xref: /freebsd/sys/kern/vfs_bio.c (revision d7e0694a595722a6839f88ec7219a981ed734b95)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * Portions of this software were developed by Konstantin Belousov
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * this file contains a new buffer I/O scheme implementing a coherent
34  * VM object and buffer cache scheme.  Pains have been taken to make
35  * sure that the performance degradation associated with schemes such
36  * as this is not realized.
37  *
38  * Author:  John S. Dyson
39  * Significant help during the development and debugging phases
40  * had been provided by David Greenman, also of the FreeBSD core team.
41  *
42  * see man buf(9) for more info.
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/conf.h>
52 #include <sys/buf.h>
53 #include <sys/devicestat.h>
54 #include <sys/eventhandler.h>
55 #include <sys/fail.h>
56 #include <sys/limits.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/kernel.h>
62 #include <sys/kthread.h>
63 #include <sys/proc.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
66 #include <sys/sysctl.h>
67 #include <sys/vmem.h>
68 #include <sys/vmmeter.h>
69 #include <sys/vnode.h>
70 #include <geom/geom.h>
71 #include <vm/vm.h>
72 #include <vm/vm_param.h>
73 #include <vm/vm_kern.h>
74 #include <vm/vm_pageout.h>
75 #include <vm/vm_page.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/vm_map.h>
79 #include "opt_compat.h"
80 #include "opt_swap.h"
81 
82 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
83 
84 struct	bio_ops bioops;		/* I/O operation notification */
85 
86 struct	buf_ops buf_ops_bio = {
87 	.bop_name	=	"buf_ops_bio",
88 	.bop_write	=	bufwrite,
89 	.bop_strategy	=	bufstrategy,
90 	.bop_sync	=	bufsync,
91 	.bop_bdflush	=	bufbdflush,
92 };
93 
94 /*
95  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
96  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
97  */
98 struct buf *buf;		/* buffer header pool */
99 caddr_t unmapped_buf;
100 
101 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
102 struct proc *bufdaemonproc;
103 
104 static int inmem(struct vnode *vp, daddr_t blkno);
105 static void vm_hold_free_pages(struct buf *bp, int newbsize);
106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
107 		vm_offset_t to);
108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
110 		vm_page_t m);
111 static void vfs_clean_pages_dirty_buf(struct buf *bp);
112 static void vfs_setdirty_locked_object(struct buf *bp);
113 static void vfs_vmio_release(struct buf *bp);
114 static int vfs_bio_clcheck(struct vnode *vp, int size,
115 		daddr_t lblkno, daddr_t blkno);
116 static int buf_flush(int);
117 static int flushbufqueues(int, int);
118 static void buf_daemon(void);
119 static void bremfreel(struct buf *bp);
120 static __inline void bd_wakeup(void);
121 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
122 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
123     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
124 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
125 #endif
126 
127 int vmiodirenable = TRUE;
128 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
129     "Use the VM system for directory writes");
130 long runningbufspace;
131 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
132     "Amount of presently outstanding async buffer io");
133 static long bufspace;
134 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
135     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
136 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
137     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
138 #else
139 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
140     "Virtual memory used for buffers");
141 #endif
142 static long unmapped_bufspace;
143 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD,
144     &unmapped_bufspace, 0,
145     "Amount of unmapped buffers, inclusive in the bufspace");
146 static long maxbufspace;
147 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
148     "Maximum allowed value of bufspace (including buf_daemon)");
149 static long bufmallocspace;
150 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
151     "Amount of malloced memory for buffers");
152 static long maxbufmallocspace;
153 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
154     "Maximum amount of malloced memory for buffers");
155 static long lobufspace;
156 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
157     "Minimum amount of buffers we want to have");
158 long hibufspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
160     "Maximum allowed value of bufspace (excluding buf_daemon)");
161 static int bufreusecnt;
162 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
163     "Number of times we have reused a buffer");
164 static int buffreekvacnt;
165 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
166     "Number of times we have freed the KVA space from some buffer");
167 static int bufdefragcnt;
168 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
169     "Number of times we have had to repeat buffer allocation to defragment");
170 static long lorunningspace;
171 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
172     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
173     "Minimum preferred space used for in-progress I/O");
174 static long hirunningspace;
175 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
176     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
177     "Maximum amount of space to use for in-progress I/O");
178 int dirtybufferflushes;
179 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
180     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
181 int bdwriteskip;
182 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
183     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
184 int altbufferflushes;
185 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
186     0, "Number of fsync flushes to limit dirty buffers");
187 static int recursiveflushes;
188 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
189     0, "Number of flushes skipped due to being recursive");
190 static int numdirtybuffers;
191 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
192     "Number of buffers that are dirty (has unwritten changes) at the moment");
193 static int lodirtybuffers;
194 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
195     "How many buffers we want to have free before bufdaemon can sleep");
196 static int hidirtybuffers;
197 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
198     "When the number of dirty buffers is considered severe");
199 int dirtybufthresh;
200 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
201     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
202 static int numfreebuffers;
203 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
204     "Number of free buffers");
205 static int lofreebuffers;
206 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
207    "XXX Unused");
208 static int hifreebuffers;
209 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
210    "XXX Complicatedly unused");
211 static int getnewbufcalls;
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
213    "Number of calls to getnewbuf");
214 static int getnewbufrestarts;
215 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
216     "Number of times getnewbuf has had to restart a buffer aquisition");
217 static int mappingrestarts;
218 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
219     "Number of times getblk has had to restart a buffer mapping for "
220     "unmapped buffer");
221 static int flushbufqtarget = 100;
222 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
223     "Amount of work to do in flushbufqueues when helping bufdaemon");
224 static long notbufdflushes;
225 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes, 0,
226     "Number of dirty buffer flushes done by the bufdaemon helpers");
227 static long barrierwrites;
228 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
229     "Number of barrier writes");
230 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
231     &unmapped_buf_allowed, 0,
232     "Permit the use of the unmapped i/o");
233 
234 /*
235  * Lock for the non-dirty bufqueues
236  */
237 static struct mtx_padalign bqclean;
238 
239 /*
240  * Lock for the dirty queue.
241  */
242 static struct mtx_padalign bqdirty;
243 
244 /*
245  * This lock synchronizes access to bd_request.
246  */
247 static struct mtx_padalign bdlock;
248 
249 /*
250  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
251  * waitrunningbufspace().
252  */
253 static struct mtx_padalign rbreqlock;
254 
255 /*
256  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
257  */
258 static struct rwlock_padalign nblock;
259 
260 /*
261  * Lock that protects bdirtywait.
262  */
263 static struct mtx_padalign bdirtylock;
264 
265 /*
266  * Wakeup point for bufdaemon, as well as indicator of whether it is already
267  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
268  * is idling.
269  */
270 static int bd_request;
271 
272 /*
273  * Request for the buf daemon to write more buffers than is indicated by
274  * lodirtybuf.  This may be necessary to push out excess dependencies or
275  * defragment the address space where a simple count of the number of dirty
276  * buffers is insufficient to characterize the demand for flushing them.
277  */
278 static int bd_speedupreq;
279 
280 /*
281  * bogus page -- for I/O to/from partially complete buffers
282  * this is a temporary solution to the problem, but it is not
283  * really that bad.  it would be better to split the buffer
284  * for input in the case of buffers partially already in memory,
285  * but the code is intricate enough already.
286  */
287 vm_page_t bogus_page;
288 
289 /*
290  * Synchronization (sleep/wakeup) variable for active buffer space requests.
291  * Set when wait starts, cleared prior to wakeup().
292  * Used in runningbufwakeup() and waitrunningbufspace().
293  */
294 static int runningbufreq;
295 
296 /*
297  * Synchronization (sleep/wakeup) variable for buffer requests.
298  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
299  * by and/or.
300  * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
301  * getnewbuf(), and getblk().
302  */
303 static volatile int needsbuffer;
304 
305 /*
306  * Synchronization for bwillwrite() waiters.
307  */
308 static int bdirtywait;
309 
310 /*
311  * Definitions for the buffer free lists.
312  */
313 #define BUFFER_QUEUES	5	/* number of free buffer queues */
314 
315 #define QUEUE_NONE	0	/* on no queue */
316 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
317 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
318 #define QUEUE_EMPTYKVA	3	/* empty buffer headers w/KVA assignment */
319 #define QUEUE_EMPTY	4	/* empty buffer headers */
320 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
321 
322 /* Queues for free buffers with various properties */
323 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
324 #ifdef INVARIANTS
325 static int bq_len[BUFFER_QUEUES];
326 #endif
327 
328 /*
329  * Single global constant for BUF_WMESG, to avoid getting multiple references.
330  * buf_wmesg is referred from macros.
331  */
332 const char *buf_wmesg = BUF_WMESG;
333 
334 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
335 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
336 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
337 
338 static int
339 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
340 {
341 	long value;
342 	int error;
343 
344 	value = *(long *)arg1;
345 	error = sysctl_handle_long(oidp, &value, 0, req);
346 	if (error != 0 || req->newptr == NULL)
347 		return (error);
348 	mtx_lock(&rbreqlock);
349 	if (arg1 == &hirunningspace) {
350 		if (value < lorunningspace)
351 			error = EINVAL;
352 		else
353 			hirunningspace = value;
354 	} else {
355 		KASSERT(arg1 == &lorunningspace,
356 		    ("%s: unknown arg1", __func__));
357 		if (value > hirunningspace)
358 			error = EINVAL;
359 		else
360 			lorunningspace = value;
361 	}
362 	mtx_unlock(&rbreqlock);
363 	return (error);
364 }
365 
366 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
367     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
368 static int
369 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
370 {
371 	long lvalue;
372 	int ivalue;
373 
374 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
375 		return (sysctl_handle_long(oidp, arg1, arg2, req));
376 	lvalue = *(long *)arg1;
377 	if (lvalue > INT_MAX)
378 		/* On overflow, still write out a long to trigger ENOMEM. */
379 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
380 	ivalue = lvalue;
381 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
382 }
383 #endif
384 
385 /*
386  *	bqlock:
387  *
388  *	Return the appropriate queue lock based on the index.
389  */
390 static inline struct mtx *
391 bqlock(int qindex)
392 {
393 
394 	if (qindex == QUEUE_DIRTY)
395 		return (struct mtx *)(&bqdirty);
396 	return (struct mtx *)(&bqclean);
397 }
398 
399 /*
400  *	bdirtywakeup:
401  *
402  *	Wakeup any bwillwrite() waiters.
403  */
404 static void
405 bdirtywakeup(void)
406 {
407 	mtx_lock(&bdirtylock);
408 	if (bdirtywait) {
409 		bdirtywait = 0;
410 		wakeup(&bdirtywait);
411 	}
412 	mtx_unlock(&bdirtylock);
413 }
414 
415 /*
416  *	bdirtysub:
417  *
418  *	Decrement the numdirtybuffers count by one and wakeup any
419  *	threads blocked in bwillwrite().
420  */
421 static void
422 bdirtysub(void)
423 {
424 
425 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
426 	    (lodirtybuffers + hidirtybuffers) / 2)
427 		bdirtywakeup();
428 }
429 
430 /*
431  *	bdirtyadd:
432  *
433  *	Increment the numdirtybuffers count by one and wakeup the buf
434  *	daemon if needed.
435  */
436 static void
437 bdirtyadd(void)
438 {
439 
440 	/*
441 	 * Only do the wakeup once as we cross the boundary.  The
442 	 * buf daemon will keep running until the condition clears.
443 	 */
444 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
445 	    (lodirtybuffers + hidirtybuffers) / 2)
446 		bd_wakeup();
447 }
448 
449 /*
450  *	bufspacewakeup:
451  *
452  *	Called when buffer space is potentially available for recovery.
453  *	getnewbuf() will block on this flag when it is unable to free
454  *	sufficient buffer space.  Buffer space becomes recoverable when
455  *	bp's get placed back in the queues.
456  */
457 
458 static __inline void
459 bufspacewakeup(void)
460 {
461 	int need_wakeup, on;
462 
463 	/*
464 	 * If someone is waiting for BUF space, wake them up.  Even
465 	 * though we haven't freed the kva space yet, the waiting
466 	 * process will be able to now.
467 	 */
468 	rw_rlock(&nblock);
469 	for (;;) {
470 		need_wakeup = 0;
471 		on = needsbuffer;
472 		if ((on & VFS_BIO_NEED_BUFSPACE) == 0)
473 			break;
474 		need_wakeup = 1;
475 		if (atomic_cmpset_rel_int(&needsbuffer, on,
476 		    on & ~VFS_BIO_NEED_BUFSPACE))
477 			break;
478 	}
479 	if (need_wakeup)
480 		wakeup(__DEVOLATILE(void *, &needsbuffer));
481 	rw_runlock(&nblock);
482 }
483 
484 /*
485  *	runningwakeup:
486  *
487  *	Wake up processes that are waiting on asynchronous writes to fall
488  *	below lorunningspace.
489  */
490 static void
491 runningwakeup(void)
492 {
493 
494 	mtx_lock(&rbreqlock);
495 	if (runningbufreq) {
496 		runningbufreq = 0;
497 		wakeup(&runningbufreq);
498 	}
499 	mtx_unlock(&rbreqlock);
500 }
501 
502 /*
503  *	runningbufwakeup:
504  *
505  *	Decrement the outstanding write count according.
506  */
507 void
508 runningbufwakeup(struct buf *bp)
509 {
510 	long space, bspace;
511 
512 	bspace = bp->b_runningbufspace;
513 	if (bspace == 0)
514 		return;
515 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
516 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
517 	    space, bspace));
518 	bp->b_runningbufspace = 0;
519 	/*
520 	 * Only acquire the lock and wakeup on the transition from exceeding
521 	 * the threshold to falling below it.
522 	 */
523 	if (space < lorunningspace)
524 		return;
525 	if (space - bspace > lorunningspace)
526 		return;
527 	runningwakeup();
528 }
529 
530 /*
531  *	bufcountadd:
532  *
533  *	Called when a buffer has been added to one of the free queues to
534  *	account for the buffer and to wakeup anyone waiting for free buffers.
535  *	This typically occurs when large amounts of metadata are being handled
536  *	by the buffer cache ( else buffer space runs out first, usually ).
537  */
538 static __inline void
539 bufcountadd(struct buf *bp)
540 {
541 	int mask, need_wakeup, old, on;
542 
543 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
544 	    ("buf %p already counted as free", bp));
545 	bp->b_flags |= B_INFREECNT;
546 	old = atomic_fetchadd_int(&numfreebuffers, 1);
547 	KASSERT(old >= 0 && old < nbuf,
548 	    ("numfreebuffers climbed to %d", old + 1));
549 	mask = VFS_BIO_NEED_ANY;
550 	if (numfreebuffers >= hifreebuffers)
551 		mask |= VFS_BIO_NEED_FREE;
552 	rw_rlock(&nblock);
553 	for (;;) {
554 		need_wakeup = 0;
555 		on = needsbuffer;
556 		if (on == 0)
557 			break;
558 		need_wakeup = 1;
559 		if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask))
560 			break;
561 	}
562 	if (need_wakeup)
563 		wakeup(__DEVOLATILE(void *, &needsbuffer));
564 	rw_runlock(&nblock);
565 }
566 
567 /*
568  *	bufcountsub:
569  *
570  *	Decrement the numfreebuffers count as needed.
571  */
572 static void
573 bufcountsub(struct buf *bp)
574 {
575 	int old;
576 
577 	/*
578 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
579 	 * delayed-write, the buffer was free and we must decrement
580 	 * numfreebuffers.
581 	 */
582 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
583 		KASSERT((bp->b_flags & B_INFREECNT) != 0,
584 		    ("buf %p not counted in numfreebuffers", bp));
585 		bp->b_flags &= ~B_INFREECNT;
586 		old = atomic_fetchadd_int(&numfreebuffers, -1);
587 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
588 	}
589 }
590 
591 /*
592  *	waitrunningbufspace()
593  *
594  *	runningbufspace is a measure of the amount of I/O currently
595  *	running.  This routine is used in async-write situations to
596  *	prevent creating huge backups of pending writes to a device.
597  *	Only asynchronous writes are governed by this function.
598  *
599  *	This does NOT turn an async write into a sync write.  It waits
600  *	for earlier writes to complete and generally returns before the
601  *	caller's write has reached the device.
602  */
603 void
604 waitrunningbufspace(void)
605 {
606 
607 	mtx_lock(&rbreqlock);
608 	while (runningbufspace > hirunningspace) {
609 		runningbufreq = 1;
610 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
611 	}
612 	mtx_unlock(&rbreqlock);
613 }
614 
615 
616 /*
617  *	vfs_buf_test_cache:
618  *
619  *	Called when a buffer is extended.  This function clears the B_CACHE
620  *	bit if the newly extended portion of the buffer does not contain
621  *	valid data.
622  */
623 static __inline
624 void
625 vfs_buf_test_cache(struct buf *bp,
626 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
627 		  vm_page_t m)
628 {
629 
630 	VM_OBJECT_ASSERT_LOCKED(m->object);
631 	if (bp->b_flags & B_CACHE) {
632 		int base = (foff + off) & PAGE_MASK;
633 		if (vm_page_is_valid(m, base, size) == 0)
634 			bp->b_flags &= ~B_CACHE;
635 	}
636 }
637 
638 /* Wake up the buffer daemon if necessary */
639 static __inline void
640 bd_wakeup(void)
641 {
642 
643 	mtx_lock(&bdlock);
644 	if (bd_request == 0) {
645 		bd_request = 1;
646 		wakeup(&bd_request);
647 	}
648 	mtx_unlock(&bdlock);
649 }
650 
651 /*
652  * bd_speedup - speedup the buffer cache flushing code
653  */
654 void
655 bd_speedup(void)
656 {
657 	int needwake;
658 
659 	mtx_lock(&bdlock);
660 	needwake = 0;
661 	if (bd_speedupreq == 0 || bd_request == 0)
662 		needwake = 1;
663 	bd_speedupreq = 1;
664 	bd_request = 1;
665 	if (needwake)
666 		wakeup(&bd_request);
667 	mtx_unlock(&bdlock);
668 }
669 
670 #ifndef NSWBUF_MIN
671 #define	NSWBUF_MIN	16
672 #endif
673 
674 #ifdef __i386__
675 #define	TRANSIENT_DENOM	5
676 #else
677 #define	TRANSIENT_DENOM 10
678 #endif
679 
680 /*
681  * Calculating buffer cache scaling values and reserve space for buffer
682  * headers.  This is called during low level kernel initialization and
683  * may be called more then once.  We CANNOT write to the memory area
684  * being reserved at this time.
685  */
686 caddr_t
687 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
688 {
689 	int tuned_nbuf;
690 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
691 
692 	/*
693 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
694 	 * PAGE_SIZE is >= 1K)
695 	 */
696 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
697 
698 	/*
699 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
700 	 * For the first 64MB of ram nominally allocate sufficient buffers to
701 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
702 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
703 	 * the buffer cache we limit the eventual kva reservation to
704 	 * maxbcache bytes.
705 	 *
706 	 * factor represents the 1/4 x ram conversion.
707 	 */
708 	if (nbuf == 0) {
709 		int factor = 4 * BKVASIZE / 1024;
710 
711 		nbuf = 50;
712 		if (physmem_est > 4096)
713 			nbuf += min((physmem_est - 4096) / factor,
714 			    65536 / factor);
715 		if (physmem_est > 65536)
716 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
717 			    32 * 1024 * 1024 / (factor * 5));
718 
719 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
720 			nbuf = maxbcache / BKVASIZE;
721 		tuned_nbuf = 1;
722 	} else
723 		tuned_nbuf = 0;
724 
725 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
726 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
727 	if (nbuf > maxbuf) {
728 		if (!tuned_nbuf)
729 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
730 			    maxbuf);
731 		nbuf = maxbuf;
732 	}
733 
734 	/*
735 	 * Ideal allocation size for the transient bio submap is 10%
736 	 * of the maximal space buffer map.  This roughly corresponds
737 	 * to the amount of the buffer mapped for typical UFS load.
738 	 *
739 	 * Clip the buffer map to reserve space for the transient
740 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
741 	 * maximum buffer map extent on the platform.
742 	 *
743 	 * The fall-back to the maxbuf in case of maxbcache unset,
744 	 * allows to not trim the buffer KVA for the architectures
745 	 * with ample KVA space.
746 	 */
747 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
748 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
749 		buf_sz = (long)nbuf * BKVASIZE;
750 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
751 		    (TRANSIENT_DENOM - 1)) {
752 			/*
753 			 * There is more KVA than memory.  Do not
754 			 * adjust buffer map size, and assign the rest
755 			 * of maxbuf to transient map.
756 			 */
757 			biotmap_sz = maxbuf_sz - buf_sz;
758 		} else {
759 			/*
760 			 * Buffer map spans all KVA we could afford on
761 			 * this platform.  Give 10% (20% on i386) of
762 			 * the buffer map to the transient bio map.
763 			 */
764 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
765 			buf_sz -= biotmap_sz;
766 		}
767 		if (biotmap_sz / INT_MAX > MAXPHYS)
768 			bio_transient_maxcnt = INT_MAX;
769 		else
770 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
771 		/*
772 		 * Artifically limit to 1024 simultaneous in-flight I/Os
773 		 * using the transient mapping.
774 		 */
775 		if (bio_transient_maxcnt > 1024)
776 			bio_transient_maxcnt = 1024;
777 		if (tuned_nbuf)
778 			nbuf = buf_sz / BKVASIZE;
779 	}
780 
781 	/*
782 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
783 	 * We have no less then 16 and no more then 256.
784 	 */
785 	nswbuf = min(nbuf / 4, 256);
786 	TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
787 	if (nswbuf < NSWBUF_MIN)
788 		nswbuf = NSWBUF_MIN;
789 
790 	/*
791 	 * Reserve space for the buffer cache buffers
792 	 */
793 	swbuf = (void *)v;
794 	v = (caddr_t)(swbuf + nswbuf);
795 	buf = (void *)v;
796 	v = (caddr_t)(buf + nbuf);
797 
798 	return(v);
799 }
800 
801 /* Initialize the buffer subsystem.  Called before use of any buffers. */
802 void
803 bufinit(void)
804 {
805 	struct buf *bp;
806 	int i;
807 
808 	mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
809 	mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
810 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
811 	rw_init(&nblock, "needsbuffer lock");
812 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
813 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
814 
815 	/* next, make a null set of free lists */
816 	for (i = 0; i < BUFFER_QUEUES; i++)
817 		TAILQ_INIT(&bufqueues[i]);
818 
819 	/* finally, initialize each buffer header and stick on empty q */
820 	for (i = 0; i < nbuf; i++) {
821 		bp = &buf[i];
822 		bzero(bp, sizeof *bp);
823 		bp->b_flags = B_INVAL | B_INFREECNT;
824 		bp->b_rcred = NOCRED;
825 		bp->b_wcred = NOCRED;
826 		bp->b_qindex = QUEUE_EMPTY;
827 		bp->b_xflags = 0;
828 		LIST_INIT(&bp->b_dep);
829 		BUF_LOCKINIT(bp);
830 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
831 #ifdef INVARIANTS
832 		bq_len[QUEUE_EMPTY]++;
833 #endif
834 	}
835 
836 	/*
837 	 * maxbufspace is the absolute maximum amount of buffer space we are
838 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
839 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
840 	 * used by most other processes.  The differential is required to
841 	 * ensure that buf_daemon is able to run when other processes might
842 	 * be blocked waiting for buffer space.
843 	 *
844 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
845 	 * this may result in KVM fragmentation which is not handled optimally
846 	 * by the system.
847 	 */
848 	maxbufspace = (long)nbuf * BKVASIZE;
849 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
850 	lobufspace = hibufspace - MAXBSIZE;
851 
852 	/*
853 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
854 	 * arbitrarily and may need further tuning. It corresponds to
855 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
856 	 * which fits with many RAID controllers' tagged queuing limits.
857 	 * The lower 1 MiB limit is the historical upper limit for
858 	 * hirunningspace.
859 	 */
860 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
861 	    16 * 1024 * 1024), 1024 * 1024);
862 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
863 
864 /*
865  * Limit the amount of malloc memory since it is wired permanently into
866  * the kernel space.  Even though this is accounted for in the buffer
867  * allocation, we don't want the malloced region to grow uncontrolled.
868  * The malloc scheme improves memory utilization significantly on average
869  * (small) directories.
870  */
871 	maxbufmallocspace = hibufspace / 20;
872 
873 /*
874  * Reduce the chance of a deadlock occuring by limiting the number
875  * of delayed-write dirty buffers we allow to stack up.
876  */
877 	hidirtybuffers = nbuf / 4 + 20;
878 	dirtybufthresh = hidirtybuffers * 9 / 10;
879 	numdirtybuffers = 0;
880 /*
881  * To support extreme low-memory systems, make sure hidirtybuffers cannot
882  * eat up all available buffer space.  This occurs when our minimum cannot
883  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
884  * BKVASIZE'd buffers.
885  */
886 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
887 		hidirtybuffers >>= 1;
888 	}
889 	lodirtybuffers = hidirtybuffers / 2;
890 
891 /*
892  * Try to keep the number of free buffers in the specified range,
893  * and give special processes (e.g. like buf_daemon) access to an
894  * emergency reserve.
895  */
896 	lofreebuffers = nbuf / 18 + 5;
897 	hifreebuffers = 2 * lofreebuffers;
898 	numfreebuffers = nbuf;
899 
900 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
901 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
902 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
903 }
904 
905 #ifdef INVARIANTS
906 static inline void
907 vfs_buf_check_mapped(struct buf *bp)
908 {
909 
910 	KASSERT((bp->b_flags & B_UNMAPPED) == 0,
911 	    ("mapped buf %p %x", bp, bp->b_flags));
912 	KASSERT(bp->b_kvabase != unmapped_buf,
913 	    ("mapped buf: b_kvabase was not updated %p", bp));
914 	KASSERT(bp->b_data != unmapped_buf,
915 	    ("mapped buf: b_data was not updated %p", bp));
916 }
917 
918 static inline void
919 vfs_buf_check_unmapped(struct buf *bp)
920 {
921 
922 	KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
923 	    ("unmapped buf %p %x", bp, bp->b_flags));
924 	KASSERT(bp->b_kvabase == unmapped_buf,
925 	    ("unmapped buf: corrupted b_kvabase %p", bp));
926 	KASSERT(bp->b_data == unmapped_buf,
927 	    ("unmapped buf: corrupted b_data %p", bp));
928 }
929 
930 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
931 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
932 #else
933 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
934 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
935 #endif
936 
937 static void
938 bpmap_qenter(struct buf *bp)
939 {
940 
941 	BUF_CHECK_MAPPED(bp);
942 
943 	/*
944 	 * bp->b_data is relative to bp->b_offset, but
945 	 * bp->b_offset may be offset into the first page.
946 	 */
947 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
948 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
949 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
950 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
951 }
952 
953 /*
954  * bfreekva() - free the kva allocation for a buffer.
955  *
956  *	Since this call frees up buffer space, we call bufspacewakeup().
957  */
958 static void
959 bfreekva(struct buf *bp)
960 {
961 
962 	if (bp->b_kvasize == 0)
963 		return;
964 
965 	atomic_add_int(&buffreekvacnt, 1);
966 	atomic_subtract_long(&bufspace, bp->b_kvasize);
967 	if ((bp->b_flags & B_UNMAPPED) == 0) {
968 		BUF_CHECK_MAPPED(bp);
969 		vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
970 		    bp->b_kvasize);
971 	} else {
972 		BUF_CHECK_UNMAPPED(bp);
973 		if ((bp->b_flags & B_KVAALLOC) != 0) {
974 			vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
975 			    bp->b_kvasize);
976 		}
977 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
978 		bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
979 	}
980 	bp->b_kvasize = 0;
981 	bufspacewakeup();
982 }
983 
984 /*
985  *	binsfree:
986  *
987  *	Insert the buffer into the appropriate free list.
988  */
989 static void
990 binsfree(struct buf *bp, int qindex)
991 {
992 	struct mtx *olock, *nlock;
993 
994 	BUF_ASSERT_XLOCKED(bp);
995 
996 	olock = bqlock(bp->b_qindex);
997 	nlock = bqlock(qindex);
998 	mtx_lock(olock);
999 	/* Handle delayed bremfree() processing. */
1000 	if (bp->b_flags & B_REMFREE)
1001 		bremfreel(bp);
1002 
1003 	if (bp->b_qindex != QUEUE_NONE)
1004 		panic("binsfree: free buffer onto another queue???");
1005 
1006 	bp->b_qindex = qindex;
1007 	if (olock != nlock) {
1008 		mtx_unlock(olock);
1009 		mtx_lock(nlock);
1010 	}
1011 	if (bp->b_flags & B_AGE)
1012 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1013 	else
1014 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1015 #ifdef INVARIANTS
1016 	bq_len[bp->b_qindex]++;
1017 #endif
1018 	mtx_unlock(nlock);
1019 
1020 	/*
1021 	 * Something we can maybe free or reuse.
1022 	 */
1023 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1024 		bufspacewakeup();
1025 
1026 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1027 		bufcountadd(bp);
1028 }
1029 
1030 /*
1031  *	bremfree:
1032  *
1033  *	Mark the buffer for removal from the appropriate free list.
1034  *
1035  */
1036 void
1037 bremfree(struct buf *bp)
1038 {
1039 
1040 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1041 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1042 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1043 	KASSERT(bp->b_qindex != QUEUE_NONE,
1044 	    ("bremfree: buffer %p not on a queue.", bp));
1045 	BUF_ASSERT_XLOCKED(bp);
1046 
1047 	bp->b_flags |= B_REMFREE;
1048 	bufcountsub(bp);
1049 }
1050 
1051 /*
1052  *	bremfreef:
1053  *
1054  *	Force an immediate removal from a free list.  Used only in nfs when
1055  *	it abuses the b_freelist pointer.
1056  */
1057 void
1058 bremfreef(struct buf *bp)
1059 {
1060 	struct mtx *qlock;
1061 
1062 	qlock = bqlock(bp->b_qindex);
1063 	mtx_lock(qlock);
1064 	bremfreel(bp);
1065 	mtx_unlock(qlock);
1066 }
1067 
1068 /*
1069  *	bremfreel:
1070  *
1071  *	Removes a buffer from the free list, must be called with the
1072  *	correct qlock held.
1073  */
1074 static void
1075 bremfreel(struct buf *bp)
1076 {
1077 
1078 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1079 	    bp, bp->b_vp, bp->b_flags);
1080 	KASSERT(bp->b_qindex != QUEUE_NONE,
1081 	    ("bremfreel: buffer %p not on a queue.", bp));
1082 	BUF_ASSERT_XLOCKED(bp);
1083 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1084 
1085 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1086 #ifdef INVARIANTS
1087 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1088 	    bp->b_qindex));
1089 	bq_len[bp->b_qindex]--;
1090 #endif
1091 	bp->b_qindex = QUEUE_NONE;
1092 	/*
1093 	 * If this was a delayed bremfree() we only need to remove the buffer
1094 	 * from the queue and return the stats are already done.
1095 	 */
1096 	if (bp->b_flags & B_REMFREE) {
1097 		bp->b_flags &= ~B_REMFREE;
1098 		return;
1099 	}
1100 	bufcountsub(bp);
1101 }
1102 
1103 /*
1104  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1105  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1106  * the buffer is valid and we do not have to do anything.
1107  */
1108 void
1109 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1110     int cnt, struct ucred * cred)
1111 {
1112 	struct buf *rabp;
1113 	int i;
1114 
1115 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1116 		if (inmem(vp, *rablkno))
1117 			continue;
1118 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1119 
1120 		if ((rabp->b_flags & B_CACHE) == 0) {
1121 			if (!TD_IS_IDLETHREAD(curthread))
1122 				curthread->td_ru.ru_inblock++;
1123 			rabp->b_flags |= B_ASYNC;
1124 			rabp->b_flags &= ~B_INVAL;
1125 			rabp->b_ioflags &= ~BIO_ERROR;
1126 			rabp->b_iocmd = BIO_READ;
1127 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1128 				rabp->b_rcred = crhold(cred);
1129 			vfs_busy_pages(rabp, 0);
1130 			BUF_KERNPROC(rabp);
1131 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1132 			bstrategy(rabp);
1133 		} else {
1134 			brelse(rabp);
1135 		}
1136 	}
1137 }
1138 
1139 /*
1140  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1141  *
1142  * Get a buffer with the specified data.  Look in the cache first.  We
1143  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1144  * is set, the buffer is valid and we do not have to do anything, see
1145  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1146  */
1147 int
1148 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1149     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1150 {
1151 	struct buf *bp;
1152 	int rv = 0, readwait = 0;
1153 
1154 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1155 	/*
1156 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1157 	 */
1158 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1159 	if (bp == NULL)
1160 		return (EBUSY);
1161 
1162 	/* if not found in cache, do some I/O */
1163 	if ((bp->b_flags & B_CACHE) == 0) {
1164 		if (!TD_IS_IDLETHREAD(curthread))
1165 			curthread->td_ru.ru_inblock++;
1166 		bp->b_iocmd = BIO_READ;
1167 		bp->b_flags &= ~B_INVAL;
1168 		bp->b_ioflags &= ~BIO_ERROR;
1169 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1170 			bp->b_rcred = crhold(cred);
1171 		vfs_busy_pages(bp, 0);
1172 		bp->b_iooffset = dbtob(bp->b_blkno);
1173 		bstrategy(bp);
1174 		++readwait;
1175 	}
1176 
1177 	breada(vp, rablkno, rabsize, cnt, cred);
1178 
1179 	if (readwait) {
1180 		rv = bufwait(bp);
1181 	}
1182 	return (rv);
1183 }
1184 
1185 /*
1186  * Write, release buffer on completion.  (Done by iodone
1187  * if async).  Do not bother writing anything if the buffer
1188  * is invalid.
1189  *
1190  * Note that we set B_CACHE here, indicating that buffer is
1191  * fully valid and thus cacheable.  This is true even of NFS
1192  * now so we set it generally.  This could be set either here
1193  * or in biodone() since the I/O is synchronous.  We put it
1194  * here.
1195  */
1196 int
1197 bufwrite(struct buf *bp)
1198 {
1199 	int oldflags;
1200 	struct vnode *vp;
1201 	long space;
1202 	int vp_md;
1203 
1204 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1205 	if (bp->b_flags & B_INVAL) {
1206 		brelse(bp);
1207 		return (0);
1208 	}
1209 
1210 	if (bp->b_flags & B_BARRIER)
1211 		barrierwrites++;
1212 
1213 	oldflags = bp->b_flags;
1214 
1215 	BUF_ASSERT_HELD(bp);
1216 
1217 	if (bp->b_pin_count > 0)
1218 		bunpin_wait(bp);
1219 
1220 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1221 	    ("FFS background buffer should not get here %p", bp));
1222 
1223 	vp = bp->b_vp;
1224 	if (vp)
1225 		vp_md = vp->v_vflag & VV_MD;
1226 	else
1227 		vp_md = 0;
1228 
1229 	/*
1230 	 * Mark the buffer clean.  Increment the bufobj write count
1231 	 * before bundirty() call, to prevent other thread from seeing
1232 	 * empty dirty list and zero counter for writes in progress,
1233 	 * falsely indicating that the bufobj is clean.
1234 	 */
1235 	bufobj_wref(bp->b_bufobj);
1236 	bundirty(bp);
1237 
1238 	bp->b_flags &= ~B_DONE;
1239 	bp->b_ioflags &= ~BIO_ERROR;
1240 	bp->b_flags |= B_CACHE;
1241 	bp->b_iocmd = BIO_WRITE;
1242 
1243 	vfs_busy_pages(bp, 1);
1244 
1245 	/*
1246 	 * Normal bwrites pipeline writes
1247 	 */
1248 	bp->b_runningbufspace = bp->b_bufsize;
1249 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1250 
1251 	if (!TD_IS_IDLETHREAD(curthread))
1252 		curthread->td_ru.ru_oublock++;
1253 	if (oldflags & B_ASYNC)
1254 		BUF_KERNPROC(bp);
1255 	bp->b_iooffset = dbtob(bp->b_blkno);
1256 	bstrategy(bp);
1257 
1258 	if ((oldflags & B_ASYNC) == 0) {
1259 		int rtval = bufwait(bp);
1260 		brelse(bp);
1261 		return (rtval);
1262 	} else if (space > hirunningspace) {
1263 		/*
1264 		 * don't allow the async write to saturate the I/O
1265 		 * system.  We will not deadlock here because
1266 		 * we are blocking waiting for I/O that is already in-progress
1267 		 * to complete. We do not block here if it is the update
1268 		 * or syncer daemon trying to clean up as that can lead
1269 		 * to deadlock.
1270 		 */
1271 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1272 			waitrunningbufspace();
1273 	}
1274 
1275 	return (0);
1276 }
1277 
1278 void
1279 bufbdflush(struct bufobj *bo, struct buf *bp)
1280 {
1281 	struct buf *nbp;
1282 
1283 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1284 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1285 		altbufferflushes++;
1286 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1287 		BO_LOCK(bo);
1288 		/*
1289 		 * Try to find a buffer to flush.
1290 		 */
1291 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1292 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1293 			    BUF_LOCK(nbp,
1294 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1295 				continue;
1296 			if (bp == nbp)
1297 				panic("bdwrite: found ourselves");
1298 			BO_UNLOCK(bo);
1299 			/* Don't countdeps with the bo lock held. */
1300 			if (buf_countdeps(nbp, 0)) {
1301 				BO_LOCK(bo);
1302 				BUF_UNLOCK(nbp);
1303 				continue;
1304 			}
1305 			if (nbp->b_flags & B_CLUSTEROK) {
1306 				vfs_bio_awrite(nbp);
1307 			} else {
1308 				bremfree(nbp);
1309 				bawrite(nbp);
1310 			}
1311 			dirtybufferflushes++;
1312 			break;
1313 		}
1314 		if (nbp == NULL)
1315 			BO_UNLOCK(bo);
1316 	}
1317 }
1318 
1319 /*
1320  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1321  * anything if the buffer is marked invalid.
1322  *
1323  * Note that since the buffer must be completely valid, we can safely
1324  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1325  * biodone() in order to prevent getblk from writing the buffer
1326  * out synchronously.
1327  */
1328 void
1329 bdwrite(struct buf *bp)
1330 {
1331 	struct thread *td = curthread;
1332 	struct vnode *vp;
1333 	struct bufobj *bo;
1334 
1335 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1336 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1337 	KASSERT((bp->b_flags & B_BARRIER) == 0,
1338 	    ("Barrier request in delayed write %p", bp));
1339 	BUF_ASSERT_HELD(bp);
1340 
1341 	if (bp->b_flags & B_INVAL) {
1342 		brelse(bp);
1343 		return;
1344 	}
1345 
1346 	/*
1347 	 * If we have too many dirty buffers, don't create any more.
1348 	 * If we are wildly over our limit, then force a complete
1349 	 * cleanup. Otherwise, just keep the situation from getting
1350 	 * out of control. Note that we have to avoid a recursive
1351 	 * disaster and not try to clean up after our own cleanup!
1352 	 */
1353 	vp = bp->b_vp;
1354 	bo = bp->b_bufobj;
1355 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1356 		td->td_pflags |= TDP_INBDFLUSH;
1357 		BO_BDFLUSH(bo, bp);
1358 		td->td_pflags &= ~TDP_INBDFLUSH;
1359 	} else
1360 		recursiveflushes++;
1361 
1362 	bdirty(bp);
1363 	/*
1364 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1365 	 * true even of NFS now.
1366 	 */
1367 	bp->b_flags |= B_CACHE;
1368 
1369 	/*
1370 	 * This bmap keeps the system from needing to do the bmap later,
1371 	 * perhaps when the system is attempting to do a sync.  Since it
1372 	 * is likely that the indirect block -- or whatever other datastructure
1373 	 * that the filesystem needs is still in memory now, it is a good
1374 	 * thing to do this.  Note also, that if the pageout daemon is
1375 	 * requesting a sync -- there might not be enough memory to do
1376 	 * the bmap then...  So, this is important to do.
1377 	 */
1378 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1379 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1380 	}
1381 
1382 	/*
1383 	 * Set the *dirty* buffer range based upon the VM system dirty
1384 	 * pages.
1385 	 *
1386 	 * Mark the buffer pages as clean.  We need to do this here to
1387 	 * satisfy the vnode_pager and the pageout daemon, so that it
1388 	 * thinks that the pages have been "cleaned".  Note that since
1389 	 * the pages are in a delayed write buffer -- the VFS layer
1390 	 * "will" see that the pages get written out on the next sync,
1391 	 * or perhaps the cluster will be completed.
1392 	 */
1393 	vfs_clean_pages_dirty_buf(bp);
1394 	bqrelse(bp);
1395 
1396 	/*
1397 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1398 	 * due to the softdep code.
1399 	 */
1400 }
1401 
1402 /*
1403  *	bdirty:
1404  *
1405  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1406  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1407  *	itself to properly update it in the dirty/clean lists.  We mark it
1408  *	B_DONE to ensure that any asynchronization of the buffer properly
1409  *	clears B_DONE ( else a panic will occur later ).
1410  *
1411  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1412  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1413  *	should only be called if the buffer is known-good.
1414  *
1415  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1416  *	count.
1417  *
1418  *	The buffer must be on QUEUE_NONE.
1419  */
1420 void
1421 bdirty(struct buf *bp)
1422 {
1423 
1424 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1425 	    bp, bp->b_vp, bp->b_flags);
1426 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1427 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1428 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1429 	BUF_ASSERT_HELD(bp);
1430 	bp->b_flags &= ~(B_RELBUF);
1431 	bp->b_iocmd = BIO_WRITE;
1432 
1433 	if ((bp->b_flags & B_DELWRI) == 0) {
1434 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1435 		reassignbuf(bp);
1436 		bdirtyadd();
1437 	}
1438 }
1439 
1440 /*
1441  *	bundirty:
1442  *
1443  *	Clear B_DELWRI for buffer.
1444  *
1445  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1446  *	count.
1447  *
1448  *	The buffer must be on QUEUE_NONE.
1449  */
1450 
1451 void
1452 bundirty(struct buf *bp)
1453 {
1454 
1455 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1456 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1457 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1458 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1459 	BUF_ASSERT_HELD(bp);
1460 
1461 	if (bp->b_flags & B_DELWRI) {
1462 		bp->b_flags &= ~B_DELWRI;
1463 		reassignbuf(bp);
1464 		bdirtysub();
1465 	}
1466 	/*
1467 	 * Since it is now being written, we can clear its deferred write flag.
1468 	 */
1469 	bp->b_flags &= ~B_DEFERRED;
1470 }
1471 
1472 /*
1473  *	bawrite:
1474  *
1475  *	Asynchronous write.  Start output on a buffer, but do not wait for
1476  *	it to complete.  The buffer is released when the output completes.
1477  *
1478  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1479  *	B_INVAL buffers.  Not us.
1480  */
1481 void
1482 bawrite(struct buf *bp)
1483 {
1484 
1485 	bp->b_flags |= B_ASYNC;
1486 	(void) bwrite(bp);
1487 }
1488 
1489 /*
1490  *	babarrierwrite:
1491  *
1492  *	Asynchronous barrier write.  Start output on a buffer, but do not
1493  *	wait for it to complete.  Place a write barrier after this write so
1494  *	that this buffer and all buffers written before it are committed to
1495  *	the disk before any buffers written after this write are committed
1496  *	to the disk.  The buffer is released when the output completes.
1497  */
1498 void
1499 babarrierwrite(struct buf *bp)
1500 {
1501 
1502 	bp->b_flags |= B_ASYNC | B_BARRIER;
1503 	(void) bwrite(bp);
1504 }
1505 
1506 /*
1507  *	bbarrierwrite:
1508  *
1509  *	Synchronous barrier write.  Start output on a buffer and wait for
1510  *	it to complete.  Place a write barrier after this write so that
1511  *	this buffer and all buffers written before it are committed to
1512  *	the disk before any buffers written after this write are committed
1513  *	to the disk.  The buffer is released when the output completes.
1514  */
1515 int
1516 bbarrierwrite(struct buf *bp)
1517 {
1518 
1519 	bp->b_flags |= B_BARRIER;
1520 	return (bwrite(bp));
1521 }
1522 
1523 /*
1524  *	bwillwrite:
1525  *
1526  *	Called prior to the locking of any vnodes when we are expecting to
1527  *	write.  We do not want to starve the buffer cache with too many
1528  *	dirty buffers so we block here.  By blocking prior to the locking
1529  *	of any vnodes we attempt to avoid the situation where a locked vnode
1530  *	prevents the various system daemons from flushing related buffers.
1531  */
1532 void
1533 bwillwrite(void)
1534 {
1535 
1536 	if (numdirtybuffers >= hidirtybuffers) {
1537 		mtx_lock(&bdirtylock);
1538 		while (numdirtybuffers >= hidirtybuffers) {
1539 			bdirtywait = 1;
1540 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1541 			    "flswai", 0);
1542 		}
1543 		mtx_unlock(&bdirtylock);
1544 	}
1545 }
1546 
1547 /*
1548  * Return true if we have too many dirty buffers.
1549  */
1550 int
1551 buf_dirty_count_severe(void)
1552 {
1553 
1554 	return(numdirtybuffers >= hidirtybuffers);
1555 }
1556 
1557 static __noinline int
1558 buf_vm_page_count_severe(void)
1559 {
1560 
1561 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1562 
1563 	return vm_page_count_severe();
1564 }
1565 
1566 /*
1567  *	brelse:
1568  *
1569  *	Release a busy buffer and, if requested, free its resources.  The
1570  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1571  *	to be accessed later as a cache entity or reused for other purposes.
1572  */
1573 void
1574 brelse(struct buf *bp)
1575 {
1576 	int qindex;
1577 
1578 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1579 	    bp, bp->b_vp, bp->b_flags);
1580 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1581 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1582 
1583 	if (BUF_LOCKRECURSED(bp)) {
1584 		/*
1585 		 * Do not process, in particular, do not handle the
1586 		 * B_INVAL/B_RELBUF and do not release to free list.
1587 		 */
1588 		BUF_UNLOCK(bp);
1589 		return;
1590 	}
1591 
1592 	if (bp->b_flags & B_MANAGED) {
1593 		bqrelse(bp);
1594 		return;
1595 	}
1596 
1597 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1598 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1599 		/*
1600 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1601 		 * pages from being scrapped.  If the error is anything
1602 		 * other than an I/O error (EIO), assume that retrying
1603 		 * is futile.
1604 		 */
1605 		bp->b_ioflags &= ~BIO_ERROR;
1606 		bdirty(bp);
1607 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1608 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1609 		/*
1610 		 * Either a failed I/O or we were asked to free or not
1611 		 * cache the buffer.
1612 		 */
1613 		bp->b_flags |= B_INVAL;
1614 		if (!LIST_EMPTY(&bp->b_dep))
1615 			buf_deallocate(bp);
1616 		if (bp->b_flags & B_DELWRI)
1617 			bdirtysub();
1618 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1619 		if ((bp->b_flags & B_VMIO) == 0) {
1620 			if (bp->b_bufsize)
1621 				allocbuf(bp, 0);
1622 			if (bp->b_vp)
1623 				brelvp(bp);
1624 		}
1625 	}
1626 
1627 	/*
1628 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1629 	 * is called with B_DELWRI set, the underlying pages may wind up
1630 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1631 	 * because pages associated with a B_DELWRI bp are marked clean.
1632 	 *
1633 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1634 	 * if B_DELWRI is set.
1635 	 *
1636 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1637 	 * on pages to return pages to the VM page queues.
1638 	 */
1639 	if (bp->b_flags & B_DELWRI)
1640 		bp->b_flags &= ~B_RELBUF;
1641 	else if (buf_vm_page_count_severe()) {
1642 		/*
1643 		 * BKGRDINPROG can only be set with the buf and bufobj
1644 		 * locks both held.  We tolerate a race to clear it here.
1645 		 */
1646 		if (!(bp->b_vflags & BV_BKGRDINPROG))
1647 			bp->b_flags |= B_RELBUF;
1648 	}
1649 
1650 	/*
1651 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1652 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1653 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1654 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1655 	 *
1656 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1657 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1658 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1659 	 *
1660 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1661 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1662 	 * the commit state and we cannot afford to lose the buffer. If the
1663 	 * buffer has a background write in progress, we need to keep it
1664 	 * around to prevent it from being reconstituted and starting a second
1665 	 * background write.
1666 	 */
1667 	if ((bp->b_flags & B_VMIO)
1668 	    && !(bp->b_vp->v_mount != NULL &&
1669 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1670 		 !vn_isdisk(bp->b_vp, NULL) &&
1671 		 (bp->b_flags & B_DELWRI))
1672 	    ) {
1673 
1674 		int i, j, resid;
1675 		vm_page_t m;
1676 		off_t foff;
1677 		vm_pindex_t poff;
1678 		vm_object_t obj;
1679 
1680 		obj = bp->b_bufobj->bo_object;
1681 
1682 		/*
1683 		 * Get the base offset and length of the buffer.  Note that
1684 		 * in the VMIO case if the buffer block size is not
1685 		 * page-aligned then b_data pointer may not be page-aligned.
1686 		 * But our b_pages[] array *IS* page aligned.
1687 		 *
1688 		 * block sizes less then DEV_BSIZE (usually 512) are not
1689 		 * supported due to the page granularity bits (m->valid,
1690 		 * m->dirty, etc...).
1691 		 *
1692 		 * See man buf(9) for more information
1693 		 */
1694 		resid = bp->b_bufsize;
1695 		foff = bp->b_offset;
1696 		for (i = 0; i < bp->b_npages; i++) {
1697 			int had_bogus = 0;
1698 
1699 			m = bp->b_pages[i];
1700 
1701 			/*
1702 			 * If we hit a bogus page, fixup *all* the bogus pages
1703 			 * now.
1704 			 */
1705 			if (m == bogus_page) {
1706 				poff = OFF_TO_IDX(bp->b_offset);
1707 				had_bogus = 1;
1708 
1709 				VM_OBJECT_RLOCK(obj);
1710 				for (j = i; j < bp->b_npages; j++) {
1711 					vm_page_t mtmp;
1712 					mtmp = bp->b_pages[j];
1713 					if (mtmp == bogus_page) {
1714 						mtmp = vm_page_lookup(obj, poff + j);
1715 						if (!mtmp) {
1716 							panic("brelse: page missing\n");
1717 						}
1718 						bp->b_pages[j] = mtmp;
1719 					}
1720 				}
1721 				VM_OBJECT_RUNLOCK(obj);
1722 
1723 				if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1724 					BUF_CHECK_MAPPED(bp);
1725 					pmap_qenter(
1726 					    trunc_page((vm_offset_t)bp->b_data),
1727 					    bp->b_pages, bp->b_npages);
1728 				}
1729 				m = bp->b_pages[i];
1730 			}
1731 			if ((bp->b_flags & B_NOCACHE) ||
1732 			    (bp->b_ioflags & BIO_ERROR &&
1733 			     bp->b_iocmd == BIO_READ)) {
1734 				int poffset = foff & PAGE_MASK;
1735 				int presid = resid > (PAGE_SIZE - poffset) ?
1736 					(PAGE_SIZE - poffset) : resid;
1737 
1738 				KASSERT(presid >= 0, ("brelse: extra page"));
1739 				VM_OBJECT_WLOCK(obj);
1740 				while (vm_page_xbusied(m)) {
1741 					vm_page_lock(m);
1742 					VM_OBJECT_WUNLOCK(obj);
1743 					vm_page_busy_sleep(m, "mbncsh");
1744 					VM_OBJECT_WLOCK(obj);
1745 				}
1746 				if (pmap_page_wired_mappings(m) == 0)
1747 					vm_page_set_invalid(m, poffset, presid);
1748 				VM_OBJECT_WUNLOCK(obj);
1749 				if (had_bogus)
1750 					printf("avoided corruption bug in bogus_page/brelse code\n");
1751 			}
1752 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1753 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1754 		}
1755 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1756 			vfs_vmio_release(bp);
1757 
1758 	} else if (bp->b_flags & B_VMIO) {
1759 
1760 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1761 			vfs_vmio_release(bp);
1762 		}
1763 
1764 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1765 		if (bp->b_bufsize != 0)
1766 			allocbuf(bp, 0);
1767 		if (bp->b_vp != NULL)
1768 			brelvp(bp);
1769 	}
1770 
1771 	/*
1772 	 * If the buffer has junk contents signal it and eventually
1773 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1774 	 * doesn't find it.
1775 	 */
1776 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1777 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1778 		bp->b_flags |= B_INVAL;
1779 	if (bp->b_flags & B_INVAL) {
1780 		if (bp->b_flags & B_DELWRI)
1781 			bundirty(bp);
1782 		if (bp->b_vp)
1783 			brelvp(bp);
1784 	}
1785 
1786 	/* buffers with no memory */
1787 	if (bp->b_bufsize == 0) {
1788 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1789 		if (bp->b_vflags & BV_BKGRDINPROG)
1790 			panic("losing buffer 1");
1791 		if (bp->b_kvasize)
1792 			qindex = QUEUE_EMPTYKVA;
1793 		else
1794 			qindex = QUEUE_EMPTY;
1795 		bp->b_flags |= B_AGE;
1796 	/* buffers with junk contents */
1797 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1798 	    (bp->b_ioflags & BIO_ERROR)) {
1799 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1800 		if (bp->b_vflags & BV_BKGRDINPROG)
1801 			panic("losing buffer 2");
1802 		qindex = QUEUE_CLEAN;
1803 		bp->b_flags |= B_AGE;
1804 	/* remaining buffers */
1805 	} else if (bp->b_flags & B_DELWRI)
1806 		qindex = QUEUE_DIRTY;
1807 	else
1808 		qindex = QUEUE_CLEAN;
1809 
1810 	binsfree(bp, qindex);
1811 
1812 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1813 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1814 		panic("brelse: not dirty");
1815 	/* unlock */
1816 	BUF_UNLOCK(bp);
1817 }
1818 
1819 /*
1820  * Release a buffer back to the appropriate queue but do not try to free
1821  * it.  The buffer is expected to be used again soon.
1822  *
1823  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1824  * biodone() to requeue an async I/O on completion.  It is also used when
1825  * known good buffers need to be requeued but we think we may need the data
1826  * again soon.
1827  *
1828  * XXX we should be able to leave the B_RELBUF hint set on completion.
1829  */
1830 void
1831 bqrelse(struct buf *bp)
1832 {
1833 	int qindex;
1834 
1835 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1836 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1837 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1838 
1839 	if (BUF_LOCKRECURSED(bp)) {
1840 		/* do not release to free list */
1841 		BUF_UNLOCK(bp);
1842 		return;
1843 	}
1844 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1845 
1846 	if (bp->b_flags & B_MANAGED) {
1847 		if (bp->b_flags & B_REMFREE)
1848 			bremfreef(bp);
1849 		goto out;
1850 	}
1851 
1852 	/* buffers with stale but valid contents */
1853 	if (bp->b_flags & B_DELWRI) {
1854 		qindex = QUEUE_DIRTY;
1855 	} else {
1856 		if ((bp->b_flags & B_DELWRI) == 0 &&
1857 		    (bp->b_xflags & BX_VNDIRTY))
1858 			panic("bqrelse: not dirty");
1859 		/*
1860 		 * BKGRDINPROG can only be set with the buf and bufobj
1861 		 * locks both held.  We tolerate a race to clear it here.
1862 		 */
1863 		if (buf_vm_page_count_severe() &&
1864 		    (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1865 			/*
1866 			 * We are too low on memory, we have to try to free
1867 			 * the buffer (most importantly: the wired pages
1868 			 * making up its backing store) *now*.
1869 			 */
1870 			brelse(bp);
1871 			return;
1872 		}
1873 		qindex = QUEUE_CLEAN;
1874 	}
1875 	binsfree(bp, qindex);
1876 
1877 out:
1878 	/* unlock */
1879 	BUF_UNLOCK(bp);
1880 }
1881 
1882 /* Give pages used by the bp back to the VM system (where possible) */
1883 static void
1884 vfs_vmio_release(struct buf *bp)
1885 {
1886 	vm_object_t obj;
1887 	vm_page_t m;
1888 	int i;
1889 
1890 	if ((bp->b_flags & B_UNMAPPED) == 0) {
1891 		BUF_CHECK_MAPPED(bp);
1892 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1893 	} else
1894 		BUF_CHECK_UNMAPPED(bp);
1895 	obj = bp->b_bufobj->bo_object;
1896 	if (obj != NULL)
1897 		VM_OBJECT_WLOCK(obj);
1898 	for (i = 0; i < bp->b_npages; i++) {
1899 		m = bp->b_pages[i];
1900 		bp->b_pages[i] = NULL;
1901 		/*
1902 		 * In order to keep page LRU ordering consistent, put
1903 		 * everything on the inactive queue.
1904 		 */
1905 		vm_page_lock(m);
1906 		vm_page_unwire(m, PQ_INACTIVE);
1907 
1908 		/*
1909 		 * Might as well free the page if we can and it has
1910 		 * no valid data.  We also free the page if the
1911 		 * buffer was used for direct I/O
1912 		 */
1913 		if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1914 			if (m->wire_count == 0 && !vm_page_busied(m))
1915 				vm_page_free(m);
1916 		} else if (bp->b_flags & B_DIRECT)
1917 			vm_page_try_to_free(m);
1918 		else if (buf_vm_page_count_severe())
1919 			vm_page_try_to_cache(m);
1920 		vm_page_unlock(m);
1921 	}
1922 	if (obj != NULL)
1923 		VM_OBJECT_WUNLOCK(obj);
1924 
1925 	if (bp->b_bufsize) {
1926 		bufspacewakeup();
1927 		bp->b_bufsize = 0;
1928 	}
1929 	bp->b_npages = 0;
1930 	bp->b_flags &= ~B_VMIO;
1931 	if (bp->b_vp)
1932 		brelvp(bp);
1933 }
1934 
1935 /*
1936  * Check to see if a block at a particular lbn is available for a clustered
1937  * write.
1938  */
1939 static int
1940 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1941 {
1942 	struct buf *bpa;
1943 	int match;
1944 
1945 	match = 0;
1946 
1947 	/* If the buf isn't in core skip it */
1948 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1949 		return (0);
1950 
1951 	/* If the buf is busy we don't want to wait for it */
1952 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1953 		return (0);
1954 
1955 	/* Only cluster with valid clusterable delayed write buffers */
1956 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1957 	    (B_DELWRI | B_CLUSTEROK))
1958 		goto done;
1959 
1960 	if (bpa->b_bufsize != size)
1961 		goto done;
1962 
1963 	/*
1964 	 * Check to see if it is in the expected place on disk and that the
1965 	 * block has been mapped.
1966 	 */
1967 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1968 		match = 1;
1969 done:
1970 	BUF_UNLOCK(bpa);
1971 	return (match);
1972 }
1973 
1974 /*
1975  *	vfs_bio_awrite:
1976  *
1977  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1978  *	This is much better then the old way of writing only one buffer at
1979  *	a time.  Note that we may not be presented with the buffers in the
1980  *	correct order, so we search for the cluster in both directions.
1981  */
1982 int
1983 vfs_bio_awrite(struct buf *bp)
1984 {
1985 	struct bufobj *bo;
1986 	int i;
1987 	int j;
1988 	daddr_t lblkno = bp->b_lblkno;
1989 	struct vnode *vp = bp->b_vp;
1990 	int ncl;
1991 	int nwritten;
1992 	int size;
1993 	int maxcl;
1994 	int gbflags;
1995 
1996 	bo = &vp->v_bufobj;
1997 	gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1998 	/*
1999 	 * right now we support clustered writing only to regular files.  If
2000 	 * we find a clusterable block we could be in the middle of a cluster
2001 	 * rather then at the beginning.
2002 	 */
2003 	if ((vp->v_type == VREG) &&
2004 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
2005 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
2006 
2007 		size = vp->v_mount->mnt_stat.f_iosize;
2008 		maxcl = MAXPHYS / size;
2009 
2010 		BO_RLOCK(bo);
2011 		for (i = 1; i < maxcl; i++)
2012 			if (vfs_bio_clcheck(vp, size, lblkno + i,
2013 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2014 				break;
2015 
2016 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2017 			if (vfs_bio_clcheck(vp, size, lblkno - j,
2018 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2019 				break;
2020 		BO_RUNLOCK(bo);
2021 		--j;
2022 		ncl = i + j;
2023 		/*
2024 		 * this is a possible cluster write
2025 		 */
2026 		if (ncl != 1) {
2027 			BUF_UNLOCK(bp);
2028 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2029 			    gbflags);
2030 			return (nwritten);
2031 		}
2032 	}
2033 	bremfree(bp);
2034 	bp->b_flags |= B_ASYNC;
2035 	/*
2036 	 * default (old) behavior, writing out only one block
2037 	 *
2038 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
2039 	 */
2040 	nwritten = bp->b_bufsize;
2041 	(void) bwrite(bp);
2042 
2043 	return (nwritten);
2044 }
2045 
2046 static void
2047 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2048 {
2049 
2050 	KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2051 	    bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2052 	if ((gbflags & GB_UNMAPPED) == 0) {
2053 		bp->b_kvabase = (caddr_t)addr;
2054 	} else if ((gbflags & GB_KVAALLOC) != 0) {
2055 		KASSERT((gbflags & GB_UNMAPPED) != 0,
2056 		    ("GB_KVAALLOC without GB_UNMAPPED"));
2057 		bp->b_kvaalloc = (caddr_t)addr;
2058 		bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2059 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2060 	}
2061 	bp->b_kvasize = maxsize;
2062 }
2063 
2064 /*
2065  * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2066  * needed.
2067  */
2068 static int
2069 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2070 {
2071 	vm_offset_t addr;
2072 
2073 	bfreekva(bp);
2074 	addr = 0;
2075 
2076 	if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2077 		/*
2078 		 * Buffer map is too fragmented.  Request the caller
2079 		 * to defragment the map.
2080 		 */
2081 		atomic_add_int(&bufdefragcnt, 1);
2082 		return (1);
2083 	}
2084 	setbufkva(bp, addr, maxsize, gbflags);
2085 	atomic_add_long(&bufspace, bp->b_kvasize);
2086 	return (0);
2087 }
2088 
2089 /*
2090  * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2091  * locked vnode is supplied.
2092  */
2093 static void
2094 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2095     int defrag)
2096 {
2097 	struct thread *td;
2098 	char *waitmsg;
2099 	int cnt, error, flags, norunbuf, wait;
2100 
2101 	mtx_assert(&bqclean, MA_OWNED);
2102 
2103 	if (defrag) {
2104 		flags = VFS_BIO_NEED_BUFSPACE;
2105 		waitmsg = "nbufkv";
2106 	} else if (bufspace >= hibufspace) {
2107 		waitmsg = "nbufbs";
2108 		flags = VFS_BIO_NEED_BUFSPACE;
2109 	} else {
2110 		waitmsg = "newbuf";
2111 		flags = VFS_BIO_NEED_ANY;
2112 	}
2113 	atomic_set_int(&needsbuffer, flags);
2114 	mtx_unlock(&bqclean);
2115 
2116 	bd_speedup();	/* heeeelp */
2117 	if ((gbflags & GB_NOWAIT_BD) != 0)
2118 		return;
2119 
2120 	td = curthread;
2121 	cnt = 0;
2122 	wait = MNT_NOWAIT;
2123 	rw_wlock(&nblock);
2124 	while ((needsbuffer & flags) != 0) {
2125 		if (vp != NULL && vp->v_type != VCHR &&
2126 		    (td->td_pflags & TDP_BUFNEED) == 0) {
2127 			rw_wunlock(&nblock);
2128 			/*
2129 			 * getblk() is called with a vnode locked, and
2130 			 * some majority of the dirty buffers may as
2131 			 * well belong to the vnode.  Flushing the
2132 			 * buffers there would make a progress that
2133 			 * cannot be achieved by the buf_daemon, that
2134 			 * cannot lock the vnode.
2135 			 */
2136 			if (cnt++ > 2)
2137 				wait = MNT_WAIT;
2138 			ASSERT_VOP_LOCKED(vp, "bufd_helper");
2139 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2140 			    vn_lock(vp, LK_TRYUPGRADE);
2141 			if (error == 0) {
2142 				/* play bufdaemon */
2143 				norunbuf = curthread_pflags_set(TDP_BUFNEED |
2144 				    TDP_NORUNNINGBUF);
2145 				VOP_FSYNC(vp, wait, td);
2146 				atomic_add_long(&notbufdflushes, 1);
2147 				curthread_pflags_restore(norunbuf);
2148 			}
2149 			rw_wlock(&nblock);
2150 			if ((needsbuffer & flags) == 0)
2151 				break;
2152 		}
2153 		error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
2154 		    (PRIBIO + 4) | slpflag, waitmsg, slptimeo);
2155 		if (error != 0)
2156 			break;
2157 	}
2158 	rw_wunlock(&nblock);
2159 }
2160 
2161 static void
2162 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2163 {
2164 
2165 	CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2166 	    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2167 	     bp->b_kvasize, bp->b_bufsize, qindex);
2168 	mtx_assert(&bqclean, MA_NOTOWNED);
2169 
2170 	/*
2171 	 * Note: we no longer distinguish between VMIO and non-VMIO
2172 	 * buffers.
2173 	 */
2174 	KASSERT((bp->b_flags & B_DELWRI) == 0,
2175 	    ("delwri buffer %p found in queue %d", bp, qindex));
2176 
2177 	if (qindex == QUEUE_CLEAN) {
2178 		if (bp->b_flags & B_VMIO) {
2179 			bp->b_flags &= ~B_ASYNC;
2180 			vfs_vmio_release(bp);
2181 		}
2182 		if (bp->b_vp != NULL)
2183 			brelvp(bp);
2184 	}
2185 
2186 	/*
2187 	 * Get the rest of the buffer freed up.  b_kva* is still valid
2188 	 * after this operation.
2189 	 */
2190 
2191 	if (bp->b_rcred != NOCRED) {
2192 		crfree(bp->b_rcred);
2193 		bp->b_rcred = NOCRED;
2194 	}
2195 	if (bp->b_wcred != NOCRED) {
2196 		crfree(bp->b_wcred);
2197 		bp->b_wcred = NOCRED;
2198 	}
2199 	if (!LIST_EMPTY(&bp->b_dep))
2200 		buf_deallocate(bp);
2201 	if (bp->b_vflags & BV_BKGRDINPROG)
2202 		panic("losing buffer 3");
2203 	KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.  qindex: %d",
2204 	    bp, bp->b_vp, qindex));
2205 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2206 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2207 
2208 	if (bp->b_bufsize)
2209 		allocbuf(bp, 0);
2210 
2211 	bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2212 	bp->b_ioflags = 0;
2213 	bp->b_xflags = 0;
2214 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
2215 	    ("buf %p still counted as free?", bp));
2216 	bp->b_vflags = 0;
2217 	bp->b_vp = NULL;
2218 	bp->b_blkno = bp->b_lblkno = 0;
2219 	bp->b_offset = NOOFFSET;
2220 	bp->b_iodone = 0;
2221 	bp->b_error = 0;
2222 	bp->b_resid = 0;
2223 	bp->b_bcount = 0;
2224 	bp->b_npages = 0;
2225 	bp->b_dirtyoff = bp->b_dirtyend = 0;
2226 	bp->b_bufobj = NULL;
2227 	bp->b_pin_count = 0;
2228 	bp->b_fsprivate1 = NULL;
2229 	bp->b_fsprivate2 = NULL;
2230 	bp->b_fsprivate3 = NULL;
2231 
2232 	LIST_INIT(&bp->b_dep);
2233 }
2234 
2235 static int flushingbufs;
2236 
2237 static struct buf *
2238 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2239 {
2240 	struct buf *bp, *nbp;
2241 	int nqindex, qindex, pass;
2242 
2243 	KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2244 
2245 	pass = 1;
2246 restart:
2247 	atomic_add_int(&getnewbufrestarts, 1);
2248 
2249 	/*
2250 	 * Setup for scan.  If we do not have enough free buffers,
2251 	 * we setup a degenerate case that immediately fails.  Note
2252 	 * that if we are specially marked process, we are allowed to
2253 	 * dip into our reserves.
2254 	 *
2255 	 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2256 	 * for the allocation of the mapped buffer.  For unmapped, the
2257 	 * easiest is to start with EMPTY outright.
2258 	 *
2259 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
2260 	 * However, there are a number of cases (defragging, reusing, ...)
2261 	 * where we cannot backup.
2262 	 */
2263 	nbp = NULL;
2264 	mtx_lock(&bqclean);
2265 	if (!defrag && unmapped) {
2266 		nqindex = QUEUE_EMPTY;
2267 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2268 	}
2269 	if (nbp == NULL) {
2270 		nqindex = QUEUE_EMPTYKVA;
2271 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2272 	}
2273 
2274 	/*
2275 	 * If no EMPTYKVA buffers and we are either defragging or
2276 	 * reusing, locate a CLEAN buffer to free or reuse.  If
2277 	 * bufspace useage is low skip this step so we can allocate a
2278 	 * new buffer.
2279 	 */
2280 	if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2281 		nqindex = QUEUE_CLEAN;
2282 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2283 	}
2284 
2285 	/*
2286 	 * If we could not find or were not allowed to reuse a CLEAN
2287 	 * buffer, check to see if it is ok to use an EMPTY buffer.
2288 	 * We can only use an EMPTY buffer if allocating its KVA would
2289 	 * not otherwise run us out of buffer space.  No KVA is needed
2290 	 * for the unmapped allocation.
2291 	 */
2292 	if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2293 	    metadata)) {
2294 		nqindex = QUEUE_EMPTY;
2295 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2296 	}
2297 
2298 	/*
2299 	 * All available buffers might be clean, retry ignoring the
2300 	 * lobufspace as the last resort.
2301 	 */
2302 	if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2303 		nqindex = QUEUE_CLEAN;
2304 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2305 	}
2306 
2307 	/*
2308 	 * Run scan, possibly freeing data and/or kva mappings on the fly
2309 	 * depending.
2310 	 */
2311 	while ((bp = nbp) != NULL) {
2312 		qindex = nqindex;
2313 
2314 		/*
2315 		 * Calculate next bp (we can only use it if we do not
2316 		 * block or do other fancy things).
2317 		 */
2318 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2319 			switch (qindex) {
2320 			case QUEUE_EMPTY:
2321 				nqindex = QUEUE_EMPTYKVA;
2322 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2323 				if (nbp != NULL)
2324 					break;
2325 				/* FALLTHROUGH */
2326 			case QUEUE_EMPTYKVA:
2327 				nqindex = QUEUE_CLEAN;
2328 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2329 				if (nbp != NULL)
2330 					break;
2331 				/* FALLTHROUGH */
2332 			case QUEUE_CLEAN:
2333 				if (metadata && pass == 1) {
2334 					pass = 2;
2335 					nqindex = QUEUE_EMPTY;
2336 					nbp = TAILQ_FIRST(
2337 					    &bufqueues[QUEUE_EMPTY]);
2338 				}
2339 				/*
2340 				 * nbp is NULL.
2341 				 */
2342 				break;
2343 			}
2344 		}
2345 		/*
2346 		 * If we are defragging then we need a buffer with
2347 		 * b_kvasize != 0.  XXX this situation should no longer
2348 		 * occur, if defrag is non-zero the buffer's b_kvasize
2349 		 * should also be non-zero at this point.  XXX
2350 		 */
2351 		if (defrag && bp->b_kvasize == 0) {
2352 			printf("Warning: defrag empty buffer %p\n", bp);
2353 			continue;
2354 		}
2355 
2356 		/*
2357 		 * Start freeing the bp.  This is somewhat involved.  nbp
2358 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2359 		 */
2360 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2361 			continue;
2362 		/*
2363 		 * BKGRDINPROG can only be set with the buf and bufobj
2364 		 * locks both held.  We tolerate a race to clear it here.
2365 		 */
2366 		if (bp->b_vflags & BV_BKGRDINPROG) {
2367 			BUF_UNLOCK(bp);
2368 			continue;
2369 		}
2370 
2371 		KASSERT(bp->b_qindex == qindex,
2372 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2373 
2374 		bremfreel(bp);
2375 		mtx_unlock(&bqclean);
2376 		/*
2377 		 * NOTE:  nbp is now entirely invalid.  We can only restart
2378 		 * the scan from this point on.
2379 		 */
2380 
2381 		getnewbuf_reuse_bp(bp, qindex);
2382 		mtx_assert(&bqclean, MA_NOTOWNED);
2383 
2384 		/*
2385 		 * If we are defragging then free the buffer.
2386 		 */
2387 		if (defrag) {
2388 			bp->b_flags |= B_INVAL;
2389 			bfreekva(bp);
2390 			brelse(bp);
2391 			defrag = 0;
2392 			goto restart;
2393 		}
2394 
2395 		/*
2396 		 * Notify any waiters for the buffer lock about
2397 		 * identity change by freeing the buffer.
2398 		 */
2399 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2400 			bp->b_flags |= B_INVAL;
2401 			bfreekva(bp);
2402 			brelse(bp);
2403 			goto restart;
2404 		}
2405 
2406 		if (metadata)
2407 			break;
2408 
2409 		/*
2410 		 * If we are overcomitted then recover the buffer and its
2411 		 * KVM space.  This occurs in rare situations when multiple
2412 		 * processes are blocked in getnewbuf() or allocbuf().
2413 		 */
2414 		if (bufspace >= hibufspace)
2415 			flushingbufs = 1;
2416 		if (flushingbufs && bp->b_kvasize != 0) {
2417 			bp->b_flags |= B_INVAL;
2418 			bfreekva(bp);
2419 			brelse(bp);
2420 			goto restart;
2421 		}
2422 		if (bufspace < lobufspace)
2423 			flushingbufs = 0;
2424 		break;
2425 	}
2426 	return (bp);
2427 }
2428 
2429 /*
2430  *	getnewbuf:
2431  *
2432  *	Find and initialize a new buffer header, freeing up existing buffers
2433  *	in the bufqueues as necessary.  The new buffer is returned locked.
2434  *
2435  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
2436  *	buffer away, the caller must set B_INVAL prior to calling brelse().
2437  *
2438  *	We block if:
2439  *		We have insufficient buffer headers
2440  *		We have insufficient buffer space
2441  *		buffer_arena is too fragmented ( space reservation fails )
2442  *		If we have to flush dirty buffers ( but we try to avoid this )
2443  */
2444 static struct buf *
2445 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2446     int gbflags)
2447 {
2448 	struct buf *bp;
2449 	int defrag, metadata;
2450 
2451 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2452 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2453 	if (!unmapped_buf_allowed)
2454 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2455 
2456 	defrag = 0;
2457 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2458 	    vp->v_type == VCHR)
2459 		metadata = 1;
2460 	else
2461 		metadata = 0;
2462 	/*
2463 	 * We can't afford to block since we might be holding a vnode lock,
2464 	 * which may prevent system daemons from running.  We deal with
2465 	 * low-memory situations by proactively returning memory and running
2466 	 * async I/O rather then sync I/O.
2467 	 */
2468 	atomic_add_int(&getnewbufcalls, 1);
2469 	atomic_subtract_int(&getnewbufrestarts, 1);
2470 restart:
2471 	bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2472 	    GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2473 	if (bp != NULL)
2474 		defrag = 0;
2475 
2476 	/*
2477 	 * If we exhausted our list, sleep as appropriate.  We may have to
2478 	 * wakeup various daemons and write out some dirty buffers.
2479 	 *
2480 	 * Generally we are sleeping due to insufficient buffer space.
2481 	 */
2482 	if (bp == NULL) {
2483 		mtx_assert(&bqclean, MA_OWNED);
2484 		getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2485 		mtx_assert(&bqclean, MA_NOTOWNED);
2486 	} else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2487 		mtx_assert(&bqclean, MA_NOTOWNED);
2488 
2489 		bfreekva(bp);
2490 		bp->b_flags |= B_UNMAPPED;
2491 		bp->b_kvabase = bp->b_data = unmapped_buf;
2492 		bp->b_kvasize = maxsize;
2493 		atomic_add_long(&bufspace, bp->b_kvasize);
2494 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2495 		atomic_add_int(&bufreusecnt, 1);
2496 	} else {
2497 		mtx_assert(&bqclean, MA_NOTOWNED);
2498 
2499 		/*
2500 		 * We finally have a valid bp.  We aren't quite out of the
2501 		 * woods, we still have to reserve kva space.  In order
2502 		 * to keep fragmentation sane we only allocate kva in
2503 		 * BKVASIZE chunks.
2504 		 */
2505 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2506 
2507 		if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2508 		    B_KVAALLOC)) == B_UNMAPPED) {
2509 			if (allocbufkva(bp, maxsize, gbflags)) {
2510 				defrag = 1;
2511 				bp->b_flags |= B_INVAL;
2512 				brelse(bp);
2513 				goto restart;
2514 			}
2515 			atomic_add_int(&bufreusecnt, 1);
2516 		} else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2517 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2518 			/*
2519 			 * If the reused buffer has KVA allocated,
2520 			 * reassign b_kvaalloc to b_kvabase.
2521 			 */
2522 			bp->b_kvabase = bp->b_kvaalloc;
2523 			bp->b_flags &= ~B_KVAALLOC;
2524 			atomic_subtract_long(&unmapped_bufspace,
2525 			    bp->b_kvasize);
2526 			atomic_add_int(&bufreusecnt, 1);
2527 		} else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2528 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2529 		    GB_KVAALLOC)) {
2530 			/*
2531 			 * The case of reused buffer already have KVA
2532 			 * mapped, but the request is for unmapped
2533 			 * buffer with KVA allocated.
2534 			 */
2535 			bp->b_kvaalloc = bp->b_kvabase;
2536 			bp->b_data = bp->b_kvabase = unmapped_buf;
2537 			bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2538 			atomic_add_long(&unmapped_bufspace,
2539 			    bp->b_kvasize);
2540 			atomic_add_int(&bufreusecnt, 1);
2541 		}
2542 		if ((gbflags & GB_UNMAPPED) == 0) {
2543 			bp->b_saveaddr = bp->b_kvabase;
2544 			bp->b_data = bp->b_saveaddr;
2545 			bp->b_flags &= ~B_UNMAPPED;
2546 			BUF_CHECK_MAPPED(bp);
2547 		}
2548 	}
2549 	return (bp);
2550 }
2551 
2552 /*
2553  *	buf_daemon:
2554  *
2555  *	buffer flushing daemon.  Buffers are normally flushed by the
2556  *	update daemon but if it cannot keep up this process starts to
2557  *	take the load in an attempt to prevent getnewbuf() from blocking.
2558  */
2559 
2560 static struct kproc_desc buf_kp = {
2561 	"bufdaemon",
2562 	buf_daemon,
2563 	&bufdaemonproc
2564 };
2565 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2566 
2567 static int
2568 buf_flush(int target)
2569 {
2570 	int flushed;
2571 
2572 	flushed = flushbufqueues(target, 0);
2573 	if (flushed == 0) {
2574 		/*
2575 		 * Could not find any buffers without rollback
2576 		 * dependencies, so just write the first one
2577 		 * in the hopes of eventually making progress.
2578 		 */
2579 		flushed = flushbufqueues(target, 1);
2580 	}
2581 	return (flushed);
2582 }
2583 
2584 static void
2585 buf_daemon()
2586 {
2587 	int lodirty;
2588 
2589 	/*
2590 	 * This process needs to be suspended prior to shutdown sync.
2591 	 */
2592 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2593 	    SHUTDOWN_PRI_LAST);
2594 
2595 	/*
2596 	 * This process is allowed to take the buffer cache to the limit
2597 	 */
2598 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2599 	mtx_lock(&bdlock);
2600 	for (;;) {
2601 		bd_request = 0;
2602 		mtx_unlock(&bdlock);
2603 
2604 		kproc_suspend_check(bufdaemonproc);
2605 		lodirty = lodirtybuffers;
2606 		if (bd_speedupreq) {
2607 			lodirty = numdirtybuffers / 2;
2608 			bd_speedupreq = 0;
2609 		}
2610 		/*
2611 		 * Do the flush.  Limit the amount of in-transit I/O we
2612 		 * allow to build up, otherwise we would completely saturate
2613 		 * the I/O system.
2614 		 */
2615 		while (numdirtybuffers > lodirty) {
2616 			if (buf_flush(numdirtybuffers - lodirty) == 0)
2617 				break;
2618 			kern_yield(PRI_USER);
2619 		}
2620 
2621 		/*
2622 		 * Only clear bd_request if we have reached our low water
2623 		 * mark.  The buf_daemon normally waits 1 second and
2624 		 * then incrementally flushes any dirty buffers that have
2625 		 * built up, within reason.
2626 		 *
2627 		 * If we were unable to hit our low water mark and couldn't
2628 		 * find any flushable buffers, we sleep for a short period
2629 		 * to avoid endless loops on unlockable buffers.
2630 		 */
2631 		mtx_lock(&bdlock);
2632 		if (numdirtybuffers <= lodirtybuffers) {
2633 			/*
2634 			 * We reached our low water mark, reset the
2635 			 * request and sleep until we are needed again.
2636 			 * The sleep is just so the suspend code works.
2637 			 */
2638 			bd_request = 0;
2639 			/*
2640 			 * Do an extra wakeup in case dirty threshold
2641 			 * changed via sysctl and the explicit transition
2642 			 * out of shortfall was missed.
2643 			 */
2644 			bdirtywakeup();
2645 			if (runningbufspace <= lorunningspace)
2646 				runningwakeup();
2647 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2648 		} else {
2649 			/*
2650 			 * We couldn't find any flushable dirty buffers but
2651 			 * still have too many dirty buffers, we
2652 			 * have to sleep and try again.  (rare)
2653 			 */
2654 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2655 		}
2656 	}
2657 }
2658 
2659 /*
2660  *	flushbufqueues:
2661  *
2662  *	Try to flush a buffer in the dirty queue.  We must be careful to
2663  *	free up B_INVAL buffers instead of write them, which NFS is
2664  *	particularly sensitive to.
2665  */
2666 static int flushwithdeps = 0;
2667 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2668     0, "Number of buffers flushed with dependecies that require rollbacks");
2669 
2670 static int
2671 flushbufqueues(int target, int flushdeps)
2672 {
2673 	struct buf *sentinel;
2674 	struct vnode *vp;
2675 	struct mount *mp;
2676 	struct buf *bp;
2677 	int hasdeps;
2678 	int flushed;
2679 	int queue;
2680 	int error;
2681 
2682 	flushed = 0;
2683 	queue = QUEUE_DIRTY;
2684 	bp = NULL;
2685 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2686 	sentinel->b_qindex = QUEUE_SENTINEL;
2687 	mtx_lock(&bqdirty);
2688 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2689 	mtx_unlock(&bqdirty);
2690 	while (flushed != target) {
2691 		maybe_yield();
2692 		mtx_lock(&bqdirty);
2693 		bp = TAILQ_NEXT(sentinel, b_freelist);
2694 		if (bp != NULL) {
2695 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2696 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2697 			    b_freelist);
2698 		} else {
2699 			mtx_unlock(&bqdirty);
2700 			break;
2701 		}
2702 		KASSERT(bp->b_qindex != QUEUE_SENTINEL,
2703 		    ("parallel calls to flushbufqueues() bp %p", bp));
2704 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2705 		mtx_unlock(&bqdirty);
2706 		if (error != 0)
2707 			continue;
2708 		if (bp->b_pin_count > 0) {
2709 			BUF_UNLOCK(bp);
2710 			continue;
2711 		}
2712 		/*
2713 		 * BKGRDINPROG can only be set with the buf and bufobj
2714 		 * locks both held.  We tolerate a race to clear it here.
2715 		 */
2716 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2717 		    (bp->b_flags & B_DELWRI) == 0) {
2718 			BUF_UNLOCK(bp);
2719 			continue;
2720 		}
2721 		if (bp->b_flags & B_INVAL) {
2722 			bremfreef(bp);
2723 			brelse(bp);
2724 			flushed++;
2725 			continue;
2726 		}
2727 
2728 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2729 			if (flushdeps == 0) {
2730 				BUF_UNLOCK(bp);
2731 				continue;
2732 			}
2733 			hasdeps = 1;
2734 		} else
2735 			hasdeps = 0;
2736 		/*
2737 		 * We must hold the lock on a vnode before writing
2738 		 * one of its buffers. Otherwise we may confuse, or
2739 		 * in the case of a snapshot vnode, deadlock the
2740 		 * system.
2741 		 *
2742 		 * The lock order here is the reverse of the normal
2743 		 * of vnode followed by buf lock.  This is ok because
2744 		 * the NOWAIT will prevent deadlock.
2745 		 */
2746 		vp = bp->b_vp;
2747 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2748 			BUF_UNLOCK(bp);
2749 			continue;
2750 		}
2751 		error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2752 		if (error == 0) {
2753 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2754 			    bp, bp->b_vp, bp->b_flags);
2755 			vfs_bio_awrite(bp);
2756 			vn_finished_write(mp);
2757 			VOP_UNLOCK(vp, 0);
2758 			flushwithdeps += hasdeps;
2759 			flushed++;
2760 			if (runningbufspace > hirunningspace)
2761 				waitrunningbufspace();
2762 			continue;
2763 		}
2764 		vn_finished_write(mp);
2765 		BUF_UNLOCK(bp);
2766 	}
2767 	mtx_lock(&bqdirty);
2768 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2769 	mtx_unlock(&bqdirty);
2770 	free(sentinel, M_TEMP);
2771 	return (flushed);
2772 }
2773 
2774 /*
2775  * Check to see if a block is currently memory resident.
2776  */
2777 struct buf *
2778 incore(struct bufobj *bo, daddr_t blkno)
2779 {
2780 	struct buf *bp;
2781 
2782 	BO_RLOCK(bo);
2783 	bp = gbincore(bo, blkno);
2784 	BO_RUNLOCK(bo);
2785 	return (bp);
2786 }
2787 
2788 /*
2789  * Returns true if no I/O is needed to access the
2790  * associated VM object.  This is like incore except
2791  * it also hunts around in the VM system for the data.
2792  */
2793 
2794 static int
2795 inmem(struct vnode * vp, daddr_t blkno)
2796 {
2797 	vm_object_t obj;
2798 	vm_offset_t toff, tinc, size;
2799 	vm_page_t m;
2800 	vm_ooffset_t off;
2801 
2802 	ASSERT_VOP_LOCKED(vp, "inmem");
2803 
2804 	if (incore(&vp->v_bufobj, blkno))
2805 		return 1;
2806 	if (vp->v_mount == NULL)
2807 		return 0;
2808 	obj = vp->v_object;
2809 	if (obj == NULL)
2810 		return (0);
2811 
2812 	size = PAGE_SIZE;
2813 	if (size > vp->v_mount->mnt_stat.f_iosize)
2814 		size = vp->v_mount->mnt_stat.f_iosize;
2815 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2816 
2817 	VM_OBJECT_RLOCK(obj);
2818 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2819 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2820 		if (!m)
2821 			goto notinmem;
2822 		tinc = size;
2823 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2824 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2825 		if (vm_page_is_valid(m,
2826 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2827 			goto notinmem;
2828 	}
2829 	VM_OBJECT_RUNLOCK(obj);
2830 	return 1;
2831 
2832 notinmem:
2833 	VM_OBJECT_RUNLOCK(obj);
2834 	return (0);
2835 }
2836 
2837 /*
2838  * Set the dirty range for a buffer based on the status of the dirty
2839  * bits in the pages comprising the buffer.  The range is limited
2840  * to the size of the buffer.
2841  *
2842  * Tell the VM system that the pages associated with this buffer
2843  * are clean.  This is used for delayed writes where the data is
2844  * going to go to disk eventually without additional VM intevention.
2845  *
2846  * Note that while we only really need to clean through to b_bcount, we
2847  * just go ahead and clean through to b_bufsize.
2848  */
2849 static void
2850 vfs_clean_pages_dirty_buf(struct buf *bp)
2851 {
2852 	vm_ooffset_t foff, noff, eoff;
2853 	vm_page_t m;
2854 	int i;
2855 
2856 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2857 		return;
2858 
2859 	foff = bp->b_offset;
2860 	KASSERT(bp->b_offset != NOOFFSET,
2861 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2862 
2863 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2864 	vfs_drain_busy_pages(bp);
2865 	vfs_setdirty_locked_object(bp);
2866 	for (i = 0; i < bp->b_npages; i++) {
2867 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2868 		eoff = noff;
2869 		if (eoff > bp->b_offset + bp->b_bufsize)
2870 			eoff = bp->b_offset + bp->b_bufsize;
2871 		m = bp->b_pages[i];
2872 		vfs_page_set_validclean(bp, foff, m);
2873 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2874 		foff = noff;
2875 	}
2876 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2877 }
2878 
2879 static void
2880 vfs_setdirty_locked_object(struct buf *bp)
2881 {
2882 	vm_object_t object;
2883 	int i;
2884 
2885 	object = bp->b_bufobj->bo_object;
2886 	VM_OBJECT_ASSERT_WLOCKED(object);
2887 
2888 	/*
2889 	 * We qualify the scan for modified pages on whether the
2890 	 * object has been flushed yet.
2891 	 */
2892 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2893 		vm_offset_t boffset;
2894 		vm_offset_t eoffset;
2895 
2896 		/*
2897 		 * test the pages to see if they have been modified directly
2898 		 * by users through the VM system.
2899 		 */
2900 		for (i = 0; i < bp->b_npages; i++)
2901 			vm_page_test_dirty(bp->b_pages[i]);
2902 
2903 		/*
2904 		 * Calculate the encompassing dirty range, boffset and eoffset,
2905 		 * (eoffset - boffset) bytes.
2906 		 */
2907 
2908 		for (i = 0; i < bp->b_npages; i++) {
2909 			if (bp->b_pages[i]->dirty)
2910 				break;
2911 		}
2912 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2913 
2914 		for (i = bp->b_npages - 1; i >= 0; --i) {
2915 			if (bp->b_pages[i]->dirty) {
2916 				break;
2917 			}
2918 		}
2919 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2920 
2921 		/*
2922 		 * Fit it to the buffer.
2923 		 */
2924 
2925 		if (eoffset > bp->b_bcount)
2926 			eoffset = bp->b_bcount;
2927 
2928 		/*
2929 		 * If we have a good dirty range, merge with the existing
2930 		 * dirty range.
2931 		 */
2932 
2933 		if (boffset < eoffset) {
2934 			if (bp->b_dirtyoff > boffset)
2935 				bp->b_dirtyoff = boffset;
2936 			if (bp->b_dirtyend < eoffset)
2937 				bp->b_dirtyend = eoffset;
2938 		}
2939 	}
2940 }
2941 
2942 /*
2943  * Allocate the KVA mapping for an existing buffer. It handles the
2944  * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2945  * KVA which is not mapped (B_KVAALLOC).
2946  */
2947 static void
2948 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2949 {
2950 	struct buf *scratch_bp;
2951 	int bsize, maxsize, need_mapping, need_kva;
2952 	off_t offset;
2953 
2954 	need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2955 	    (gbflags & GB_UNMAPPED) == 0;
2956 	need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2957 	    (gbflags & GB_KVAALLOC) != 0;
2958 	if (!need_mapping && !need_kva)
2959 		return;
2960 
2961 	BUF_CHECK_UNMAPPED(bp);
2962 
2963 	if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2964 		/*
2965 		 * Buffer is not mapped, but the KVA was already
2966 		 * reserved at the time of the instantiation.  Use the
2967 		 * allocated space.
2968 		 */
2969 		bp->b_flags &= ~B_KVAALLOC;
2970 		KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2971 		bp->b_kvabase = bp->b_kvaalloc;
2972 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
2973 		goto has_addr;
2974 	}
2975 
2976 	/*
2977 	 * Calculate the amount of the address space we would reserve
2978 	 * if the buffer was mapped.
2979 	 */
2980 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
2981 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
2982 	offset = blkno * bsize;
2983 	maxsize = size + (offset & PAGE_MASK);
2984 	maxsize = imax(maxsize, bsize);
2985 
2986 mapping_loop:
2987 	if (allocbufkva(bp, maxsize, gbflags)) {
2988 		/*
2989 		 * Request defragmentation. getnewbuf() returns us the
2990 		 * allocated space by the scratch buffer KVA.
2991 		 */
2992 		scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
2993 		    (GB_UNMAPPED | GB_KVAALLOC));
2994 		if (scratch_bp == NULL) {
2995 			if ((gbflags & GB_NOWAIT_BD) != 0) {
2996 				/*
2997 				 * XXXKIB: defragmentation cannot
2998 				 * succeed, not sure what else to do.
2999 				 */
3000 				panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
3001 			}
3002 			atomic_add_int(&mappingrestarts, 1);
3003 			goto mapping_loop;
3004 		}
3005 		KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
3006 		    ("scratch bp !B_KVAALLOC %p", scratch_bp));
3007 		setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
3008 		    scratch_bp->b_kvasize, gbflags);
3009 
3010 		/* Get rid of the scratch buffer. */
3011 		scratch_bp->b_kvasize = 0;
3012 		scratch_bp->b_flags |= B_INVAL;
3013 		scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
3014 		brelse(scratch_bp);
3015 	}
3016 	if (!need_mapping)
3017 		return;
3018 
3019 has_addr:
3020 	bp->b_saveaddr = bp->b_kvabase;
3021 	bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
3022 	bp->b_flags &= ~B_UNMAPPED;
3023 	BUF_CHECK_MAPPED(bp);
3024 	bpmap_qenter(bp);
3025 }
3026 
3027 /*
3028  *	getblk:
3029  *
3030  *	Get a block given a specified block and offset into a file/device.
3031  *	The buffers B_DONE bit will be cleared on return, making it almost
3032  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3033  *	return.  The caller should clear B_INVAL prior to initiating a
3034  *	READ.
3035  *
3036  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3037  *	an existing buffer.
3038  *
3039  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3040  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3041  *	and then cleared based on the backing VM.  If the previous buffer is
3042  *	non-0-sized but invalid, B_CACHE will be cleared.
3043  *
3044  *	If getblk() must create a new buffer, the new buffer is returned with
3045  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3046  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3047  *	backing VM.
3048  *
3049  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3050  *	B_CACHE bit is clear.
3051  *
3052  *	What this means, basically, is that the caller should use B_CACHE to
3053  *	determine whether the buffer is fully valid or not and should clear
3054  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3055  *	the buffer by loading its data area with something, the caller needs
3056  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3057  *	the caller should set B_CACHE ( as an optimization ), else the caller
3058  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3059  *	a write attempt or if it was a successfull read.  If the caller
3060  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3061  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3062  */
3063 struct buf *
3064 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3065     int flags)
3066 {
3067 	struct buf *bp;
3068 	struct bufobj *bo;
3069 	int bsize, error, maxsize, vmio;
3070 	off_t offset;
3071 
3072 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3073 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3074 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3075 	ASSERT_VOP_LOCKED(vp, "getblk");
3076 	if (size > MAXBSIZE)
3077 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
3078 	if (!unmapped_buf_allowed)
3079 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3080 
3081 	bo = &vp->v_bufobj;
3082 loop:
3083 	BO_RLOCK(bo);
3084 	bp = gbincore(bo, blkno);
3085 	if (bp != NULL) {
3086 		int lockflags;
3087 		/*
3088 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3089 		 * it must be on a queue.
3090 		 */
3091 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3092 
3093 		if (flags & GB_LOCK_NOWAIT)
3094 			lockflags |= LK_NOWAIT;
3095 
3096 		error = BUF_TIMELOCK(bp, lockflags,
3097 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3098 
3099 		/*
3100 		 * If we slept and got the lock we have to restart in case
3101 		 * the buffer changed identities.
3102 		 */
3103 		if (error == ENOLCK)
3104 			goto loop;
3105 		/* We timed out or were interrupted. */
3106 		else if (error)
3107 			return (NULL);
3108 		/* If recursed, assume caller knows the rules. */
3109 		else if (BUF_LOCKRECURSED(bp))
3110 			goto end;
3111 
3112 		/*
3113 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3114 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3115 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3116 		 * backing VM cache.
3117 		 */
3118 		if (bp->b_flags & B_INVAL)
3119 			bp->b_flags &= ~B_CACHE;
3120 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3121 			bp->b_flags |= B_CACHE;
3122 		if (bp->b_flags & B_MANAGED)
3123 			MPASS(bp->b_qindex == QUEUE_NONE);
3124 		else
3125 			bremfree(bp);
3126 
3127 		/*
3128 		 * check for size inconsistencies for non-VMIO case.
3129 		 */
3130 		if (bp->b_bcount != size) {
3131 			if ((bp->b_flags & B_VMIO) == 0 ||
3132 			    (size > bp->b_kvasize)) {
3133 				if (bp->b_flags & B_DELWRI) {
3134 					/*
3135 					 * If buffer is pinned and caller does
3136 					 * not want sleep  waiting for it to be
3137 					 * unpinned, bail out
3138 					 * */
3139 					if (bp->b_pin_count > 0) {
3140 						if (flags & GB_LOCK_NOWAIT) {
3141 							bqrelse(bp);
3142 							return (NULL);
3143 						} else {
3144 							bunpin_wait(bp);
3145 						}
3146 					}
3147 					bp->b_flags |= B_NOCACHE;
3148 					bwrite(bp);
3149 				} else {
3150 					if (LIST_EMPTY(&bp->b_dep)) {
3151 						bp->b_flags |= B_RELBUF;
3152 						brelse(bp);
3153 					} else {
3154 						bp->b_flags |= B_NOCACHE;
3155 						bwrite(bp);
3156 					}
3157 				}
3158 				goto loop;
3159 			}
3160 		}
3161 
3162 		/*
3163 		 * Handle the case of unmapped buffer which should
3164 		 * become mapped, or the buffer for which KVA
3165 		 * reservation is requested.
3166 		 */
3167 		bp_unmapped_get_kva(bp, blkno, size, flags);
3168 
3169 		/*
3170 		 * If the size is inconsistant in the VMIO case, we can resize
3171 		 * the buffer.  This might lead to B_CACHE getting set or
3172 		 * cleared.  If the size has not changed, B_CACHE remains
3173 		 * unchanged from its previous state.
3174 		 */
3175 		if (bp->b_bcount != size)
3176 			allocbuf(bp, size);
3177 
3178 		KASSERT(bp->b_offset != NOOFFSET,
3179 		    ("getblk: no buffer offset"));
3180 
3181 		/*
3182 		 * A buffer with B_DELWRI set and B_CACHE clear must
3183 		 * be committed before we can return the buffer in
3184 		 * order to prevent the caller from issuing a read
3185 		 * ( due to B_CACHE not being set ) and overwriting
3186 		 * it.
3187 		 *
3188 		 * Most callers, including NFS and FFS, need this to
3189 		 * operate properly either because they assume they
3190 		 * can issue a read if B_CACHE is not set, or because
3191 		 * ( for example ) an uncached B_DELWRI might loop due
3192 		 * to softupdates re-dirtying the buffer.  In the latter
3193 		 * case, B_CACHE is set after the first write completes,
3194 		 * preventing further loops.
3195 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3196 		 * above while extending the buffer, we cannot allow the
3197 		 * buffer to remain with B_CACHE set after the write
3198 		 * completes or it will represent a corrupt state.  To
3199 		 * deal with this we set B_NOCACHE to scrap the buffer
3200 		 * after the write.
3201 		 *
3202 		 * We might be able to do something fancy, like setting
3203 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3204 		 * so the below call doesn't set B_CACHE, but that gets real
3205 		 * confusing.  This is much easier.
3206 		 */
3207 
3208 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3209 			bp->b_flags |= B_NOCACHE;
3210 			bwrite(bp);
3211 			goto loop;
3212 		}
3213 		bp->b_flags &= ~B_DONE;
3214 	} else {
3215 		/*
3216 		 * Buffer is not in-core, create new buffer.  The buffer
3217 		 * returned by getnewbuf() is locked.  Note that the returned
3218 		 * buffer is also considered valid (not marked B_INVAL).
3219 		 */
3220 		BO_RUNLOCK(bo);
3221 		/*
3222 		 * If the user does not want us to create the buffer, bail out
3223 		 * here.
3224 		 */
3225 		if (flags & GB_NOCREAT)
3226 			return NULL;
3227 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3228 			return NULL;
3229 
3230 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3231 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3232 		offset = blkno * bsize;
3233 		vmio = vp->v_object != NULL;
3234 		if (vmio) {
3235 			maxsize = size + (offset & PAGE_MASK);
3236 		} else {
3237 			maxsize = size;
3238 			/* Do not allow non-VMIO notmapped buffers. */
3239 			flags &= ~GB_UNMAPPED;
3240 		}
3241 		maxsize = imax(maxsize, bsize);
3242 
3243 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3244 		if (bp == NULL) {
3245 			if (slpflag || slptimeo)
3246 				return NULL;
3247 			goto loop;
3248 		}
3249 
3250 		/*
3251 		 * This code is used to make sure that a buffer is not
3252 		 * created while the getnewbuf routine is blocked.
3253 		 * This can be a problem whether the vnode is locked or not.
3254 		 * If the buffer is created out from under us, we have to
3255 		 * throw away the one we just created.
3256 		 *
3257 		 * Note: this must occur before we associate the buffer
3258 		 * with the vp especially considering limitations in
3259 		 * the splay tree implementation when dealing with duplicate
3260 		 * lblkno's.
3261 		 */
3262 		BO_LOCK(bo);
3263 		if (gbincore(bo, blkno)) {
3264 			BO_UNLOCK(bo);
3265 			bp->b_flags |= B_INVAL;
3266 			brelse(bp);
3267 			goto loop;
3268 		}
3269 
3270 		/*
3271 		 * Insert the buffer into the hash, so that it can
3272 		 * be found by incore.
3273 		 */
3274 		bp->b_blkno = bp->b_lblkno = blkno;
3275 		bp->b_offset = offset;
3276 		bgetvp(vp, bp);
3277 		BO_UNLOCK(bo);
3278 
3279 		/*
3280 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3281 		 * buffer size starts out as 0, B_CACHE will be set by
3282 		 * allocbuf() for the VMIO case prior to it testing the
3283 		 * backing store for validity.
3284 		 */
3285 
3286 		if (vmio) {
3287 			bp->b_flags |= B_VMIO;
3288 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3289 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3290 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3291 		} else {
3292 			bp->b_flags &= ~B_VMIO;
3293 			KASSERT(bp->b_bufobj->bo_object == NULL,
3294 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3295 			    bp, bp->b_bufobj->bo_object));
3296 			BUF_CHECK_MAPPED(bp);
3297 		}
3298 
3299 		allocbuf(bp, size);
3300 		bp->b_flags &= ~B_DONE;
3301 	}
3302 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3303 	BUF_ASSERT_HELD(bp);
3304 end:
3305 	KASSERT(bp->b_bufobj == bo,
3306 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3307 	return (bp);
3308 }
3309 
3310 /*
3311  * Get an empty, disassociated buffer of given size.  The buffer is initially
3312  * set to B_INVAL.
3313  */
3314 struct buf *
3315 geteblk(int size, int flags)
3316 {
3317 	struct buf *bp;
3318 	int maxsize;
3319 
3320 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3321 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3322 		if ((flags & GB_NOWAIT_BD) &&
3323 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3324 			return (NULL);
3325 	}
3326 	allocbuf(bp, size);
3327 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3328 	BUF_ASSERT_HELD(bp);
3329 	return (bp);
3330 }
3331 
3332 
3333 /*
3334  * This code constitutes the buffer memory from either anonymous system
3335  * memory (in the case of non-VMIO operations) or from an associated
3336  * VM object (in the case of VMIO operations).  This code is able to
3337  * resize a buffer up or down.
3338  *
3339  * Note that this code is tricky, and has many complications to resolve
3340  * deadlock or inconsistant data situations.  Tread lightly!!!
3341  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3342  * the caller.  Calling this code willy nilly can result in the loss of data.
3343  *
3344  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3345  * B_CACHE for the non-VMIO case.
3346  */
3347 
3348 int
3349 allocbuf(struct buf *bp, int size)
3350 {
3351 	int newbsize, mbsize;
3352 	int i;
3353 
3354 	BUF_ASSERT_HELD(bp);
3355 
3356 	if (bp->b_kvasize < size)
3357 		panic("allocbuf: buffer too small");
3358 
3359 	if ((bp->b_flags & B_VMIO) == 0) {
3360 		caddr_t origbuf;
3361 		int origbufsize;
3362 		/*
3363 		 * Just get anonymous memory from the kernel.  Don't
3364 		 * mess with B_CACHE.
3365 		 */
3366 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3367 		if (bp->b_flags & B_MALLOC)
3368 			newbsize = mbsize;
3369 		else
3370 			newbsize = round_page(size);
3371 
3372 		if (newbsize < bp->b_bufsize) {
3373 			/*
3374 			 * malloced buffers are not shrunk
3375 			 */
3376 			if (bp->b_flags & B_MALLOC) {
3377 				if (newbsize) {
3378 					bp->b_bcount = size;
3379 				} else {
3380 					free(bp->b_data, M_BIOBUF);
3381 					if (bp->b_bufsize) {
3382 						atomic_subtract_long(
3383 						    &bufmallocspace,
3384 						    bp->b_bufsize);
3385 						bufspacewakeup();
3386 						bp->b_bufsize = 0;
3387 					}
3388 					bp->b_saveaddr = bp->b_kvabase;
3389 					bp->b_data = bp->b_saveaddr;
3390 					bp->b_bcount = 0;
3391 					bp->b_flags &= ~B_MALLOC;
3392 				}
3393 				return 1;
3394 			}
3395 			vm_hold_free_pages(bp, newbsize);
3396 		} else if (newbsize > bp->b_bufsize) {
3397 			/*
3398 			 * We only use malloced memory on the first allocation.
3399 			 * and revert to page-allocated memory when the buffer
3400 			 * grows.
3401 			 */
3402 			/*
3403 			 * There is a potential smp race here that could lead
3404 			 * to bufmallocspace slightly passing the max.  It
3405 			 * is probably extremely rare and not worth worrying
3406 			 * over.
3407 			 */
3408 			if ( (bufmallocspace < maxbufmallocspace) &&
3409 				(bp->b_bufsize == 0) &&
3410 				(mbsize <= PAGE_SIZE/2)) {
3411 
3412 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3413 				bp->b_bufsize = mbsize;
3414 				bp->b_bcount = size;
3415 				bp->b_flags |= B_MALLOC;
3416 				atomic_add_long(&bufmallocspace, mbsize);
3417 				return 1;
3418 			}
3419 			origbuf = NULL;
3420 			origbufsize = 0;
3421 			/*
3422 			 * If the buffer is growing on its other-than-first allocation,
3423 			 * then we revert to the page-allocation scheme.
3424 			 */
3425 			if (bp->b_flags & B_MALLOC) {
3426 				origbuf = bp->b_data;
3427 				origbufsize = bp->b_bufsize;
3428 				bp->b_data = bp->b_kvabase;
3429 				if (bp->b_bufsize) {
3430 					atomic_subtract_long(&bufmallocspace,
3431 					    bp->b_bufsize);
3432 					bufspacewakeup();
3433 					bp->b_bufsize = 0;
3434 				}
3435 				bp->b_flags &= ~B_MALLOC;
3436 				newbsize = round_page(newbsize);
3437 			}
3438 			vm_hold_load_pages(
3439 			    bp,
3440 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
3441 			    (vm_offset_t) bp->b_data + newbsize);
3442 			if (origbuf) {
3443 				bcopy(origbuf, bp->b_data, origbufsize);
3444 				free(origbuf, M_BIOBUF);
3445 			}
3446 		}
3447 	} else {
3448 		int desiredpages;
3449 
3450 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3451 		desiredpages = (size == 0) ? 0 :
3452 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3453 
3454 		if (bp->b_flags & B_MALLOC)
3455 			panic("allocbuf: VMIO buffer can't be malloced");
3456 		/*
3457 		 * Set B_CACHE initially if buffer is 0 length or will become
3458 		 * 0-length.
3459 		 */
3460 		if (size == 0 || bp->b_bufsize == 0)
3461 			bp->b_flags |= B_CACHE;
3462 
3463 		if (newbsize < bp->b_bufsize) {
3464 			/*
3465 			 * DEV_BSIZE aligned new buffer size is less then the
3466 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3467 			 * if we have to remove any pages.
3468 			 */
3469 			if (desiredpages < bp->b_npages) {
3470 				vm_page_t m;
3471 
3472 				if ((bp->b_flags & B_UNMAPPED) == 0) {
3473 					BUF_CHECK_MAPPED(bp);
3474 					pmap_qremove((vm_offset_t)trunc_page(
3475 					    (vm_offset_t)bp->b_data) +
3476 					    (desiredpages << PAGE_SHIFT),
3477 					    (bp->b_npages - desiredpages));
3478 				} else
3479 					BUF_CHECK_UNMAPPED(bp);
3480 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3481 				for (i = desiredpages; i < bp->b_npages; i++) {
3482 					/*
3483 					 * the page is not freed here -- it
3484 					 * is the responsibility of
3485 					 * vnode_pager_setsize
3486 					 */
3487 					m = bp->b_pages[i];
3488 					KASSERT(m != bogus_page,
3489 					    ("allocbuf: bogus page found"));
3490 					while (vm_page_sleep_if_busy(m,
3491 					    "biodep"))
3492 						continue;
3493 
3494 					bp->b_pages[i] = NULL;
3495 					vm_page_lock(m);
3496 					vm_page_unwire(m, PQ_INACTIVE);
3497 					vm_page_unlock(m);
3498 				}
3499 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3500 				bp->b_npages = desiredpages;
3501 			}
3502 		} else if (size > bp->b_bcount) {
3503 			/*
3504 			 * We are growing the buffer, possibly in a
3505 			 * byte-granular fashion.
3506 			 */
3507 			vm_object_t obj;
3508 			vm_offset_t toff;
3509 			vm_offset_t tinc;
3510 
3511 			/*
3512 			 * Step 1, bring in the VM pages from the object,
3513 			 * allocating them if necessary.  We must clear
3514 			 * B_CACHE if these pages are not valid for the
3515 			 * range covered by the buffer.
3516 			 */
3517 
3518 			obj = bp->b_bufobj->bo_object;
3519 
3520 			VM_OBJECT_WLOCK(obj);
3521 			while (bp->b_npages < desiredpages) {
3522 				vm_page_t m;
3523 
3524 				/*
3525 				 * We must allocate system pages since blocking
3526 				 * here could interfere with paging I/O, no
3527 				 * matter which process we are.
3528 				 *
3529 				 * Only exclusive busy can be tested here.
3530 				 * Blocking on shared busy might lead to
3531 				 * deadlocks once allocbuf() is called after
3532 				 * pages are vfs_busy_pages().
3533 				 */
3534 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3535 				    bp->b_npages, VM_ALLOC_NOBUSY |
3536 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3537 				    VM_ALLOC_IGN_SBUSY |
3538 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3539 				if (m->valid == 0)
3540 					bp->b_flags &= ~B_CACHE;
3541 				bp->b_pages[bp->b_npages] = m;
3542 				++bp->b_npages;
3543 			}
3544 
3545 			/*
3546 			 * Step 2.  We've loaded the pages into the buffer,
3547 			 * we have to figure out if we can still have B_CACHE
3548 			 * set.  Note that B_CACHE is set according to the
3549 			 * byte-granular range ( bcount and size ), new the
3550 			 * aligned range ( newbsize ).
3551 			 *
3552 			 * The VM test is against m->valid, which is DEV_BSIZE
3553 			 * aligned.  Needless to say, the validity of the data
3554 			 * needs to also be DEV_BSIZE aligned.  Note that this
3555 			 * fails with NFS if the server or some other client
3556 			 * extends the file's EOF.  If our buffer is resized,
3557 			 * B_CACHE may remain set! XXX
3558 			 */
3559 
3560 			toff = bp->b_bcount;
3561 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3562 
3563 			while ((bp->b_flags & B_CACHE) && toff < size) {
3564 				vm_pindex_t pi;
3565 
3566 				if (tinc > (size - toff))
3567 					tinc = size - toff;
3568 
3569 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3570 				    PAGE_SHIFT;
3571 
3572 				vfs_buf_test_cache(
3573 				    bp,
3574 				    bp->b_offset,
3575 				    toff,
3576 				    tinc,
3577 				    bp->b_pages[pi]
3578 				);
3579 				toff += tinc;
3580 				tinc = PAGE_SIZE;
3581 			}
3582 			VM_OBJECT_WUNLOCK(obj);
3583 
3584 			/*
3585 			 * Step 3, fixup the KVM pmap.
3586 			 */
3587 			if ((bp->b_flags & B_UNMAPPED) == 0)
3588 				bpmap_qenter(bp);
3589 			else
3590 				BUF_CHECK_UNMAPPED(bp);
3591 		}
3592 	}
3593 	if (newbsize < bp->b_bufsize)
3594 		bufspacewakeup();
3595 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3596 	bp->b_bcount = size;		/* requested buffer size	*/
3597 	return 1;
3598 }
3599 
3600 extern int inflight_transient_maps;
3601 
3602 void
3603 biodone(struct bio *bp)
3604 {
3605 	struct mtx *mtxp;
3606 	void (*done)(struct bio *);
3607 	vm_offset_t start, end;
3608 
3609 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3610 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3611 		bp->bio_flags |= BIO_UNMAPPED;
3612 		start = trunc_page((vm_offset_t)bp->bio_data);
3613 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3614 		pmap_qremove(start, OFF_TO_IDX(end - start));
3615 		vmem_free(transient_arena, start, end - start);
3616 		atomic_add_int(&inflight_transient_maps, -1);
3617 	}
3618 	done = bp->bio_done;
3619 	if (done == NULL) {
3620 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
3621 		mtx_lock(mtxp);
3622 		bp->bio_flags |= BIO_DONE;
3623 		wakeup(bp);
3624 		mtx_unlock(mtxp);
3625 	} else {
3626 		bp->bio_flags |= BIO_DONE;
3627 		done(bp);
3628 	}
3629 }
3630 
3631 /*
3632  * Wait for a BIO to finish.
3633  */
3634 int
3635 biowait(struct bio *bp, const char *wchan)
3636 {
3637 	struct mtx *mtxp;
3638 
3639 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3640 	mtx_lock(mtxp);
3641 	while ((bp->bio_flags & BIO_DONE) == 0)
3642 		msleep(bp, mtxp, PRIBIO, wchan, 0);
3643 	mtx_unlock(mtxp);
3644 	if (bp->bio_error != 0)
3645 		return (bp->bio_error);
3646 	if (!(bp->bio_flags & BIO_ERROR))
3647 		return (0);
3648 	return (EIO);
3649 }
3650 
3651 void
3652 biofinish(struct bio *bp, struct devstat *stat, int error)
3653 {
3654 
3655 	if (error) {
3656 		bp->bio_error = error;
3657 		bp->bio_flags |= BIO_ERROR;
3658 	}
3659 	if (stat != NULL)
3660 		devstat_end_transaction_bio(stat, bp);
3661 	biodone(bp);
3662 }
3663 
3664 /*
3665  *	bufwait:
3666  *
3667  *	Wait for buffer I/O completion, returning error status.  The buffer
3668  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3669  *	error and cleared.
3670  */
3671 int
3672 bufwait(struct buf *bp)
3673 {
3674 	if (bp->b_iocmd == BIO_READ)
3675 		bwait(bp, PRIBIO, "biord");
3676 	else
3677 		bwait(bp, PRIBIO, "biowr");
3678 	if (bp->b_flags & B_EINTR) {
3679 		bp->b_flags &= ~B_EINTR;
3680 		return (EINTR);
3681 	}
3682 	if (bp->b_ioflags & BIO_ERROR) {
3683 		return (bp->b_error ? bp->b_error : EIO);
3684 	} else {
3685 		return (0);
3686 	}
3687 }
3688 
3689  /*
3690   * Call back function from struct bio back up to struct buf.
3691   */
3692 static void
3693 bufdonebio(struct bio *bip)
3694 {
3695 	struct buf *bp;
3696 
3697 	bp = bip->bio_caller2;
3698 	bp->b_resid = bip->bio_resid;
3699 	bp->b_ioflags = bip->bio_flags;
3700 	bp->b_error = bip->bio_error;
3701 	if (bp->b_error)
3702 		bp->b_ioflags |= BIO_ERROR;
3703 	bufdone(bp);
3704 	g_destroy_bio(bip);
3705 }
3706 
3707 void
3708 dev_strategy(struct cdev *dev, struct buf *bp)
3709 {
3710 	struct cdevsw *csw;
3711 	int ref;
3712 
3713 	KASSERT(dev->si_refcount > 0,
3714 	    ("dev_strategy on un-referenced struct cdev *(%s) %p",
3715 	    devtoname(dev), dev));
3716 
3717 	csw = dev_refthread(dev, &ref);
3718 	dev_strategy_csw(dev, csw, bp);
3719 	dev_relthread(dev, ref);
3720 }
3721 
3722 void
3723 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3724 {
3725 	struct bio *bip;
3726 
3727 	KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3728 	    ("b_iocmd botch"));
3729 	KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3730 	    dev->si_threadcount > 0,
3731 	    ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3732 	    dev));
3733 	if (csw == NULL) {
3734 		bp->b_error = ENXIO;
3735 		bp->b_ioflags = BIO_ERROR;
3736 		bufdone(bp);
3737 		return;
3738 	}
3739 	for (;;) {
3740 		bip = g_new_bio();
3741 		if (bip != NULL)
3742 			break;
3743 		/* Try again later */
3744 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3745 	}
3746 	bip->bio_cmd = bp->b_iocmd;
3747 	bip->bio_offset = bp->b_iooffset;
3748 	bip->bio_length = bp->b_bcount;
3749 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3750 	bdata2bio(bp, bip);
3751 	bip->bio_done = bufdonebio;
3752 	bip->bio_caller2 = bp;
3753 	bip->bio_dev = dev;
3754 	(*csw->d_strategy)(bip);
3755 }
3756 
3757 /*
3758  *	bufdone:
3759  *
3760  *	Finish I/O on a buffer, optionally calling a completion function.
3761  *	This is usually called from an interrupt so process blocking is
3762  *	not allowed.
3763  *
3764  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3765  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3766  *	assuming B_INVAL is clear.
3767  *
3768  *	For the VMIO case, we set B_CACHE if the op was a read and no
3769  *	read error occured, or if the op was a write.  B_CACHE is never
3770  *	set if the buffer is invalid or otherwise uncacheable.
3771  *
3772  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3773  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3774  *	in the biodone routine.
3775  */
3776 void
3777 bufdone(struct buf *bp)
3778 {
3779 	struct bufobj *dropobj;
3780 	void    (*biodone)(struct buf *);
3781 
3782 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3783 	dropobj = NULL;
3784 
3785 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3786 	BUF_ASSERT_HELD(bp);
3787 
3788 	runningbufwakeup(bp);
3789 	if (bp->b_iocmd == BIO_WRITE)
3790 		dropobj = bp->b_bufobj;
3791 	/* call optional completion function if requested */
3792 	if (bp->b_iodone != NULL) {
3793 		biodone = bp->b_iodone;
3794 		bp->b_iodone = NULL;
3795 		(*biodone) (bp);
3796 		if (dropobj)
3797 			bufobj_wdrop(dropobj);
3798 		return;
3799 	}
3800 
3801 	bufdone_finish(bp);
3802 
3803 	if (dropobj)
3804 		bufobj_wdrop(dropobj);
3805 }
3806 
3807 void
3808 bufdone_finish(struct buf *bp)
3809 {
3810 	BUF_ASSERT_HELD(bp);
3811 
3812 	if (!LIST_EMPTY(&bp->b_dep))
3813 		buf_complete(bp);
3814 
3815 	if (bp->b_flags & B_VMIO) {
3816 		vm_ooffset_t foff;
3817 		vm_page_t m;
3818 		vm_object_t obj;
3819 		struct vnode *vp;
3820 		int bogus, i, iosize;
3821 
3822 		obj = bp->b_bufobj->bo_object;
3823 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3824 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3825 		    obj->paging_in_progress, bp->b_npages));
3826 
3827 		vp = bp->b_vp;
3828 		KASSERT(vp->v_holdcnt > 0,
3829 		    ("biodone_finish: vnode %p has zero hold count", vp));
3830 		KASSERT(vp->v_object != NULL,
3831 		    ("biodone_finish: vnode %p has no vm_object", vp));
3832 
3833 		foff = bp->b_offset;
3834 		KASSERT(bp->b_offset != NOOFFSET,
3835 		    ("biodone_finish: bp %p has no buffer offset", bp));
3836 
3837 		/*
3838 		 * Set B_CACHE if the op was a normal read and no error
3839 		 * occured.  B_CACHE is set for writes in the b*write()
3840 		 * routines.
3841 		 */
3842 		iosize = bp->b_bcount - bp->b_resid;
3843 		if (bp->b_iocmd == BIO_READ &&
3844 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3845 		    !(bp->b_ioflags & BIO_ERROR)) {
3846 			bp->b_flags |= B_CACHE;
3847 		}
3848 		bogus = 0;
3849 		VM_OBJECT_WLOCK(obj);
3850 		for (i = 0; i < bp->b_npages; i++) {
3851 			int bogusflag = 0;
3852 			int resid;
3853 
3854 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3855 			if (resid > iosize)
3856 				resid = iosize;
3857 
3858 			/*
3859 			 * cleanup bogus pages, restoring the originals
3860 			 */
3861 			m = bp->b_pages[i];
3862 			if (m == bogus_page) {
3863 				bogus = bogusflag = 1;
3864 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3865 				if (m == NULL)
3866 					panic("biodone: page disappeared!");
3867 				bp->b_pages[i] = m;
3868 			}
3869 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3870 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3871 			    (intmax_t)foff, (uintmax_t)m->pindex));
3872 
3873 			/*
3874 			 * In the write case, the valid and clean bits are
3875 			 * already changed correctly ( see bdwrite() ), so we
3876 			 * only need to do this here in the read case.
3877 			 */
3878 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3879 				KASSERT((m->dirty & vm_page_bits(foff &
3880 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3881 				    " page %p has unexpected dirty bits", m));
3882 				vfs_page_set_valid(bp, foff, m);
3883 			}
3884 
3885 			vm_page_sunbusy(m);
3886 			vm_object_pip_subtract(obj, 1);
3887 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3888 			iosize -= resid;
3889 		}
3890 		vm_object_pip_wakeupn(obj, 0);
3891 		VM_OBJECT_WUNLOCK(obj);
3892 		if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3893 			BUF_CHECK_MAPPED(bp);
3894 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3895 			    bp->b_pages, bp->b_npages);
3896 		}
3897 	}
3898 
3899 	/*
3900 	 * For asynchronous completions, release the buffer now. The brelse
3901 	 * will do a wakeup there if necessary - so no need to do a wakeup
3902 	 * here in the async case. The sync case always needs to do a wakeup.
3903 	 */
3904 
3905 	if (bp->b_flags & B_ASYNC) {
3906 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3907 			brelse(bp);
3908 		else
3909 			bqrelse(bp);
3910 	} else
3911 		bdone(bp);
3912 }
3913 
3914 /*
3915  * This routine is called in lieu of iodone in the case of
3916  * incomplete I/O.  This keeps the busy status for pages
3917  * consistant.
3918  */
3919 void
3920 vfs_unbusy_pages(struct buf *bp)
3921 {
3922 	int i;
3923 	vm_object_t obj;
3924 	vm_page_t m;
3925 
3926 	runningbufwakeup(bp);
3927 	if (!(bp->b_flags & B_VMIO))
3928 		return;
3929 
3930 	obj = bp->b_bufobj->bo_object;
3931 	VM_OBJECT_WLOCK(obj);
3932 	for (i = 0; i < bp->b_npages; i++) {
3933 		m = bp->b_pages[i];
3934 		if (m == bogus_page) {
3935 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3936 			if (!m)
3937 				panic("vfs_unbusy_pages: page missing\n");
3938 			bp->b_pages[i] = m;
3939 			if ((bp->b_flags & B_UNMAPPED) == 0) {
3940 				BUF_CHECK_MAPPED(bp);
3941 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3942 				    bp->b_pages, bp->b_npages);
3943 			} else
3944 				BUF_CHECK_UNMAPPED(bp);
3945 		}
3946 		vm_object_pip_subtract(obj, 1);
3947 		vm_page_sunbusy(m);
3948 	}
3949 	vm_object_pip_wakeupn(obj, 0);
3950 	VM_OBJECT_WUNLOCK(obj);
3951 }
3952 
3953 /*
3954  * vfs_page_set_valid:
3955  *
3956  *	Set the valid bits in a page based on the supplied offset.   The
3957  *	range is restricted to the buffer's size.
3958  *
3959  *	This routine is typically called after a read completes.
3960  */
3961 static void
3962 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3963 {
3964 	vm_ooffset_t eoff;
3965 
3966 	/*
3967 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3968 	 * page boundary and eoff is not greater than the end of the buffer.
3969 	 * The end of the buffer, in this case, is our file EOF, not the
3970 	 * allocation size of the buffer.
3971 	 */
3972 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3973 	if (eoff > bp->b_offset + bp->b_bcount)
3974 		eoff = bp->b_offset + bp->b_bcount;
3975 
3976 	/*
3977 	 * Set valid range.  This is typically the entire buffer and thus the
3978 	 * entire page.
3979 	 */
3980 	if (eoff > off)
3981 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3982 }
3983 
3984 /*
3985  * vfs_page_set_validclean:
3986  *
3987  *	Set the valid bits and clear the dirty bits in a page based on the
3988  *	supplied offset.   The range is restricted to the buffer's size.
3989  */
3990 static void
3991 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3992 {
3993 	vm_ooffset_t soff, eoff;
3994 
3995 	/*
3996 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3997 	 * page boundry or cross the end of the buffer.  The end of the
3998 	 * buffer, in this case, is our file EOF, not the allocation size
3999 	 * of the buffer.
4000 	 */
4001 	soff = off;
4002 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4003 	if (eoff > bp->b_offset + bp->b_bcount)
4004 		eoff = bp->b_offset + bp->b_bcount;
4005 
4006 	/*
4007 	 * Set valid range.  This is typically the entire buffer and thus the
4008 	 * entire page.
4009 	 */
4010 	if (eoff > soff) {
4011 		vm_page_set_validclean(
4012 		    m,
4013 		   (vm_offset_t) (soff & PAGE_MASK),
4014 		   (vm_offset_t) (eoff - soff)
4015 		);
4016 	}
4017 }
4018 
4019 /*
4020  * Ensure that all buffer pages are not exclusive busied.  If any page is
4021  * exclusive busy, drain it.
4022  */
4023 void
4024 vfs_drain_busy_pages(struct buf *bp)
4025 {
4026 	vm_page_t m;
4027 	int i, last_busied;
4028 
4029 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4030 	last_busied = 0;
4031 	for (i = 0; i < bp->b_npages; i++) {
4032 		m = bp->b_pages[i];
4033 		if (vm_page_xbusied(m)) {
4034 			for (; last_busied < i; last_busied++)
4035 				vm_page_sbusy(bp->b_pages[last_busied]);
4036 			while (vm_page_xbusied(m)) {
4037 				vm_page_lock(m);
4038 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4039 				vm_page_busy_sleep(m, "vbpage");
4040 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4041 			}
4042 		}
4043 	}
4044 	for (i = 0; i < last_busied; i++)
4045 		vm_page_sunbusy(bp->b_pages[i]);
4046 }
4047 
4048 /*
4049  * This routine is called before a device strategy routine.
4050  * It is used to tell the VM system that paging I/O is in
4051  * progress, and treat the pages associated with the buffer
4052  * almost as being exclusive busy.  Also the object paging_in_progress
4053  * flag is handled to make sure that the object doesn't become
4054  * inconsistant.
4055  *
4056  * Since I/O has not been initiated yet, certain buffer flags
4057  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4058  * and should be ignored.
4059  */
4060 void
4061 vfs_busy_pages(struct buf *bp, int clear_modify)
4062 {
4063 	int i, bogus;
4064 	vm_object_t obj;
4065 	vm_ooffset_t foff;
4066 	vm_page_t m;
4067 
4068 	if (!(bp->b_flags & B_VMIO))
4069 		return;
4070 
4071 	obj = bp->b_bufobj->bo_object;
4072 	foff = bp->b_offset;
4073 	KASSERT(bp->b_offset != NOOFFSET,
4074 	    ("vfs_busy_pages: no buffer offset"));
4075 	VM_OBJECT_WLOCK(obj);
4076 	vfs_drain_busy_pages(bp);
4077 	if (bp->b_bufsize != 0)
4078 		vfs_setdirty_locked_object(bp);
4079 	bogus = 0;
4080 	for (i = 0; i < bp->b_npages; i++) {
4081 		m = bp->b_pages[i];
4082 
4083 		if ((bp->b_flags & B_CLUSTER) == 0) {
4084 			vm_object_pip_add(obj, 1);
4085 			vm_page_sbusy(m);
4086 		}
4087 		/*
4088 		 * When readying a buffer for a read ( i.e
4089 		 * clear_modify == 0 ), it is important to do
4090 		 * bogus_page replacement for valid pages in
4091 		 * partially instantiated buffers.  Partially
4092 		 * instantiated buffers can, in turn, occur when
4093 		 * reconstituting a buffer from its VM backing store
4094 		 * base.  We only have to do this if B_CACHE is
4095 		 * clear ( which causes the I/O to occur in the
4096 		 * first place ).  The replacement prevents the read
4097 		 * I/O from overwriting potentially dirty VM-backed
4098 		 * pages.  XXX bogus page replacement is, uh, bogus.
4099 		 * It may not work properly with small-block devices.
4100 		 * We need to find a better way.
4101 		 */
4102 		if (clear_modify) {
4103 			pmap_remove_write(m);
4104 			vfs_page_set_validclean(bp, foff, m);
4105 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4106 		    (bp->b_flags & B_CACHE) == 0) {
4107 			bp->b_pages[i] = bogus_page;
4108 			bogus++;
4109 		}
4110 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4111 	}
4112 	VM_OBJECT_WUNLOCK(obj);
4113 	if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4114 		BUF_CHECK_MAPPED(bp);
4115 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4116 		    bp->b_pages, bp->b_npages);
4117 	}
4118 }
4119 
4120 /*
4121  *	vfs_bio_set_valid:
4122  *
4123  *	Set the range within the buffer to valid.  The range is
4124  *	relative to the beginning of the buffer, b_offset.  Note that
4125  *	b_offset itself may be offset from the beginning of the first
4126  *	page.
4127  */
4128 void
4129 vfs_bio_set_valid(struct buf *bp, int base, int size)
4130 {
4131 	int i, n;
4132 	vm_page_t m;
4133 
4134 	if (!(bp->b_flags & B_VMIO))
4135 		return;
4136 
4137 	/*
4138 	 * Fixup base to be relative to beginning of first page.
4139 	 * Set initial n to be the maximum number of bytes in the
4140 	 * first page that can be validated.
4141 	 */
4142 	base += (bp->b_offset & PAGE_MASK);
4143 	n = PAGE_SIZE - (base & PAGE_MASK);
4144 
4145 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4146 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4147 		m = bp->b_pages[i];
4148 		if (n > size)
4149 			n = size;
4150 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4151 		base += n;
4152 		size -= n;
4153 		n = PAGE_SIZE;
4154 	}
4155 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4156 }
4157 
4158 /*
4159  *	vfs_bio_clrbuf:
4160  *
4161  *	If the specified buffer is a non-VMIO buffer, clear the entire
4162  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4163  *	validate only the previously invalid portions of the buffer.
4164  *	This routine essentially fakes an I/O, so we need to clear
4165  *	BIO_ERROR and B_INVAL.
4166  *
4167  *	Note that while we only theoretically need to clear through b_bcount,
4168  *	we go ahead and clear through b_bufsize.
4169  */
4170 void
4171 vfs_bio_clrbuf(struct buf *bp)
4172 {
4173 	int i, j, mask, sa, ea, slide;
4174 
4175 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4176 		clrbuf(bp);
4177 		return;
4178 	}
4179 	bp->b_flags &= ~B_INVAL;
4180 	bp->b_ioflags &= ~BIO_ERROR;
4181 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4182 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4183 	    (bp->b_offset & PAGE_MASK) == 0) {
4184 		if (bp->b_pages[0] == bogus_page)
4185 			goto unlock;
4186 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4187 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4188 		if ((bp->b_pages[0]->valid & mask) == mask)
4189 			goto unlock;
4190 		if ((bp->b_pages[0]->valid & mask) == 0) {
4191 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4192 			bp->b_pages[0]->valid |= mask;
4193 			goto unlock;
4194 		}
4195 	}
4196 	sa = bp->b_offset & PAGE_MASK;
4197 	slide = 0;
4198 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4199 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4200 		ea = slide & PAGE_MASK;
4201 		if (ea == 0)
4202 			ea = PAGE_SIZE;
4203 		if (bp->b_pages[i] == bogus_page)
4204 			continue;
4205 		j = sa / DEV_BSIZE;
4206 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4207 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4208 		if ((bp->b_pages[i]->valid & mask) == mask)
4209 			continue;
4210 		if ((bp->b_pages[i]->valid & mask) == 0)
4211 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4212 		else {
4213 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4214 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4215 					pmap_zero_page_area(bp->b_pages[i],
4216 					    sa, DEV_BSIZE);
4217 				}
4218 			}
4219 		}
4220 		bp->b_pages[i]->valid |= mask;
4221 	}
4222 unlock:
4223 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4224 	bp->b_resid = 0;
4225 }
4226 
4227 void
4228 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4229 {
4230 	vm_page_t m;
4231 	int i, n;
4232 
4233 	if ((bp->b_flags & B_UNMAPPED) == 0) {
4234 		BUF_CHECK_MAPPED(bp);
4235 		bzero(bp->b_data + base, size);
4236 	} else {
4237 		BUF_CHECK_UNMAPPED(bp);
4238 		n = PAGE_SIZE - (base & PAGE_MASK);
4239 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4240 			m = bp->b_pages[i];
4241 			if (n > size)
4242 				n = size;
4243 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4244 			base += n;
4245 			size -= n;
4246 			n = PAGE_SIZE;
4247 		}
4248 	}
4249 }
4250 
4251 /*
4252  * vm_hold_load_pages and vm_hold_free_pages get pages into
4253  * a buffers address space.  The pages are anonymous and are
4254  * not associated with a file object.
4255  */
4256 static void
4257 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4258 {
4259 	vm_offset_t pg;
4260 	vm_page_t p;
4261 	int index;
4262 
4263 	BUF_CHECK_MAPPED(bp);
4264 
4265 	to = round_page(to);
4266 	from = round_page(from);
4267 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4268 
4269 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4270 tryagain:
4271 		/*
4272 		 * note: must allocate system pages since blocking here
4273 		 * could interfere with paging I/O, no matter which
4274 		 * process we are.
4275 		 */
4276 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4277 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4278 		if (p == NULL) {
4279 			VM_WAIT;
4280 			goto tryagain;
4281 		}
4282 		pmap_qenter(pg, &p, 1);
4283 		bp->b_pages[index] = p;
4284 	}
4285 	bp->b_npages = index;
4286 }
4287 
4288 /* Return pages associated with this buf to the vm system */
4289 static void
4290 vm_hold_free_pages(struct buf *bp, int newbsize)
4291 {
4292 	vm_offset_t from;
4293 	vm_page_t p;
4294 	int index, newnpages;
4295 
4296 	BUF_CHECK_MAPPED(bp);
4297 
4298 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4299 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4300 	if (bp->b_npages > newnpages)
4301 		pmap_qremove(from, bp->b_npages - newnpages);
4302 	for (index = newnpages; index < bp->b_npages; index++) {
4303 		p = bp->b_pages[index];
4304 		bp->b_pages[index] = NULL;
4305 		if (vm_page_sbusied(p))
4306 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4307 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4308 		p->wire_count--;
4309 		vm_page_free(p);
4310 		atomic_subtract_int(&vm_cnt.v_wire_count, 1);
4311 	}
4312 	bp->b_npages = newnpages;
4313 }
4314 
4315 /*
4316  * Map an IO request into kernel virtual address space.
4317  *
4318  * All requests are (re)mapped into kernel VA space.
4319  * Notice that we use b_bufsize for the size of the buffer
4320  * to be mapped.  b_bcount might be modified by the driver.
4321  *
4322  * Note that even if the caller determines that the address space should
4323  * be valid, a race or a smaller-file mapped into a larger space may
4324  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4325  * check the return value.
4326  */
4327 int
4328 vmapbuf(struct buf *bp, int mapbuf)
4329 {
4330 	caddr_t kva;
4331 	vm_prot_t prot;
4332 	int pidx;
4333 
4334 	if (bp->b_bufsize < 0)
4335 		return (-1);
4336 	prot = VM_PROT_READ;
4337 	if (bp->b_iocmd == BIO_READ)
4338 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4339 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4340 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4341 	    btoc(MAXPHYS))) < 0)
4342 		return (-1);
4343 	bp->b_npages = pidx;
4344 	if (mapbuf || !unmapped_buf_allowed) {
4345 		pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4346 		kva = bp->b_saveaddr;
4347 		bp->b_saveaddr = bp->b_data;
4348 		bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4349 		bp->b_flags &= ~B_UNMAPPED;
4350 	} else {
4351 		bp->b_flags |= B_UNMAPPED;
4352 		bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4353 		bp->b_saveaddr = bp->b_data;
4354 		bp->b_data = unmapped_buf;
4355 	}
4356 	return(0);
4357 }
4358 
4359 /*
4360  * Free the io map PTEs associated with this IO operation.
4361  * We also invalidate the TLB entries and restore the original b_addr.
4362  */
4363 void
4364 vunmapbuf(struct buf *bp)
4365 {
4366 	int npages;
4367 
4368 	npages = bp->b_npages;
4369 	if (bp->b_flags & B_UNMAPPED)
4370 		bp->b_flags &= ~B_UNMAPPED;
4371 	else
4372 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4373 	vm_page_unhold_pages(bp->b_pages, npages);
4374 
4375 	bp->b_data = bp->b_saveaddr;
4376 }
4377 
4378 void
4379 bdone(struct buf *bp)
4380 {
4381 	struct mtx *mtxp;
4382 
4383 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4384 	mtx_lock(mtxp);
4385 	bp->b_flags |= B_DONE;
4386 	wakeup(bp);
4387 	mtx_unlock(mtxp);
4388 }
4389 
4390 void
4391 bwait(struct buf *bp, u_char pri, const char *wchan)
4392 {
4393 	struct mtx *mtxp;
4394 
4395 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4396 	mtx_lock(mtxp);
4397 	while ((bp->b_flags & B_DONE) == 0)
4398 		msleep(bp, mtxp, pri, wchan, 0);
4399 	mtx_unlock(mtxp);
4400 }
4401 
4402 int
4403 bufsync(struct bufobj *bo, int waitfor)
4404 {
4405 
4406 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4407 }
4408 
4409 void
4410 bufstrategy(struct bufobj *bo, struct buf *bp)
4411 {
4412 	int i = 0;
4413 	struct vnode *vp;
4414 
4415 	vp = bp->b_vp;
4416 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4417 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4418 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4419 	i = VOP_STRATEGY(vp, bp);
4420 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4421 }
4422 
4423 void
4424 bufobj_wrefl(struct bufobj *bo)
4425 {
4426 
4427 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4428 	ASSERT_BO_WLOCKED(bo);
4429 	bo->bo_numoutput++;
4430 }
4431 
4432 void
4433 bufobj_wref(struct bufobj *bo)
4434 {
4435 
4436 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4437 	BO_LOCK(bo);
4438 	bo->bo_numoutput++;
4439 	BO_UNLOCK(bo);
4440 }
4441 
4442 void
4443 bufobj_wdrop(struct bufobj *bo)
4444 {
4445 
4446 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4447 	BO_LOCK(bo);
4448 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4449 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4450 		bo->bo_flag &= ~BO_WWAIT;
4451 		wakeup(&bo->bo_numoutput);
4452 	}
4453 	BO_UNLOCK(bo);
4454 }
4455 
4456 int
4457 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4458 {
4459 	int error;
4460 
4461 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4462 	ASSERT_BO_WLOCKED(bo);
4463 	error = 0;
4464 	while (bo->bo_numoutput) {
4465 		bo->bo_flag |= BO_WWAIT;
4466 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4467 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4468 		if (error)
4469 			break;
4470 	}
4471 	return (error);
4472 }
4473 
4474 void
4475 bpin(struct buf *bp)
4476 {
4477 	struct mtx *mtxp;
4478 
4479 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4480 	mtx_lock(mtxp);
4481 	bp->b_pin_count++;
4482 	mtx_unlock(mtxp);
4483 }
4484 
4485 void
4486 bunpin(struct buf *bp)
4487 {
4488 	struct mtx *mtxp;
4489 
4490 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4491 	mtx_lock(mtxp);
4492 	if (--bp->b_pin_count == 0)
4493 		wakeup(bp);
4494 	mtx_unlock(mtxp);
4495 }
4496 
4497 void
4498 bunpin_wait(struct buf *bp)
4499 {
4500 	struct mtx *mtxp;
4501 
4502 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4503 	mtx_lock(mtxp);
4504 	while (bp->b_pin_count > 0)
4505 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4506 	mtx_unlock(mtxp);
4507 }
4508 
4509 /*
4510  * Set bio_data or bio_ma for struct bio from the struct buf.
4511  */
4512 void
4513 bdata2bio(struct buf *bp, struct bio *bip)
4514 {
4515 
4516 	if ((bp->b_flags & B_UNMAPPED) != 0) {
4517 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4518 		bip->bio_ma = bp->b_pages;
4519 		bip->bio_ma_n = bp->b_npages;
4520 		bip->bio_data = unmapped_buf;
4521 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4522 		bip->bio_flags |= BIO_UNMAPPED;
4523 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4524 		    PAGE_SIZE == bp->b_npages,
4525 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4526 		    (long long)bip->bio_length, bip->bio_ma_n));
4527 	} else {
4528 		bip->bio_data = bp->b_data;
4529 		bip->bio_ma = NULL;
4530 	}
4531 }
4532 
4533 #include "opt_ddb.h"
4534 #ifdef DDB
4535 #include <ddb/ddb.h>
4536 
4537 /* DDB command to show buffer data */
4538 DB_SHOW_COMMAND(buffer, db_show_buffer)
4539 {
4540 	/* get args */
4541 	struct buf *bp = (struct buf *)addr;
4542 
4543 	if (!have_addr) {
4544 		db_printf("usage: show buffer <addr>\n");
4545 		return;
4546 	}
4547 
4548 	db_printf("buf at %p\n", bp);
4549 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4550 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4551 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4552 	db_printf(
4553 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4554 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4555 	    "b_dep = %p\n",
4556 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4557 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4558 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4559 	if (bp->b_npages) {
4560 		int i;
4561 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4562 		for (i = 0; i < bp->b_npages; i++) {
4563 			vm_page_t m;
4564 			m = bp->b_pages[i];
4565 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4566 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4567 			if ((i + 1) < bp->b_npages)
4568 				db_printf(",");
4569 		}
4570 		db_printf("\n");
4571 	}
4572 	db_printf(" ");
4573 	BUF_LOCKPRINTINFO(bp);
4574 }
4575 
4576 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4577 {
4578 	struct buf *bp;
4579 	int i;
4580 
4581 	for (i = 0; i < nbuf; i++) {
4582 		bp = &buf[i];
4583 		if (BUF_ISLOCKED(bp)) {
4584 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4585 			db_printf("\n");
4586 		}
4587 	}
4588 }
4589 
4590 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4591 {
4592 	struct vnode *vp;
4593 	struct buf *bp;
4594 
4595 	if (!have_addr) {
4596 		db_printf("usage: show vnodebufs <addr>\n");
4597 		return;
4598 	}
4599 	vp = (struct vnode *)addr;
4600 	db_printf("Clean buffers:\n");
4601 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4602 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4603 		db_printf("\n");
4604 	}
4605 	db_printf("Dirty buffers:\n");
4606 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4607 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4608 		db_printf("\n");
4609 	}
4610 }
4611 
4612 DB_COMMAND(countfreebufs, db_coundfreebufs)
4613 {
4614 	struct buf *bp;
4615 	int i, used = 0, nfree = 0;
4616 
4617 	if (have_addr) {
4618 		db_printf("usage: countfreebufs\n");
4619 		return;
4620 	}
4621 
4622 	for (i = 0; i < nbuf; i++) {
4623 		bp = &buf[i];
4624 		if ((bp->b_flags & B_INFREECNT) != 0)
4625 			nfree++;
4626 		else
4627 			used++;
4628 	}
4629 
4630 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4631 	    nfree + used);
4632 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4633 }
4634 #endif /* DDB */
4635