xref: /freebsd/sys/kern/vfs_bio.c (revision 0f2bd1e89db1a2f09268edea21e0ead329e092df)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * this file contains a new buffer I/O scheme implementing a coherent
30  * VM object and buffer cache scheme.  Pains have been taken to make
31  * sure that the performance degradation associated with schemes such
32  * as this is not realized.
33  *
34  * Author:  John S. Dyson
35  * Significant help during the development and debugging phases
36  * had been provided by David Greenman, also of the FreeBSD core team.
37  *
38  * see man buf(9) for more info.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bio.h>
47 #include <sys/conf.h>
48 #include <sys/buf.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
51 #include <sys/fail.h>
52 #include <sys/limits.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mount.h>
56 #include <sys/mutex.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
59 #include <sys/proc.h>
60 #include <sys/resourcevar.h>
61 #include <sys/sysctl.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
64 #include <geom/geom.h>
65 #include <vm/vm.h>
66 #include <vm/vm_param.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_pageout.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_extern.h>
72 #include <vm/vm_map.h>
73 #include "opt_compat.h"
74 #include "opt_directio.h"
75 #include "opt_swap.h"
76 
77 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
78 
79 struct	bio_ops bioops;		/* I/O operation notification */
80 
81 struct	buf_ops buf_ops_bio = {
82 	.bop_name	=	"buf_ops_bio",
83 	.bop_write	=	bufwrite,
84 	.bop_strategy	=	bufstrategy,
85 	.bop_sync	=	bufsync,
86 	.bop_bdflush	=	bufbdflush,
87 };
88 
89 /*
90  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
91  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
92  */
93 struct buf *buf;		/* buffer header pool */
94 
95 static struct proc *bufdaemonproc;
96 
97 static int inmem(struct vnode *vp, daddr_t blkno);
98 static void vm_hold_free_pages(struct buf *bp, int newbsize);
99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
100 		vm_offset_t to);
101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
103 		vm_page_t m);
104 static void vfs_drain_busy_pages(struct buf *bp);
105 static void vfs_clean_pages_dirty_buf(struct buf *bp);
106 static void vfs_setdirty_locked_object(struct buf *bp);
107 static void vfs_vmio_release(struct buf *bp);
108 static int vfs_bio_clcheck(struct vnode *vp, int size,
109 		daddr_t lblkno, daddr_t blkno);
110 static int buf_do_flush(struct vnode *vp);
111 static int flushbufqueues(struct vnode *, int, int);
112 static void buf_daemon(void);
113 static void bremfreel(struct buf *bp);
114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
115     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
117 #endif
118 
119 int vmiodirenable = TRUE;
120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
121     "Use the VM system for directory writes");
122 long runningbufspace;
123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
124     "Amount of presently outstanding async buffer io");
125 static long bufspace;
126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
127     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
129     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
130 #else
131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
132     "Virtual memory used for buffers");
133 #endif
134 static long maxbufspace;
135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
136     "Maximum allowed value of bufspace (including buf_daemon)");
137 static long bufmallocspace;
138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
139     "Amount of malloced memory for buffers");
140 static long maxbufmallocspace;
141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
142     "Maximum amount of malloced memory for buffers");
143 static long lobufspace;
144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
145     "Minimum amount of buffers we want to have");
146 long hibufspace;
147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
148     "Maximum allowed value of bufspace (excluding buf_daemon)");
149 static int bufreusecnt;
150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
151     "Number of times we have reused a buffer");
152 static int buffreekvacnt;
153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
154     "Number of times we have freed the KVA space from some buffer");
155 static int bufdefragcnt;
156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
157     "Number of times we have had to repeat buffer allocation to defragment");
158 static long lorunningspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
160     "Minimum preferred space used for in-progress I/O");
161 static long hirunningspace;
162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
163     "Maximum amount of space to use for in-progress I/O");
164 int dirtybufferflushes;
165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
166     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
167 int bdwriteskip;
168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
169     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
170 int altbufferflushes;
171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
172     0, "Number of fsync flushes to limit dirty buffers");
173 static int recursiveflushes;
174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
175     0, "Number of flushes skipped due to being recursive");
176 static int numdirtybuffers;
177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
178     "Number of buffers that are dirty (has unwritten changes) at the moment");
179 static int lodirtybuffers;
180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
181     "How many buffers we want to have free before bufdaemon can sleep");
182 static int hidirtybuffers;
183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
184     "When the number of dirty buffers is considered severe");
185 int dirtybufthresh;
186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
187     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
188 static int numfreebuffers;
189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
190     "Number of free buffers");
191 static int lofreebuffers;
192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
193    "XXX Unused");
194 static int hifreebuffers;
195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
196    "XXX Complicatedly unused");
197 static int getnewbufcalls;
198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
199    "Number of calls to getnewbuf");
200 static int getnewbufrestarts;
201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
202     "Number of times getnewbuf has had to restart a buffer aquisition");
203 static int flushbufqtarget = 100;
204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
205     "Amount of work to do in flushbufqueues when helping bufdaemon");
206 static long notbufdflashes;
207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, &notbufdflashes, 0,
208     "Number of dirty buffer flushes done by the bufdaemon helpers");
209 
210 /*
211  * Wakeup point for bufdaemon, as well as indicator of whether it is already
212  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
213  * is idling.
214  */
215 static int bd_request;
216 
217 /*
218  * Request for the buf daemon to write more buffers than is indicated by
219  * lodirtybuf.  This may be necessary to push out excess dependencies or
220  * defragment the address space where a simple count of the number of dirty
221  * buffers is insufficient to characterize the demand for flushing them.
222  */
223 static int bd_speedupreq;
224 
225 /*
226  * This lock synchronizes access to bd_request.
227  */
228 static struct mtx bdlock;
229 
230 /*
231  * bogus page -- for I/O to/from partially complete buffers
232  * this is a temporary solution to the problem, but it is not
233  * really that bad.  it would be better to split the buffer
234  * for input in the case of buffers partially already in memory,
235  * but the code is intricate enough already.
236  */
237 vm_page_t bogus_page;
238 
239 /*
240  * Synchronization (sleep/wakeup) variable for active buffer space requests.
241  * Set when wait starts, cleared prior to wakeup().
242  * Used in runningbufwakeup() and waitrunningbufspace().
243  */
244 static int runningbufreq;
245 
246 /*
247  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
248  * waitrunningbufspace().
249  */
250 static struct mtx rbreqlock;
251 
252 /*
253  * Synchronization (sleep/wakeup) variable for buffer requests.
254  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
255  * by and/or.
256  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
257  * getnewbuf(), and getblk().
258  */
259 static int needsbuffer;
260 
261 /*
262  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
263  */
264 static struct mtx nblock;
265 
266 /*
267  * Definitions for the buffer free lists.
268  */
269 #define BUFFER_QUEUES	6	/* number of free buffer queues */
270 
271 #define QUEUE_NONE	0	/* on no queue */
272 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
273 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
274 #define QUEUE_DIRTY_GIANT 3	/* B_DELWRI buffers that need giant */
275 #define QUEUE_EMPTYKVA	4	/* empty buffer headers w/KVA assignment */
276 #define QUEUE_EMPTY	5	/* empty buffer headers */
277 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
278 
279 /* Queues for free buffers with various properties */
280 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
281 
282 /* Lock for the bufqueues */
283 static struct mtx bqlock;
284 
285 /*
286  * Single global constant for BUF_WMESG, to avoid getting multiple references.
287  * buf_wmesg is referred from macros.
288  */
289 const char *buf_wmesg = BUF_WMESG;
290 
291 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
292 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
293 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
294 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
295 
296 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
297     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
298 static int
299 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
300 {
301 	long lvalue;
302 	int ivalue;
303 
304 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
305 		return (sysctl_handle_long(oidp, arg1, arg2, req));
306 	lvalue = *(long *)arg1;
307 	if (lvalue > INT_MAX)
308 		/* On overflow, still write out a long to trigger ENOMEM. */
309 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
310 	ivalue = lvalue;
311 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
312 }
313 #endif
314 
315 #ifdef DIRECTIO
316 extern void ffs_rawread_setup(void);
317 #endif /* DIRECTIO */
318 /*
319  *	numdirtywakeup:
320  *
321  *	If someone is blocked due to there being too many dirty buffers,
322  *	and numdirtybuffers is now reasonable, wake them up.
323  */
324 
325 static __inline void
326 numdirtywakeup(int level)
327 {
328 
329 	if (numdirtybuffers <= level) {
330 		mtx_lock(&nblock);
331 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
332 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
333 			wakeup(&needsbuffer);
334 		}
335 		mtx_unlock(&nblock);
336 	}
337 }
338 
339 /*
340  *	bufspacewakeup:
341  *
342  *	Called when buffer space is potentially available for recovery.
343  *	getnewbuf() will block on this flag when it is unable to free
344  *	sufficient buffer space.  Buffer space becomes recoverable when
345  *	bp's get placed back in the queues.
346  */
347 
348 static __inline void
349 bufspacewakeup(void)
350 {
351 
352 	/*
353 	 * If someone is waiting for BUF space, wake them up.  Even
354 	 * though we haven't freed the kva space yet, the waiting
355 	 * process will be able to now.
356 	 */
357 	mtx_lock(&nblock);
358 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
359 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
360 		wakeup(&needsbuffer);
361 	}
362 	mtx_unlock(&nblock);
363 }
364 
365 /*
366  * runningbufwakeup() - in-progress I/O accounting.
367  *
368  */
369 void
370 runningbufwakeup(struct buf *bp)
371 {
372 
373 	if (bp->b_runningbufspace) {
374 		atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
375 		bp->b_runningbufspace = 0;
376 		mtx_lock(&rbreqlock);
377 		if (runningbufreq && runningbufspace <= lorunningspace) {
378 			runningbufreq = 0;
379 			wakeup(&runningbufreq);
380 		}
381 		mtx_unlock(&rbreqlock);
382 	}
383 }
384 
385 /*
386  *	bufcountwakeup:
387  *
388  *	Called when a buffer has been added to one of the free queues to
389  *	account for the buffer and to wakeup anyone waiting for free buffers.
390  *	This typically occurs when large amounts of metadata are being handled
391  *	by the buffer cache ( else buffer space runs out first, usually ).
392  */
393 
394 static __inline void
395 bufcountwakeup(struct buf *bp)
396 {
397 	int old;
398 
399 	KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
400 	    ("buf %p already counted as free", bp));
401 	if (bp->b_bufobj != NULL)
402 		mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
403 	bp->b_vflags |= BV_INFREECNT;
404 	old = atomic_fetchadd_int(&numfreebuffers, 1);
405 	KASSERT(old >= 0 && old < nbuf,
406 	    ("numfreebuffers climbed to %d", old + 1));
407 	mtx_lock(&nblock);
408 	if (needsbuffer) {
409 		needsbuffer &= ~VFS_BIO_NEED_ANY;
410 		if (numfreebuffers >= hifreebuffers)
411 			needsbuffer &= ~VFS_BIO_NEED_FREE;
412 		wakeup(&needsbuffer);
413 	}
414 	mtx_unlock(&nblock);
415 }
416 
417 /*
418  *	waitrunningbufspace()
419  *
420  *	runningbufspace is a measure of the amount of I/O currently
421  *	running.  This routine is used in async-write situations to
422  *	prevent creating huge backups of pending writes to a device.
423  *	Only asynchronous writes are governed by this function.
424  *
425  *	Reads will adjust runningbufspace, but will not block based on it.
426  *	The read load has a side effect of reducing the allowed write load.
427  *
428  *	This does NOT turn an async write into a sync write.  It waits
429  *	for earlier writes to complete and generally returns before the
430  *	caller's write has reached the device.
431  */
432 void
433 waitrunningbufspace(void)
434 {
435 
436 	mtx_lock(&rbreqlock);
437 	while (runningbufspace > hirunningspace) {
438 		++runningbufreq;
439 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
440 	}
441 	mtx_unlock(&rbreqlock);
442 }
443 
444 
445 /*
446  *	vfs_buf_test_cache:
447  *
448  *	Called when a buffer is extended.  This function clears the B_CACHE
449  *	bit if the newly extended portion of the buffer does not contain
450  *	valid data.
451  */
452 static __inline
453 void
454 vfs_buf_test_cache(struct buf *bp,
455 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
456 		  vm_page_t m)
457 {
458 
459 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
460 	if (bp->b_flags & B_CACHE) {
461 		int base = (foff + off) & PAGE_MASK;
462 		if (vm_page_is_valid(m, base, size) == 0)
463 			bp->b_flags &= ~B_CACHE;
464 	}
465 }
466 
467 /* Wake up the buffer daemon if necessary */
468 static __inline
469 void
470 bd_wakeup(int dirtybuflevel)
471 {
472 
473 	mtx_lock(&bdlock);
474 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
475 		bd_request = 1;
476 		wakeup(&bd_request);
477 	}
478 	mtx_unlock(&bdlock);
479 }
480 
481 /*
482  * bd_speedup - speedup the buffer cache flushing code
483  */
484 
485 void
486 bd_speedup(void)
487 {
488 	int needwake;
489 
490 	mtx_lock(&bdlock);
491 	needwake = 0;
492 	if (bd_speedupreq == 0 || bd_request == 0)
493 		needwake = 1;
494 	bd_speedupreq = 1;
495 	bd_request = 1;
496 	if (needwake)
497 		wakeup(&bd_request);
498 	mtx_unlock(&bdlock);
499 }
500 
501 /*
502  * Calculating buffer cache scaling values and reserve space for buffer
503  * headers.  This is called during low level kernel initialization and
504  * may be called more then once.  We CANNOT write to the memory area
505  * being reserved at this time.
506  */
507 caddr_t
508 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
509 {
510 	int tuned_nbuf;
511 	long maxbuf;
512 
513 	/*
514 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
515 	 * PAGE_SIZE is >= 1K)
516 	 */
517 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
518 
519 	/*
520 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
521 	 * For the first 64MB of ram nominally allocate sufficient buffers to
522 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
523 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
524 	 * the buffer cache we limit the eventual kva reservation to
525 	 * maxbcache bytes.
526 	 *
527 	 * factor represents the 1/4 x ram conversion.
528 	 */
529 	if (nbuf == 0) {
530 		int factor = 4 * BKVASIZE / 1024;
531 
532 		nbuf = 50;
533 		if (physmem_est > 4096)
534 			nbuf += min((physmem_est - 4096) / factor,
535 			    65536 / factor);
536 		if (physmem_est > 65536)
537 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
538 
539 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
540 			nbuf = maxbcache / BKVASIZE;
541 		tuned_nbuf = 1;
542 	} else
543 		tuned_nbuf = 0;
544 
545 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
546 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
547 	if (nbuf > maxbuf) {
548 		if (!tuned_nbuf)
549 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
550 			    maxbuf);
551 		nbuf = maxbuf;
552 	}
553 
554 	/*
555 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
556 	 * We have no less then 16 and no more then 256.
557 	 */
558 	nswbuf = max(min(nbuf/4, 256), 16);
559 #ifdef NSWBUF_MIN
560 	if (nswbuf < NSWBUF_MIN)
561 		nswbuf = NSWBUF_MIN;
562 #endif
563 #ifdef DIRECTIO
564 	ffs_rawread_setup();
565 #endif
566 
567 	/*
568 	 * Reserve space for the buffer cache buffers
569 	 */
570 	swbuf = (void *)v;
571 	v = (caddr_t)(swbuf + nswbuf);
572 	buf = (void *)v;
573 	v = (caddr_t)(buf + nbuf);
574 
575 	return(v);
576 }
577 
578 /* Initialize the buffer subsystem.  Called before use of any buffers. */
579 void
580 bufinit(void)
581 {
582 	struct buf *bp;
583 	int i;
584 
585 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
586 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
587 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
588 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
589 
590 	/* next, make a null set of free lists */
591 	for (i = 0; i < BUFFER_QUEUES; i++)
592 		TAILQ_INIT(&bufqueues[i]);
593 
594 	/* finally, initialize each buffer header and stick on empty q */
595 	for (i = 0; i < nbuf; i++) {
596 		bp = &buf[i];
597 		bzero(bp, sizeof *bp);
598 		bp->b_flags = B_INVAL;	/* we're just an empty header */
599 		bp->b_rcred = NOCRED;
600 		bp->b_wcred = NOCRED;
601 		bp->b_qindex = QUEUE_EMPTY;
602 		bp->b_vflags = BV_INFREECNT;	/* buf is counted as free */
603 		bp->b_xflags = 0;
604 		LIST_INIT(&bp->b_dep);
605 		BUF_LOCKINIT(bp);
606 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
607 	}
608 
609 	/*
610 	 * maxbufspace is the absolute maximum amount of buffer space we are
611 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
612 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
613 	 * used by most other processes.  The differential is required to
614 	 * ensure that buf_daemon is able to run when other processes might
615 	 * be blocked waiting for buffer space.
616 	 *
617 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
618 	 * this may result in KVM fragmentation which is not handled optimally
619 	 * by the system.
620 	 */
621 	maxbufspace = (long)nbuf * BKVASIZE;
622 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
623 	lobufspace = hibufspace - MAXBSIZE;
624 
625 	/*
626 	 * Note: The 16 MB upper limit for hirunningspace was chosen
627 	 * arbitrarily and may need further tuning. It corresponds to
628 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
629 	 * which fits with many RAID controllers' tagged queuing limits.
630 	 * The lower 1 MB limit is the historical upper limit for
631 	 * hirunningspace.
632 	 */
633 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
634 	    16 * 1024 * 1024), 1024 * 1024);
635 	lorunningspace = roundup(hirunningspace / 2, MAXBSIZE);
636 
637 /*
638  * Limit the amount of malloc memory since it is wired permanently into
639  * the kernel space.  Even though this is accounted for in the buffer
640  * allocation, we don't want the malloced region to grow uncontrolled.
641  * The malloc scheme improves memory utilization significantly on average
642  * (small) directories.
643  */
644 	maxbufmallocspace = hibufspace / 20;
645 
646 /*
647  * Reduce the chance of a deadlock occuring by limiting the number
648  * of delayed-write dirty buffers we allow to stack up.
649  */
650 	hidirtybuffers = nbuf / 4 + 20;
651 	dirtybufthresh = hidirtybuffers * 9 / 10;
652 	numdirtybuffers = 0;
653 /*
654  * To support extreme low-memory systems, make sure hidirtybuffers cannot
655  * eat up all available buffer space.  This occurs when our minimum cannot
656  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
657  * BKVASIZE'd (8K) buffers.
658  */
659 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
660 		hidirtybuffers >>= 1;
661 	}
662 	lodirtybuffers = hidirtybuffers / 2;
663 
664 /*
665  * Try to keep the number of free buffers in the specified range,
666  * and give special processes (e.g. like buf_daemon) access to an
667  * emergency reserve.
668  */
669 	lofreebuffers = nbuf / 18 + 5;
670 	hifreebuffers = 2 * lofreebuffers;
671 	numfreebuffers = nbuf;
672 
673 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
674 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
675 }
676 
677 /*
678  * bfreekva() - free the kva allocation for a buffer.
679  *
680  *	Since this call frees up buffer space, we call bufspacewakeup().
681  */
682 static void
683 bfreekva(struct buf *bp)
684 {
685 
686 	if (bp->b_kvasize) {
687 		atomic_add_int(&buffreekvacnt, 1);
688 		atomic_subtract_long(&bufspace, bp->b_kvasize);
689 		vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
690 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
691 		bp->b_kvasize = 0;
692 		bufspacewakeup();
693 	}
694 }
695 
696 /*
697  *	bremfree:
698  *
699  *	Mark the buffer for removal from the appropriate free list in brelse.
700  *
701  */
702 void
703 bremfree(struct buf *bp)
704 {
705 	int old;
706 
707 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
708 	KASSERT((bp->b_flags & B_REMFREE) == 0,
709 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
710 	KASSERT(bp->b_qindex != QUEUE_NONE,
711 	    ("bremfree: buffer %p not on a queue.", bp));
712 	BUF_ASSERT_HELD(bp);
713 
714 	bp->b_flags |= B_REMFREE;
715 	/* Fixup numfreebuffers count.  */
716 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
717 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
718 		    ("buf %p not counted in numfreebuffers", bp));
719 		if (bp->b_bufobj != NULL)
720 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
721 		bp->b_vflags &= ~BV_INFREECNT;
722 		old = atomic_fetchadd_int(&numfreebuffers, -1);
723 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
724 	}
725 }
726 
727 /*
728  *	bremfreef:
729  *
730  *	Force an immediate removal from a free list.  Used only in nfs when
731  *	it abuses the b_freelist pointer.
732  */
733 void
734 bremfreef(struct buf *bp)
735 {
736 	mtx_lock(&bqlock);
737 	bremfreel(bp);
738 	mtx_unlock(&bqlock);
739 }
740 
741 /*
742  *	bremfreel:
743  *
744  *	Removes a buffer from the free list, must be called with the
745  *	bqlock held.
746  */
747 static void
748 bremfreel(struct buf *bp)
749 {
750 	int old;
751 
752 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
753 	    bp, bp->b_vp, bp->b_flags);
754 	KASSERT(bp->b_qindex != QUEUE_NONE,
755 	    ("bremfreel: buffer %p not on a queue.", bp));
756 	BUF_ASSERT_HELD(bp);
757 	mtx_assert(&bqlock, MA_OWNED);
758 
759 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
760 	bp->b_qindex = QUEUE_NONE;
761 	/*
762 	 * If this was a delayed bremfree() we only need to remove the buffer
763 	 * from the queue and return the stats are already done.
764 	 */
765 	if (bp->b_flags & B_REMFREE) {
766 		bp->b_flags &= ~B_REMFREE;
767 		return;
768 	}
769 	/*
770 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
771 	 * delayed-write, the buffer was free and we must decrement
772 	 * numfreebuffers.
773 	 */
774 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
775 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
776 		    ("buf %p not counted in numfreebuffers", bp));
777 		if (bp->b_bufobj != NULL)
778 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
779 		bp->b_vflags &= ~BV_INFREECNT;
780 		old = atomic_fetchadd_int(&numfreebuffers, -1);
781 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
782 	}
783 }
784 
785 
786 /*
787  * Get a buffer with the specified data.  Look in the cache first.  We
788  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
789  * is set, the buffer is valid and we do not have to do anything ( see
790  * getblk() ).  This is really just a special case of breadn().
791  */
792 int
793 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
794     struct buf **bpp)
795 {
796 
797 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
798 }
799 
800 /*
801  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
802  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
803  * the buffer is valid and we do not have to do anything.
804  */
805 void
806 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
807     int cnt, struct ucred * cred)
808 {
809 	struct buf *rabp;
810 	int i;
811 
812 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
813 		if (inmem(vp, *rablkno))
814 			continue;
815 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
816 
817 		if ((rabp->b_flags & B_CACHE) == 0) {
818 			if (!TD_IS_IDLETHREAD(curthread))
819 				curthread->td_ru.ru_inblock++;
820 			rabp->b_flags |= B_ASYNC;
821 			rabp->b_flags &= ~B_INVAL;
822 			rabp->b_ioflags &= ~BIO_ERROR;
823 			rabp->b_iocmd = BIO_READ;
824 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
825 				rabp->b_rcred = crhold(cred);
826 			vfs_busy_pages(rabp, 0);
827 			BUF_KERNPROC(rabp);
828 			rabp->b_iooffset = dbtob(rabp->b_blkno);
829 			bstrategy(rabp);
830 		} else {
831 			brelse(rabp);
832 		}
833 	}
834 }
835 
836 /*
837  * Operates like bread, but also starts asynchronous I/O on
838  * read-ahead blocks.
839  */
840 int
841 breadn(struct vnode * vp, daddr_t blkno, int size,
842     daddr_t * rablkno, int *rabsize,
843     int cnt, struct ucred * cred, struct buf **bpp)
844 {
845 	struct buf *bp;
846 	int rv = 0, readwait = 0;
847 
848 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
849 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
850 
851 	/* if not found in cache, do some I/O */
852 	if ((bp->b_flags & B_CACHE) == 0) {
853 		if (!TD_IS_IDLETHREAD(curthread))
854 			curthread->td_ru.ru_inblock++;
855 		bp->b_iocmd = BIO_READ;
856 		bp->b_flags &= ~B_INVAL;
857 		bp->b_ioflags &= ~BIO_ERROR;
858 		if (bp->b_rcred == NOCRED && cred != NOCRED)
859 			bp->b_rcred = crhold(cred);
860 		vfs_busy_pages(bp, 0);
861 		bp->b_iooffset = dbtob(bp->b_blkno);
862 		bstrategy(bp);
863 		++readwait;
864 	}
865 
866 	breada(vp, rablkno, rabsize, cnt, cred);
867 
868 	if (readwait) {
869 		rv = bufwait(bp);
870 	}
871 	return (rv);
872 }
873 
874 /*
875  * Write, release buffer on completion.  (Done by iodone
876  * if async).  Do not bother writing anything if the buffer
877  * is invalid.
878  *
879  * Note that we set B_CACHE here, indicating that buffer is
880  * fully valid and thus cacheable.  This is true even of NFS
881  * now so we set it generally.  This could be set either here
882  * or in biodone() since the I/O is synchronous.  We put it
883  * here.
884  */
885 int
886 bufwrite(struct buf *bp)
887 {
888 	int oldflags;
889 	struct vnode *vp;
890 	int vp_md;
891 
892 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
893 	if (bp->b_flags & B_INVAL) {
894 		brelse(bp);
895 		return (0);
896 	}
897 
898 	oldflags = bp->b_flags;
899 
900 	BUF_ASSERT_HELD(bp);
901 
902 	if (bp->b_pin_count > 0)
903 		bunpin_wait(bp);
904 
905 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
906 	    ("FFS background buffer should not get here %p", bp));
907 
908 	vp = bp->b_vp;
909 	if (vp)
910 		vp_md = vp->v_vflag & VV_MD;
911 	else
912 		vp_md = 0;
913 
914 	/* Mark the buffer clean */
915 	bundirty(bp);
916 
917 	bp->b_flags &= ~B_DONE;
918 	bp->b_ioflags &= ~BIO_ERROR;
919 	bp->b_flags |= B_CACHE;
920 	bp->b_iocmd = BIO_WRITE;
921 
922 	bufobj_wref(bp->b_bufobj);
923 	vfs_busy_pages(bp, 1);
924 
925 	/*
926 	 * Normal bwrites pipeline writes
927 	 */
928 	bp->b_runningbufspace = bp->b_bufsize;
929 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
930 
931 	if (!TD_IS_IDLETHREAD(curthread))
932 		curthread->td_ru.ru_oublock++;
933 	if (oldflags & B_ASYNC)
934 		BUF_KERNPROC(bp);
935 	bp->b_iooffset = dbtob(bp->b_blkno);
936 	bstrategy(bp);
937 
938 	if ((oldflags & B_ASYNC) == 0) {
939 		int rtval = bufwait(bp);
940 		brelse(bp);
941 		return (rtval);
942 	} else {
943 		/*
944 		 * don't allow the async write to saturate the I/O
945 		 * system.  We will not deadlock here because
946 		 * we are blocking waiting for I/O that is already in-progress
947 		 * to complete. We do not block here if it is the update
948 		 * or syncer daemon trying to clean up as that can lead
949 		 * to deadlock.
950 		 */
951 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
952 			waitrunningbufspace();
953 	}
954 
955 	return (0);
956 }
957 
958 void
959 bufbdflush(struct bufobj *bo, struct buf *bp)
960 {
961 	struct buf *nbp;
962 
963 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
964 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
965 		altbufferflushes++;
966 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
967 		BO_LOCK(bo);
968 		/*
969 		 * Try to find a buffer to flush.
970 		 */
971 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
972 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
973 			    BUF_LOCK(nbp,
974 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
975 				continue;
976 			if (bp == nbp)
977 				panic("bdwrite: found ourselves");
978 			BO_UNLOCK(bo);
979 			/* Don't countdeps with the bo lock held. */
980 			if (buf_countdeps(nbp, 0)) {
981 				BO_LOCK(bo);
982 				BUF_UNLOCK(nbp);
983 				continue;
984 			}
985 			if (nbp->b_flags & B_CLUSTEROK) {
986 				vfs_bio_awrite(nbp);
987 			} else {
988 				bremfree(nbp);
989 				bawrite(nbp);
990 			}
991 			dirtybufferflushes++;
992 			break;
993 		}
994 		if (nbp == NULL)
995 			BO_UNLOCK(bo);
996 	}
997 }
998 
999 /*
1000  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1001  * anything if the buffer is marked invalid.
1002  *
1003  * Note that since the buffer must be completely valid, we can safely
1004  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1005  * biodone() in order to prevent getblk from writing the buffer
1006  * out synchronously.
1007  */
1008 void
1009 bdwrite(struct buf *bp)
1010 {
1011 	struct thread *td = curthread;
1012 	struct vnode *vp;
1013 	struct bufobj *bo;
1014 
1015 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1016 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1017 	BUF_ASSERT_HELD(bp);
1018 
1019 	if (bp->b_flags & B_INVAL) {
1020 		brelse(bp);
1021 		return;
1022 	}
1023 
1024 	/*
1025 	 * If we have too many dirty buffers, don't create any more.
1026 	 * If we are wildly over our limit, then force a complete
1027 	 * cleanup. Otherwise, just keep the situation from getting
1028 	 * out of control. Note that we have to avoid a recursive
1029 	 * disaster and not try to clean up after our own cleanup!
1030 	 */
1031 	vp = bp->b_vp;
1032 	bo = bp->b_bufobj;
1033 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1034 		td->td_pflags |= TDP_INBDFLUSH;
1035 		BO_BDFLUSH(bo, bp);
1036 		td->td_pflags &= ~TDP_INBDFLUSH;
1037 	} else
1038 		recursiveflushes++;
1039 
1040 	bdirty(bp);
1041 	/*
1042 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1043 	 * true even of NFS now.
1044 	 */
1045 	bp->b_flags |= B_CACHE;
1046 
1047 	/*
1048 	 * This bmap keeps the system from needing to do the bmap later,
1049 	 * perhaps when the system is attempting to do a sync.  Since it
1050 	 * is likely that the indirect block -- or whatever other datastructure
1051 	 * that the filesystem needs is still in memory now, it is a good
1052 	 * thing to do this.  Note also, that if the pageout daemon is
1053 	 * requesting a sync -- there might not be enough memory to do
1054 	 * the bmap then...  So, this is important to do.
1055 	 */
1056 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1057 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1058 	}
1059 
1060 	/*
1061 	 * Set the *dirty* buffer range based upon the VM system dirty
1062 	 * pages.
1063 	 *
1064 	 * Mark the buffer pages as clean.  We need to do this here to
1065 	 * satisfy the vnode_pager and the pageout daemon, so that it
1066 	 * thinks that the pages have been "cleaned".  Note that since
1067 	 * the pages are in a delayed write buffer -- the VFS layer
1068 	 * "will" see that the pages get written out on the next sync,
1069 	 * or perhaps the cluster will be completed.
1070 	 */
1071 	vfs_clean_pages_dirty_buf(bp);
1072 	bqrelse(bp);
1073 
1074 	/*
1075 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
1076 	 * buffers (midpoint between our recovery point and our stall
1077 	 * point).
1078 	 */
1079 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1080 
1081 	/*
1082 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1083 	 * due to the softdep code.
1084 	 */
1085 }
1086 
1087 /*
1088  *	bdirty:
1089  *
1090  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1091  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1092  *	itself to properly update it in the dirty/clean lists.  We mark it
1093  *	B_DONE to ensure that any asynchronization of the buffer properly
1094  *	clears B_DONE ( else a panic will occur later ).
1095  *
1096  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1097  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1098  *	should only be called if the buffer is known-good.
1099  *
1100  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1101  *	count.
1102  *
1103  *	The buffer must be on QUEUE_NONE.
1104  */
1105 void
1106 bdirty(struct buf *bp)
1107 {
1108 
1109 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1110 	    bp, bp->b_vp, bp->b_flags);
1111 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1112 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1113 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1114 	BUF_ASSERT_HELD(bp);
1115 	bp->b_flags &= ~(B_RELBUF);
1116 	bp->b_iocmd = BIO_WRITE;
1117 
1118 	if ((bp->b_flags & B_DELWRI) == 0) {
1119 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1120 		reassignbuf(bp);
1121 		atomic_add_int(&numdirtybuffers, 1);
1122 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1123 	}
1124 }
1125 
1126 /*
1127  *	bundirty:
1128  *
1129  *	Clear B_DELWRI for buffer.
1130  *
1131  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1132  *	count.
1133  *
1134  *	The buffer must be on QUEUE_NONE.
1135  */
1136 
1137 void
1138 bundirty(struct buf *bp)
1139 {
1140 
1141 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1142 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1143 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1144 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1145 	BUF_ASSERT_HELD(bp);
1146 
1147 	if (bp->b_flags & B_DELWRI) {
1148 		bp->b_flags &= ~B_DELWRI;
1149 		reassignbuf(bp);
1150 		atomic_subtract_int(&numdirtybuffers, 1);
1151 		numdirtywakeup(lodirtybuffers);
1152 	}
1153 	/*
1154 	 * Since it is now being written, we can clear its deferred write flag.
1155 	 */
1156 	bp->b_flags &= ~B_DEFERRED;
1157 }
1158 
1159 /*
1160  *	bawrite:
1161  *
1162  *	Asynchronous write.  Start output on a buffer, but do not wait for
1163  *	it to complete.  The buffer is released when the output completes.
1164  *
1165  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1166  *	B_INVAL buffers.  Not us.
1167  */
1168 void
1169 bawrite(struct buf *bp)
1170 {
1171 
1172 	bp->b_flags |= B_ASYNC;
1173 	(void) bwrite(bp);
1174 }
1175 
1176 /*
1177  *	bwillwrite:
1178  *
1179  *	Called prior to the locking of any vnodes when we are expecting to
1180  *	write.  We do not want to starve the buffer cache with too many
1181  *	dirty buffers so we block here.  By blocking prior to the locking
1182  *	of any vnodes we attempt to avoid the situation where a locked vnode
1183  *	prevents the various system daemons from flushing related buffers.
1184  */
1185 
1186 void
1187 bwillwrite(void)
1188 {
1189 
1190 	if (numdirtybuffers >= hidirtybuffers) {
1191 		mtx_lock(&nblock);
1192 		while (numdirtybuffers >= hidirtybuffers) {
1193 			bd_wakeup(1);
1194 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1195 			msleep(&needsbuffer, &nblock,
1196 			    (PRIBIO + 4), "flswai", 0);
1197 		}
1198 		mtx_unlock(&nblock);
1199 	}
1200 }
1201 
1202 /*
1203  * Return true if we have too many dirty buffers.
1204  */
1205 int
1206 buf_dirty_count_severe(void)
1207 {
1208 
1209 	return(numdirtybuffers >= hidirtybuffers);
1210 }
1211 
1212 static __noinline int
1213 buf_vm_page_count_severe(void)
1214 {
1215 
1216 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1217 
1218 	return vm_page_count_severe();
1219 }
1220 
1221 /*
1222  *	brelse:
1223  *
1224  *	Release a busy buffer and, if requested, free its resources.  The
1225  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1226  *	to be accessed later as a cache entity or reused for other purposes.
1227  */
1228 void
1229 brelse(struct buf *bp)
1230 {
1231 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1232 	    bp, bp->b_vp, bp->b_flags);
1233 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1234 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1235 
1236 	if (bp->b_flags & B_MANAGED) {
1237 		bqrelse(bp);
1238 		return;
1239 	}
1240 
1241 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1242 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1243 		/*
1244 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1245 		 * pages from being scrapped.  If the error is anything
1246 		 * other than an I/O error (EIO), assume that retrying
1247 		 * is futile.
1248 		 */
1249 		bp->b_ioflags &= ~BIO_ERROR;
1250 		bdirty(bp);
1251 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1252 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1253 		/*
1254 		 * Either a failed I/O or we were asked to free or not
1255 		 * cache the buffer.
1256 		 */
1257 		bp->b_flags |= B_INVAL;
1258 		if (!LIST_EMPTY(&bp->b_dep))
1259 			buf_deallocate(bp);
1260 		if (bp->b_flags & B_DELWRI) {
1261 			atomic_subtract_int(&numdirtybuffers, 1);
1262 			numdirtywakeup(lodirtybuffers);
1263 		}
1264 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1265 		if ((bp->b_flags & B_VMIO) == 0) {
1266 			if (bp->b_bufsize)
1267 				allocbuf(bp, 0);
1268 			if (bp->b_vp)
1269 				brelvp(bp);
1270 		}
1271 	}
1272 
1273 	/*
1274 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1275 	 * is called with B_DELWRI set, the underlying pages may wind up
1276 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1277 	 * because pages associated with a B_DELWRI bp are marked clean.
1278 	 *
1279 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1280 	 * if B_DELWRI is set.
1281 	 *
1282 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1283 	 * on pages to return pages to the VM page queues.
1284 	 */
1285 	if (bp->b_flags & B_DELWRI)
1286 		bp->b_flags &= ~B_RELBUF;
1287 	else if (buf_vm_page_count_severe()) {
1288 		/*
1289 		 * The locking of the BO_LOCK is not necessary since
1290 		 * BKGRDINPROG cannot be set while we hold the buf
1291 		 * lock, it can only be cleared if it is already
1292 		 * pending.
1293 		 */
1294 		if (bp->b_vp) {
1295 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1296 				bp->b_flags |= B_RELBUF;
1297 		} else
1298 			bp->b_flags |= B_RELBUF;
1299 	}
1300 
1301 	/*
1302 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1303 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1304 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1305 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1306 	 *
1307 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1308 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1309 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1310 	 *
1311 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1312 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1313 	 * the commit state and we cannot afford to lose the buffer. If the
1314 	 * buffer has a background write in progress, we need to keep it
1315 	 * around to prevent it from being reconstituted and starting a second
1316 	 * background write.
1317 	 */
1318 	if ((bp->b_flags & B_VMIO)
1319 	    && !(bp->b_vp->v_mount != NULL &&
1320 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1321 		 !vn_isdisk(bp->b_vp, NULL) &&
1322 		 (bp->b_flags & B_DELWRI))
1323 	    ) {
1324 
1325 		int i, j, resid;
1326 		vm_page_t m;
1327 		off_t foff;
1328 		vm_pindex_t poff;
1329 		vm_object_t obj;
1330 
1331 		obj = bp->b_bufobj->bo_object;
1332 
1333 		/*
1334 		 * Get the base offset and length of the buffer.  Note that
1335 		 * in the VMIO case if the buffer block size is not
1336 		 * page-aligned then b_data pointer may not be page-aligned.
1337 		 * But our b_pages[] array *IS* page aligned.
1338 		 *
1339 		 * block sizes less then DEV_BSIZE (usually 512) are not
1340 		 * supported due to the page granularity bits (m->valid,
1341 		 * m->dirty, etc...).
1342 		 *
1343 		 * See man buf(9) for more information
1344 		 */
1345 		resid = bp->b_bufsize;
1346 		foff = bp->b_offset;
1347 		VM_OBJECT_LOCK(obj);
1348 		for (i = 0; i < bp->b_npages; i++) {
1349 			int had_bogus = 0;
1350 
1351 			m = bp->b_pages[i];
1352 
1353 			/*
1354 			 * If we hit a bogus page, fixup *all* the bogus pages
1355 			 * now.
1356 			 */
1357 			if (m == bogus_page) {
1358 				poff = OFF_TO_IDX(bp->b_offset);
1359 				had_bogus = 1;
1360 
1361 				for (j = i; j < bp->b_npages; j++) {
1362 					vm_page_t mtmp;
1363 					mtmp = bp->b_pages[j];
1364 					if (mtmp == bogus_page) {
1365 						mtmp = vm_page_lookup(obj, poff + j);
1366 						if (!mtmp) {
1367 							panic("brelse: page missing\n");
1368 						}
1369 						bp->b_pages[j] = mtmp;
1370 					}
1371 				}
1372 
1373 				if ((bp->b_flags & B_INVAL) == 0) {
1374 					pmap_qenter(
1375 					    trunc_page((vm_offset_t)bp->b_data),
1376 					    bp->b_pages, bp->b_npages);
1377 				}
1378 				m = bp->b_pages[i];
1379 			}
1380 			if ((bp->b_flags & B_NOCACHE) ||
1381 			    (bp->b_ioflags & BIO_ERROR &&
1382 			     bp->b_iocmd == BIO_READ)) {
1383 				int poffset = foff & PAGE_MASK;
1384 				int presid = resid > (PAGE_SIZE - poffset) ?
1385 					(PAGE_SIZE - poffset) : resid;
1386 
1387 				KASSERT(presid >= 0, ("brelse: extra page"));
1388 				vm_page_set_invalid(m, poffset, presid);
1389 				if (had_bogus)
1390 					printf("avoided corruption bug in bogus_page/brelse code\n");
1391 			}
1392 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1393 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1394 		}
1395 		VM_OBJECT_UNLOCK(obj);
1396 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1397 			vfs_vmio_release(bp);
1398 
1399 	} else if (bp->b_flags & B_VMIO) {
1400 
1401 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1402 			vfs_vmio_release(bp);
1403 		}
1404 
1405 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1406 		if (bp->b_bufsize != 0)
1407 			allocbuf(bp, 0);
1408 		if (bp->b_vp != NULL)
1409 			brelvp(bp);
1410 	}
1411 
1412 	if (BUF_LOCKRECURSED(bp)) {
1413 		/* do not release to free list */
1414 		BUF_UNLOCK(bp);
1415 		return;
1416 	}
1417 
1418 	/* enqueue */
1419 	mtx_lock(&bqlock);
1420 	/* Handle delayed bremfree() processing. */
1421 	if (bp->b_flags & B_REMFREE) {
1422 		struct bufobj *bo;
1423 
1424 		bo = bp->b_bufobj;
1425 		if (bo != NULL)
1426 			BO_LOCK(bo);
1427 		bremfreel(bp);
1428 		if (bo != NULL)
1429 			BO_UNLOCK(bo);
1430 	}
1431 	if (bp->b_qindex != QUEUE_NONE)
1432 		panic("brelse: free buffer onto another queue???");
1433 
1434 	/*
1435 	 * If the buffer has junk contents signal it and eventually
1436 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1437 	 * doesn't find it.
1438 	 */
1439 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1440 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1441 		bp->b_flags |= B_INVAL;
1442 	if (bp->b_flags & B_INVAL) {
1443 		if (bp->b_flags & B_DELWRI)
1444 			bundirty(bp);
1445 		if (bp->b_vp)
1446 			brelvp(bp);
1447 	}
1448 
1449 	/* buffers with no memory */
1450 	if (bp->b_bufsize == 0) {
1451 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1452 		if (bp->b_vflags & BV_BKGRDINPROG)
1453 			panic("losing buffer 1");
1454 		if (bp->b_kvasize) {
1455 			bp->b_qindex = QUEUE_EMPTYKVA;
1456 		} else {
1457 			bp->b_qindex = QUEUE_EMPTY;
1458 		}
1459 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1460 	/* buffers with junk contents */
1461 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1462 	    (bp->b_ioflags & BIO_ERROR)) {
1463 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1464 		if (bp->b_vflags & BV_BKGRDINPROG)
1465 			panic("losing buffer 2");
1466 		bp->b_qindex = QUEUE_CLEAN;
1467 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1468 	/* remaining buffers */
1469 	} else {
1470 		if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1471 		    (B_DELWRI|B_NEEDSGIANT))
1472 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1473 		else if (bp->b_flags & B_DELWRI)
1474 			bp->b_qindex = QUEUE_DIRTY;
1475 		else
1476 			bp->b_qindex = QUEUE_CLEAN;
1477 		if (bp->b_flags & B_AGE)
1478 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1479 		else
1480 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1481 	}
1482 	mtx_unlock(&bqlock);
1483 
1484 	/*
1485 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1486 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1487 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1488 	 * if B_INVAL is set ).
1489 	 */
1490 
1491 	if (!(bp->b_flags & B_DELWRI)) {
1492 		struct bufobj *bo;
1493 
1494 		bo = bp->b_bufobj;
1495 		if (bo != NULL)
1496 			BO_LOCK(bo);
1497 		bufcountwakeup(bp);
1498 		if (bo != NULL)
1499 			BO_UNLOCK(bo);
1500 	}
1501 
1502 	/*
1503 	 * Something we can maybe free or reuse
1504 	 */
1505 	if (bp->b_bufsize || bp->b_kvasize)
1506 		bufspacewakeup();
1507 
1508 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1509 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1510 		panic("brelse: not dirty");
1511 	/* unlock */
1512 	BUF_UNLOCK(bp);
1513 }
1514 
1515 /*
1516  * Release a buffer back to the appropriate queue but do not try to free
1517  * it.  The buffer is expected to be used again soon.
1518  *
1519  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1520  * biodone() to requeue an async I/O on completion.  It is also used when
1521  * known good buffers need to be requeued but we think we may need the data
1522  * again soon.
1523  *
1524  * XXX we should be able to leave the B_RELBUF hint set on completion.
1525  */
1526 void
1527 bqrelse(struct buf *bp)
1528 {
1529 	struct bufobj *bo;
1530 
1531 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1532 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1533 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1534 
1535 	if (BUF_LOCKRECURSED(bp)) {
1536 		/* do not release to free list */
1537 		BUF_UNLOCK(bp);
1538 		return;
1539 	}
1540 
1541 	bo = bp->b_bufobj;
1542 	if (bp->b_flags & B_MANAGED) {
1543 		if (bp->b_flags & B_REMFREE) {
1544 			mtx_lock(&bqlock);
1545 			if (bo != NULL)
1546 				BO_LOCK(bo);
1547 			bremfreel(bp);
1548 			if (bo != NULL)
1549 				BO_UNLOCK(bo);
1550 			mtx_unlock(&bqlock);
1551 		}
1552 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1553 		BUF_UNLOCK(bp);
1554 		return;
1555 	}
1556 
1557 	mtx_lock(&bqlock);
1558 	/* Handle delayed bremfree() processing. */
1559 	if (bp->b_flags & B_REMFREE) {
1560 		if (bo != NULL)
1561 			BO_LOCK(bo);
1562 		bremfreel(bp);
1563 		if (bo != NULL)
1564 			BO_UNLOCK(bo);
1565 	}
1566 	if (bp->b_qindex != QUEUE_NONE)
1567 		panic("bqrelse: free buffer onto another queue???");
1568 	/* buffers with stale but valid contents */
1569 	if (bp->b_flags & B_DELWRI) {
1570 		if (bp->b_flags & B_NEEDSGIANT)
1571 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1572 		else
1573 			bp->b_qindex = QUEUE_DIRTY;
1574 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1575 	} else {
1576 		/*
1577 		 * The locking of the BO_LOCK for checking of the
1578 		 * BV_BKGRDINPROG is not necessary since the
1579 		 * BV_BKGRDINPROG cannot be set while we hold the buf
1580 		 * lock, it can only be cleared if it is already
1581 		 * pending.
1582 		 */
1583 		if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1584 			bp->b_qindex = QUEUE_CLEAN;
1585 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1586 			    b_freelist);
1587 		} else {
1588 			/*
1589 			 * We are too low on memory, we have to try to free
1590 			 * the buffer (most importantly: the wired pages
1591 			 * making up its backing store) *now*.
1592 			 */
1593 			mtx_unlock(&bqlock);
1594 			brelse(bp);
1595 			return;
1596 		}
1597 	}
1598 	mtx_unlock(&bqlock);
1599 
1600 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) {
1601 		if (bo != NULL)
1602 			BO_LOCK(bo);
1603 		bufcountwakeup(bp);
1604 		if (bo != NULL)
1605 			BO_UNLOCK(bo);
1606 	}
1607 
1608 	/*
1609 	 * Something we can maybe free or reuse.
1610 	 */
1611 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1612 		bufspacewakeup();
1613 
1614 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1615 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1616 		panic("bqrelse: not dirty");
1617 	/* unlock */
1618 	BUF_UNLOCK(bp);
1619 }
1620 
1621 /* Give pages used by the bp back to the VM system (where possible) */
1622 static void
1623 vfs_vmio_release(struct buf *bp)
1624 {
1625 	int i;
1626 	vm_page_t m;
1627 
1628 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1629 	for (i = 0; i < bp->b_npages; i++) {
1630 		m = bp->b_pages[i];
1631 		bp->b_pages[i] = NULL;
1632 		/*
1633 		 * In order to keep page LRU ordering consistent, put
1634 		 * everything on the inactive queue.
1635 		 */
1636 		vm_page_lock(m);
1637 		vm_page_unwire(m, 0);
1638 		/*
1639 		 * We don't mess with busy pages, it is
1640 		 * the responsibility of the process that
1641 		 * busied the pages to deal with them.
1642 		 */
1643 		if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 &&
1644 		    m->wire_count == 0) {
1645 			/*
1646 			 * Might as well free the page if we can and it has
1647 			 * no valid data.  We also free the page if the
1648 			 * buffer was used for direct I/O
1649 			 */
1650 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1651 			    m->hold_count == 0) {
1652 				vm_page_free(m);
1653 			} else if (bp->b_flags & B_DIRECT) {
1654 				vm_page_try_to_free(m);
1655 			} else if (buf_vm_page_count_severe()) {
1656 				vm_page_try_to_cache(m);
1657 			}
1658 		}
1659 		vm_page_unlock(m);
1660 	}
1661 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1662 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1663 
1664 	if (bp->b_bufsize) {
1665 		bufspacewakeup();
1666 		bp->b_bufsize = 0;
1667 	}
1668 	bp->b_npages = 0;
1669 	bp->b_flags &= ~B_VMIO;
1670 	if (bp->b_vp)
1671 		brelvp(bp);
1672 }
1673 
1674 /*
1675  * Check to see if a block at a particular lbn is available for a clustered
1676  * write.
1677  */
1678 static int
1679 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1680 {
1681 	struct buf *bpa;
1682 	int match;
1683 
1684 	match = 0;
1685 
1686 	/* If the buf isn't in core skip it */
1687 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1688 		return (0);
1689 
1690 	/* If the buf is busy we don't want to wait for it */
1691 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1692 		return (0);
1693 
1694 	/* Only cluster with valid clusterable delayed write buffers */
1695 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1696 	    (B_DELWRI | B_CLUSTEROK))
1697 		goto done;
1698 
1699 	if (bpa->b_bufsize != size)
1700 		goto done;
1701 
1702 	/*
1703 	 * Check to see if it is in the expected place on disk and that the
1704 	 * block has been mapped.
1705 	 */
1706 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1707 		match = 1;
1708 done:
1709 	BUF_UNLOCK(bpa);
1710 	return (match);
1711 }
1712 
1713 /*
1714  *	vfs_bio_awrite:
1715  *
1716  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1717  *	This is much better then the old way of writing only one buffer at
1718  *	a time.  Note that we may not be presented with the buffers in the
1719  *	correct order, so we search for the cluster in both directions.
1720  */
1721 int
1722 vfs_bio_awrite(struct buf *bp)
1723 {
1724 	struct bufobj *bo;
1725 	int i;
1726 	int j;
1727 	daddr_t lblkno = bp->b_lblkno;
1728 	struct vnode *vp = bp->b_vp;
1729 	int ncl;
1730 	int nwritten;
1731 	int size;
1732 	int maxcl;
1733 
1734 	bo = &vp->v_bufobj;
1735 	/*
1736 	 * right now we support clustered writing only to regular files.  If
1737 	 * we find a clusterable block we could be in the middle of a cluster
1738 	 * rather then at the beginning.
1739 	 */
1740 	if ((vp->v_type == VREG) &&
1741 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1742 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1743 
1744 		size = vp->v_mount->mnt_stat.f_iosize;
1745 		maxcl = MAXPHYS / size;
1746 
1747 		BO_LOCK(bo);
1748 		for (i = 1; i < maxcl; i++)
1749 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1750 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1751 				break;
1752 
1753 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1754 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1755 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1756 				break;
1757 		BO_UNLOCK(bo);
1758 		--j;
1759 		ncl = i + j;
1760 		/*
1761 		 * this is a possible cluster write
1762 		 */
1763 		if (ncl != 1) {
1764 			BUF_UNLOCK(bp);
1765 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1766 			return nwritten;
1767 		}
1768 	}
1769 	bremfree(bp);
1770 	bp->b_flags |= B_ASYNC;
1771 	/*
1772 	 * default (old) behavior, writing out only one block
1773 	 *
1774 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1775 	 */
1776 	nwritten = bp->b_bufsize;
1777 	(void) bwrite(bp);
1778 
1779 	return nwritten;
1780 }
1781 
1782 /*
1783  *	getnewbuf:
1784  *
1785  *	Find and initialize a new buffer header, freeing up existing buffers
1786  *	in the bufqueues as necessary.  The new buffer is returned locked.
1787  *
1788  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1789  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1790  *
1791  *	We block if:
1792  *		We have insufficient buffer headers
1793  *		We have insufficient buffer space
1794  *		buffer_map is too fragmented ( space reservation fails )
1795  *		If we have to flush dirty buffers ( but we try to avoid this )
1796  *
1797  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1798  *	Instead we ask the buf daemon to do it for us.  We attempt to
1799  *	avoid piecemeal wakeups of the pageout daemon.
1800  */
1801 
1802 static struct buf *
1803 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1804     int gbflags)
1805 {
1806 	struct thread *td;
1807 	struct buf *bp;
1808 	struct buf *nbp;
1809 	int defrag = 0;
1810 	int nqindex;
1811 	static int flushingbufs;
1812 
1813 	td = curthread;
1814 	/*
1815 	 * We can't afford to block since we might be holding a vnode lock,
1816 	 * which may prevent system daemons from running.  We deal with
1817 	 * low-memory situations by proactively returning memory and running
1818 	 * async I/O rather then sync I/O.
1819 	 */
1820 	atomic_add_int(&getnewbufcalls, 1);
1821 	atomic_subtract_int(&getnewbufrestarts, 1);
1822 restart:
1823 	atomic_add_int(&getnewbufrestarts, 1);
1824 
1825 	/*
1826 	 * Setup for scan.  If we do not have enough free buffers,
1827 	 * we setup a degenerate case that immediately fails.  Note
1828 	 * that if we are specially marked process, we are allowed to
1829 	 * dip into our reserves.
1830 	 *
1831 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1832 	 *
1833 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1834 	 * However, there are a number of cases (defragging, reusing, ...)
1835 	 * where we cannot backup.
1836 	 */
1837 	mtx_lock(&bqlock);
1838 	nqindex = QUEUE_EMPTYKVA;
1839 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1840 
1841 	if (nbp == NULL) {
1842 		/*
1843 		 * If no EMPTYKVA buffers and we are either
1844 		 * defragging or reusing, locate a CLEAN buffer
1845 		 * to free or reuse.  If bufspace useage is low
1846 		 * skip this step so we can allocate a new buffer.
1847 		 */
1848 		if (defrag || bufspace >= lobufspace) {
1849 			nqindex = QUEUE_CLEAN;
1850 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1851 		}
1852 
1853 		/*
1854 		 * If we could not find or were not allowed to reuse a
1855 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1856 		 * buffer.  We can only use an EMPTY buffer if allocating
1857 		 * its KVA would not otherwise run us out of buffer space.
1858 		 */
1859 		if (nbp == NULL && defrag == 0 &&
1860 		    bufspace + maxsize < hibufspace) {
1861 			nqindex = QUEUE_EMPTY;
1862 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1863 		}
1864 	}
1865 
1866 	/*
1867 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1868 	 * depending.
1869 	 */
1870 
1871 	while ((bp = nbp) != NULL) {
1872 		int qindex = nqindex;
1873 
1874 		/*
1875 		 * Calculate next bp ( we can only use it if we do not block
1876 		 * or do other fancy things ).
1877 		 */
1878 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1879 			switch(qindex) {
1880 			case QUEUE_EMPTY:
1881 				nqindex = QUEUE_EMPTYKVA;
1882 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1883 					break;
1884 				/* FALLTHROUGH */
1885 			case QUEUE_EMPTYKVA:
1886 				nqindex = QUEUE_CLEAN;
1887 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1888 					break;
1889 				/* FALLTHROUGH */
1890 			case QUEUE_CLEAN:
1891 				/*
1892 				 * nbp is NULL.
1893 				 */
1894 				break;
1895 			}
1896 		}
1897 		/*
1898 		 * If we are defragging then we need a buffer with
1899 		 * b_kvasize != 0.  XXX this situation should no longer
1900 		 * occur, if defrag is non-zero the buffer's b_kvasize
1901 		 * should also be non-zero at this point.  XXX
1902 		 */
1903 		if (defrag && bp->b_kvasize == 0) {
1904 			printf("Warning: defrag empty buffer %p\n", bp);
1905 			continue;
1906 		}
1907 
1908 		/*
1909 		 * Start freeing the bp.  This is somewhat involved.  nbp
1910 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1911 		 */
1912 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1913 			continue;
1914 		if (bp->b_vp) {
1915 			BO_LOCK(bp->b_bufobj);
1916 			if (bp->b_vflags & BV_BKGRDINPROG) {
1917 				BO_UNLOCK(bp->b_bufobj);
1918 				BUF_UNLOCK(bp);
1919 				continue;
1920 			}
1921 			BO_UNLOCK(bp->b_bufobj);
1922 		}
1923 		CTR6(KTR_BUF,
1924 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1925 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1926 		    bp->b_kvasize, bp->b_bufsize, qindex);
1927 
1928 		/*
1929 		 * Sanity Checks
1930 		 */
1931 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1932 
1933 		/*
1934 		 * Note: we no longer distinguish between VMIO and non-VMIO
1935 		 * buffers.
1936 		 */
1937 
1938 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1939 
1940 		if (bp->b_bufobj != NULL)
1941 			BO_LOCK(bp->b_bufobj);
1942 		bremfreel(bp);
1943 		if (bp->b_bufobj != NULL)
1944 			BO_UNLOCK(bp->b_bufobj);
1945 		mtx_unlock(&bqlock);
1946 
1947 		if (qindex == QUEUE_CLEAN) {
1948 			if (bp->b_flags & B_VMIO) {
1949 				bp->b_flags &= ~B_ASYNC;
1950 				vfs_vmio_release(bp);
1951 			}
1952 			if (bp->b_vp)
1953 				brelvp(bp);
1954 		}
1955 
1956 		/*
1957 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1958 		 * the scan from this point on.
1959 		 *
1960 		 * Get the rest of the buffer freed up.  b_kva* is still
1961 		 * valid after this operation.
1962 		 */
1963 
1964 		if (bp->b_rcred != NOCRED) {
1965 			crfree(bp->b_rcred);
1966 			bp->b_rcred = NOCRED;
1967 		}
1968 		if (bp->b_wcred != NOCRED) {
1969 			crfree(bp->b_wcred);
1970 			bp->b_wcred = NOCRED;
1971 		}
1972 		if (!LIST_EMPTY(&bp->b_dep))
1973 			buf_deallocate(bp);
1974 		if (bp->b_vflags & BV_BKGRDINPROG)
1975 			panic("losing buffer 3");
1976 		KASSERT(bp->b_vp == NULL,
1977 		    ("bp: %p still has vnode %p.  qindex: %d",
1978 		    bp, bp->b_vp, qindex));
1979 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1980 		   ("bp: %p still on a buffer list. xflags %X",
1981 		    bp, bp->b_xflags));
1982 
1983 		if (bp->b_bufsize)
1984 			allocbuf(bp, 0);
1985 
1986 		bp->b_flags = 0;
1987 		bp->b_ioflags = 0;
1988 		bp->b_xflags = 0;
1989 		KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
1990 		    ("buf %p still counted as free?", bp));
1991 		bp->b_vflags = 0;
1992 		bp->b_vp = NULL;
1993 		bp->b_blkno = bp->b_lblkno = 0;
1994 		bp->b_offset = NOOFFSET;
1995 		bp->b_iodone = 0;
1996 		bp->b_error = 0;
1997 		bp->b_resid = 0;
1998 		bp->b_bcount = 0;
1999 		bp->b_npages = 0;
2000 		bp->b_dirtyoff = bp->b_dirtyend = 0;
2001 		bp->b_bufobj = NULL;
2002 		bp->b_pin_count = 0;
2003 		bp->b_fsprivate1 = NULL;
2004 		bp->b_fsprivate2 = NULL;
2005 		bp->b_fsprivate3 = NULL;
2006 
2007 		LIST_INIT(&bp->b_dep);
2008 
2009 		/*
2010 		 * If we are defragging then free the buffer.
2011 		 */
2012 		if (defrag) {
2013 			bp->b_flags |= B_INVAL;
2014 			bfreekva(bp);
2015 			brelse(bp);
2016 			defrag = 0;
2017 			goto restart;
2018 		}
2019 
2020 		/*
2021 		 * Notify any waiters for the buffer lock about
2022 		 * identity change by freeing the buffer.
2023 		 */
2024 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2025 			bp->b_flags |= B_INVAL;
2026 			bfreekva(bp);
2027 			brelse(bp);
2028 			goto restart;
2029 		}
2030 
2031 		/*
2032 		 * If we are overcomitted then recover the buffer and its
2033 		 * KVM space.  This occurs in rare situations when multiple
2034 		 * processes are blocked in getnewbuf() or allocbuf().
2035 		 */
2036 		if (bufspace >= hibufspace)
2037 			flushingbufs = 1;
2038 		if (flushingbufs && bp->b_kvasize != 0) {
2039 			bp->b_flags |= B_INVAL;
2040 			bfreekva(bp);
2041 			brelse(bp);
2042 			goto restart;
2043 		}
2044 		if (bufspace < lobufspace)
2045 			flushingbufs = 0;
2046 		break;
2047 	}
2048 
2049 	/*
2050 	 * If we exhausted our list, sleep as appropriate.  We may have to
2051 	 * wakeup various daemons and write out some dirty buffers.
2052 	 *
2053 	 * Generally we are sleeping due to insufficient buffer space.
2054 	 */
2055 
2056 	if (bp == NULL) {
2057 		int flags, norunbuf;
2058 		char *waitmsg;
2059 		int fl;
2060 
2061 		if (defrag) {
2062 			flags = VFS_BIO_NEED_BUFSPACE;
2063 			waitmsg = "nbufkv";
2064 		} else if (bufspace >= hibufspace) {
2065 			waitmsg = "nbufbs";
2066 			flags = VFS_BIO_NEED_BUFSPACE;
2067 		} else {
2068 			waitmsg = "newbuf";
2069 			flags = VFS_BIO_NEED_ANY;
2070 		}
2071 		mtx_lock(&nblock);
2072 		needsbuffer |= flags;
2073 		mtx_unlock(&nblock);
2074 		mtx_unlock(&bqlock);
2075 
2076 		bd_speedup();	/* heeeelp */
2077 		if (gbflags & GB_NOWAIT_BD)
2078 			return (NULL);
2079 
2080 		mtx_lock(&nblock);
2081 		while (needsbuffer & flags) {
2082 			if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2083 				mtx_unlock(&nblock);
2084 				/*
2085 				 * getblk() is called with a vnode
2086 				 * locked, and some majority of the
2087 				 * dirty buffers may as well belong to
2088 				 * the vnode. Flushing the buffers
2089 				 * there would make a progress that
2090 				 * cannot be achieved by the
2091 				 * buf_daemon, that cannot lock the
2092 				 * vnode.
2093 				 */
2094 				norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2095 				    (td->td_pflags & TDP_NORUNNINGBUF);
2096 				/* play bufdaemon */
2097 				td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2098 				fl = buf_do_flush(vp);
2099 				td->td_pflags &= norunbuf;
2100 				mtx_lock(&nblock);
2101 				if (fl != 0)
2102 					continue;
2103 				if ((needsbuffer & flags) == 0)
2104 					break;
2105 			}
2106 			if (msleep(&needsbuffer, &nblock,
2107 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2108 				mtx_unlock(&nblock);
2109 				return (NULL);
2110 			}
2111 		}
2112 		mtx_unlock(&nblock);
2113 	} else {
2114 		/*
2115 		 * We finally have a valid bp.  We aren't quite out of the
2116 		 * woods, we still have to reserve kva space.  In order
2117 		 * to keep fragmentation sane we only allocate kva in
2118 		 * BKVASIZE chunks.
2119 		 */
2120 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2121 
2122 		if (maxsize != bp->b_kvasize) {
2123 			vm_offset_t addr = 0;
2124 
2125 			bfreekva(bp);
2126 
2127 			vm_map_lock(buffer_map);
2128 			if (vm_map_findspace(buffer_map,
2129 				vm_map_min(buffer_map), maxsize, &addr)) {
2130 				/*
2131 				 * Uh oh.  Buffer map is to fragmented.  We
2132 				 * must defragment the map.
2133 				 */
2134 				atomic_add_int(&bufdefragcnt, 1);
2135 				vm_map_unlock(buffer_map);
2136 				defrag = 1;
2137 				bp->b_flags |= B_INVAL;
2138 				brelse(bp);
2139 				goto restart;
2140 			}
2141 			if (addr) {
2142 				vm_map_insert(buffer_map, NULL, 0,
2143 					addr, addr + maxsize,
2144 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2145 
2146 				bp->b_kvabase = (caddr_t) addr;
2147 				bp->b_kvasize = maxsize;
2148 				atomic_add_long(&bufspace, bp->b_kvasize);
2149 				atomic_add_int(&bufreusecnt, 1);
2150 			}
2151 			vm_map_unlock(buffer_map);
2152 		}
2153 		bp->b_saveaddr = bp->b_kvabase;
2154 		bp->b_data = bp->b_saveaddr;
2155 	}
2156 	return(bp);
2157 }
2158 
2159 /*
2160  *	buf_daemon:
2161  *
2162  *	buffer flushing daemon.  Buffers are normally flushed by the
2163  *	update daemon but if it cannot keep up this process starts to
2164  *	take the load in an attempt to prevent getnewbuf() from blocking.
2165  */
2166 
2167 static struct kproc_desc buf_kp = {
2168 	"bufdaemon",
2169 	buf_daemon,
2170 	&bufdaemonproc
2171 };
2172 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2173 
2174 static int
2175 buf_do_flush(struct vnode *vp)
2176 {
2177 	int flushed;
2178 
2179 	flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2180 	/* The list empty check here is slightly racy */
2181 	if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2182 		mtx_lock(&Giant);
2183 		flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0);
2184 		mtx_unlock(&Giant);
2185 	}
2186 	if (flushed == 0) {
2187 		/*
2188 		 * Could not find any buffers without rollback
2189 		 * dependencies, so just write the first one
2190 		 * in the hopes of eventually making progress.
2191 		 */
2192 		flushbufqueues(vp, QUEUE_DIRTY, 1);
2193 		if (!TAILQ_EMPTY(
2194 			    &bufqueues[QUEUE_DIRTY_GIANT])) {
2195 			mtx_lock(&Giant);
2196 			flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1);
2197 			mtx_unlock(&Giant);
2198 		}
2199 	}
2200 	return (flushed);
2201 }
2202 
2203 static void
2204 buf_daemon()
2205 {
2206 	int lodirtysave;
2207 
2208 	/*
2209 	 * This process needs to be suspended prior to shutdown sync.
2210 	 */
2211 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2212 	    SHUTDOWN_PRI_LAST);
2213 
2214 	/*
2215 	 * This process is allowed to take the buffer cache to the limit
2216 	 */
2217 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2218 	mtx_lock(&bdlock);
2219 	for (;;) {
2220 		bd_request = 0;
2221 		mtx_unlock(&bdlock);
2222 
2223 		kproc_suspend_check(bufdaemonproc);
2224 		lodirtysave = lodirtybuffers;
2225 		if (bd_speedupreq) {
2226 			lodirtybuffers = numdirtybuffers / 2;
2227 			bd_speedupreq = 0;
2228 		}
2229 		/*
2230 		 * Do the flush.  Limit the amount of in-transit I/O we
2231 		 * allow to build up, otherwise we would completely saturate
2232 		 * the I/O system.  Wakeup any waiting processes before we
2233 		 * normally would so they can run in parallel with our drain.
2234 		 */
2235 		while (numdirtybuffers > lodirtybuffers) {
2236 			if (buf_do_flush(NULL) == 0)
2237 				break;
2238 			uio_yield();
2239 		}
2240 		lodirtybuffers = lodirtysave;
2241 
2242 		/*
2243 		 * Only clear bd_request if we have reached our low water
2244 		 * mark.  The buf_daemon normally waits 1 second and
2245 		 * then incrementally flushes any dirty buffers that have
2246 		 * built up, within reason.
2247 		 *
2248 		 * If we were unable to hit our low water mark and couldn't
2249 		 * find any flushable buffers, we sleep half a second.
2250 		 * Otherwise we loop immediately.
2251 		 */
2252 		mtx_lock(&bdlock);
2253 		if (numdirtybuffers <= lodirtybuffers) {
2254 			/*
2255 			 * We reached our low water mark, reset the
2256 			 * request and sleep until we are needed again.
2257 			 * The sleep is just so the suspend code works.
2258 			 */
2259 			bd_request = 0;
2260 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2261 		} else {
2262 			/*
2263 			 * We couldn't find any flushable dirty buffers but
2264 			 * still have too many dirty buffers, we
2265 			 * have to sleep and try again.  (rare)
2266 			 */
2267 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2268 		}
2269 	}
2270 }
2271 
2272 /*
2273  *	flushbufqueues:
2274  *
2275  *	Try to flush a buffer in the dirty queue.  We must be careful to
2276  *	free up B_INVAL buffers instead of write them, which NFS is
2277  *	particularly sensitive to.
2278  */
2279 static int flushwithdeps = 0;
2280 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2281     0, "Number of buffers flushed with dependecies that require rollbacks");
2282 
2283 static int
2284 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2285 {
2286 	struct buf *sentinel;
2287 	struct vnode *vp;
2288 	struct mount *mp;
2289 	struct buf *bp;
2290 	int hasdeps;
2291 	int flushed;
2292 	int target;
2293 
2294 	if (lvp == NULL) {
2295 		target = numdirtybuffers - lodirtybuffers;
2296 		if (flushdeps && target > 2)
2297 			target /= 2;
2298 	} else
2299 		target = flushbufqtarget;
2300 	flushed = 0;
2301 	bp = NULL;
2302 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2303 	sentinel->b_qindex = QUEUE_SENTINEL;
2304 	mtx_lock(&bqlock);
2305 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2306 	while (flushed != target) {
2307 		bp = TAILQ_NEXT(sentinel, b_freelist);
2308 		if (bp != NULL) {
2309 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2310 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2311 			    b_freelist);
2312 		} else
2313 			break;
2314 		/*
2315 		 * Skip sentinels inserted by other invocations of the
2316 		 * flushbufqueues(), taking care to not reorder them.
2317 		 */
2318 		if (bp->b_qindex == QUEUE_SENTINEL)
2319 			continue;
2320 		/*
2321 		 * Only flush the buffers that belong to the
2322 		 * vnode locked by the curthread.
2323 		 */
2324 		if (lvp != NULL && bp->b_vp != lvp)
2325 			continue;
2326 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2327 			continue;
2328 		if (bp->b_pin_count > 0) {
2329 			BUF_UNLOCK(bp);
2330 			continue;
2331 		}
2332 		BO_LOCK(bp->b_bufobj);
2333 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2334 		    (bp->b_flags & B_DELWRI) == 0) {
2335 			BO_UNLOCK(bp->b_bufobj);
2336 			BUF_UNLOCK(bp);
2337 			continue;
2338 		}
2339 		BO_UNLOCK(bp->b_bufobj);
2340 		if (bp->b_flags & B_INVAL) {
2341 			bremfreel(bp);
2342 			mtx_unlock(&bqlock);
2343 			brelse(bp);
2344 			flushed++;
2345 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2346 			mtx_lock(&bqlock);
2347 			continue;
2348 		}
2349 
2350 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2351 			if (flushdeps == 0) {
2352 				BUF_UNLOCK(bp);
2353 				continue;
2354 			}
2355 			hasdeps = 1;
2356 		} else
2357 			hasdeps = 0;
2358 		/*
2359 		 * We must hold the lock on a vnode before writing
2360 		 * one of its buffers. Otherwise we may confuse, or
2361 		 * in the case of a snapshot vnode, deadlock the
2362 		 * system.
2363 		 *
2364 		 * The lock order here is the reverse of the normal
2365 		 * of vnode followed by buf lock.  This is ok because
2366 		 * the NOWAIT will prevent deadlock.
2367 		 */
2368 		vp = bp->b_vp;
2369 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2370 			BUF_UNLOCK(bp);
2371 			continue;
2372 		}
2373 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2374 			mtx_unlock(&bqlock);
2375 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2376 			    bp, bp->b_vp, bp->b_flags);
2377 			if (curproc == bufdaemonproc)
2378 				vfs_bio_awrite(bp);
2379 			else {
2380 				bremfree(bp);
2381 				bwrite(bp);
2382 				notbufdflashes++;
2383 			}
2384 			vn_finished_write(mp);
2385 			VOP_UNLOCK(vp, 0);
2386 			flushwithdeps += hasdeps;
2387 			flushed++;
2388 
2389 			/*
2390 			 * Sleeping on runningbufspace while holding
2391 			 * vnode lock leads to deadlock.
2392 			 */
2393 			if (curproc == bufdaemonproc)
2394 				waitrunningbufspace();
2395 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2396 			mtx_lock(&bqlock);
2397 			continue;
2398 		}
2399 		vn_finished_write(mp);
2400 		BUF_UNLOCK(bp);
2401 	}
2402 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2403 	mtx_unlock(&bqlock);
2404 	free(sentinel, M_TEMP);
2405 	return (flushed);
2406 }
2407 
2408 /*
2409  * Check to see if a block is currently memory resident.
2410  */
2411 struct buf *
2412 incore(struct bufobj *bo, daddr_t blkno)
2413 {
2414 	struct buf *bp;
2415 
2416 	BO_LOCK(bo);
2417 	bp = gbincore(bo, blkno);
2418 	BO_UNLOCK(bo);
2419 	return (bp);
2420 }
2421 
2422 /*
2423  * Returns true if no I/O is needed to access the
2424  * associated VM object.  This is like incore except
2425  * it also hunts around in the VM system for the data.
2426  */
2427 
2428 static int
2429 inmem(struct vnode * vp, daddr_t blkno)
2430 {
2431 	vm_object_t obj;
2432 	vm_offset_t toff, tinc, size;
2433 	vm_page_t m;
2434 	vm_ooffset_t off;
2435 
2436 	ASSERT_VOP_LOCKED(vp, "inmem");
2437 
2438 	if (incore(&vp->v_bufobj, blkno))
2439 		return 1;
2440 	if (vp->v_mount == NULL)
2441 		return 0;
2442 	obj = vp->v_object;
2443 	if (obj == NULL)
2444 		return (0);
2445 
2446 	size = PAGE_SIZE;
2447 	if (size > vp->v_mount->mnt_stat.f_iosize)
2448 		size = vp->v_mount->mnt_stat.f_iosize;
2449 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2450 
2451 	VM_OBJECT_LOCK(obj);
2452 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2453 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2454 		if (!m)
2455 			goto notinmem;
2456 		tinc = size;
2457 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2458 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2459 		if (vm_page_is_valid(m,
2460 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2461 			goto notinmem;
2462 	}
2463 	VM_OBJECT_UNLOCK(obj);
2464 	return 1;
2465 
2466 notinmem:
2467 	VM_OBJECT_UNLOCK(obj);
2468 	return (0);
2469 }
2470 
2471 /*
2472  * Set the dirty range for a buffer based on the status of the dirty
2473  * bits in the pages comprising the buffer.  The range is limited
2474  * to the size of the buffer.
2475  *
2476  * Tell the VM system that the pages associated with this buffer
2477  * are clean.  This is used for delayed writes where the data is
2478  * going to go to disk eventually without additional VM intevention.
2479  *
2480  * Note that while we only really need to clean through to b_bcount, we
2481  * just go ahead and clean through to b_bufsize.
2482  */
2483 static void
2484 vfs_clean_pages_dirty_buf(struct buf *bp)
2485 {
2486 	vm_ooffset_t foff, noff, eoff;
2487 	vm_page_t m;
2488 	int i;
2489 
2490 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2491 		return;
2492 
2493 	foff = bp->b_offset;
2494 	KASSERT(bp->b_offset != NOOFFSET,
2495 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2496 
2497 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2498 	vfs_drain_busy_pages(bp);
2499 	vfs_setdirty_locked_object(bp);
2500 	for (i = 0; i < bp->b_npages; i++) {
2501 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2502 		eoff = noff;
2503 		if (eoff > bp->b_offset + bp->b_bufsize)
2504 			eoff = bp->b_offset + bp->b_bufsize;
2505 		m = bp->b_pages[i];
2506 		vfs_page_set_validclean(bp, foff, m);
2507 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2508 		foff = noff;
2509 	}
2510 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2511 }
2512 
2513 static void
2514 vfs_setdirty_locked_object(struct buf *bp)
2515 {
2516 	vm_object_t object;
2517 	int i;
2518 
2519 	object = bp->b_bufobj->bo_object;
2520 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2521 
2522 	/*
2523 	 * We qualify the scan for modified pages on whether the
2524 	 * object has been flushed yet.
2525 	 */
2526 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2527 		vm_offset_t boffset;
2528 		vm_offset_t eoffset;
2529 
2530 		/*
2531 		 * test the pages to see if they have been modified directly
2532 		 * by users through the VM system.
2533 		 */
2534 		for (i = 0; i < bp->b_npages; i++)
2535 			vm_page_test_dirty(bp->b_pages[i]);
2536 
2537 		/*
2538 		 * Calculate the encompassing dirty range, boffset and eoffset,
2539 		 * (eoffset - boffset) bytes.
2540 		 */
2541 
2542 		for (i = 0; i < bp->b_npages; i++) {
2543 			if (bp->b_pages[i]->dirty)
2544 				break;
2545 		}
2546 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2547 
2548 		for (i = bp->b_npages - 1; i >= 0; --i) {
2549 			if (bp->b_pages[i]->dirty) {
2550 				break;
2551 			}
2552 		}
2553 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2554 
2555 		/*
2556 		 * Fit it to the buffer.
2557 		 */
2558 
2559 		if (eoffset > bp->b_bcount)
2560 			eoffset = bp->b_bcount;
2561 
2562 		/*
2563 		 * If we have a good dirty range, merge with the existing
2564 		 * dirty range.
2565 		 */
2566 
2567 		if (boffset < eoffset) {
2568 			if (bp->b_dirtyoff > boffset)
2569 				bp->b_dirtyoff = boffset;
2570 			if (bp->b_dirtyend < eoffset)
2571 				bp->b_dirtyend = eoffset;
2572 		}
2573 	}
2574 }
2575 
2576 /*
2577  *	getblk:
2578  *
2579  *	Get a block given a specified block and offset into a file/device.
2580  *	The buffers B_DONE bit will be cleared on return, making it almost
2581  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2582  *	return.  The caller should clear B_INVAL prior to initiating a
2583  *	READ.
2584  *
2585  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2586  *	an existing buffer.
2587  *
2588  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2589  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2590  *	and then cleared based on the backing VM.  If the previous buffer is
2591  *	non-0-sized but invalid, B_CACHE will be cleared.
2592  *
2593  *	If getblk() must create a new buffer, the new buffer is returned with
2594  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2595  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2596  *	backing VM.
2597  *
2598  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2599  *	B_CACHE bit is clear.
2600  *
2601  *	What this means, basically, is that the caller should use B_CACHE to
2602  *	determine whether the buffer is fully valid or not and should clear
2603  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2604  *	the buffer by loading its data area with something, the caller needs
2605  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2606  *	the caller should set B_CACHE ( as an optimization ), else the caller
2607  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2608  *	a write attempt or if it was a successfull read.  If the caller
2609  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2610  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2611  */
2612 struct buf *
2613 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2614     int flags)
2615 {
2616 	struct buf *bp;
2617 	struct bufobj *bo;
2618 	int error;
2619 
2620 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2621 	ASSERT_VOP_LOCKED(vp, "getblk");
2622 	if (size > MAXBSIZE)
2623 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2624 
2625 	bo = &vp->v_bufobj;
2626 loop:
2627 	/*
2628 	 * Block if we are low on buffers.   Certain processes are allowed
2629 	 * to completely exhaust the buffer cache.
2630          *
2631          * If this check ever becomes a bottleneck it may be better to
2632          * move it into the else, when gbincore() fails.  At the moment
2633          * it isn't a problem.
2634 	 *
2635 	 * XXX remove if 0 sections (clean this up after its proven)
2636          */
2637 	if (numfreebuffers == 0) {
2638 		if (TD_IS_IDLETHREAD(curthread))
2639 			return NULL;
2640 		mtx_lock(&nblock);
2641 		needsbuffer |= VFS_BIO_NEED_ANY;
2642 		mtx_unlock(&nblock);
2643 	}
2644 
2645 	BO_LOCK(bo);
2646 	bp = gbincore(bo, blkno);
2647 	if (bp != NULL) {
2648 		int lockflags;
2649 		/*
2650 		 * Buffer is in-core.  If the buffer is not busy, it must
2651 		 * be on a queue.
2652 		 */
2653 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2654 
2655 		if (flags & GB_LOCK_NOWAIT)
2656 			lockflags |= LK_NOWAIT;
2657 
2658 		error = BUF_TIMELOCK(bp, lockflags,
2659 		    BO_MTX(bo), "getblk", slpflag, slptimeo);
2660 
2661 		/*
2662 		 * If we slept and got the lock we have to restart in case
2663 		 * the buffer changed identities.
2664 		 */
2665 		if (error == ENOLCK)
2666 			goto loop;
2667 		/* We timed out or were interrupted. */
2668 		else if (error)
2669 			return (NULL);
2670 
2671 		/*
2672 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2673 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2674 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2675 		 * backing VM cache.
2676 		 */
2677 		if (bp->b_flags & B_INVAL)
2678 			bp->b_flags &= ~B_CACHE;
2679 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2680 			bp->b_flags |= B_CACHE;
2681 		BO_LOCK(bo);
2682 		bremfree(bp);
2683 		BO_UNLOCK(bo);
2684 
2685 		/*
2686 		 * check for size inconsistancies for non-VMIO case.
2687 		 */
2688 
2689 		if (bp->b_bcount != size) {
2690 			if ((bp->b_flags & B_VMIO) == 0 ||
2691 			    (size > bp->b_kvasize)) {
2692 				if (bp->b_flags & B_DELWRI) {
2693 					/*
2694 					 * If buffer is pinned and caller does
2695 					 * not want sleep  waiting for it to be
2696 					 * unpinned, bail out
2697 					 * */
2698 					if (bp->b_pin_count > 0) {
2699 						if (flags & GB_LOCK_NOWAIT) {
2700 							bqrelse(bp);
2701 							return (NULL);
2702 						} else {
2703 							bunpin_wait(bp);
2704 						}
2705 					}
2706 					bp->b_flags |= B_NOCACHE;
2707 					bwrite(bp);
2708 				} else {
2709 					if (LIST_EMPTY(&bp->b_dep)) {
2710 						bp->b_flags |= B_RELBUF;
2711 						brelse(bp);
2712 					} else {
2713 						bp->b_flags |= B_NOCACHE;
2714 						bwrite(bp);
2715 					}
2716 				}
2717 				goto loop;
2718 			}
2719 		}
2720 
2721 		/*
2722 		 * If the size is inconsistant in the VMIO case, we can resize
2723 		 * the buffer.  This might lead to B_CACHE getting set or
2724 		 * cleared.  If the size has not changed, B_CACHE remains
2725 		 * unchanged from its previous state.
2726 		 */
2727 
2728 		if (bp->b_bcount != size)
2729 			allocbuf(bp, size);
2730 
2731 		KASSERT(bp->b_offset != NOOFFSET,
2732 		    ("getblk: no buffer offset"));
2733 
2734 		/*
2735 		 * A buffer with B_DELWRI set and B_CACHE clear must
2736 		 * be committed before we can return the buffer in
2737 		 * order to prevent the caller from issuing a read
2738 		 * ( due to B_CACHE not being set ) and overwriting
2739 		 * it.
2740 		 *
2741 		 * Most callers, including NFS and FFS, need this to
2742 		 * operate properly either because they assume they
2743 		 * can issue a read if B_CACHE is not set, or because
2744 		 * ( for example ) an uncached B_DELWRI might loop due
2745 		 * to softupdates re-dirtying the buffer.  In the latter
2746 		 * case, B_CACHE is set after the first write completes,
2747 		 * preventing further loops.
2748 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2749 		 * above while extending the buffer, we cannot allow the
2750 		 * buffer to remain with B_CACHE set after the write
2751 		 * completes or it will represent a corrupt state.  To
2752 		 * deal with this we set B_NOCACHE to scrap the buffer
2753 		 * after the write.
2754 		 *
2755 		 * We might be able to do something fancy, like setting
2756 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2757 		 * so the below call doesn't set B_CACHE, but that gets real
2758 		 * confusing.  This is much easier.
2759 		 */
2760 
2761 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2762 			bp->b_flags |= B_NOCACHE;
2763 			bwrite(bp);
2764 			goto loop;
2765 		}
2766 		bp->b_flags &= ~B_DONE;
2767 	} else {
2768 		int bsize, maxsize, vmio;
2769 		off_t offset;
2770 
2771 		/*
2772 		 * Buffer is not in-core, create new buffer.  The buffer
2773 		 * returned by getnewbuf() is locked.  Note that the returned
2774 		 * buffer is also considered valid (not marked B_INVAL).
2775 		 */
2776 		BO_UNLOCK(bo);
2777 		/*
2778 		 * If the user does not want us to create the buffer, bail out
2779 		 * here.
2780 		 */
2781 		if (flags & GB_NOCREAT)
2782 			return NULL;
2783 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
2784 		offset = blkno * bsize;
2785 		vmio = vp->v_object != NULL;
2786 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2787 		maxsize = imax(maxsize, bsize);
2788 
2789 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2790 		if (bp == NULL) {
2791 			if (slpflag || slptimeo)
2792 				return NULL;
2793 			goto loop;
2794 		}
2795 
2796 		/*
2797 		 * This code is used to make sure that a buffer is not
2798 		 * created while the getnewbuf routine is blocked.
2799 		 * This can be a problem whether the vnode is locked or not.
2800 		 * If the buffer is created out from under us, we have to
2801 		 * throw away the one we just created.
2802 		 *
2803 		 * Note: this must occur before we associate the buffer
2804 		 * with the vp especially considering limitations in
2805 		 * the splay tree implementation when dealing with duplicate
2806 		 * lblkno's.
2807 		 */
2808 		BO_LOCK(bo);
2809 		if (gbincore(bo, blkno)) {
2810 			BO_UNLOCK(bo);
2811 			bp->b_flags |= B_INVAL;
2812 			brelse(bp);
2813 			goto loop;
2814 		}
2815 
2816 		/*
2817 		 * Insert the buffer into the hash, so that it can
2818 		 * be found by incore.
2819 		 */
2820 		bp->b_blkno = bp->b_lblkno = blkno;
2821 		bp->b_offset = offset;
2822 		bgetvp(vp, bp);
2823 		BO_UNLOCK(bo);
2824 
2825 		/*
2826 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2827 		 * buffer size starts out as 0, B_CACHE will be set by
2828 		 * allocbuf() for the VMIO case prior to it testing the
2829 		 * backing store for validity.
2830 		 */
2831 
2832 		if (vmio) {
2833 			bp->b_flags |= B_VMIO;
2834 #if defined(VFS_BIO_DEBUG)
2835 			if (vn_canvmio(vp) != TRUE)
2836 				printf("getblk: VMIO on vnode type %d\n",
2837 					vp->v_type);
2838 #endif
2839 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2840 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2841 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2842 		} else {
2843 			bp->b_flags &= ~B_VMIO;
2844 			KASSERT(bp->b_bufobj->bo_object == NULL,
2845 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2846 			    bp, bp->b_bufobj->bo_object));
2847 		}
2848 
2849 		allocbuf(bp, size);
2850 		bp->b_flags &= ~B_DONE;
2851 	}
2852 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2853 	BUF_ASSERT_HELD(bp);
2854 	KASSERT(bp->b_bufobj == bo,
2855 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2856 	return (bp);
2857 }
2858 
2859 /*
2860  * Get an empty, disassociated buffer of given size.  The buffer is initially
2861  * set to B_INVAL.
2862  */
2863 struct buf *
2864 geteblk(int size, int flags)
2865 {
2866 	struct buf *bp;
2867 	int maxsize;
2868 
2869 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2870 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2871 		if ((flags & GB_NOWAIT_BD) &&
2872 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
2873 			return (NULL);
2874 	}
2875 	allocbuf(bp, size);
2876 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2877 	BUF_ASSERT_HELD(bp);
2878 	return (bp);
2879 }
2880 
2881 
2882 /*
2883  * This code constitutes the buffer memory from either anonymous system
2884  * memory (in the case of non-VMIO operations) or from an associated
2885  * VM object (in the case of VMIO operations).  This code is able to
2886  * resize a buffer up or down.
2887  *
2888  * Note that this code is tricky, and has many complications to resolve
2889  * deadlock or inconsistant data situations.  Tread lightly!!!
2890  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2891  * the caller.  Calling this code willy nilly can result in the loss of data.
2892  *
2893  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2894  * B_CACHE for the non-VMIO case.
2895  */
2896 
2897 int
2898 allocbuf(struct buf *bp, int size)
2899 {
2900 	int newbsize, mbsize;
2901 	int i;
2902 
2903 	BUF_ASSERT_HELD(bp);
2904 
2905 	if (bp->b_kvasize < size)
2906 		panic("allocbuf: buffer too small");
2907 
2908 	if ((bp->b_flags & B_VMIO) == 0) {
2909 		caddr_t origbuf;
2910 		int origbufsize;
2911 		/*
2912 		 * Just get anonymous memory from the kernel.  Don't
2913 		 * mess with B_CACHE.
2914 		 */
2915 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2916 		if (bp->b_flags & B_MALLOC)
2917 			newbsize = mbsize;
2918 		else
2919 			newbsize = round_page(size);
2920 
2921 		if (newbsize < bp->b_bufsize) {
2922 			/*
2923 			 * malloced buffers are not shrunk
2924 			 */
2925 			if (bp->b_flags & B_MALLOC) {
2926 				if (newbsize) {
2927 					bp->b_bcount = size;
2928 				} else {
2929 					free(bp->b_data, M_BIOBUF);
2930 					if (bp->b_bufsize) {
2931 						atomic_subtract_long(
2932 						    &bufmallocspace,
2933 						    bp->b_bufsize);
2934 						bufspacewakeup();
2935 						bp->b_bufsize = 0;
2936 					}
2937 					bp->b_saveaddr = bp->b_kvabase;
2938 					bp->b_data = bp->b_saveaddr;
2939 					bp->b_bcount = 0;
2940 					bp->b_flags &= ~B_MALLOC;
2941 				}
2942 				return 1;
2943 			}
2944 			vm_hold_free_pages(bp, newbsize);
2945 		} else if (newbsize > bp->b_bufsize) {
2946 			/*
2947 			 * We only use malloced memory on the first allocation.
2948 			 * and revert to page-allocated memory when the buffer
2949 			 * grows.
2950 			 */
2951 			/*
2952 			 * There is a potential smp race here that could lead
2953 			 * to bufmallocspace slightly passing the max.  It
2954 			 * is probably extremely rare and not worth worrying
2955 			 * over.
2956 			 */
2957 			if ( (bufmallocspace < maxbufmallocspace) &&
2958 				(bp->b_bufsize == 0) &&
2959 				(mbsize <= PAGE_SIZE/2)) {
2960 
2961 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2962 				bp->b_bufsize = mbsize;
2963 				bp->b_bcount = size;
2964 				bp->b_flags |= B_MALLOC;
2965 				atomic_add_long(&bufmallocspace, mbsize);
2966 				return 1;
2967 			}
2968 			origbuf = NULL;
2969 			origbufsize = 0;
2970 			/*
2971 			 * If the buffer is growing on its other-than-first allocation,
2972 			 * then we revert to the page-allocation scheme.
2973 			 */
2974 			if (bp->b_flags & B_MALLOC) {
2975 				origbuf = bp->b_data;
2976 				origbufsize = bp->b_bufsize;
2977 				bp->b_data = bp->b_kvabase;
2978 				if (bp->b_bufsize) {
2979 					atomic_subtract_long(&bufmallocspace,
2980 					    bp->b_bufsize);
2981 					bufspacewakeup();
2982 					bp->b_bufsize = 0;
2983 				}
2984 				bp->b_flags &= ~B_MALLOC;
2985 				newbsize = round_page(newbsize);
2986 			}
2987 			vm_hold_load_pages(
2988 			    bp,
2989 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2990 			    (vm_offset_t) bp->b_data + newbsize);
2991 			if (origbuf) {
2992 				bcopy(origbuf, bp->b_data, origbufsize);
2993 				free(origbuf, M_BIOBUF);
2994 			}
2995 		}
2996 	} else {
2997 		int desiredpages;
2998 
2999 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3000 		desiredpages = (size == 0) ? 0 :
3001 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3002 
3003 		if (bp->b_flags & B_MALLOC)
3004 			panic("allocbuf: VMIO buffer can't be malloced");
3005 		/*
3006 		 * Set B_CACHE initially if buffer is 0 length or will become
3007 		 * 0-length.
3008 		 */
3009 		if (size == 0 || bp->b_bufsize == 0)
3010 			bp->b_flags |= B_CACHE;
3011 
3012 		if (newbsize < bp->b_bufsize) {
3013 			/*
3014 			 * DEV_BSIZE aligned new buffer size is less then the
3015 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3016 			 * if we have to remove any pages.
3017 			 */
3018 			if (desiredpages < bp->b_npages) {
3019 				vm_page_t m;
3020 
3021 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3022 				for (i = desiredpages; i < bp->b_npages; i++) {
3023 					/*
3024 					 * the page is not freed here -- it
3025 					 * is the responsibility of
3026 					 * vnode_pager_setsize
3027 					 */
3028 					m = bp->b_pages[i];
3029 					KASSERT(m != bogus_page,
3030 					    ("allocbuf: bogus page found"));
3031 					while (vm_page_sleep_if_busy(m, TRUE,
3032 					    "biodep"))
3033 						continue;
3034 
3035 					bp->b_pages[i] = NULL;
3036 					vm_page_lock(m);
3037 					vm_page_unwire(m, 0);
3038 					vm_page_unlock(m);
3039 				}
3040 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3041 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3042 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
3043 				bp->b_npages = desiredpages;
3044 			}
3045 		} else if (size > bp->b_bcount) {
3046 			/*
3047 			 * We are growing the buffer, possibly in a
3048 			 * byte-granular fashion.
3049 			 */
3050 			vm_object_t obj;
3051 			vm_offset_t toff;
3052 			vm_offset_t tinc;
3053 
3054 			/*
3055 			 * Step 1, bring in the VM pages from the object,
3056 			 * allocating them if necessary.  We must clear
3057 			 * B_CACHE if these pages are not valid for the
3058 			 * range covered by the buffer.
3059 			 */
3060 
3061 			obj = bp->b_bufobj->bo_object;
3062 
3063 			VM_OBJECT_LOCK(obj);
3064 			while (bp->b_npages < desiredpages) {
3065 				vm_page_t m;
3066 
3067 				/*
3068 				 * We must allocate system pages since blocking
3069 				 * here could intefere with paging I/O, no
3070 				 * matter which process we are.
3071 				 *
3072 				 * We can only test VPO_BUSY here.  Blocking on
3073 				 * m->busy might lead to a deadlock:
3074 				 *  vm_fault->getpages->cluster_read->allocbuf
3075 				 * Thus, we specify VM_ALLOC_IGN_SBUSY.
3076 				 */
3077 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3078 				    bp->b_npages, VM_ALLOC_NOBUSY |
3079 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3080 				    VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY |
3081 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3082 				if (m->valid == 0)
3083 					bp->b_flags &= ~B_CACHE;
3084 				bp->b_pages[bp->b_npages] = m;
3085 				++bp->b_npages;
3086 			}
3087 
3088 			/*
3089 			 * Step 2.  We've loaded the pages into the buffer,
3090 			 * we have to figure out if we can still have B_CACHE
3091 			 * set.  Note that B_CACHE is set according to the
3092 			 * byte-granular range ( bcount and size ), new the
3093 			 * aligned range ( newbsize ).
3094 			 *
3095 			 * The VM test is against m->valid, which is DEV_BSIZE
3096 			 * aligned.  Needless to say, the validity of the data
3097 			 * needs to also be DEV_BSIZE aligned.  Note that this
3098 			 * fails with NFS if the server or some other client
3099 			 * extends the file's EOF.  If our buffer is resized,
3100 			 * B_CACHE may remain set! XXX
3101 			 */
3102 
3103 			toff = bp->b_bcount;
3104 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3105 
3106 			while ((bp->b_flags & B_CACHE) && toff < size) {
3107 				vm_pindex_t pi;
3108 
3109 				if (tinc > (size - toff))
3110 					tinc = size - toff;
3111 
3112 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3113 				    PAGE_SHIFT;
3114 
3115 				vfs_buf_test_cache(
3116 				    bp,
3117 				    bp->b_offset,
3118 				    toff,
3119 				    tinc,
3120 				    bp->b_pages[pi]
3121 				);
3122 				toff += tinc;
3123 				tinc = PAGE_SIZE;
3124 			}
3125 			VM_OBJECT_UNLOCK(obj);
3126 
3127 			/*
3128 			 * Step 3, fixup the KVM pmap.  Remember that
3129 			 * bp->b_data is relative to bp->b_offset, but
3130 			 * bp->b_offset may be offset into the first page.
3131 			 */
3132 
3133 			bp->b_data = (caddr_t)
3134 			    trunc_page((vm_offset_t)bp->b_data);
3135 			pmap_qenter(
3136 			    (vm_offset_t)bp->b_data,
3137 			    bp->b_pages,
3138 			    bp->b_npages
3139 			);
3140 
3141 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3142 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
3143 		}
3144 	}
3145 	if (newbsize < bp->b_bufsize)
3146 		bufspacewakeup();
3147 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3148 	bp->b_bcount = size;		/* requested buffer size	*/
3149 	return 1;
3150 }
3151 
3152 void
3153 biodone(struct bio *bp)
3154 {
3155 	struct mtx *mtxp;
3156 	void (*done)(struct bio *);
3157 
3158 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3159 	mtx_lock(mtxp);
3160 	bp->bio_flags |= BIO_DONE;
3161 	done = bp->bio_done;
3162 	if (done == NULL)
3163 		wakeup(bp);
3164 	mtx_unlock(mtxp);
3165 	if (done != NULL)
3166 		done(bp);
3167 }
3168 
3169 /*
3170  * Wait for a BIO to finish.
3171  *
3172  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3173  * case is not yet clear.
3174  */
3175 int
3176 biowait(struct bio *bp, const char *wchan)
3177 {
3178 	struct mtx *mtxp;
3179 
3180 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3181 	mtx_lock(mtxp);
3182 	while ((bp->bio_flags & BIO_DONE) == 0)
3183 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3184 	mtx_unlock(mtxp);
3185 	if (bp->bio_error != 0)
3186 		return (bp->bio_error);
3187 	if (!(bp->bio_flags & BIO_ERROR))
3188 		return (0);
3189 	return (EIO);
3190 }
3191 
3192 void
3193 biofinish(struct bio *bp, struct devstat *stat, int error)
3194 {
3195 
3196 	if (error) {
3197 		bp->bio_error = error;
3198 		bp->bio_flags |= BIO_ERROR;
3199 	}
3200 	if (stat != NULL)
3201 		devstat_end_transaction_bio(stat, bp);
3202 	biodone(bp);
3203 }
3204 
3205 /*
3206  *	bufwait:
3207  *
3208  *	Wait for buffer I/O completion, returning error status.  The buffer
3209  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3210  *	error and cleared.
3211  */
3212 int
3213 bufwait(struct buf *bp)
3214 {
3215 	if (bp->b_iocmd == BIO_READ)
3216 		bwait(bp, PRIBIO, "biord");
3217 	else
3218 		bwait(bp, PRIBIO, "biowr");
3219 	if (bp->b_flags & B_EINTR) {
3220 		bp->b_flags &= ~B_EINTR;
3221 		return (EINTR);
3222 	}
3223 	if (bp->b_ioflags & BIO_ERROR) {
3224 		return (bp->b_error ? bp->b_error : EIO);
3225 	} else {
3226 		return (0);
3227 	}
3228 }
3229 
3230  /*
3231   * Call back function from struct bio back up to struct buf.
3232   */
3233 static void
3234 bufdonebio(struct bio *bip)
3235 {
3236 	struct buf *bp;
3237 
3238 	bp = bip->bio_caller2;
3239 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3240 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3241 	bp->b_ioflags = bip->bio_flags;
3242 	bp->b_error = bip->bio_error;
3243 	if (bp->b_error)
3244 		bp->b_ioflags |= BIO_ERROR;
3245 	bufdone(bp);
3246 	g_destroy_bio(bip);
3247 }
3248 
3249 void
3250 dev_strategy(struct cdev *dev, struct buf *bp)
3251 {
3252 	struct cdevsw *csw;
3253 	struct bio *bip;
3254 	int ref;
3255 
3256 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3257 		panic("b_iocmd botch");
3258 	for (;;) {
3259 		bip = g_new_bio();
3260 		if (bip != NULL)
3261 			break;
3262 		/* Try again later */
3263 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3264 	}
3265 	bip->bio_cmd = bp->b_iocmd;
3266 	bip->bio_offset = bp->b_iooffset;
3267 	bip->bio_length = bp->b_bcount;
3268 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3269 	bip->bio_data = bp->b_data;
3270 	bip->bio_done = bufdonebio;
3271 	bip->bio_caller2 = bp;
3272 	bip->bio_dev = dev;
3273 	KASSERT(dev->si_refcount > 0,
3274 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3275 	    devtoname(dev)));
3276 	csw = dev_refthread(dev, &ref);
3277 	if (csw == NULL) {
3278 		g_destroy_bio(bip);
3279 		bp->b_error = ENXIO;
3280 		bp->b_ioflags = BIO_ERROR;
3281 		bufdone(bp);
3282 		return;
3283 	}
3284 	(*csw->d_strategy)(bip);
3285 	dev_relthread(dev, ref);
3286 }
3287 
3288 /*
3289  *	bufdone:
3290  *
3291  *	Finish I/O on a buffer, optionally calling a completion function.
3292  *	This is usually called from an interrupt so process blocking is
3293  *	not allowed.
3294  *
3295  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3296  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3297  *	assuming B_INVAL is clear.
3298  *
3299  *	For the VMIO case, we set B_CACHE if the op was a read and no
3300  *	read error occured, or if the op was a write.  B_CACHE is never
3301  *	set if the buffer is invalid or otherwise uncacheable.
3302  *
3303  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3304  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3305  *	in the biodone routine.
3306  */
3307 void
3308 bufdone(struct buf *bp)
3309 {
3310 	struct bufobj *dropobj;
3311 	void    (*biodone)(struct buf *);
3312 
3313 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3314 	dropobj = NULL;
3315 
3316 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3317 	BUF_ASSERT_HELD(bp);
3318 
3319 	runningbufwakeup(bp);
3320 	if (bp->b_iocmd == BIO_WRITE)
3321 		dropobj = bp->b_bufobj;
3322 	/* call optional completion function if requested */
3323 	if (bp->b_iodone != NULL) {
3324 		biodone = bp->b_iodone;
3325 		bp->b_iodone = NULL;
3326 		(*biodone) (bp);
3327 		if (dropobj)
3328 			bufobj_wdrop(dropobj);
3329 		return;
3330 	}
3331 
3332 	bufdone_finish(bp);
3333 
3334 	if (dropobj)
3335 		bufobj_wdrop(dropobj);
3336 }
3337 
3338 void
3339 bufdone_finish(struct buf *bp)
3340 {
3341 	BUF_ASSERT_HELD(bp);
3342 
3343 	if (!LIST_EMPTY(&bp->b_dep))
3344 		buf_complete(bp);
3345 
3346 	if (bp->b_flags & B_VMIO) {
3347 		int i;
3348 		vm_ooffset_t foff;
3349 		vm_page_t m;
3350 		vm_object_t obj;
3351 		int bogus, iosize;
3352 		struct vnode *vp = bp->b_vp;
3353 
3354 		obj = bp->b_bufobj->bo_object;
3355 
3356 #if defined(VFS_BIO_DEBUG)
3357 		mp_fixme("usecount and vflag accessed without locks.");
3358 		if (vp->v_usecount == 0) {
3359 			panic("biodone: zero vnode ref count");
3360 		}
3361 
3362 		KASSERT(vp->v_object != NULL,
3363 			("biodone: vnode %p has no vm_object", vp));
3364 #endif
3365 
3366 		foff = bp->b_offset;
3367 		KASSERT(bp->b_offset != NOOFFSET,
3368 		    ("biodone: no buffer offset"));
3369 
3370 		VM_OBJECT_LOCK(obj);
3371 #if defined(VFS_BIO_DEBUG)
3372 		if (obj->paging_in_progress < bp->b_npages) {
3373 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3374 			    obj->paging_in_progress, bp->b_npages);
3375 		}
3376 #endif
3377 
3378 		/*
3379 		 * Set B_CACHE if the op was a normal read and no error
3380 		 * occured.  B_CACHE is set for writes in the b*write()
3381 		 * routines.
3382 		 */
3383 		iosize = bp->b_bcount - bp->b_resid;
3384 		if (bp->b_iocmd == BIO_READ &&
3385 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3386 		    !(bp->b_ioflags & BIO_ERROR)) {
3387 			bp->b_flags |= B_CACHE;
3388 		}
3389 		bogus = 0;
3390 		for (i = 0; i < bp->b_npages; i++) {
3391 			int bogusflag = 0;
3392 			int resid;
3393 
3394 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3395 			if (resid > iosize)
3396 				resid = iosize;
3397 
3398 			/*
3399 			 * cleanup bogus pages, restoring the originals
3400 			 */
3401 			m = bp->b_pages[i];
3402 			if (m == bogus_page) {
3403 				bogus = bogusflag = 1;
3404 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3405 				if (m == NULL)
3406 					panic("biodone: page disappeared!");
3407 				bp->b_pages[i] = m;
3408 			}
3409 #if defined(VFS_BIO_DEBUG)
3410 			if (OFF_TO_IDX(foff) != m->pindex) {
3411 				printf(
3412 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3413 				    (intmax_t)foff, (uintmax_t)m->pindex);
3414 			}
3415 #endif
3416 
3417 			/*
3418 			 * In the write case, the valid and clean bits are
3419 			 * already changed correctly ( see bdwrite() ), so we
3420 			 * only need to do this here in the read case.
3421 			 */
3422 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3423 				KASSERT((m->dirty & vm_page_bits(foff &
3424 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3425 				    " page %p has unexpected dirty bits", m));
3426 				vfs_page_set_valid(bp, foff, m);
3427 			}
3428 
3429 			/*
3430 			 * when debugging new filesystems or buffer I/O methods, this
3431 			 * is the most common error that pops up.  if you see this, you
3432 			 * have not set the page busy flag correctly!!!
3433 			 */
3434 			if (m->busy == 0) {
3435 				printf("biodone: page busy < 0, "
3436 				    "pindex: %d, foff: 0x(%x,%x), "
3437 				    "resid: %d, index: %d\n",
3438 				    (int) m->pindex, (int)(foff >> 32),
3439 						(int) foff & 0xffffffff, resid, i);
3440 				if (!vn_isdisk(vp, NULL))
3441 					printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3442 					    (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3443 					    (intmax_t) bp->b_lblkno,
3444 					    bp->b_flags, bp->b_npages);
3445 				else
3446 					printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3447 					    (intmax_t) bp->b_lblkno,
3448 					    bp->b_flags, bp->b_npages);
3449 				printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3450 				    (u_long)m->valid, (u_long)m->dirty,
3451 				    m->wire_count);
3452 				panic("biodone: page busy < 0\n");
3453 			}
3454 			vm_page_io_finish(m);
3455 			vm_object_pip_subtract(obj, 1);
3456 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3457 			iosize -= resid;
3458 		}
3459 		vm_object_pip_wakeupn(obj, 0);
3460 		VM_OBJECT_UNLOCK(obj);
3461 		if (bogus)
3462 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3463 			    bp->b_pages, bp->b_npages);
3464 	}
3465 
3466 	/*
3467 	 * For asynchronous completions, release the buffer now. The brelse
3468 	 * will do a wakeup there if necessary - so no need to do a wakeup
3469 	 * here in the async case. The sync case always needs to do a wakeup.
3470 	 */
3471 
3472 	if (bp->b_flags & B_ASYNC) {
3473 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3474 			brelse(bp);
3475 		else
3476 			bqrelse(bp);
3477 	} else
3478 		bdone(bp);
3479 }
3480 
3481 /*
3482  * This routine is called in lieu of iodone in the case of
3483  * incomplete I/O.  This keeps the busy status for pages
3484  * consistant.
3485  */
3486 void
3487 vfs_unbusy_pages(struct buf *bp)
3488 {
3489 	int i;
3490 	vm_object_t obj;
3491 	vm_page_t m;
3492 
3493 	runningbufwakeup(bp);
3494 	if (!(bp->b_flags & B_VMIO))
3495 		return;
3496 
3497 	obj = bp->b_bufobj->bo_object;
3498 	VM_OBJECT_LOCK(obj);
3499 	for (i = 0; i < bp->b_npages; i++) {
3500 		m = bp->b_pages[i];
3501 		if (m == bogus_page) {
3502 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3503 			if (!m)
3504 				panic("vfs_unbusy_pages: page missing\n");
3505 			bp->b_pages[i] = m;
3506 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3507 			    bp->b_pages, bp->b_npages);
3508 		}
3509 		vm_object_pip_subtract(obj, 1);
3510 		vm_page_io_finish(m);
3511 	}
3512 	vm_object_pip_wakeupn(obj, 0);
3513 	VM_OBJECT_UNLOCK(obj);
3514 }
3515 
3516 /*
3517  * vfs_page_set_valid:
3518  *
3519  *	Set the valid bits in a page based on the supplied offset.   The
3520  *	range is restricted to the buffer's size.
3521  *
3522  *	This routine is typically called after a read completes.
3523  */
3524 static void
3525 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3526 {
3527 	vm_ooffset_t eoff;
3528 
3529 	/*
3530 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3531 	 * page boundary and eoff is not greater than the end of the buffer.
3532 	 * The end of the buffer, in this case, is our file EOF, not the
3533 	 * allocation size of the buffer.
3534 	 */
3535 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3536 	if (eoff > bp->b_offset + bp->b_bcount)
3537 		eoff = bp->b_offset + bp->b_bcount;
3538 
3539 	/*
3540 	 * Set valid range.  This is typically the entire buffer and thus the
3541 	 * entire page.
3542 	 */
3543 	if (eoff > off)
3544 		vm_page_set_valid(m, off & PAGE_MASK, eoff - off);
3545 }
3546 
3547 /*
3548  * vfs_page_set_validclean:
3549  *
3550  *	Set the valid bits and clear the dirty bits in a page based on the
3551  *	supplied offset.   The range is restricted to the buffer's size.
3552  */
3553 static void
3554 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3555 {
3556 	vm_ooffset_t soff, eoff;
3557 
3558 	/*
3559 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3560 	 * page boundry or cross the end of the buffer.  The end of the
3561 	 * buffer, in this case, is our file EOF, not the allocation size
3562 	 * of the buffer.
3563 	 */
3564 	soff = off;
3565 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3566 	if (eoff > bp->b_offset + bp->b_bcount)
3567 		eoff = bp->b_offset + bp->b_bcount;
3568 
3569 	/*
3570 	 * Set valid range.  This is typically the entire buffer and thus the
3571 	 * entire page.
3572 	 */
3573 	if (eoff > soff) {
3574 		vm_page_set_validclean(
3575 		    m,
3576 		   (vm_offset_t) (soff & PAGE_MASK),
3577 		   (vm_offset_t) (eoff - soff)
3578 		);
3579 	}
3580 }
3581 
3582 /*
3583  * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3584  * any page is busy, drain the flag.
3585  */
3586 static void
3587 vfs_drain_busy_pages(struct buf *bp)
3588 {
3589 	vm_page_t m;
3590 	int i, last_busied;
3591 
3592 	VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED);
3593 	last_busied = 0;
3594 	for (i = 0; i < bp->b_npages; i++) {
3595 		m = bp->b_pages[i];
3596 		if ((m->oflags & VPO_BUSY) != 0) {
3597 			for (; last_busied < i; last_busied++)
3598 				vm_page_busy(bp->b_pages[last_busied]);
3599 			while ((m->oflags & VPO_BUSY) != 0)
3600 				vm_page_sleep(m, "vbpage");
3601 		}
3602 	}
3603 	for (i = 0; i < last_busied; i++)
3604 		vm_page_wakeup(bp->b_pages[i]);
3605 }
3606 
3607 /*
3608  * This routine is called before a device strategy routine.
3609  * It is used to tell the VM system that paging I/O is in
3610  * progress, and treat the pages associated with the buffer
3611  * almost as being VPO_BUSY.  Also the object paging_in_progress
3612  * flag is handled to make sure that the object doesn't become
3613  * inconsistant.
3614  *
3615  * Since I/O has not been initiated yet, certain buffer flags
3616  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3617  * and should be ignored.
3618  */
3619 void
3620 vfs_busy_pages(struct buf *bp, int clear_modify)
3621 {
3622 	int i, bogus;
3623 	vm_object_t obj;
3624 	vm_ooffset_t foff;
3625 	vm_page_t m;
3626 
3627 	if (!(bp->b_flags & B_VMIO))
3628 		return;
3629 
3630 	obj = bp->b_bufobj->bo_object;
3631 	foff = bp->b_offset;
3632 	KASSERT(bp->b_offset != NOOFFSET,
3633 	    ("vfs_busy_pages: no buffer offset"));
3634 	VM_OBJECT_LOCK(obj);
3635 	vfs_drain_busy_pages(bp);
3636 	if (bp->b_bufsize != 0)
3637 		vfs_setdirty_locked_object(bp);
3638 	bogus = 0;
3639 	for (i = 0; i < bp->b_npages; i++) {
3640 		m = bp->b_pages[i];
3641 
3642 		if ((bp->b_flags & B_CLUSTER) == 0) {
3643 			vm_object_pip_add(obj, 1);
3644 			vm_page_io_start(m);
3645 		}
3646 		/*
3647 		 * When readying a buffer for a read ( i.e
3648 		 * clear_modify == 0 ), it is important to do
3649 		 * bogus_page replacement for valid pages in
3650 		 * partially instantiated buffers.  Partially
3651 		 * instantiated buffers can, in turn, occur when
3652 		 * reconstituting a buffer from its VM backing store
3653 		 * base.  We only have to do this if B_CACHE is
3654 		 * clear ( which causes the I/O to occur in the
3655 		 * first place ).  The replacement prevents the read
3656 		 * I/O from overwriting potentially dirty VM-backed
3657 		 * pages.  XXX bogus page replacement is, uh, bogus.
3658 		 * It may not work properly with small-block devices.
3659 		 * We need to find a better way.
3660 		 */
3661 		if (clear_modify) {
3662 			pmap_remove_write(m);
3663 			vfs_page_set_validclean(bp, foff, m);
3664 		} else if (m->valid == VM_PAGE_BITS_ALL &&
3665 		    (bp->b_flags & B_CACHE) == 0) {
3666 			bp->b_pages[i] = bogus_page;
3667 			bogus++;
3668 		}
3669 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3670 	}
3671 	VM_OBJECT_UNLOCK(obj);
3672 	if (bogus)
3673 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3674 		    bp->b_pages, bp->b_npages);
3675 }
3676 
3677 /*
3678  *	vfs_bio_set_valid:
3679  *
3680  *	Set the range within the buffer to valid.  The range is
3681  *	relative to the beginning of the buffer, b_offset.  Note that
3682  *	b_offset itself may be offset from the beginning of the first
3683  *	page.
3684  */
3685 void
3686 vfs_bio_set_valid(struct buf *bp, int base, int size)
3687 {
3688 	int i, n;
3689 	vm_page_t m;
3690 
3691 	if (!(bp->b_flags & B_VMIO))
3692 		return;
3693 
3694 	/*
3695 	 * Fixup base to be relative to beginning of first page.
3696 	 * Set initial n to be the maximum number of bytes in the
3697 	 * first page that can be validated.
3698 	 */
3699 	base += (bp->b_offset & PAGE_MASK);
3700 	n = PAGE_SIZE - (base & PAGE_MASK);
3701 
3702 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3703 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3704 		m = bp->b_pages[i];
3705 		if (n > size)
3706 			n = size;
3707 		vm_page_set_valid(m, base & PAGE_MASK, n);
3708 		base += n;
3709 		size -= n;
3710 		n = PAGE_SIZE;
3711 	}
3712 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3713 }
3714 
3715 /*
3716  *	vfs_bio_clrbuf:
3717  *
3718  *	If the specified buffer is a non-VMIO buffer, clear the entire
3719  *	buffer.  If the specified buffer is a VMIO buffer, clear and
3720  *	validate only the previously invalid portions of the buffer.
3721  *	This routine essentially fakes an I/O, so we need to clear
3722  *	BIO_ERROR and B_INVAL.
3723  *
3724  *	Note that while we only theoretically need to clear through b_bcount,
3725  *	we go ahead and clear through b_bufsize.
3726  */
3727 void
3728 vfs_bio_clrbuf(struct buf *bp)
3729 {
3730 	int i, j, mask;
3731 	caddr_t sa, ea;
3732 
3733 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3734 		clrbuf(bp);
3735 		return;
3736 	}
3737 	bp->b_flags &= ~B_INVAL;
3738 	bp->b_ioflags &= ~BIO_ERROR;
3739 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3740 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3741 	    (bp->b_offset & PAGE_MASK) == 0) {
3742 		if (bp->b_pages[0] == bogus_page)
3743 			goto unlock;
3744 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3745 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3746 		if ((bp->b_pages[0]->valid & mask) == mask)
3747 			goto unlock;
3748 		if ((bp->b_pages[0]->valid & mask) == 0) {
3749 			bzero(bp->b_data, bp->b_bufsize);
3750 			bp->b_pages[0]->valid |= mask;
3751 			goto unlock;
3752 		}
3753 	}
3754 	ea = sa = bp->b_data;
3755 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3756 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3757 		ea = (caddr_t)(vm_offset_t)ulmin(
3758 		    (u_long)(vm_offset_t)ea,
3759 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3760 		if (bp->b_pages[i] == bogus_page)
3761 			continue;
3762 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3763 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3764 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3765 		if ((bp->b_pages[i]->valid & mask) == mask)
3766 			continue;
3767 		if ((bp->b_pages[i]->valid & mask) == 0)
3768 			bzero(sa, ea - sa);
3769 		else {
3770 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3771 				if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3772 					bzero(sa, DEV_BSIZE);
3773 			}
3774 		}
3775 		bp->b_pages[i]->valid |= mask;
3776 	}
3777 unlock:
3778 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3779 	bp->b_resid = 0;
3780 }
3781 
3782 /*
3783  * vm_hold_load_pages and vm_hold_free_pages get pages into
3784  * a buffers address space.  The pages are anonymous and are
3785  * not associated with a file object.
3786  */
3787 static void
3788 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3789 {
3790 	vm_offset_t pg;
3791 	vm_page_t p;
3792 	int index;
3793 
3794 	to = round_page(to);
3795 	from = round_page(from);
3796 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3797 
3798 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3799 tryagain:
3800 		/*
3801 		 * note: must allocate system pages since blocking here
3802 		 * could interfere with paging I/O, no matter which
3803 		 * process we are.
3804 		 */
3805 		p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ |
3806 		    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3807 		    VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
3808 		if (!p) {
3809 			VM_WAIT;
3810 			goto tryagain;
3811 		}
3812 		pmap_qenter(pg, &p, 1);
3813 		bp->b_pages[index] = p;
3814 	}
3815 	bp->b_npages = index;
3816 }
3817 
3818 /* Return pages associated with this buf to the vm system */
3819 static void
3820 vm_hold_free_pages(struct buf *bp, int newbsize)
3821 {
3822 	vm_offset_t from;
3823 	vm_page_t p;
3824 	int index, newnpages;
3825 
3826 	from = round_page((vm_offset_t)bp->b_data + newbsize);
3827 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3828 	if (bp->b_npages > newnpages)
3829 		pmap_qremove(from, bp->b_npages - newnpages);
3830 	for (index = newnpages; index < bp->b_npages; index++) {
3831 		p = bp->b_pages[index];
3832 		bp->b_pages[index] = NULL;
3833 		if (p->busy != 0)
3834 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3835 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
3836 		p->wire_count--;
3837 		vm_page_free(p);
3838 		atomic_subtract_int(&cnt.v_wire_count, 1);
3839 	}
3840 	bp->b_npages = newnpages;
3841 }
3842 
3843 /*
3844  * Map an IO request into kernel virtual address space.
3845  *
3846  * All requests are (re)mapped into kernel VA space.
3847  * Notice that we use b_bufsize for the size of the buffer
3848  * to be mapped.  b_bcount might be modified by the driver.
3849  *
3850  * Note that even if the caller determines that the address space should
3851  * be valid, a race or a smaller-file mapped into a larger space may
3852  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3853  * check the return value.
3854  */
3855 int
3856 vmapbuf(struct buf *bp)
3857 {
3858 	caddr_t addr, kva;
3859 	vm_prot_t prot;
3860 	int pidx, i;
3861 	struct vm_page *m;
3862 	struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3863 
3864 	if (bp->b_bufsize < 0)
3865 		return (-1);
3866 	prot = VM_PROT_READ;
3867 	if (bp->b_iocmd == BIO_READ)
3868 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3869 	for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3870 	     addr < bp->b_data + bp->b_bufsize;
3871 	     addr += PAGE_SIZE, pidx++) {
3872 		/*
3873 		 * Do the vm_fault if needed; do the copy-on-write thing
3874 		 * when reading stuff off device into memory.
3875 		 *
3876 		 * NOTE! Must use pmap_extract() because addr may be in
3877 		 * the userland address space, and kextract is only guarenteed
3878 		 * to work for the kernland address space (see: sparc64 port).
3879 		 */
3880 retry:
3881 		if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3882 		    prot) < 0) {
3883 			for (i = 0; i < pidx; ++i) {
3884 				vm_page_lock(bp->b_pages[i]);
3885 				vm_page_unhold(bp->b_pages[i]);
3886 				vm_page_unlock(bp->b_pages[i]);
3887 				bp->b_pages[i] = NULL;
3888 			}
3889 			return(-1);
3890 		}
3891 		m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3892 		if (m == NULL)
3893 			goto retry;
3894 		bp->b_pages[pidx] = m;
3895 	}
3896 	if (pidx > btoc(MAXPHYS))
3897 		panic("vmapbuf: mapped more than MAXPHYS");
3898 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3899 
3900 	kva = bp->b_saveaddr;
3901 	bp->b_npages = pidx;
3902 	bp->b_saveaddr = bp->b_data;
3903 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3904 	return(0);
3905 }
3906 
3907 /*
3908  * Free the io map PTEs associated with this IO operation.
3909  * We also invalidate the TLB entries and restore the original b_addr.
3910  */
3911 void
3912 vunmapbuf(struct buf *bp)
3913 {
3914 	int pidx;
3915 	int npages;
3916 
3917 	npages = bp->b_npages;
3918 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3919 	for (pidx = 0; pidx < npages; pidx++) {
3920 		vm_page_lock(bp->b_pages[pidx]);
3921 		vm_page_unhold(bp->b_pages[pidx]);
3922 		vm_page_unlock(bp->b_pages[pidx]);
3923 	}
3924 
3925 	bp->b_data = bp->b_saveaddr;
3926 }
3927 
3928 void
3929 bdone(struct buf *bp)
3930 {
3931 	struct mtx *mtxp;
3932 
3933 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3934 	mtx_lock(mtxp);
3935 	bp->b_flags |= B_DONE;
3936 	wakeup(bp);
3937 	mtx_unlock(mtxp);
3938 }
3939 
3940 void
3941 bwait(struct buf *bp, u_char pri, const char *wchan)
3942 {
3943 	struct mtx *mtxp;
3944 
3945 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3946 	mtx_lock(mtxp);
3947 	while ((bp->b_flags & B_DONE) == 0)
3948 		msleep(bp, mtxp, pri, wchan, 0);
3949 	mtx_unlock(mtxp);
3950 }
3951 
3952 int
3953 bufsync(struct bufobj *bo, int waitfor)
3954 {
3955 
3956 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3957 }
3958 
3959 void
3960 bufstrategy(struct bufobj *bo, struct buf *bp)
3961 {
3962 	int i = 0;
3963 	struct vnode *vp;
3964 
3965 	vp = bp->b_vp;
3966 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3967 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3968 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3969 	i = VOP_STRATEGY(vp, bp);
3970 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3971 }
3972 
3973 void
3974 bufobj_wrefl(struct bufobj *bo)
3975 {
3976 
3977 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3978 	ASSERT_BO_LOCKED(bo);
3979 	bo->bo_numoutput++;
3980 }
3981 
3982 void
3983 bufobj_wref(struct bufobj *bo)
3984 {
3985 
3986 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3987 	BO_LOCK(bo);
3988 	bo->bo_numoutput++;
3989 	BO_UNLOCK(bo);
3990 }
3991 
3992 void
3993 bufobj_wdrop(struct bufobj *bo)
3994 {
3995 
3996 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3997 	BO_LOCK(bo);
3998 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3999 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4000 		bo->bo_flag &= ~BO_WWAIT;
4001 		wakeup(&bo->bo_numoutput);
4002 	}
4003 	BO_UNLOCK(bo);
4004 }
4005 
4006 int
4007 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4008 {
4009 	int error;
4010 
4011 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4012 	ASSERT_BO_LOCKED(bo);
4013 	error = 0;
4014 	while (bo->bo_numoutput) {
4015 		bo->bo_flag |= BO_WWAIT;
4016 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
4017 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4018 		if (error)
4019 			break;
4020 	}
4021 	return (error);
4022 }
4023 
4024 void
4025 bpin(struct buf *bp)
4026 {
4027 	struct mtx *mtxp;
4028 
4029 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4030 	mtx_lock(mtxp);
4031 	bp->b_pin_count++;
4032 	mtx_unlock(mtxp);
4033 }
4034 
4035 void
4036 bunpin(struct buf *bp)
4037 {
4038 	struct mtx *mtxp;
4039 
4040 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4041 	mtx_lock(mtxp);
4042 	if (--bp->b_pin_count == 0)
4043 		wakeup(bp);
4044 	mtx_unlock(mtxp);
4045 }
4046 
4047 void
4048 bunpin_wait(struct buf *bp)
4049 {
4050 	struct mtx *mtxp;
4051 
4052 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4053 	mtx_lock(mtxp);
4054 	while (bp->b_pin_count > 0)
4055 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4056 	mtx_unlock(mtxp);
4057 }
4058 
4059 #include "opt_ddb.h"
4060 #ifdef DDB
4061 #include <ddb/ddb.h>
4062 
4063 /* DDB command to show buffer data */
4064 DB_SHOW_COMMAND(buffer, db_show_buffer)
4065 {
4066 	/* get args */
4067 	struct buf *bp = (struct buf *)addr;
4068 
4069 	if (!have_addr) {
4070 		db_printf("usage: show buffer <addr>\n");
4071 		return;
4072 	}
4073 
4074 	db_printf("buf at %p\n", bp);
4075 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4076 	db_printf(
4077 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4078 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n",
4079 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4080 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4081 	    bp->b_dep.lh_first);
4082 	if (bp->b_npages) {
4083 		int i;
4084 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4085 		for (i = 0; i < bp->b_npages; i++) {
4086 			vm_page_t m;
4087 			m = bp->b_pages[i];
4088 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4089 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4090 			if ((i + 1) < bp->b_npages)
4091 				db_printf(",");
4092 		}
4093 		db_printf("\n");
4094 	}
4095 	db_printf(" ");
4096 	lockmgr_printinfo(&bp->b_lock);
4097 }
4098 
4099 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4100 {
4101 	struct buf *bp;
4102 	int i;
4103 
4104 	for (i = 0; i < nbuf; i++) {
4105 		bp = &buf[i];
4106 		if (BUF_ISLOCKED(bp)) {
4107 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4108 			db_printf("\n");
4109 		}
4110 	}
4111 }
4112 
4113 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4114 {
4115 	struct vnode *vp;
4116 	struct buf *bp;
4117 
4118 	if (!have_addr) {
4119 		db_printf("usage: show vnodebufs <addr>\n");
4120 		return;
4121 	}
4122 	vp = (struct vnode *)addr;
4123 	db_printf("Clean buffers:\n");
4124 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4125 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4126 		db_printf("\n");
4127 	}
4128 	db_printf("Dirty buffers:\n");
4129 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4130 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4131 		db_printf("\n");
4132 	}
4133 }
4134 
4135 DB_COMMAND(countfreebufs, db_coundfreebufs)
4136 {
4137 	struct buf *bp;
4138 	int i, used = 0, nfree = 0;
4139 
4140 	if (have_addr) {
4141 		db_printf("usage: countfreebufs\n");
4142 		return;
4143 	}
4144 
4145 	for (i = 0; i < nbuf; i++) {
4146 		bp = &buf[i];
4147 		if ((bp->b_vflags & BV_INFREECNT) != 0)
4148 			nfree++;
4149 		else
4150 			used++;
4151 	}
4152 
4153 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4154 	    nfree + used);
4155 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4156 }
4157 #endif /* DDB */
4158