	 */
1340 	vp = bp->b_vp;
1341 	bo = bp->b_bufobj;
1342 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1343 		td->td_pflags |= TDP_INBDFLUSH;
1344 		BO_BDFLUSH(bo, bp);
1345 		td->td_pflags &= ~TDP_INBDFLUSH;
1346 	} else
1347 		recursiveflushes++;
1348 
1349 	bdirty(bp);
1350 	/*
1351 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1352 	 * true even of NFS now.
1353 	 */
1354 	bp->b_flags |= B_CACHE;
1355 
1356 	/*
1357 	 * This bmap keeps the system from needing to do the bmap later,
1358 	 * perhaps when the system is attempting to do a sync.  Since it
1359 	 * is likely that the indirect block -- or whatever other datastructure
1360 	 * that the filesystem needs is still in memory now, it is a good
1361 	 * thing to do this.  Note also, that if the pageout daemon is
1362 	 * requesting a sync -- there might not be enough memory to do
1363 	 * the bmap then...  So, this is important to do.
1364 	 */
1365 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1366 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1367 	}
1368 
1369 	/*
1370 	 * Set the *dirty* buffer range based upon the VM system dirty
1371 	 * pages.
1372 	 *
1373 	 * Mark the buffer pages as clean.  We need to do this here to
1374 	 * satisfy the vnode_pager and the pageout daemon, so that it
1375 	 * thinks that the pages have been "cleaned".  Note that since
1376 	 * the pages are in a delayed write buffer -- the VFS layer
1377 	 * "will" see that the pages get written out on the next sync,
1378 	 * or perhaps the cluster will be completed.
1379 	 */
1380 	vfs_clean_pages_dirty_buf(bp);
1381 	bqrelse(bp);
1382 
1383 	/*
1384 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1385 	 * due to the softdep code.
1386 	 */
1387 }
1388 
1389 /*
1390  *	bdirty:
1391  *
1392  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1393  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1394  *	itself to properly update it in the dirty/clean lists.  We mark it
1395  *	B_DONE to ensure that any asynchronization of the buffer properly
1396  *	clears B_DONE ( else a panic will occur later ).
1397  *
1398  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1399  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1400  *	should only be called if the buffer is known-good.
1401  *
1402  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1403  *	count.
1404  *
1405  *	The buffer must be on QUEUE_NONE.
1406  */
1407 void
1408 bdirty(struct buf *bp)
1409 {
1410 
1411 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1412 	    bp, bp->b_vp, bp->b_flags);
1413 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1414 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1415 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1416 	BUF_ASSERT_HELD(bp);
1417 	bp->b_flags &= ~(B_RELBUF);
1418 	bp->b_iocmd = BIO_WRITE;
1419 
1420 	if ((bp->b_flags & B_DELWRI) == 0) {
1421 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1422 		reassignbuf(bp);
1423 		bdirtyadd();
1424 	}
1425 }
1426 
1427 /*
1428  *	bundirty:
1429  *
1430  *	Clear B_DELWRI for buffer.
1431  *
1432  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1433  *	count.
1434  *
1435  *	The buffer must be on QUEUE_NONE.
1436  */
1437 
1438 void
1439 bundirty(struct buf *bp)
1440 {
1441 
1442 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1443 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1444 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1445 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1446 	BUF_ASSERT_HELD(bp);
1447 
1448 	if (bp->b_flags & B_DELWRI) {
1449 		bp->b_flags &= ~B_DELWRI;
1450 		reassignbuf(bp);
1451 		bdirtysub();
1452 	}
1453 	/*
1454 	 * Since it is now being written, we can clear its deferred write flag.
1455 	 */
1456 	bp->b_flags &= ~B_DEFERRED;
1457 }
1458 
1459 /*
1460  *	bawrite:
1461  *
1462  *	Asynchronous write.  Start output on a buffer, but do not wait for
1463  *	it to complete.  The buffer is released when the output completes.
1464  *
1465  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1466  *	B_INVAL buffers.  Not us.
1467  */
1468 void
1469 bawrite(struct buf *bp)
1470 {
1471 
1472 	bp->b_flags |= B_ASYNC;
1473 	(void) bwrite(bp);
1474 }
1475 
1476 /*
1477  *	babarrierwrite:
1478  *
1479  *	Asynchronous barrier write.  Start output on a buffer, but do not
1480  *	wait for it to complete.  Place a write barrier after this write so
1481  *	that this buffer and all buffers written before it are committed to
1482  *	the disk before any buffers written after this write are committed
1483  *	to the disk.  The buffer is released when the output completes.
1484  */
1485 void
1486 babarrierwrite(struct buf *bp)
1487 {
1488 
1489 	bp->b_flags |= B_ASYNC | B_BARRIER;
1490 	(void) bwrite(bp);
1491 }
1492 
1493 /*
1494  *	bbarrierwrite:
1495  *
1496  *	Synchronous barrier write.  Start output on a buffer and wait for
1497  *	it to complete.  Place a write barrier after this write so that
1498  *	this buffer and all buffers written before it are committed to
1499  *	the disk before any buffers written after this write are committed
1500  *	to the disk.  The buffer is released when the output completes.
1501  */
1502 int
1503 bbarrierwrite(struct buf *bp)
1504 {
1505 
1506 	bp->b_flags |= B_BARRIER;
1507 	return (bwrite(bp));
1508 }
1509 
1510 /*
1511  *	bwillwrite:
1512  *
1513  *	Called prior to the locking of any vnodes when we are expecting to
1514  *	write.  We do not want to starve the buffer cache with too many
1515  *	dirty buffers so we block here.  By blocking prior to the locking
1516  *	of any vnodes we attempt to avoid the situation where a locked vnode
1517  *	prevents the various system daemons from flushing related buffers.
1518  */
1519 void
1520 bwillwrite(void)
1521 {
1522 
1523 	if (numdirtybuffers >= hidirtybuffers) {
1524 		mtx_lock(&bdirtylock);
1525 		while (numdirtybuffers >= hidirtybuffers) {
1526 			bdirtywait = 1;
1527 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1528 			    "flswai", 0);
1529 		}
1530 		mtx_unlock(&bdirtylock);
1531 	}
1532 }
1533 
1534 /*
1535  * Return true if we have too many dirty buffers.
1536  */
1537 int
1538 buf_dirty_count_severe(void)
1539 {
1540 
1541 	return(numdirtybuffers >= hidirtybuffers);
1542 }
1543 
1544 static __noinline int
1545 buf_vm_page_count_severe(void)
1546 {
1547 
1548 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1549 
1550 	return vm_page_count_severe();
1551 }
1552 
1553 /*
1554  *	brelse:
1555  *
1556  *	Release a busy buffer and, if requested, free its resources.  The
1557  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1558  *	to be accessed later as a cache entity or reused for other purposes.
1559  */
1560 void
1561 brelse(struct buf *bp)
1562 {
1563 	int qindex;
1564 
1565 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1566 	    bp, bp->b_vp, bp->b_flags);
1567 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1568 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1569 
1570 	if (BUF_LOCKRECURSED(bp)) {
1571 		/*
1572 		 * Do not process, in particular, do not handle the
1573 		 * B_INVAL/B_RELBUF and do not release to free list.
1574 		 */
1575 		BUF_UNLOCK(bp);
1576 		return;
1577 	}
1578 
1579 	if (bp->b_flags & B_MANAGED) {
1580 		bqrelse(bp);
1581 		return;
1582 	}
1583 
1584 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1585 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1586 		/*
1587 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1588 		 * pages from being scrapped.  If the error is anything
1589 		 * other than an I/O error (EIO), assume that retrying
1590 		 * is futile.
1591 		 */
1592 		bp->b_ioflags &= ~BIO_ERROR;
1593 		bdirty(bp);
1594 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1595 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1596 		/*
1597 		 * Either a failed I/O or we were asked to free or not
1598 		 * cache the buffer.
1599 		 */
1600 		bp->b_flags |= B_INVAL;
1601 		if (!LIST_EMPTY(&bp->b_dep))
1602 			buf_deallocate(bp);
1603 		if (bp->b_flags & B_DELWRI)
1604 			bdirtysub();
1605 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1606 		if ((bp->b_flags & B_VMIO) == 0) {
1607 			if (bp->b_bufsize)
1608 				allocbuf(bp, 0);
1609 			if (bp->b_vp)
1610 				brelvp(bp);
1611 		}
1612 	}
1613 
1614 	/*
1615 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1616 	 * is called with B_DELWRI set, the underlying pages may wind up
1617 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1618 	 * because pages associated with a B_DELWRI bp are marked clean.
1619 	 *
1620 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1621 	 * if B_DELWRI is set.
1622 	 *
1623 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1624 	 * on pages to return pages to the VM page queues.
1625 	 */
1626 	if (bp->b_flags & B_DELWRI)
1627 		bp->b_flags &= ~B_RELBUF;
1628 	else if (buf_vm_page_count_severe()) {
1629 		/*
1630 		 * BKGRDINPROG can only be set with the buf and bufobj
1631 		 * locks both held.  We tolerate a race to clear it here.
1632 		 */
1633 		if (!(bp->b_vflags & BV_BKGRDINPROG))
1634 			bp->b_flags |= B_RELBUF;
1635 	}
1636 
1637 	/*
1638 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1639 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1640 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1641 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1642 	 *
1643 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1644 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1645 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1646 	 *
1647 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1648 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1649 	 * the commit state and we cannot afford to lose the buffer. If the
1650 	 * buffer has a background write in progress, we need to keep it
1651 	 * around to prevent it from being reconstituted and starting a second
1652 	 * background write.
1653 	 */
1654 	if ((bp->b_flags & B_VMIO)
1655 	    && !(bp->b_vp->v_mount != NULL &&
1656 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1657 		 !vn_isdisk(bp->b_vp, NULL) &&
1658 		 (bp->b_flags & B_DELWRI))
1659 	    ) {
1660 
1661 		int i, j, resid;
1662 		vm_page_t m;
1663 		off_t foff;
1664 		vm_pindex_t poff;
1665 		vm_object_t obj;
1666 
1667 		obj = bp->b_bufobj->bo_object;
1668 
1669 		/*
1670 		 * Get the base offset and length of the buffer.  Note that
1671 		 * in the VMIO case if the buffer block size is not
1672 		 * page-aligned then b_data pointer may not be page-aligned.
1673 		 * But our b_pages[] array *IS* page aligned.
1674 		 *
1675 		 * block sizes less then DEV_BSIZE (usually 512) are not
1676 		 * supported due to the page granularity bits (m->valid,
1677 		 * m->dirty, etc...).
1678 		 *
1679 		 * See man buf(9) for more information
1680 		 */
1681 		resid = bp->b_bufsize;
1682 		foff = bp->b_offset;
1683 		for (i = 0; i < bp->b_npages; i++) {
1684 			int had_bogus = 0;
1685 
1686 			m = bp->b_pages[i];
1687 
1688 			/*
1689 			 * If we hit a bogus page, fixup *all* the bogus pages
1690 			 * now.
1691 			 */
1692 			if (m == bogus_page) {
1693 				poff = OFF_TO_IDX(bp->b_offset);
1694 				had_bogus = 1;
1695 
1696 				VM_OBJECT_RLOCK(obj);
1697 				for (j = i; j < bp->b_npages; j++) {
1698 					vm_page_t mtmp;
1699 					mtmp = bp->b_pages[j];
1700 					if (mtmp == bogus_page) {
1701 						mtmp = vm_page_lookup(obj, poff + j);
1702 						if (!mtmp) {
1703 							panic("brelse: page missing\n");
1704 						}
1705 						bp->b_pages[j] = mtmp;
1706 					}
1707 				}
1708 				VM_OBJECT_RUNLOCK(obj);
1709 
1710 				if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1711 					BUF_CHECK_MAPPED(bp);
1712 					pmap_qenter(
1713 					    trunc_page((vm_offset_t)bp->b_data),
1714 					    bp->b_pages, bp->b_npages);
1715 				}
1716 				m = bp->b_pages[i];
1717 			}
1718 			if ((bp->b_flags & B_NOCACHE) ||
1719 			    (bp->b_ioflags & BIO_ERROR &&
1720 			     bp->b_iocmd == BIO_READ)) {
1721 				int poffset = foff & PAGE_MASK;
1722 				int presid = resid > (PAGE_SIZE - poffset) ?
1723 					(PAGE_SIZE - poffset) : resid;
1724 
1725 				KASSERT(presid >= 0, ("brelse: extra page"));
1726 				VM_OBJECT_WLOCK(obj);
1727 				while (vm_page_xbusied(m)) {
1728 					vm_page_lock(m);
1729 					VM_OBJECT_WUNLOCK(obj);
1730 					vm_page_busy_sleep(m, "mbncsh");
1731 					VM_OBJECT_WLOCK(obj);
1732 				}
1733 				if (pmap_page_wired_mappings(m) == 0)
1734 					vm_page_set_invalid(m, poffset, presid);
1735 				VM_OBJECT_WUNLOCK(obj);
1736 				if (had_bogus)
1737 					printf("avoided corruption bug in bogus_page/brelse code\n");
1738 			}
1739 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1740 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1741 		}
1742 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1743 			vfs_vmio_release(bp);
1744 
1745 	} else if (bp->b_flags & B_VMIO) {
1746 
1747 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1748 			vfs_vmio_release(bp);
1749 		}
1750 
1751 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1752 		if (bp->b_bufsize != 0)
1753 			allocbuf(bp, 0);
1754 		if (bp->b_vp != NULL)
1755 			brelvp(bp);
1756 	}
1757 
1758 	/*
1759 	 * If the buffer has junk contents signal it and eventually
1760 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1761 	 * doesn't find it.
1762 	 */
1763 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1764 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1765 		bp->b_flags |= B_INVAL;
1766 	if (bp->b_flags & B_INVAL) {
1767 		if (bp->b_flags & B_DELWRI)
1768 			bundirty(bp);
1769 		if (bp->b_vp)
1770 			brelvp(bp);
1771 	}
1772 
1773 	/* buffers with no memory */
1774 	if (bp->b_bufsize == 0) {
1775 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1776 		if (bp->b_vflags & BV_BKGRDINPROG)
1777 			panic("losing buffer 1");
1778 		if (bp->b_kvasize)
1779 			qindex = QUEUE_EMPTYKVA;
1780 		else
1781 			qindex = QUEUE_EMPTY;
1782 		bp->b_flags |= B_AGE;
1783 	/* buffers with junk contents */
1784 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1785 	    (bp->b_ioflags & BIO_ERROR)) {
1786 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1787 		if (bp->b_vflags & BV_BKGRDINPROG)
1788 			panic("losing buffer 2");
1789 		qindex = QUEUE_CLEAN;
1790 		bp->b_flags |= B_AGE;
1791 	/* remaining buffers */
1792 	} else if (bp->b_flags & B_DELWRI)
1793 		qindex = QUEUE_DIRTY;
1794 	else
1795 		qindex = QUEUE_CLEAN;
1796 
1797 	binsfree(bp, qindex);
1798 
1799 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1800 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1801 		panic("brelse: not dirty");
1802 	/* unlock */
1803 	BUF_UNLOCK(bp);
1804 }
1805 
1806 /*
1807  * Release a buffer back to the appropriate queue but do not try to free
1808  * it.  The buffer is expected to be used again soon.
1809  *
1810  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1811  * biodone() to requeue an async I/O on completion.  It is also used when
1812  * known good buffers need to be requeued but we think we may need the data
1813  * again soon.
1814  *
1815  * XXX we should be able to leave the B_RELBUF hint set on completion.
1816  */
1817 void
1818 bqrelse(struct buf *bp)
1819 {
1820 	int qindex;
1821 
1822 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1823 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1824 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1825 
1826 	if (BUF_LOCKRECURSED(bp)) {
1827 		/* do not release to free list */
1828 		BUF_UNLOCK(bp);
1829 		return;
1830 	}
1831 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1832 
1833 	if (bp->b_flags & B_MANAGED) {
1834 		if (bp->b_flags & B_REMFREE)
1835 			bremfreef(bp);
1836 		goto out;
1837 	}
1838 
1839 	/* buffers with stale but valid contents */
1840 	if (bp->b_flags & B_DELWRI) {
1841 		qindex = QUEUE_DIRTY;
1842 	} else {
1843 		if ((bp->b_flags & B_DELWRI) == 0 &&
1844 		    (bp->b_xflags & BX_VNDIRTY))
1845 			panic("bqrelse: not dirty");
1846 		/*
1847 		 * BKGRDINPROG can only be set with the buf and bufobj
1848 		 * locks both held.  We tolerate a race to clear it here.
1849 		 */
1850 		if (buf_vm_page_count_severe() &&
1851 		    (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1852 			/*
1853 			 * We are too low on memory, we have to try to free
1854 			 * the buffer (most importantly: the wired pages
1855 			 * making up its backing store) *now*.
1856 			 */
1857 			brelse(bp);
1858 			return;
1859 		}
1860 		qindex = QUEUE_CLEAN;
1861 	}
1862 	binsfree(bp, qindex);
1863 
1864 out:
1865 	/* unlock */
1866 	BUF_UNLOCK(bp);
1867 }
1868 
1869 /* Give pages used by the bp back to the VM system (where possible) */
1870 static void
1871 vfs_vmio_release(struct buf *bp)
1872 {
1873 	int i;
1874 	vm_page_t m;
1875 
1876 	if ((bp->b_flags & B_UNMAPPED) == 0) {
1877 		BUF_CHECK_MAPPED(bp);
1878 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1879 	} else
1880 		BUF_CHECK_UNMAPPED(bp);
1881 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
1882 	for (i = 0; i < bp->b_npages; i++) {
1883 		m = bp->b_pages[i];
1884 		bp->b_pages[i] = NULL;
1885 		/*
1886 		 * In order to keep page LRU ordering consistent, put
1887 		 * everything on the inactive queue.
1888 		 */
1889 		vm_page_lock(m);
1890 		vm_page_unwire(m, 0);
1891 
1892 		/*
1893 		 * Might as well free the page if we can and it has
1894 		 * no valid data.  We also free the page if the
1895 		 * buffer was used for direct I/O
1896 		 */
1897 		if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1898 			if (m->wire_count == 0 && !vm_page_busied(m))
1899 				vm_page_free(m);
1900 		} else if (bp->b_flags & B_DIRECT)
1901 			vm_page_try_to_free(m);
1902 		else if (buf_vm_page_count_severe())
1903 			vm_page_try_to_cache(m);
1904 		vm_page_unlock(m);
1905 	}
1906 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
1907 
1908 	if (bp->b_bufsize) {
1909 		bufspacewakeup();
1910 		bp->b_bufsize = 0;
1911 	}
1912 	bp->b_npages = 0;
1913 	bp->b_flags &= ~B_VMIO;
1914 	if (bp->b_vp)
1915 		brelvp(bp);
1916 }
1917 
1918 /*
1919  * Check to see if a block at a particular lbn is available for a clustered
1920  * write.
1921  */
1922 static int
1923 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1924 {
1925 	struct buf *bpa;
1926 	int match;
1927 
1928 	match = 0;
1929 
1930 	/* If the buf isn't in core skip it */
1931 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1932 		return (0);
1933 
1934 	/* If the buf is busy we don't want to wait for it */
1935 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1936 		return (0);
1937 
1938 	/* Only cluster with valid clusterable delayed write buffers */
1939 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1940 	    (B_DELWRI | B_CLUSTEROK))
1941 		goto done;
1942 
1943 	if (bpa->b_bufsize != size)
1944 		goto done;
1945 
1946 	/*
1947 	 * Check to see if it is in the expected place on disk and that the
1948 	 * block has been mapped.
1949 	 */
1950 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1951 		match = 1;
1952 done:
1953 	BUF_UNLOCK(bpa);
1954 	return (match);
1955 }
1956 
1957 /*
1958  *	vfs_bio_awrite:
1959  *
1960  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1961  *	This is much better then the old way of writing only one buffer at
1962  *	a time.  Note that we may not be presented with the buffers in the
1963  *	correct order, so we search for the cluster in both directions.
1964  */
1965 int
1966 vfs_bio_awrite(struct buf *bp)
1967 {
1968 	struct bufobj *bo;
1969 	int i;
1970 	int j;
1971 	daddr_t lblkno = bp->b_lblkno;
1972 	struct vnode *vp = bp->b_vp;
1973 	int ncl;
1974 	int nwritten;
1975 	int size;
1976 	int maxcl;
1977 	int gbflags;
1978 
1979 	bo = &vp->v_bufobj;
1980 	gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1981 	/*
1982 	 * right now we support clustered writing only to regular files.  If
1983 	 * we find a clusterable block we could be in the middle of a cluster
1984 	 * rather then at the beginning.
1985 	 */
1986 	if ((vp->v_type == VREG) &&
1987 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1988 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1989 
1990 		size = vp->v_mount->mnt_stat.f_iosize;
1991 		maxcl = MAXPHYS / size;
1992 
1993 		BO_RLOCK(bo);
1994 		for (i = 1; i < maxcl; i++)
1995 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1996 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1997 				break;
1998 
1999 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2000 			if (vfs_bio_clcheck(vp, size, lblkno - j,
2001 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2002 				break;
2003 		BO_RUNLOCK(bo);
2004 		--j;
2005 		ncl = i + j;
2006 		/*
2007 		 * this is a possible cluster write
2008 		 */
2009 		if (ncl != 1) {
2010 			BUF_UNLOCK(bp);
2011 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2012 			    gbflags);
2013 			return (nwritten);
2014 		}
2015 	}
2016 	bremfree(bp);
2017 	bp->b_flags |= B_ASYNC;
2018 	/*
2019 	 * default (old) behavior, writing out only one block
2020 	 *
2021 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
2022 	 */
2023 	nwritten = bp->b_bufsize;
2024 	(void) bwrite(bp);
2025 
2026 	return (nwritten);
2027 }
2028 
2029 static void
2030 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2031 {
2032 
2033 	KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2034 	    bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2035 	if ((gbflags & GB_UNMAPPED) == 0) {
2036 		bp->b_kvabase = (caddr_t)addr;
2037 	} else if ((gbflags & GB_KVAALLOC) != 0) {
2038 		KASSERT((gbflags & GB_UNMAPPED) != 0,
2039 		    ("GB_KVAALLOC without GB_UNMAPPED"));
2040 		bp->b_kvaalloc = (caddr_t)addr;
2041 		bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2042 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2043 	}
2044 	bp->b_kvasize = maxsize;
2045 }
2046 
2047 /*
2048  * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2049  * needed.
2050  */
2051 static int
2052 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2053 {
2054 	vm_offset_t addr;
2055 
2056 	bfreekva(bp);
2057 	addr = 0;
2058 
2059 	if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2060 		/*
2061 		 * Buffer map is too fragmented.  Request the caller
2062 		 * to defragment the map.
2063 		 */
2064 		atomic_add_int(&bufdefragcnt, 1);
2065 		return (1);
2066 	}
2067 	setbufkva(bp, addr, maxsize, gbflags);
2068 	atomic_add_long(&bufspace, bp->b_kvasize);
2069 	return (0);
2070 }
2071 
2072 /*
2073  * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2074  * locked vnode is supplied.
2075  */
2076 static void
2077 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2078     int defrag)
2079 {
2080 	struct thread *td;
2081 	char *waitmsg;
2082 	int cnt, error, flags, norunbuf, wait;
2083 
2084 	mtx_assert(&bqclean, MA_OWNED);
2085 
2086 	if (defrag) {
2087 		flags = VFS_BIO_NEED_BUFSPACE;
2088 		waitmsg = "nbufkv";
2089 	} else if (bufspace >= hibufspace) {
2090 		waitmsg = "nbufbs";
2091 		flags = VFS_BIO_NEED_BUFSPACE;
2092 	} else {
2093 		waitmsg = "newbuf";
2094 		flags = VFS_BIO_NEED_ANY;
2095 	}
2096 	mtx_lock(&nblock);
2097 	needsbuffer |= flags;
2098 	mtx_unlock(&nblock);
2099 	mtx_unlock(&bqclean);
2100 
2101 	bd_speedup();	/* heeeelp */
2102 	if ((gbflags & GB_NOWAIT_BD) != 0)
2103 		return;
2104 
2105 	td = curthread;
2106 	cnt = 0;
2107 	wait = MNT_NOWAIT;
2108 	mtx_lock(&nblock);
2109 	while (needsbuffer & flags) {
2110 		if (vp != NULL && vp->v_type != VCHR &&
2111 		    (td->td_pflags & TDP_BUFNEED) == 0) {
2112 			mtx_unlock(&nblock);
2113 
2114 			/*
2115 			 * getblk() is called with a vnode locked, and
2116 			 * some majority of the dirty buffers may as
2117 			 * well belong to the vnode.  Flushing the
2118 			 * buffers there would make a progress that
2119 			 * cannot be achieved by the buf_daemon, that
2120 			 * cannot lock the vnode.
2121 			 */
2122 			if (cnt++ > 2)
2123 				wait = MNT_WAIT;
2124 			ASSERT_VOP_LOCKED(vp, "bufd_helper");
2125 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2126 			    vn_lock(vp, LK_TRYUPGRADE);
2127 			if (error == 0) {
2128 				/* play bufdaemon */
2129 				norunbuf = curthread_pflags_set(TDP_BUFNEED |
2130 				    TDP_NORUNNINGBUF);
2131 				VOP_FSYNC(vp, wait, td);
2132 				atomic_add_long(&notbufdflushes, 1);
2133 				curthread_pflags_restore(norunbuf);
2134 			}
2135 			mtx_lock(&nblock);
2136 			if ((needsbuffer & flags) == 0)
2137 				break;
2138 		}
2139 		if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag,
2140 		    waitmsg, slptimeo))
2141 			break;
2142 	}
2143 	mtx_unlock(&nblock);
2144 }
2145 
2146 static void
2147 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2148 {
2149 
2150 	CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2151 	    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2152 	     bp->b_kvasize, bp->b_bufsize, qindex);
2153 	mtx_assert(&bqclean, MA_NOTOWNED);
2154 
2155 	/*
2156 	 * Note: we no longer distinguish between VMIO and non-VMIO
2157 	 * buffers.
2158 	 */
2159 	KASSERT((bp->b_flags & B_DELWRI) == 0,
2160 	    ("delwri buffer %p found in queue %d", bp, qindex));
2161 
2162 	if (qindex == QUEUE_CLEAN) {
2163 		if (bp->b_flags & B_VMIO) {
2164 			bp->b_flags &= ~B_ASYNC;
2165 			vfs_vmio_release(bp);
2166 		}
2167 		if (bp->b_vp != NULL)
2168 			brelvp(bp);
2169 	}
2170 
2171 	/*
2172 	 * Get the rest of the buffer freed up.  b_kva* is still valid
2173 	 * after this operation.
2174 	 */
2175 
2176 	if (bp->b_rcred != NOCRED) {
2177 		crfree(bp->b_rcred);
2178 		bp->b_rcred = NOCRED;
2179 	}
2180 	if (bp->b_wcred != NOCRED) {
2181 		crfree(bp->b_wcred);
2182 		bp->b_wcred = NOCRED;
2183 	}
2184 	if (!LIST_EMPTY(&bp->b_dep))
2185 		buf_deallocate(bp);
2186 	if (bp->b_vflags & BV_BKGRDINPROG)
2187 		panic("losing buffer 3");
2188 	KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.  qindex: %d",
2189 	    bp, bp->b_vp, qindex));
2190 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2191 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2192 
2193 	if (bp->b_bufsize)
2194 		allocbuf(bp, 0);
2195 
2196 	bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2197 	bp->b_ioflags = 0;
2198 	bp->b_xflags = 0;
2199 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
2200 	    ("buf %p still counted as free?", bp));
2201 	bp->b_vflags = 0;
2202 	bp->b_vp = NULL;
2203 	bp->b_blkno = bp->b_lblkno = 0;
2204 	bp->b_offset = NOOFFSET;
2205 	bp->b_iodone = 0;
2206 	bp->b_error = 0;
2207 	bp->b_resid = 0;
2208 	bp->b_bcount = 0;
2209 	bp->b_npages = 0;
2210 	bp->b_dirtyoff = bp->b_dirtyend = 0;
2211 	bp->b_bufobj = NULL;
2212 	bp->b_pin_count = 0;
2213 	bp->b_fsprivate1 = NULL;
2214 	bp->b_fsprivate2 = NULL;
2215 	bp->b_fsprivate3 = NULL;
2216 
2217 	LIST_INIT(&bp->b_dep);
2218 }
2219 
2220 static int flushingbufs;
2221 
2222 static struct buf *
2223 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2224 {
2225 	struct buf *bp, *nbp;
2226 	int nqindex, qindex, pass;
2227 
2228 	KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2229 
2230 	pass = 1;
2231 restart:
2232 	atomic_add_int(&getnewbufrestarts, 1);
2233 
2234 	/*
2235 	 * Setup for scan.  If we do not have enough free buffers,
2236 	 * we setup a degenerate case that immediately fails.  Note
2237 	 * that if we are specially marked process, we are allowed to
2238 	 * dip into our reserves.
2239 	 *
2240 	 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2241 	 * for the allocation of the mapped buffer.  For unmapped, the
2242 	 * easiest is to start with EMPTY outright.
2243 	 *
2244 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
2245 	 * However, there are a number of cases (defragging, reusing, ...)
2246 	 * where we cannot backup.
2247 	 */
2248 	nbp = NULL;
2249 	mtx_lock(&bqclean);
2250 	if (!defrag && unmapped) {
2251 		nqindex = QUEUE_EMPTY;
2252 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2253 	}
2254 	if (nbp == NULL) {
2255 		nqindex = QUEUE_EMPTYKVA;
2256 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2257 	}
2258 
2259 	/*
2260 	 * If no EMPTYKVA buffers and we are either defragging or
2261 	 * reusing, locate a CLEAN buffer to free or reuse.  If
2262 	 * bufspace useage is low skip this step so we can allocate a
2263 	 * new buffer.
2264 	 */
2265 	if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2266 		nqindex = QUEUE_CLEAN;
2267 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2268 	}
2269 
2270 	/*
2271 	 * If we could not find or were not allowed to reuse a CLEAN
2272 	 * buffer, check to see if it is ok to use an EMPTY buffer.
2273 	 * We can only use an EMPTY buffer if allocating its KVA would
2274 	 * not otherwise run us out of buffer space.  No KVA is needed
2275 	 * for the unmapped allocation.
2276 	 */
2277 	if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2278 	    metadata)) {
2279 		nqindex = QUEUE_EMPTY;
2280 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2281 	}
2282 
2283 	/*
2284 	 * All available buffers might be clean, retry ignoring the
2285 	 * lobufspace as the last resort.
2286 	 */
2287 	if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2288 		nqindex = QUEUE_CLEAN;
2289 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2290 	}
2291 
2292 	/*
2293 	 * Run scan, possibly freeing data and/or kva mappings on the fly
2294 	 * depending.
2295 	 */
2296 	while ((bp = nbp) != NULL) {
2297 		qindex = nqindex;
2298 
2299 		/*
2300 		 * Calculate next bp (we can only use it if we do not
2301 		 * block or do other fancy things).
2302 		 */
2303 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2304 			switch (qindex) {
2305 			case QUEUE_EMPTY:
2306 				nqindex = QUEUE_EMPTYKVA;
2307 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2308 				if (nbp != NULL)
2309 					break;
2310 				/* FALLTHROUGH */
2311 			case QUEUE_EMPTYKVA:
2312 				nqindex = QUEUE_CLEAN;
2313 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2314 				if (nbp != NULL)
2315 					break;
2316 				/* FALLTHROUGH */
2317 			case QUEUE_CLEAN:
2318 				if (metadata && pass == 1) {
2319 					pass = 2;
2320 					nqindex = QUEUE_EMPTY;
2321 					nbp = TAILQ_FIRST(
2322 					    &bufqueues[QUEUE_EMPTY]);
2323 				}
2324 				/*
2325 				 * nbp is NULL.
2326 				 */
2327 				break;
2328 			}
2329 		}
2330 		/*
2331 		 * If we are defragging then we need a buffer with
2332 		 * b_kvasize != 0.  XXX this situation should no longer
2333 		 * occur, if defrag is non-zero the buffer's b_kvasize
2334 		 * should also be non-zero at this point.  XXX
2335 		 */
2336 		if (defrag && bp->b_kvasize == 0) {
2337 			printf("Warning: defrag empty buffer %p\n", bp);
2338 			continue;
2339 		}
2340 
2341 		/*
2342 		 * Start freeing the bp.  This is somewhat involved.  nbp
2343 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2344 		 */
2345 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2346 			continue;
2347 		/*
2348 		 * BKGRDINPROG can only be set with the buf and bufobj
2349 		 * locks both held.  We tolerate a race to clear it here.
2350 		 */
2351 		if (bp->b_vflags & BV_BKGRDINPROG) {
2352 			BUF_UNLOCK(bp);
2353 			continue;
2354 		}
2355 
2356 		KASSERT(bp->b_qindex == qindex,
2357 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2358 
2359 		bremfreel(bp);
2360 		mtx_unlock(&bqclean);
2361 		/*
2362 		 * NOTE:  nbp is now entirely invalid.  We can only restart
2363 		 * the scan from this point on.
2364 		 */
2365 
2366 		getnewbuf_reuse_bp(bp, qindex);
2367 		mtx_assert(&bqclean, MA_NOTOWNED);
2368 
2369 		/*
2370 		 * If we are defragging then free the buffer.
2371 		 */
2372 		if (defrag) {
2373 			bp->b_flags |= B_INVAL;
2374 			bfreekva(bp);
2375 			brelse(bp);
2376 			defrag = 0;
2377 			goto restart;
2378 		}
2379 
2380 		/*
2381 		 * Notify any waiters for the buffer lock about
2382 		 * identity change by freeing the buffer.
2383 		 */
2384 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2385 			bp->b_flags |= B_INVAL;
2386 			bfreekva(bp);
2387 			brelse(bp);
2388 			goto restart;
2389 		}
2390 
2391 		if (metadata)
2392 			break;
2393 
2394 		/*
2395 		 * If we are overcomitted then recover the buffer and its
2396 		 * KVM space.  This occurs in rare situations when multiple
2397 		 * processes are blocked in getnewbuf() or allocbuf().
2398 		 */
2399 		if (bufspace >= hibufspace)
2400 			flushingbufs = 1;
2401 		if (flushingbufs && bp->b_kvasize != 0) {
2402 			bp->b_flags |= B_INVAL;
2403 			bfreekva(bp);
2404 			brelse(bp);
2405 			goto restart;
2406 		}
2407 		if (bufspace < lobufspace)
2408 			flushingbufs = 0;
2409 		break;
2410 	}
2411 	return (bp);
2412 }
2413 
2414 /*
2415  *	getnewbuf:
2416  *
2417  *	Find and initialize a new buffer header, freeing up existing buffers
2418  *	in the bufqueues as necessary.  The new buffer is returned locked.
2419  *
2420  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
2421  *	buffer away, the caller must set B_INVAL prior to calling brelse().
2422  *
2423  *	We block if:
2424  *		We have insufficient buffer headers
2425  *		We have insufficient buffer space
2426  *		buffer_arena is too fragmented ( space reservation fails )
2427  *		If we have to flush dirty buffers ( but we try to avoid this )
2428  */
2429 static struct buf *
2430 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2431     int gbflags)
2432 {
2433 	struct buf *bp;
2434 	int defrag, metadata;
2435 
2436 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2437 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2438 	if (!unmapped_buf_allowed)
2439 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2440 
2441 	defrag = 0;
2442 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2443 	    vp->v_type == VCHR)
2444 		metadata = 1;
2445 	else
2446 		metadata = 0;
2447 	/*
2448 	 * We can't afford to block since we might be holding a vnode lock,
2449 	 * which may prevent system daemons from running.  We deal with
2450 	 * low-memory situations by proactively returning memory and running
2451 	 * async I/O rather then sync I/O.
2452 	 */
2453 	atomic_add_int(&getnewbufcalls, 1);
2454 	atomic_subtract_int(&getnewbufrestarts, 1);
2455 restart:
2456 	bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2457 	    GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2458 	if (bp != NULL)
2459 		defrag = 0;
2460 
2461 	/*
2462 	 * If we exhausted our list, sleep as appropriate.  We may have to
2463 	 * wakeup various daemons and write out some dirty buffers.
2464 	 *
2465 	 * Generally we are sleeping due to insufficient buffer space.
2466 	 */
2467 	if (bp == NULL) {
2468 		mtx_assert(&bqclean, MA_OWNED);
2469 		getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2470 		mtx_assert(&bqclean, MA_NOTOWNED);
2471 	} else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2472 		mtx_assert(&bqclean, MA_NOTOWNED);
2473 
2474 		bfreekva(bp);
2475 		bp->b_flags |= B_UNMAPPED;
2476 		bp->b_kvabase = bp->b_data = unmapped_buf;
2477 		bp->b_kvasize = maxsize;
2478 		atomic_add_long(&bufspace, bp->b_kvasize);
2479 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2480 		atomic_add_int(&bufreusecnt, 1);
2481 	} else {
2482 		mtx_assert(&bqclean, MA_NOTOWNED);
2483 
2484 		/*
2485 		 * We finally have a valid bp.  We aren't quite out of the
2486 		 * woods, we still have to reserve kva space.  In order
2487 		 * to keep fragmentation sane we only allocate kva in
2488 		 * BKVASIZE chunks.
2489 		 */
2490 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2491 
2492 		if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2493 		    B_KVAALLOC)) == B_UNMAPPED) {
2494 			if (allocbufkva(bp, maxsize, gbflags)) {
2495 				defrag = 1;
2496 				bp->b_flags |= B_INVAL;
2497 				brelse(bp);
2498 				goto restart;
2499 			}
2500 			atomic_add_int(&bufreusecnt, 1);
2501 		} else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2502 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2503 			/*
2504 			 * If the reused buffer has KVA allocated,
2505 			 * reassign b_kvaalloc to b_kvabase.
2506 			 */
2507 			bp->b_kvabase = bp->b_kvaalloc;
2508 			bp->b_flags &= ~B_KVAALLOC;
2509 			atomic_subtract_long(&unmapped_bufspace,
2510 			    bp->b_kvasize);
2511 			atomic_add_int(&bufreusecnt, 1);
2512 		} else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2513 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2514 		    GB_KVAALLOC)) {
2515 			/*
2516 			 * The case of reused buffer already have KVA
2517 			 * mapped, but the request is for unmapped
2518 			 * buffer with KVA allocated.
2519 			 */
2520 			bp->b_kvaalloc = bp->b_kvabase;
2521 			bp->b_data = bp->b_kvabase = unmapped_buf;
2522 			bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2523 			atomic_add_long(&unmapped_bufspace,
2524 			    bp->b_kvasize);
2525 			atomic_add_int(&bufreusecnt, 1);
2526 		}
2527 		if ((gbflags & GB_UNMAPPED) == 0) {
2528 			bp->b_saveaddr = bp->b_kvabase;
2529 			bp->b_data = bp->b_saveaddr;
2530 			bp->b_flags &= ~B_UNMAPPED;
2531 			BUF_CHECK_MAPPED(bp);
2532 		}
2533 	}
2534 	return (bp);
2535 }
2536 
2537 /*
2538  *	buf_daemon:
2539  *
2540  *	buffer flushing daemon.  Buffers are normally flushed by the
2541  *	update daemon but if it cannot keep up this process starts to
2542  *	take the load in an attempt to prevent getnewbuf() from blocking.
2543  */
2544 
2545 static struct kproc_desc buf_kp = {
2546 	"bufdaemon",
2547 	buf_daemon,
2548 	&bufdaemonproc
2549 };
2550 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2551 
2552 static int
2553 buf_flush(int target)
2554 {
2555 	int flushed;
2556 
2557 	flushed = flushbufqueues(target, 0);
2558 	if (flushed == 0) {
2559 		/*
2560 		 * Could not find any buffers without rollback
2561 		 * dependencies, so just write the first one
2562 		 * in the hopes of eventually making progress.
2563 		 */
2564 		flushed = flushbufqueues(target, 1);
2565 	}
2566 	return (flushed);
2567 }
2568 
2569 static void
2570 buf_daemon()
2571 {
2572 	int lodirty;
2573 
2574 	/*
2575 	 * This process needs to be suspended prior to shutdown sync.
2576 	 */
2577 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2578 	    SHUTDOWN_PRI_LAST);
2579 
2580 	/*
2581 	 * This process is allowed to take the buffer cache to the limit
2582 	 */
2583 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2584 	mtx_lock(&bdlock);
2585 	for (;;) {
2586 		bd_request = 0;
2587 		mtx_unlock(&bdlock);
2588 
2589 		kproc_suspend_check(bufdaemonproc);
2590 		lodirty = lodirtybuffers;
2591 		if (bd_speedupreq) {
2592 			lodirty = numdirtybuffers / 2;
2593 			bd_speedupreq = 0;
2594 		}
2595 		/*
2596 		 * Do the flush.  Limit the amount of in-transit I/O we
2597 		 * allow to build up, otherwise we would completely saturate
2598 		 * the I/O system.
2599 		 */
2600 		while (numdirtybuffers > lodirty) {
2601 			if (buf_flush(numdirtybuffers - lodirty) == 0)
2602 				break;
2603 			kern_yield(PRI_USER);
2604 		}
2605 
2606 		/*
2607 		 * Only clear bd_request if we have reached our low water
2608 		 * mark.  The buf_daemon normally waits 1 second and
2609 		 * then incrementally flushes any dirty buffers that have
2610 		 * built up, within reason.
2611 		 *
2612 		 * If we were unable to hit our low water mark and couldn't
2613 		 * find any flushable buffers, we sleep for a short period
2614 		 * to avoid endless loops on unlockable buffers.
2615 		 */
2616 		mtx_lock(&bdlock);
2617 		if (numdirtybuffers <= lodirtybuffers) {
2618 			/*
2619 			 * We reached our low water mark, reset the
2620 			 * request and sleep until we are needed again.
2621 			 * The sleep is just so the suspend code works.
2622 			 */
2623 			bd_request = 0;
2624 			/*
2625 			 * Do an extra wakeup in case dirty threshold
2626 			 * changed via sysctl and the explicit transition
2627 			 * out of shortfall was missed.
2628 			 */
2629 			bdirtywakeup();
2630 			if (runningbufspace <= lorunningspace)
2631 				runningwakeup();
2632 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2633 		} else {
2634 			/*
2635 			 * We couldn't find any flushable dirty buffers but
2636 			 * still have too many dirty buffers, we
2637 			 * have to sleep and try again.  (rare)
2638 			 */
2639 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2640 		}
2641 	}
2642 }
2643 
2644 /*
2645  *	flushbufqueues:
2646  *
2647  *	Try to flush a buffer in the dirty queue.  We must be careful to
2648  *	free up B_INVAL buffers instead of write them, which NFS is
2649  *	particularly sensitive to.
2650  */
2651 static int flushwithdeps = 0;
2652 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2653     0, "Number of buffers flushed with dependecies that require rollbacks");
2654 
2655 static int
2656 flushbufqueues(int target, int flushdeps)
2657 {
2658 	struct buf *sentinel;
2659 	struct vnode *vp;
2660 	struct mount *mp;
2661 	struct buf *bp;
2662 	int hasdeps;
2663 	int flushed;
2664 	int queue;
2665 	int error;
2666 
2667 	flushed = 0;
2668 	queue = QUEUE_DIRTY;
2669 	bp = NULL;
2670 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2671 	sentinel->b_qindex = QUEUE_SENTINEL;
2672 	mtx_lock(&bqdirty);
2673 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2674 	mtx_unlock(&bqdirty);
2675 	while (flushed != target) {
2676 		maybe_yield();
2677 		mtx_lock(&bqdirty);
2678 		bp = TAILQ_NEXT(sentinel, b_freelist);
2679 		if (bp != NULL) {
2680 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2681 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2682 			    b_freelist);
2683 		} else {
2684 			mtx_unlock(&bqdirty);
2685 			break;
2686 		}
2687 		KASSERT(bp->b_qindex != QUEUE_SENTINEL,
2688 		    ("parallel calls to flushbufqueues() bp %p", bp));
2689 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2690 		mtx_unlock(&bqdirty);
2691 		if (error != 0)
2692 			continue;
2693 		if (bp->b_pin_count > 0) {
2694 			BUF_UNLOCK(bp);
2695 			continue;
2696 		}
2697 		/*
2698 		 * BKGRDINPROG can only be set with the buf and bufobj
2699 		 * locks both held.  We tolerate a race to clear it here.
2700 		 */
2701 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2702 		    (bp->b_flags & B_DELWRI) == 0) {
2703 			BUF_UNLOCK(bp);
2704 			continue;
2705 		}
2706 		if (bp->b_flags & B_INVAL) {
2707 			bremfreef(bp);
2708 			brelse(bp);
2709 			flushed++;
2710 			continue;
2711 		}
2712 
2713 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2714 			if (flushdeps == 0) {
2715 				BUF_UNLOCK(bp);
2716 				continue;
2717 			}
2718 			hasdeps = 1;
2719 		} else
2720 			hasdeps = 0;
2721 		/*
2722 		 * We must hold the lock on a vnode before writing
2723 		 * one of its buffers. Otherwise we may confuse, or
2724 		 * in the case of a snapshot vnode, deadlock the
2725 		 * system.
2726 		 *
2727 		 * The lock order here is the reverse of the normal
2728 		 * of vnode followed by buf lock.  This is ok because
2729 		 * the NOWAIT will prevent deadlock.
2730 		 */
2731 		vp = bp->b_vp;
2732 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2733 			BUF_UNLOCK(bp);
2734 			continue;
2735 		}
2736 		error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2737 		if (error == 0) {
2738 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2739 			    bp, bp->b_vp, bp->b_flags);
2740 			vfs_bio_awrite(bp);
2741 			vn_finished_write(mp);
2742 			VOP_UNLOCK(vp, 0);
2743 			flushwithdeps += hasdeps;
2744 			flushed++;
2745 			if (runningbufspace > hirunningspace)
2746 				waitrunningbufspace();
2747 			continue;
2748 		}
2749 		vn_finished_write(mp);
2750 		BUF_UNLOCK(bp);
2751 	}
2752 	mtx_lock(&bqdirty);
2753 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2754 	mtx_unlock(&bqdirty);
2755 	free(sentinel, M_TEMP);
2756 	return (flushed);
2757 }
2758 
2759 /*
2760  * Check to see if a block is currently memory resident.
2761  */
2762 struct buf *
2763 incore(struct bufobj *bo, daddr_t blkno)
2764 {
2765 	struct buf *bp;
2766 
2767 	BO_RLOCK(bo);
2768 	bp = gbincore(bo, blkno);
2769 	BO_RUNLOCK(bo);
2770 	return (bp);
2771 }
2772 
2773 /*
2774  * Returns true if no I/O is needed to access the
2775  * associated VM object.  This is like incore except
2776  * it also hunts around in the VM system for the data.
2777  */
2778 
2779 static int
2780 inmem(struct vnode * vp, daddr_t blkno)
2781 {
2782 	vm_object_t obj;
2783 	vm_offset_t toff, tinc, size;
2784 	vm_page_t m;
2785 	vm_ooffset_t off;
2786 
2787 	ASSERT_VOP_LOCKED(vp, "inmem");
2788 
2789 	if (incore(&vp->v_bufobj, blkno))
2790 		return 1;
2791 	if (vp->v_mount == NULL)
2792 		return 0;
2793 	obj = vp->v_object;
2794 	if (obj == NULL)
2795 		return (0);
2796 
2797 	size = PAGE_SIZE;
2798 	if (size > vp->v_mount->mnt_stat.f_iosize)
2799 		size = vp->v_mount->mnt_stat.f_iosize;
2800 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2801 
2802 	VM_OBJECT_RLOCK(obj);
2803 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2804 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2805 		if (!m)
2806 			goto notinmem;
2807 		tinc = size;
2808 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2809 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2810 		if (vm_page_is_valid(m,
2811 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2812 			goto notinmem;
2813 	}
2814 	VM_OBJECT_RUNLOCK(obj);
2815 	return 1;
2816 
2817 notinmem:
2818 	VM_OBJECT_RUNLOCK(obj);
2819 	return (0);
2820 }
2821 
2822 /*
2823  * Set the dirty range for a buffer based on the status of the dirty
2824  * bits in the pages comprising the buffer.  The range is limited
2825  * to the size of the buffer.
2826  *
2827  * Tell the VM system that the pages associated with this buffer
2828  * are clean.  This is used for delayed writes where the data is
2829  * going to go to disk eventually without additional VM intevention.
2830  *
2831  * Note that while we only really need to clean through to b_bcount, we
2832  * just go ahead and clean through to b_bufsize.
2833  */
2834 static void
2835 vfs_clean_pages_dirty_buf(struct buf *bp)
2836 {
2837 	vm_ooffset_t foff, noff, eoff;
2838 	vm_page_t m;
2839 	int i;
2840 
2841 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2842 		return;
2843 
2844 	foff = bp->b_offset;
2845 	KASSERT(bp->b_offset != NOOFFSET,
2846 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2847 
2848 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2849 	vfs_drain_busy_pages(bp);
2850 	vfs_setdirty_locked_object(bp);
2851 	for (i = 0; i < bp->b_npages; i++) {
2852 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2853 		eoff = noff;
2854 		if (eoff > bp->b_offset + bp->b_bufsize)
2855 			eoff = bp->b_offset + bp->b_bufsize;
2856 		m = bp->b_pages[i];
2857 		vfs_page_set_validclean(bp, foff, m);
2858 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2859 		foff = noff;
2860 	}
2861 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2862 }
2863 
2864 static void
2865 vfs_setdirty_locked_object(struct buf *bp)
2866 {
2867 	vm_object_t object;
2868 	int i;
2869 
2870 	object = bp->b_bufobj->bo_object;
2871 	VM_OBJECT_ASSERT_WLOCKED(object);
2872 
2873 	/*
2874 	 * We qualify the scan for modified pages on whether the
2875 	 * object has been flushed yet.
2876 	 */
2877 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2878 		vm_offset_t boffset;
2879 		vm_offset_t eoffset;
2880 
2881 		/*
2882 		 * test the pages to see if they have been modified directly
2883 		 * by users through the VM system.
2884 		 */
2885 		for (i = 0; i < bp->b_npages; i++)
2886 			vm_page_test_dirty(bp->b_pages[i]);
2887 
2888 		/*
2889 		 * Calculate the encompassing dirty range, boffset and eoffset,
2890 		 * (eoffset - boffset) bytes.
2891 		 */
2892 
2893 		for (i = 0; i < bp->b_npages; i++) {
2894 			if (bp->b_pages[i]->dirty)
2895 				break;
2896 		}
2897 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2898 
2899 		for (i = bp->b_npages - 1; i >= 0; --i) {
2900 			if (bp->b_pages[i]->dirty) {
2901 				break;
2902 			}
2903 		}
2904 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2905 
2906 		/*
2907 		 * Fit it to the buffer.
2908 		 */
2909 
2910 		if (eoffset > bp->b_bcount)
2911 			eoffset = bp->b_bcount;
2912 
2913 		/*
2914 		 * If we have a good dirty range, merge with the existing
2915 		 * dirty range.
2916 		 */
2917 
2918 		if (boffset < eoffset) {
2919 			if (bp->b_dirtyoff > boffset)
2920 				bp->b_dirtyoff = boffset;
2921 			if (bp->b_dirtyend < eoffset)
2922 				bp->b_dirtyend = eoffset;
2923 		}
2924 	}
2925 }
2926 
2927 /*
2928  * Allocate the KVA mapping for an existing buffer. It handles the
2929  * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2930  * KVA which is not mapped (B_KVAALLOC).
2931  */
2932 static void
2933 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2934 {
2935 	struct buf *scratch_bp;
2936 	int bsize, maxsize, need_mapping, need_kva;
2937 	off_t offset;
2938 
2939 	need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2940 	    (gbflags & GB_UNMAPPED) == 0;
2941 	need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2942 	    (gbflags & GB_KVAALLOC) != 0;
2943 	if (!need_mapping && !need_kva)
2944 		return;
2945 
2946 	BUF_CHECK_UNMAPPED(bp);
2947 
2948 	if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2949 		/*
2950 		 * Buffer is not mapped, but the KVA was already
2951 		 * reserved at the time of the instantiation.  Use the
2952 		 * allocated space.
2953 		 */
2954 		bp->b_flags &= ~B_KVAALLOC;
2955 		KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2956 		bp->b_kvabase = bp->b_kvaalloc;
2957 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
2958 		goto has_addr;
2959 	}
2960 
2961 	/*
2962 	 * Calculate the amount of the address space we would reserve
2963 	 * if the buffer was mapped.
2964 	 */
2965 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
2966 	offset = blkno * bsize;
2967 	maxsize = size + (offset & PAGE_MASK);
2968 	maxsize = imax(maxsize, bsize);
2969 
2970 mapping_loop:
2971 	if (allocbufkva(bp, maxsize, gbflags)) {
2972 		/*
2973 		 * Request defragmentation. getnewbuf() returns us the
2974 		 * allocated space by the scratch buffer KVA.
2975 		 */
2976 		scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
2977 		    (GB_UNMAPPED | GB_KVAALLOC));
2978 		if (scratch_bp == NULL) {
2979 			if ((gbflags & GB_NOWAIT_BD) != 0) {
2980 				/*
2981 				 * XXXKIB: defragmentation cannot
2982 				 * succeed, not sure what else to do.
2983 				 */
2984 				panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
2985 			}
2986 			atomic_add_int(&mappingrestarts, 1);
2987 			goto mapping_loop;
2988 		}
2989 		KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
2990 		    ("scratch bp !B_KVAALLOC %p", scratch_bp));
2991 		setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
2992 		    scratch_bp->b_kvasize, gbflags);
2993 
2994 		/* Get rid of the scratch buffer. */
2995 		scratch_bp->b_kvasize = 0;
2996 		scratch_bp->b_flags |= B_INVAL;
2997 		scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
2998 		brelse(scratch_bp);
2999 	}
3000 	if (!need_mapping)
3001 		return;
3002 
3003 has_addr:
3004 	bp->b_saveaddr = bp->b_kvabase;
3005 	bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
3006 	bp->b_flags &= ~B_UNMAPPED;
3007 	BUF_CHECK_MAPPED(bp);
3008 	bpmap_qenter(bp);
3009 }
3010 
3011 /*
3012  *	getblk:
3013  *
3014  *	Get a block given a specified block and offset into a file/device.
3015  *	The buffers B_DONE bit will be cleared on return, making it almost
3016  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3017  *	return.  The caller should clear B_INVAL prior to initiating a
3018  *	READ.
3019  *
3020  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3021  *	an existing buffer.
3022  *
3023  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3024  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3025  *	and then cleared based on the backing VM.  If the previous buffer is
3026  *	non-0-sized but invalid, B_CACHE will be cleared.
3027  *
3028  *	If getblk() must create a new buffer, the new buffer is returned with
3029  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3030  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3031  *	backing VM.
3032  *
3033  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3034  *	B_CACHE bit is clear.
3035  *
3036  *	What this means, basically, is that the caller should use B_CACHE to
3037  *	determine whether the buffer is fully valid or not and should clear
3038  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3039  *	the buffer by loading its data area with something, the caller needs
3040  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3041  *	the caller should set B_CACHE ( as an optimization ), else the caller
3042  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3043  *	a write attempt or if it was a successfull read.  If the caller
3044  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3045  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3046  */
3047 struct buf *
3048 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3049     int flags)
3050 {
3051 	struct buf *bp;
3052 	struct bufobj *bo;
3053 	int bsize, error, maxsize, vmio;
3054 	off_t offset;
3055 
3056 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3057 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3058 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3059 	ASSERT_VOP_LOCKED(vp, "getblk");
3060 	if (size > MAXBSIZE)
3061 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
3062 	if (!unmapped_buf_allowed)
3063 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3064 
3065 	bo = &vp->v_bufobj;
3066 loop:
3067 	BO_RLOCK(bo);
3068 	bp = gbincore(bo, blkno);
3069 	if (bp != NULL) {
3070 		int lockflags;
3071 		/*
3072 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3073 		 * it must be on a queue.
3074 		 */
3075 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3076 
3077 		if (flags & GB_LOCK_NOWAIT)
3078 			lockflags |= LK_NOWAIT;
3079 
3080 		error = BUF_TIMELOCK(bp, lockflags,
3081 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3082 
3083 		/*
3084 		 * If we slept and got the lock we have to restart in case
3085 		 * the buffer changed identities.
3086 		 */
3087 		if (error == ENOLCK)
3088 			goto loop;
3089 		/* We timed out or were interrupted. */
3090 		else if (error)
3091 			return (NULL);
3092 		/* If recursed, assume caller knows the rules. */
3093 		else if (BUF_LOCKRECURSED(bp))
3094 			goto end;
3095 
3096 		/*
3097 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3098 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3099 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3100 		 * backing VM cache.
3101 		 */
3102 		if (bp->b_flags & B_INVAL)
3103 			bp->b_flags &= ~B_CACHE;
3104 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3105 			bp->b_flags |= B_CACHE;
3106 		if (bp->b_flags & B_MANAGED)
3107 			MPASS(bp->b_qindex == QUEUE_NONE);
3108 		else
3109 			bremfree(bp);
3110 
3111 		/*
3112 		 * check for size inconsistencies for non-VMIO case.
3113 		 */
3114 		if (bp->b_bcount != size) {
3115 			if ((bp->b_flags & B_VMIO) == 0 ||
3116 			    (size > bp->b_kvasize)) {
3117 				if (bp->b_flags & B_DELWRI) {
3118 					/*
3119 					 * If buffer is pinned and caller does
3120 					 * not want sleep  waiting for it to be
3121 					 * unpinned, bail out
3122 					 * */
3123 					if (bp->b_pin_count > 0) {
3124 						if (flags & GB_LOCK_NOWAIT) {
3125 							bqrelse(bp);
3126 							return (NULL);
3127 						} else {
3128 							bunpin_wait(bp);
3129 						}
3130 					}
3131 					bp->b_flags |= B_NOCACHE;
3132 					bwrite(bp);
3133 				} else {
3134 					if (LIST_EMPTY(&bp->b_dep)) {
3135 						bp->b_flags |= B_RELBUF;
3136 						brelse(bp);
3137 					} else {
3138 						bp->b_flags |= B_NOCACHE;
3139 						bwrite(bp);
3140 					}
3141 				}
3142 				goto loop;
3143 			}
3144 		}
3145 
3146 		/*
3147 		 * Handle the case of unmapped buffer which should
3148 		 * become mapped, or the buffer for which KVA
3149 		 * reservation is requested.
3150 		 */
3151 		bp_unmapped_get_kva(bp, blkno, size, flags);
3152 
3153 		/*
3154 		 * If the size is inconsistant in the VMIO case, we can resize
3155 		 * the buffer.  This might lead to B_CACHE getting set or
3156 		 * cleared.  If the size has not changed, B_CACHE remains
3157 		 * unchanged from its previous state.
3158 		 */
3159 		if (bp->b_bcount != size)
3160 			allocbuf(bp, size);
3161 
3162 		KASSERT(bp->b_offset != NOOFFSET,
3163 		    ("getblk: no buffer offset"));
3164 
3165 		/*
3166 		 * A buffer with B_DELWRI set and B_CACHE clear must
3167 		 * be committed before we can return the buffer in
3168 		 * order to prevent the caller from issuing a read
3169 		 * ( due to B_CACHE not being set ) and overwriting
3170 		 * it.
3171 		 *
3172 		 * Most callers, including NFS and FFS, need this to
3173 		 * operate properly either because they assume they
3174 		 * can issue a read if B_CACHE is not set, or because
3175 		 * ( for example ) an uncached B_DELWRI might loop due
3176 		 * to softupdates re-dirtying the buffer.  In the latter
3177 		 * case, B_CACHE is set after the first write completes,
3178 		 * preventing further loops.
3179 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3180 		 * above while extending the buffer, we cannot allow the
3181 		 * buffer to remain with B_CACHE set after the write
3182 		 * completes or it will represent a corrupt state.  To
3183 		 * deal with this we set B_NOCACHE to scrap the buffer
3184 		 * after the write.
3185 		 *
3186 		 * We might be able to do something fancy, like setting
3187 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3188 		 * so the below call doesn't set B_CACHE, but that gets real
3189 		 * confusing.  This is much easier.
3190 		 */
3191 
3192 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3193 			bp->b_flags |= B_NOCACHE;
3194 			bwrite(bp);
3195 			goto loop;
3196 		}
3197 		bp->b_flags &= ~B_DONE;
3198 	} else {
3199 		/*
3200 		 * Buffer is not in-core, create new buffer.  The buffer
3201 		 * returned by getnewbuf() is locked.  Note that the returned
3202 		 * buffer is also considered valid (not marked B_INVAL).
3203 		 */
3204 		BO_RUNLOCK(bo);
3205 		/*
3206 		 * If the user does not want us to create the buffer, bail out
3207 		 * here.
3208 		 */
3209 		if (flags & GB_NOCREAT)
3210 			return NULL;
3211 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3212 			return NULL;
3213 
3214 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3215 		offset = blkno * bsize;
3216 		vmio = vp->v_object != NULL;
3217 		if (vmio) {
3218 			maxsize = size + (offset & PAGE_MASK);
3219 		} else {
3220 			maxsize = size;
3221 			/* Do not allow non-VMIO notmapped buffers. */
3222 			flags &= ~GB_UNMAPPED;
3223 		}
3224 		maxsize = imax(maxsize, bsize);
3225 
3226 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3227 		if (bp == NULL) {
3228 			if (slpflag || slptimeo)
3229 				return NULL;
3230 			goto loop;
3231 		}
3232 
3233 		/*
3234 		 * This code is used to make sure that a buffer is not
3235 		 * created while the getnewbuf routine is blocked.
3236 		 * This can be a problem whether the vnode is locked or not.
3237 		 * If the buffer is created out from under us, we have to
3238 		 * throw away the one we just created.
3239 		 *
3240 		 * Note: this must occur before we associate the buffer
3241 		 * with the vp especially considering limitations in
3242 		 * the splay tree implementation when dealing with duplicate
3243 		 * lblkno's.
3244 		 */
3245 		BO_LOCK(bo);
3246 		if (gbincore(bo, blkno)) {
3247 			BO_UNLOCK(bo);
3248 			bp->b_flags |= B_INVAL;
3249 			brelse(bp);
3250 			goto loop;
3251 		}
3252 
3253 		/*
3254 		 * Insert the buffer into the hash, so that it can
3255 		 * be found by incore.
3256 		 */
3257 		bp->b_blkno = bp->b_lblkno = blkno;
3258 		bp->b_offset = offset;
3259 		bgetvp(vp, bp);
3260 		BO_UNLOCK(bo);
3261 
3262 		/*
3263 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3264 		 * buffer size starts out as 0, B_CACHE will be set by
3265 		 * allocbuf() for the VMIO case prior to it testing the
3266 		 * backing store for validity.
3267 		 */
3268 
3269 		if (vmio) {
3270 			bp->b_flags |= B_VMIO;
3271 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3272 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3273 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3274 		} else {
3275 			bp->b_flags &= ~B_VMIO;
3276 			KASSERT(bp->b_bufobj->bo_object == NULL,
3277 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3278 			    bp, bp->b_bufobj->bo_object));
3279 			BUF_CHECK_MAPPED(bp);
3280 		}
3281 
3282 		allocbuf(bp, size);
3283 		bp->b_flags &= ~B_DONE;
3284 	}
3285 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3286 	BUF_ASSERT_HELD(bp);
3287 end:
3288 	KASSERT(bp->b_bufobj == bo,
3289 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3290 	return (bp);
3291 }
3292 
3293 /*
3294  * Get an empty, disassociated buffer of given size.  The buffer is initially
3295  * set to B_INVAL.
3296  */
3297 struct buf *
3298 geteblk(int size, int flags)
3299 {
3300 	struct buf *bp;
3301 	int maxsize;
3302 
3303 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3304 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3305 		if ((flags & GB_NOWAIT_BD) &&
3306 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3307 			return (NULL);
3308 	}
3309 	allocbuf(bp, size);
3310 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3311 	BUF_ASSERT_HELD(bp);
3312 	return (bp);
3313 }
3314 
3315 
3316 /*
3317  * This code constitutes the buffer memory from either anonymous system
3318  * memory (in the case of non-VMIO operations) or from an associated
3319  * VM object (in the case of VMIO operations).  This code is able to
3320  * resize a buffer up or down.
3321  *
3322  * Note that this code is tricky, and has many complications to resolve
3323  * deadlock or inconsistant data situations.  Tread lightly!!!
3324  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3325  * the caller.  Calling this code willy nilly can result in the loss of data.
3326  *
3327  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3328  * B_CACHE for the non-VMIO case.
3329  */
3330 
3331 int
3332 allocbuf(struct buf *bp, int size)
3333 {
3334 	int newbsize, mbsize;
3335 	int i;
3336 
3337 	BUF_ASSERT_HELD(bp);
3338 
3339 	if (bp->b_kvasize < size)
3340 		panic("allocbuf: buffer too small");
3341 
3342 	if ((bp->b_flags & B_VMIO) == 0) {
3343 		caddr_t origbuf;
3344 		int origbufsize;
3345 		/*
3346 		 * Just get anonymous memory from the kernel.  Don't
3347 		 * mess with B_CACHE.
3348 		 */
3349 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3350 		if (bp->b_flags & B_MALLOC)
3351 			newbsize = mbsize;
3352 		else
3353 			newbsize = round_page(size);
3354 
3355 		if (newbsize < bp->b_bufsize) {
3356 			/*
3357 			 * malloced buffers are not shrunk
3358 			 */
3359 			if (bp->b_flags & B_MALLOC) {
3360 				if (newbsize) {
3361 					bp->b_bcount = size;
3362 				} else {
3363 					free(bp->b_data, M_BIOBUF);
3364 					if (bp->b_bufsize) {
3365 						atomic_subtract_long(
3366 						    &bufmallocspace,
3367 						    bp->b_bufsize);
3368 						bufspacewakeup();
3369 						bp->b_bufsize = 0;
3370 					}
3371 					bp->b_saveaddr = bp->b_kvabase;
3372 					bp->b_data = bp->b_saveaddr;
3373 					bp->b_bcount = 0;
3374 					bp->b_flags &= ~B_MALLOC;
3375 				}
3376 				return 1;
3377 			}
3378 			vm_hold_free_pages(bp, newbsize);
3379 		} else if (newbsize > bp->b_bufsize) {
3380 			/*
3381 			 * We only use malloced memory on the first allocation.
3382 			 * and revert to page-allocated memory when the buffer
3383 			 * grows.
3384 			 */
3385 			/*
3386 			 * There is a potential smp race here that could lead
3387 			 * to bufmallocspace slightly passing the max.  It
3388 			 * is probably extremely rare and not worth worrying
3389 			 * over.
3390 			 */
3391 			if ( (bufmallocspace < maxbufmallocspace) &&
3392 				(bp->b_bufsize == 0) &&
3393 				(mbsize <= PAGE_SIZE/2)) {
3394 
3395 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3396 				bp->b_bufsize = mbsize;
3397 				bp->b_bcount = size;
3398 				bp->b_flags |= B_MALLOC;
3399 				atomic_add_long(&bufmallocspace, mbsize);
3400 				return 1;
3401 			}
3402 			origbuf = NULL;
3403 			origbufsize = 0;
3404 			/*
3405 			 * If the buffer is growing on its other-than-first allocation,
3406 			 * then we revert to the page-allocation scheme.
3407 			 */
3408 			if (bp->b_flags & B_MALLOC) {
3409 				origbuf = bp->b_data;
3410 				origbufsize = bp->b_bufsize;
3411 				bp->b_data = bp->b_kvabase;
3412 				if (bp->b_bufsize) {
3413 					atomic_subtract_long(&bufmallocspace,
3414 					    bp->b_bufsize);
3415 					bufspacewakeup();
3416 					bp->b_bufsize = 0;
3417 				}
3418 				bp->b_flags &= ~B_MALLOC;
3419 				newbsize = round_page(newbsize);
3420 			}
3421 			vm_hold_load_pages(
3422 			    bp,
3423 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
3424 			    (vm_offset_t) bp->b_data + newbsize);
3425 			if (origbuf) {
3426 				bcopy(origbuf, bp->b_data, origbufsize);
3427 				free(origbuf, M_BIOBUF);
3428 			}
3429 		}
3430 	} else {
3431 		int desiredpages;
3432 
3433 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3434 		desiredpages = (size == 0) ? 0 :
3435 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3436 
3437 		if (bp->b_flags & B_MALLOC)
3438 			panic("allocbuf: VMIO buffer can't be malloced");
3439 		/*
3440 		 * Set B_CACHE initially if buffer is 0 length or will become
3441 		 * 0-length.
3442 		 */
3443 		if (size == 0 || bp->b_bufsize == 0)
3444 			bp->b_flags |= B_CACHE;
3445 
3446 		if (newbsize < bp->b_bufsize) {
3447 			/*
3448 			 * DEV_BSIZE aligned new buffer size is less then the
3449 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3450 			 * if we have to remove any pages.
3451 			 */
3452 			if (desiredpages < bp->b_npages) {
3453 				vm_page_t m;
3454 
3455 				if ((bp->b_flags & B_UNMAPPED) == 0) {
3456 					BUF_CHECK_MAPPED(bp);
3457 					pmap_qremove((vm_offset_t)trunc_page(
3458 					    (vm_offset_t)bp->b_data) +
3459 					    (desiredpages << PAGE_SHIFT),
3460 					    (bp->b_npages - desiredpages));
3461 				} else
3462 					BUF_CHECK_UNMAPPED(bp);
3463 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3464 				for (i = desiredpages; i < bp->b_npages; i++) {
3465 					/*
3466 					 * the page is not freed here -- it
3467 					 * is the responsibility of
3468 					 * vnode_pager_setsize
3469 					 */
3470 					m = bp->b_pages[i];
3471 					KASSERT(m != bogus_page,
3472 					    ("allocbuf: bogus page found"));
3473 					while (vm_page_sleep_if_busy(m,
3474 					    "biodep"))
3475 						continue;
3476 
3477 					bp->b_pages[i] = NULL;
3478 					vm_page_lock(m);
3479 					vm_page_unwire(m, 0);
3480 					vm_page_unlock(m);
3481 				}
3482 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3483 				bp->b_npages = desiredpages;
3484 			}
3485 		} else if (size > bp->b_bcount) {
3486 			/*
3487 			 * We are growing the buffer, possibly in a
3488 			 * byte-granular fashion.
3489 			 */
3490 			vm_object_t obj;
3491 			vm_offset_t toff;
3492 			vm_offset_t tinc;
3493 
3494 			/*
3495 			 * Step 1, bring in the VM pages from the object,
3496 			 * allocating them if necessary.  We must clear
3497 			 * B_CACHE if these pages are not valid for the
3498 			 * range covered by the buffer.
3499 			 */
3500 
3501 			obj = bp->b_bufobj->bo_object;
3502 
3503 			VM_OBJECT_WLOCK(obj);
3504 			while (bp->b_npages < desiredpages) {
3505 				vm_page_t m;
3506 
3507 				/*
3508 				 * We must allocate system pages since blocking
3509 				 * here could interfere with paging I/O, no
3510 				 * matter which process we are.
3511 				 *
3512 				 * Only exclusive busy can be tested here.
3513 				 * Blocking on shared busy might lead to
3514 				 * deadlocks once allocbuf() is called after
3515 				 * pages are vfs_busy_pages().
3516 				 */
3517 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3518 				    bp->b_npages, VM_ALLOC_NOBUSY |
3519 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3520 				    VM_ALLOC_IGN_SBUSY |
3521 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3522 				if (m->valid == 0)
3523 					bp->b_flags &= ~B_CACHE;
3524 				bp->b_pages[bp->b_npages] = m;
3525 				++bp->b_npages;
3526 			}
3527 
3528 			/*
3529 			 * Step 2.  We've loaded the pages into the buffer,
3530 			 * we have to figure out if we can still have B_CACHE
3531 			 * set.  Note that B_CACHE is set according to the
3532 			 * byte-granular range ( bcount and size ), new the
3533 			 * aligned range ( newbsize ).
3534 			 *
3535 			 * The VM test is against m->valid, which is DEV_BSIZE
3536 			 * aligned.  Needless to say, the validity of the data
3537 			 * needs to also be DEV_BSIZE aligned.  Note that this
3538 			 * fails with NFS if the server or some other client
3539 			 * extends the file's EOF.  If our buffer is resized,
3540 			 * B_CACHE may remain set! XXX
3541 			 */
3542 
3543 			toff = bp->b_bcount;
3544 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3545 
3546 			while ((bp->b_flags & B_CACHE) && toff < size) {
3547 				vm_pindex_t pi;
3548 
3549 				if (tinc > (size - toff))
3550 					tinc = size - toff;
3551 
3552 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3553 				    PAGE_SHIFT;
3554 
3555 				vfs_buf_test_cache(
3556 				    bp,
3557 				    bp->b_offset,
3558 				    toff,
3559 				    tinc,
3560 				    bp->b_pages[pi]
3561 				);
3562 				toff += tinc;
3563 				tinc = PAGE_SIZE;
3564 			}
3565 			VM_OBJECT_WUNLOCK(obj);
3566 
3567 			/*
3568 			 * Step 3, fixup the KVM pmap.
3569 			 */
3570 			if ((bp->b_flags & B_UNMAPPED) == 0)
3571 				bpmap_qenter(bp);
3572 			else
3573 				BUF_CHECK_UNMAPPED(bp);
3574 		}
3575 	}
3576 	if (newbsize < bp->b_bufsize)
3577 		bufspacewakeup();
3578 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3579 	bp->b_bcount = size;		/* requested buffer size	*/
3580 	return 1;
3581 }
3582 
3583 extern int inflight_transient_maps;
3584 
3585 void
3586 biodone(struct bio *bp)
3587 {
3588 	struct mtx *mtxp;
3589 	void (*done)(struct bio *);
3590 	vm_offset_t start, end;
3591 
3592 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3593 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3594 		bp->bio_flags |= BIO_UNMAPPED;
3595 		start = trunc_page((vm_offset_t)bp->bio_data);
3596 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3597 		pmap_qremove(start, OFF_TO_IDX(end - start));
3598 		vmem_free(transient_arena, start, end - start);
3599 		atomic_add_int(&inflight_transient_maps, -1);
3600 	}
3601 	done = bp->bio_done;
3602 	if (done == NULL) {
3603 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
3604 		mtx_lock(mtxp);
3605 		bp->bio_flags |= BIO_DONE;
3606 		wakeup(bp);
3607 		mtx_unlock(mtxp);
3608 	} else {
3609 		bp->bio_flags |= BIO_DONE;
3610 		done(bp);
3611 	}
3612 }
3613 
3614 /*
3615  * Wait for a BIO to finish.
3616  */
3617 int
3618 biowait(struct bio *bp, const char *wchan)
3619 {
3620 	struct mtx *mtxp;
3621 
3622 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3623 	mtx_lock(mtxp);
3624 	while ((bp->bio_flags & BIO_DONE) == 0)
3625 		msleep(bp, mtxp, PRIBIO, wchan, 0);
3626 	mtx_unlock(mtxp);
3627 	if (bp->bio_error != 0)
3628 		return (bp->bio_error);
3629 	if (!(bp->bio_flags & BIO_ERROR))
3630 		return (0);
3631 	return (EIO);
3632 }
3633 
3634 void
3635 biofinish(struct bio *bp, struct devstat *stat, int error)
3636 {
3637 
3638 	if (error) {
3639 		bp->bio_error = error;
3640 		bp->bio_flags |= BIO_ERROR;
3641 	}
3642 	if (stat != NULL)
3643 		devstat_end_transaction_bio(stat, bp);
3644 	biodone(bp);
3645 }
3646 
3647 /*
3648  *	bufwait:
3649  *
3650  *	Wait for buffer I/O completion, returning error status.  The buffer
3651  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3652  *	error and cleared.
3653  */
3654 int
3655 bufwait(struct buf *bp)
3656 {
3657 	if (bp->b_iocmd == BIO_READ)
3658 		bwait(bp, PRIBIO, "biord");
3659 	else
3660 		bwait(bp, PRIBIO, "biowr");
3661 	if (bp->b_flags & B_EINTR) {
3662 		bp->b_flags &= ~B_EINTR;
3663 		return (EINTR);
3664 	}
3665 	if (bp->b_ioflags & BIO_ERROR) {
3666 		return (bp->b_error ? bp->b_error : EIO);
3667 	} else {
3668 		return (0);
3669 	}
3670 }
3671 
3672  /*
3673   * Call back function from struct bio back up to struct buf.
3674   */
3675 static void
3676 bufdonebio(struct bio *bip)
3677 {
3678 	struct buf *bp;
3679 
3680 	bp = bip->bio_caller2;
3681 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3682 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3683 	bp->b_ioflags = bip->bio_flags;
3684 	bp->b_error = bip->bio_error;
3685 	if (bp->b_error)
3686 		bp->b_ioflags |= BIO_ERROR;
3687 	bufdone(bp);
3688 	g_destroy_bio(bip);
3689 }
3690 
3691 void
3692 dev_strategy(struct cdev *dev, struct buf *bp)
3693 {
3694 	struct cdevsw *csw;
3695 	int ref;
3696 
3697 	KASSERT(dev->si_refcount > 0,
3698 	    ("dev_strategy on un-referenced struct cdev *(%s) %p",
3699 	    devtoname(dev), dev));
3700 
3701 	csw = dev_refthread(dev, &ref);
3702 	dev_strategy_csw(dev, csw, bp);
3703 	dev_relthread(dev, ref);
3704 }
3705 
3706 void
3707 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3708 {
3709 	struct bio *bip;
3710 
3711 	KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3712 	    ("b_iocmd botch"));
3713 	KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3714 	    dev->si_threadcount > 0,
3715 	    ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3716 	    dev));
3717 	if (csw == NULL) {
3718 		bp->b_error = ENXIO;
3719 		bp->b_ioflags = BIO_ERROR;
3720 		bufdone(bp);
3721 		return;
3722 	}
3723 	for (;;) {
3724 		bip = g_new_bio();
3725 		if (bip != NULL)
3726 			break;
3727 		/* Try again later */
3728 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3729 	}
3730 	bip->bio_cmd = bp->b_iocmd;
3731 	bip->bio_offset = bp->b_iooffset;
3732 	bip->bio_length = bp->b_bcount;
3733 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3734 	bdata2bio(bp, bip);
3735 	bip->bio_done = bufdonebio;
3736 	bip->bio_caller2 = bp;
3737 	bip->bio_dev = dev;
3738 	(*csw->d_strategy)(bip);
3739 }
3740 
3741 /*
3742  *	bufdone:
3743  *
3744  *	Finish I/O on a buffer, optionally calling a completion function.
3745  *	This is usually called from an interrupt so process blocking is
3746  *	not allowed.
3747  *
3748  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3749  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3750  *	assuming B_INVAL is clear.
3751  *
3752  *	For the VMIO case, we set B_CACHE if the op was a read and no
3753  *	read error occured, or if the op was a write.  B_CACHE is never
3754  *	set if the buffer is invalid or otherwise uncacheable.
3755  *
3756  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3757  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3758  *	in the biodone routine.
3759  */
3760 void
3761 bufdone(struct buf *bp)
3762 {
3763 	struct bufobj *dropobj;
3764 	void    (*biodone)(struct buf *);
3765 
3766 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3767 	dropobj = NULL;
3768 
3769 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3770 	BUF_ASSERT_HELD(bp);
3771 
3772 	runningbufwakeup(bp);
3773 	if (bp->b_iocmd == BIO_WRITE)
3774 		dropobj = bp->b_bufobj;
3775 	/* call optional completion function if requested */
3776 	if (bp->b_iodone != NULL) {
3777 		biodone = bp->b_iodone;
3778 		bp->b_iodone = NULL;
3779 		(*biodone) (bp);
3780 		if (dropobj)
3781 			bufobj_wdrop(dropobj);
3782 		return;
3783 	}
3784 
3785 	bufdone_finish(bp);
3786 
3787 	if (dropobj)
3788 		bufobj_wdrop(dropobj);
3789 }
3790 
3791 void
3792 bufdone_finish(struct buf *bp)
3793 {
3794 	BUF_ASSERT_HELD(bp);
3795 
3796 	if (!LIST_EMPTY(&bp->b_dep))
3797 		buf_complete(bp);
3798 
3799 	if (bp->b_flags & B_VMIO) {
3800 		vm_ooffset_t foff;
3801 		vm_page_t m;
3802 		vm_object_t obj;
3803 		struct vnode *vp;
3804 		int bogus, i, iosize;
3805 
3806 		obj = bp->b_bufobj->bo_object;
3807 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3808 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3809 		    obj->paging_in_progress, bp->b_npages));
3810 
3811 		vp = bp->b_vp;
3812 		KASSERT(vp->v_holdcnt > 0,
3813 		    ("biodone_finish: vnode %p has zero hold count", vp));
3814 		KASSERT(vp->v_object != NULL,
3815 		    ("biodone_finish: vnode %p has no vm_object", vp));
3816 
3817 		foff = bp->b_offset;
3818 		KASSERT(bp->b_offset != NOOFFSET,
3819 		    ("biodone_finish: bp %p has no buffer offset", bp));
3820 
3821 		/*
3822 		 * Set B_CACHE if the op was a normal read and no error
3823 		 * occured.  B_CACHE is set for writes in the b*write()
3824 		 * routines.
3825 		 */
3826 		iosize = bp->b_bcount - bp->b_resid;
3827 		if (bp->b_iocmd == BIO_READ &&
3828 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3829 		    !(bp->b_ioflags & BIO_ERROR)) {
3830 			bp->b_flags |= B_CACHE;
3831 		}
3832 		bogus = 0;
3833 		VM_OBJECT_WLOCK(obj);
3834 		for (i = 0; i < bp->b_npages; i++) {
3835 			int bogusflag = 0;
3836 			int resid;
3837 
3838 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3839 			if (resid > iosize)
3840 				resid = iosize;
3841 
3842 			/*
3843 			 * cleanup bogus pages, restoring the originals
3844 			 */
3845 			m = bp->b_pages[i];
3846 			if (m == bogus_page) {
3847 				bogus = bogusflag = 1;
3848 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3849 				if (m == NULL)
3850 					panic("biodone: page disappeared!");
3851 				bp->b_pages[i] = m;
3852 			}
3853 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3854 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3855 			    (intmax_t)foff, (uintmax_t)m->pindex));
3856 
3857 			/*
3858 			 * In the write case, the valid and clean bits are
3859 			 * already changed correctly ( see bdwrite() ), so we
3860 			 * only need to do this here in the read case.
3861 			 */
3862 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3863 				KASSERT((m->dirty & vm_page_bits(foff &
3864 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3865 				    " page %p has unexpected dirty bits", m));
3866 				vfs_page_set_valid(bp, foff, m);
3867 			}
3868 
3869 			vm_page_sunbusy(m);
3870 			vm_object_pip_subtract(obj, 1);
3871 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3872 			iosize -= resid;
3873 		}
3874 		vm_object_pip_wakeupn(obj, 0);
3875 		VM_OBJECT_WUNLOCK(obj);
3876 		if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3877 			BUF_CHECK_MAPPED(bp);
3878 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3879 			    bp->b_pages, bp->b_npages);
3880 		}
3881 	}
3882 
3883 	/*
3884 	 * For asynchronous completions, release the buffer now. The brelse
3885 	 * will do a wakeup there if necessary - so no need to do a wakeup
3886 	 * here in the async case. The sync case always needs to do a wakeup.
3887 	 */
3888 
3889 	if (bp->b_flags & B_ASYNC) {
3890 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3891 			brelse(bp);
3892 		else
3893 			bqrelse(bp);
3894 	} else
3895 		bdone(bp);
3896 }
3897 
3898 /*
3899  * This routine is called in lieu of iodone in the case of
3900  * incomplete I/O.  This keeps the busy status for pages
3901  * consistant.
3902  */
3903 void
3904 vfs_unbusy_pages(struct buf *bp)
3905 {
3906 	int i;
3907 	vm_object_t obj;
3908 	vm_page_t m;
3909 
3910 	runningbufwakeup(bp);
3911 	if (!(bp->b_flags & B_VMIO))
3912 		return;
3913 
3914 	obj = bp->b_bufobj->bo_object;
3915 	VM_OBJECT_WLOCK(obj);
3916 	for (i = 0; i < bp->b_npages; i++) {
3917 		m = bp->b_pages[i];
3918 		if (m == bogus_page) {
3919 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3920 			if (!m)
3921 				panic("vfs_unbusy_pages: page missing\n");
3922 			bp->b_pages[i] = m;
3923 			if ((bp->b_flags & B_UNMAPPED) == 0) {
3924 				BUF_CHECK_MAPPED(bp);
3925 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3926 				    bp->b_pages, bp->b_npages);
3927 			} else
3928 				BUF_CHECK_UNMAPPED(bp);
3929 		}
3930 		vm_object_pip_subtract(obj, 1);
3931 		vm_page_sunbusy(m);
3932 	}
3933 	vm_object_pip_wakeupn(obj, 0);
3934 	VM_OBJECT_WUNLOCK(obj);
3935 }
3936 
3937 /*
3938  * vfs_page_set_valid:
3939  *
3940  *	Set the valid bits in a page based on the supplied offset.   The
3941  *	range is restricted to the buffer's size.
3942  *
3943  *	This routine is typically called after a read completes.
3944  */
3945 static void
3946 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3947 {
3948 	vm_ooffset_t eoff;
3949 
3950 	/*
3951 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3952 	 * page boundary and eoff is not greater than the end of the buffer.
3953 	 * The end of the buffer, in this case, is our file EOF, not the
3954 	 * allocation size of the buffer.
3955 	 */
3956 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3957 	if (eoff > bp->b_offset + bp->b_bcount)
3958 		eoff = bp->b_offset + bp->b_bcount;
3959 
3960 	/*
3961 	 * Set valid range.  This is typically the entire buffer and thus the
3962 	 * entire page.
3963 	 */
3964 	if (eoff > off)
3965 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3966 }
3967 
3968 /*
3969  * vfs_page_set_validclean:
3970  *
3971  *	Set the valid bits and clear the dirty bits in a page based on the
3972  *	supplied offset.   The range is restricted to the buffer's size.
3973  */
3974 static void
3975 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3976 {
3977 	vm_ooffset_t soff, eoff;
3978 
3979 	/*
3980 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3981 	 * page boundry or cross the end of the buffer.  The end of the
3982 	 * buffer, in this case, is our file EOF, not the allocation size
3983 	 * of the buffer.
3984 	 */
3985 	soff = off;
3986 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3987 	if (eoff > bp->b_offset + bp->b_bcount)
3988 		eoff = bp->b_offset + bp->b_bcount;
3989 
3990 	/*
3991 	 * Set valid range.  This is typically the entire buffer and thus the
3992 	 * entire page.
3993 	 */
3994 	if (eoff > soff) {
3995 		vm_page_set_validclean(
3996 		    m,
3997 		   (vm_offset_t) (soff & PAGE_MASK),
3998 		   (vm_offset_t) (eoff - soff)
3999 		);
4000 	}
4001 }
4002 
4003 /*
4004  * Ensure that all buffer pages are not exclusive busied.  If any page is
4005  * exclusive busy, drain it.
4006  */
4007 void
4008 vfs_drain_busy_pages(struct buf *bp)
4009 {
4010 	vm_page_t m;
4011 	int i, last_busied;
4012 
4013 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4014 	last_busied = 0;
4015 	for (i = 0; i < bp->b_npages; i++) {
4016 		m = bp->b_pages[i];
4017 		if (vm_page_xbusied(m)) {
4018 			for (; last_busied < i; last_busied++)
4019 				vm_page_sbusy(bp->b_pages[last_busied]);
4020 			while (vm_page_xbusied(m)) {
4021 				vm_page_lock(m);
4022 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4023 				vm_page_busy_sleep(m, "vbpage");
4024 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4025 			}
4026 		}
4027 	}
4028 	for (i = 0; i < last_busied; i++)
4029 		vm_page_sunbusy(bp->b_pages[i]);
4030 }
4031 
4032 /*
4033  * This routine is called before a device strategy routine.
4034  * It is used to tell the VM system that paging I/O is in
4035  * progress, and treat the pages associated with the buffer
4036  * almost as being exclusive busy.  Also the object paging_in_progress
4037  * flag is handled to make sure that the object doesn't become
4038  * inconsistant.
4039  *
4040  * Since I/O has not been initiated yet, certain buffer flags
4041  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4042  * and should be ignored.
4043  */
4044 void
4045 vfs_busy_pages(struct buf *bp, int clear_modify)
4046 {
4047 	int i, bogus;
4048 	vm_object_t obj;
4049 	vm_ooffset_t foff;
4050 	vm_page_t m;
4051 
4052 	if (!(bp->b_flags & B_VMIO))
4053 		return;
4054 
4055 	obj = bp->b_bufobj->bo_object;
4056 	foff = bp->b_offset;
4057 	KASSERT(bp->b_offset != NOOFFSET,
4058 	    ("vfs_busy_pages: no buffer offset"));
4059 	VM_OBJECT_WLOCK(obj);
4060 	vfs_drain_busy_pages(bp);
4061 	if (bp->b_bufsize != 0)
4062 		vfs_setdirty_locked_object(bp);
4063 	bogus = 0;
4064 	for (i = 0; i < bp->b_npages; i++) {
4065 		m = bp->b_pages[i];
4066 
4067 		if ((bp->b_flags & B_CLUSTER) == 0) {
4068 			vm_object_pip_add(obj, 1);
4069 			vm_page_sbusy(m);
4070 		}
4071 		/*
4072 		 * When readying a buffer for a read ( i.e
4073 		 * clear_modify == 0 ), it is important to do
4074 		 * bogus_page replacement for valid pages in
4075 		 * partially instantiated buffers.  Partially
4076 		 * instantiated buffers can, in turn, occur when
4077 		 * reconstituting a buffer from its VM backing store
4078 		 * base.  We only have to do this if B_CACHE is
4079 		 * clear ( which causes the I/O to occur in the
4080 		 * first place ).  The replacement prevents the read
4081 		 * I/O from overwriting potentially dirty VM-backed
4082 		 * pages.  XXX bogus page replacement is, uh, bogus.
4083 		 * It may not work properly with small-block devices.
4084 		 * We need to find a better way.
4085 		 */
4086 		if (clear_modify) {
4087 			pmap_remove_write(m);
4088 			vfs_page_set_validclean(bp, foff, m);
4089 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4090 		    (bp->b_flags & B_CACHE) == 0) {
4091 			bp->b_pages[i] = bogus_page;
4092 			bogus++;
4093 		}
4094 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4095 	}
4096 	VM_OBJECT_WUNLOCK(obj);
4097 	if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4098 		BUF_CHECK_MAPPED(bp);
4099 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4100 		    bp->b_pages, bp->b_npages);
4101 	}
4102 }
4103 
4104 /*
4105  *	vfs_bio_set_valid:
4106  *
4107  *	Set the range within the buffer to valid.  The range is
4108  *	relative to the beginning of the buffer, b_offset.  Note that
4109  *	b_offset itself may be offset from the beginning of the first
4110  *	page.
4111  */
4112 void
4113 vfs_bio_set_valid(struct buf *bp, int base, int size)
4114 {
4115 	int i, n;
4116 	vm_page_t m;
4117 
4118 	if (!(bp->b_flags & B_VMIO))
4119 		return;
4120 
4121 	/*
4122 	 * Fixup base to be relative to beginning of first page.
4123 	 * Set initial n to be the maximum number of bytes in the
4124 	 * first page that can be validated.
4125 	 */
4126 	base += (bp->b_offset & PAGE_MASK);
4127 	n = PAGE_SIZE - (base & PAGE_MASK);
4128 
4129 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4130 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4131 		m = bp->b_pages[i];
4132 		if (n > size)
4133 			n = size;
4134 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4135 		base += n;
4136 		size -= n;
4137 		n = PAGE_SIZE;
4138 	}
4139 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4140 }
4141 
4142 /*
4143  *	vfs_bio_clrbuf:
4144  *
4145  *	If the specified buffer is a non-VMIO buffer, clear the entire
4146  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4147  *	validate only the previously invalid portions of the buffer.
4148  *	This routine essentially fakes an I/O, so we need to clear
4149  *	BIO_ERROR and B_INVAL.
4150  *
4151  *	Note that while we only theoretically need to clear through b_bcount,
4152  *	we go ahead and clear through b_bufsize.
4153  */
4154 void
4155 vfs_bio_clrbuf(struct buf *bp)
4156 {
4157 	int i, j, mask, sa, ea, slide;
4158 
4159 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4160 		clrbuf(bp);
4161 		return;
4162 	}
4163 	bp->b_flags &= ~B_INVAL;
4164 	bp->b_ioflags &= ~BIO_ERROR;
4165 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4166 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4167 	    (bp->b_offset & PAGE_MASK) == 0) {
4168 		if (bp->b_pages[0] == bogus_page)
4169 			goto unlock;
4170 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4171 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4172 		if ((bp->b_pages[0]->valid & mask) == mask)
4173 			goto unlock;
4174 		if ((bp->b_pages[0]->valid & mask) == 0) {
4175 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4176 			bp->b_pages[0]->valid |= mask;
4177 			goto unlock;
4178 		}
4179 	}
4180 	sa = bp->b_offset & PAGE_MASK;
4181 	slide = 0;
4182 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4183 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4184 		ea = slide & PAGE_MASK;
4185 		if (ea == 0)
4186 			ea = PAGE_SIZE;
4187 		if (bp->b_pages[i] == bogus_page)
4188 			continue;
4189 		j = sa / DEV_BSIZE;
4190 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4191 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4192 		if ((bp->b_pages[i]->valid & mask) == mask)
4193 			continue;
4194 		if ((bp->b_pages[i]->valid & mask) == 0)
4195 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4196 		else {
4197 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4198 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4199 					pmap_zero_page_area(bp->b_pages[i],
4200 					    sa, DEV_BSIZE);
4201 				}
4202 			}
4203 		}
4204 		bp->b_pages[i]->valid |= mask;
4205 	}
4206 unlock:
4207 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4208 	bp->b_resid = 0;
4209 }
4210 
4211 void
4212 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4213 {
4214 	vm_page_t m;
4215 	int i, n;
4216 
4217 	if ((bp->b_flags & B_UNMAPPED) == 0) {
4218 		BUF_CHECK_MAPPED(bp);
4219 		bzero(bp->b_data + base, size);
4220 	} else {
4221 		BUF_CHECK_UNMAPPED(bp);
4222 		n = PAGE_SIZE - (base & PAGE_MASK);
4223 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4224 			m = bp->b_pages[i];
4225 			if (n > size)
4226 				n = size;
4227 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4228 			base += n;
4229 			size -= n;
4230 			n = PAGE_SIZE;
4231 		}
4232 	}
4233 }
4234 
4235 /*
4236  * vm_hold_load_pages and vm_hold_free_pages get pages into
4237  * a buffers address space.  The pages are anonymous and are
4238  * not associated with a file object.
4239  */
4240 static void
4241 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4242 {
4243 	vm_offset_t pg;
4244 	vm_page_t p;
4245 	int index;
4246 
4247 	BUF_CHECK_MAPPED(bp);
4248 
4249 	to = round_page(to);
4250 	from = round_page(from);
4251 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4252 
4253 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4254 tryagain:
4255 		/*
4256 		 * note: must allocate system pages since blocking here
4257 		 * could interfere with paging I/O, no matter which
4258 		 * process we are.
4259 		 */
4260 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4261 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4262 		if (p == NULL) {
4263 			VM_WAIT;
4264 			goto tryagain;
4265 		}
4266 		pmap_qenter(pg, &p, 1);
4267 		bp->b_pages[index] = p;
4268 	}
4269 	bp->b_npages = index;
4270 }
4271 
4272 /* Return pages associated with this buf to the vm system */
4273 static void
4274 vm_hold_free_pages(struct buf *bp, int newbsize)
4275 {
4276 	vm_offset_t from;
4277 	vm_page_t p;
4278 	int index, newnpages;
4279 
4280 	BUF_CHECK_MAPPED(bp);
4281 
4282 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4283 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4284 	if (bp->b_npages > newnpages)
4285 		pmap_qremove(from, bp->b_npages - newnpages);
4286 	for (index = newnpages; index < bp->b_npages; index++) {
4287 		p = bp->b_pages[index];
4288 		bp->b_pages[index] = NULL;
4289 		if (vm_page_sbusied(p))
4290 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4291 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4292 		p->wire_count--;
4293 		vm_page_free(p);
4294 		atomic_subtract_int(&cnt.v_wire_count, 1);
4295 	}
4296 	bp->b_npages = newnpages;
4297 }
4298 
4299 /*
4300  * Map an IO request into kernel virtual address space.
4301  *
4302  * All requests are (re)mapped into kernel VA space.
4303  * Notice that we use b_bufsize for the size of the buffer
4304  * to be mapped.  b_bcount might be modified by the driver.
4305  *
4306  * Note that even if the caller determines that the address space should
4307  * be valid, a race or a smaller-file mapped into a larger space may
4308  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4309  * check the return value.
4310  */
4311 int
4312 vmapbuf(struct buf *bp, int mapbuf)
4313 {
4314 	caddr_t kva;
4315 	vm_prot_t prot;
4316 	int pidx;
4317 
4318 	if (bp->b_bufsize < 0)
4319 		return (-1);
4320 	prot = VM_PROT_READ;
4321 	if (bp->b_iocmd == BIO_READ)
4322 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4323 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4324 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4325 	    btoc(MAXPHYS))) < 0)
4326 		return (-1);
4327 	bp->b_npages = pidx;
4328 	if (mapbuf || !unmapped_buf_allowed) {
4329 		pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4330 		kva = bp->b_saveaddr;
4331 		bp->b_saveaddr = bp->b_data;
4332 		bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4333 		bp->b_flags &= ~B_UNMAPPED;
4334 	} else {
4335 		bp->b_flags |= B_UNMAPPED;
4336 		bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4337 		bp->b_saveaddr = bp->b_data;
4338 		bp->b_data = unmapped_buf;
4339 	}
4340 	return(0);
4341 }
4342 
4343 /*
4344  * Free the io map PTEs associated with this IO operation.
4345  * We also invalidate the TLB entries and restore the original b_addr.
4346  */
4347 void
4348 vunmapbuf(struct buf *bp)
4349 {
4350 	int npages;
4351 
4352 	npages = bp->b_npages;
4353 	if (bp->b_flags & B_UNMAPPED)
4354 		bp->b_flags &= ~B_UNMAPPED;
4355 	else
4356 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4357 	vm_page_unhold_pages(bp->b_pages, npages);
4358 
4359 	bp->b_data = bp->b_saveaddr;
4360 }
4361 
4362 void
4363 bdone(struct buf *bp)
4364 {
4365 	struct mtx *mtxp;
4366 
4367 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4368 	mtx_lock(mtxp);
4369 	bp->b_flags |= B_DONE;
4370 	wakeup(bp);
4371 	mtx_unlock(mtxp);
4372 }
4373 
4374 void
4375 bwait(struct buf *bp, u_char pri, const char *wchan)
4376 {
4377 	struct mtx *mtxp;
4378 
4379 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4380 	mtx_lock(mtxp);
4381 	while ((bp->b_flags & B_DONE) == 0)
4382 		msleep(bp, mtxp, pri, wchan, 0);
4383 	mtx_unlock(mtxp);
4384 }
4385 
4386 int
4387 bufsync(struct bufobj *bo, int waitfor)
4388 {
4389 
4390 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4391 }
4392 
4393 void
4394 bufstrategy(struct bufobj *bo, struct buf *bp)
4395 {
4396 	int i = 0;
4397 	struct vnode *vp;
4398 
4399 	vp = bp->b_vp;
4400 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4401 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4402 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4403 	i = VOP_STRATEGY(vp, bp);
4404 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4405 }
4406 
4407 void
4408 bufobj_wrefl(struct bufobj *bo)
4409 {
4410 
4411 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4412 	ASSERT_BO_WLOCKED(bo);
4413 	bo->bo_numoutput++;
4414 }
4415 
4416 void
4417 bufobj_wref(struct bufobj *bo)
4418 {
4419 
4420 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4421 	BO_LOCK(bo);
4422 	bo->bo_numoutput++;
4423 	BO_UNLOCK(bo);
4424 }
4425 
4426 void
4427 bufobj_wdrop(struct bufobj *bo)
4428 {
4429 
4430 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4431 	BO_LOCK(bo);
4432 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4433 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4434 		bo->bo_flag &= ~BO_WWAIT;
4435 		wakeup(&bo->bo_numoutput);
4436 	}
4437 	BO_UNLOCK(bo);
4438 }
4439 
4440 int
4441 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4442 {
4443 	int error;
4444 
4445 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4446 	ASSERT_BO_WLOCKED(bo);
4447 	error = 0;
4448 	while (bo->bo_numoutput) {
4449 		bo->bo_flag |= BO_WWAIT;
4450 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4451 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4452 		if (error)
4453 			break;
4454 	}
4455 	return (error);
4456 }
4457 
4458 void
4459 bpin(struct buf *bp)
4460 {
4461 	struct mtx *mtxp;
4462 
4463 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4464 	mtx_lock(mtxp);
4465 	bp->b_pin_count++;
4466 	mtx_unlock(mtxp);
4467 }
4468 
4469 void
4470 bunpin(struct buf *bp)
4471 {
4472 	struct mtx *mtxp;
4473 
4474 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4475 	mtx_lock(mtxp);
4476 	if (--bp->b_pin_count == 0)
4477 		wakeup(bp);
4478 	mtx_unlock(mtxp);
4479 }
4480 
4481 void
4482 bunpin_wait(struct buf *bp)
4483 {
4484 	struct mtx *mtxp;
4485 
4486 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4487 	mtx_lock(mtxp);
4488 	while (bp->b_pin_count > 0)
4489 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4490 	mtx_unlock(mtxp);
4491 }
4492 
4493 /*
4494  * Set bio_data or bio_ma for struct bio from the struct buf.
4495  */
4496 void
4497 bdata2bio(struct buf *bp, struct bio *bip)
4498 {
4499 
4500 	if ((bp->b_flags & B_UNMAPPED) != 0) {
4501 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4502 		bip->bio_ma = bp->b_pages;
4503 		bip->bio_ma_n = bp->b_npages;
4504 		bip->bio_data = unmapped_buf;
4505 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4506 		bip->bio_flags |= BIO_UNMAPPED;
4507 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4508 		    PAGE_SIZE == bp->b_npages,
4509 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4510 		    (long long)bip->bio_length, bip->bio_ma_n));
4511 	} else {
4512 		bip->bio_data = bp->b_data;
4513 		bip->bio_ma = NULL;
4514 	}
4515 }
4516 
4517 #include "opt_ddb.h"
4518 #ifdef DDB
4519 #include <ddb/ddb.h>
4520 
4521 /* DDB command to show buffer data */
4522 DB_SHOW_COMMAND(buffer, db_show_buffer)
4523 {
4524 	/* get args */
4525 	struct buf *bp = (struct buf *)addr;
4526 
4527 	if (!have_addr) {
4528 		db_printf("usage: show buffer <addr>\n");
4529 		return;
4530 	}
4531 
4532 	db_printf("buf at %p\n", bp);
4533 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4534 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4535 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4536 	db_printf(
4537 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4538 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4539 	    "b_dep = %p\n",
4540 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4541 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4542 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4543 	if (bp->b_npages) {
4544 		int i;
4545 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4546 		for (i = 0; i < bp->b_npages; i++) {
4547 			vm_page_t m;
4548 			m = bp->b_pages[i];
4549 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4550 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4551 			if ((i + 1) < bp->b_npages)
4552 				db_printf(",");
4553 		}
4554 		db_printf("\n");
4555 	}
4556 	db_printf(" ");
4557 	BUF_LOCKPRINTINFO(bp);
4558 }
4559 
4560 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4561 {
4562 	struct buf *bp;
4563 	int i;
4564 
4565 	for (i = 0; i < nbuf; i++) {
4566 		bp = &buf[i];
4567 		if (BUF_ISLOCKED(bp)) {
4568 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4569 			db_printf("\n");
4570 		}
4571 	}
4572 }
4573 
4574 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4575 {
4576 	struct vnode *vp;
4577 	struct buf *bp;
4578 
4579 	if (!have_addr) {
4580 		db_printf("usage: show vnodebufs <addr>\n");
4581 		return;
4582 	}
4583 	vp = (struct vnode *)addr;
4584 	db_printf("Clean buffers:\n");
4585 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4586 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4587 		db_printf("\n");
4588 	}
4589 	db_printf("Dirty buffers:\n");
4590 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4591 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4592 		db_printf("\n");
4593 	}
4594 }
4595 
4596 DB_COMMAND(countfreebufs, db_coundfreebufs)
4597 {
4598 	struct buf *bp;
4599 	int i, used = 0, nfree = 0;
4600 
4601 	if (have_addr) {
4602 		db_printf("usage: countfreebufs\n");
4603 		return;
4604 	}
4605 
4606 	for (i = 0; i < nbuf; i++) {
4607 		bp = &buf[i];
4608 		if ((bp->b_flags & B_INFREECNT) != 0)
4609 			nfree++;
4610 		else
4611 			used++;
4612 	}
4613 
4614 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4615 	    nfree + used);
4616 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4617 }
4618 #endif /* DDB */
4619