xref: /freebsd/sys/kern/vfs_bio.c (revision f1f890804985a1043da42a5def13c79dc005f5e9)
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/rwlock.h>
62 #include <sys/sysctl.h>
63 #include <sys/vmmeter.h>
64 #include <sys/vnode.h>
65 #include <geom/geom.h>
66 #include <vm/vm.h>
67 #include <vm/vm_param.h>
68 #include <vm/vm_kern.h>
69 #include <vm/vm_pageout.h>
70 #include <vm/vm_page.h>
71 #include <vm/vm_object.h>
72 #include <vm/vm_extern.h>
73 #include <vm/vm_map.h>
74 #include "opt_compat.h"
75 #include "opt_directio.h"
76 #include "opt_swap.h"
77 
78 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
79 
80 struct	bio_ops bioops;		/* I/O operation notification */
81 
82 struct	buf_ops buf_ops_bio = {
83 	.bop_name	=	"buf_ops_bio",
84 	.bop_write	=	bufwrite,
85 	.bop_strategy	=	bufstrategy,
86 	.bop_sync	=	bufsync,
87 	.bop_bdflush	=	bufbdflush,
88 };
89 
90 /*
91  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
92  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
93  */
94 struct buf *buf;		/* buffer header pool */
95 
96 static struct proc *bufdaemonproc;
97 
98 static int inmem(struct vnode *vp, daddr_t blkno);
99 static void vm_hold_free_pages(struct buf *bp, int newbsize);
100 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
101 		vm_offset_t to);
102 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
103 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
104 		vm_page_t m);
105 static void vfs_drain_busy_pages(struct buf *bp);
106 static void vfs_clean_pages_dirty_buf(struct buf *bp);
107 static void vfs_setdirty_locked_object(struct buf *bp);
108 static void vfs_vmio_release(struct buf *bp);
109 static int vfs_bio_clcheck(struct vnode *vp, int size,
110 		daddr_t lblkno, daddr_t blkno);
111 static int buf_do_flush(struct vnode *vp);
112 static int flushbufqueues(struct vnode *, int, int);
113 static void buf_daemon(void);
114 static void bremfreel(struct buf *bp);
115 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
116     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
117 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
118 #endif
119 
120 int vmiodirenable = TRUE;
121 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
122     "Use the VM system for directory writes");
123 long runningbufspace;
124 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
125     "Amount of presently outstanding async buffer io");
126 static long bufspace;
127 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
128     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
129 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
130     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
131 #else
132 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
133     "Virtual memory used for buffers");
134 #endif
135 static long maxbufspace;
136 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
137     "Maximum allowed value of bufspace (including buf_daemon)");
138 static long bufmallocspace;
139 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
140     "Amount of malloced memory for buffers");
141 static long maxbufmallocspace;
142 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
143     "Maximum amount of malloced memory for buffers");
144 static long lobufspace;
145 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
146     "Minimum amount of buffers we want to have");
147 long hibufspace;
148 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
149     "Maximum allowed value of bufspace (excluding buf_daemon)");
150 static int bufreusecnt;
151 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
152     "Number of times we have reused a buffer");
153 static int buffreekvacnt;
154 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
155     "Number of times we have freed the KVA space from some buffer");
156 static int bufdefragcnt;
157 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
158     "Number of times we have had to repeat buffer allocation to defragment");
159 static long lorunningspace;
160 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
161     "Minimum preferred space used for in-progress I/O");
162 static long hirunningspace;
163 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
164     "Maximum amount of space to use for in-progress I/O");
165 int dirtybufferflushes;
166 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
167     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
168 int bdwriteskip;
169 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
170     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
171 int altbufferflushes;
172 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
173     0, "Number of fsync flushes to limit dirty buffers");
174 static int recursiveflushes;
175 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
176     0, "Number of flushes skipped due to being recursive");
177 static int numdirtybuffers;
178 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
179     "Number of buffers that are dirty (has unwritten changes) at the moment");
180 static int lodirtybuffers;
181 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
182     "How many buffers we want to have free before bufdaemon can sleep");
183 static int hidirtybuffers;
184 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
185     "When the number of dirty buffers is considered severe");
186 int dirtybufthresh;
187 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
188     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
189 static int numfreebuffers;
190 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
191     "Number of free buffers");
192 static int lofreebuffers;
193 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
194    "XXX Unused");
195 static int hifreebuffers;
196 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
197    "XXX Complicatedly unused");
198 static int getnewbufcalls;
199 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
200    "Number of calls to getnewbuf");
201 static int getnewbufrestarts;
202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
203     "Number of times getnewbuf has had to restart a buffer aquisition");
204 static int flushbufqtarget = 100;
205 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
206     "Amount of work to do in flushbufqueues when helping bufdaemon");
207 static long notbufdflashes;
208 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, &notbufdflashes, 0,
209     "Number of dirty buffer flushes done by the bufdaemon helpers");
210 static long barrierwrites;
211 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
212     "Number of barrier writes");
213 
214 /*
215  * Wakeup point for bufdaemon, as well as indicator of whether it is already
216  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
217  * is idling.
218  */
219 static int bd_request;
220 
221 /*
222  * Request for the buf daemon to write more buffers than is indicated by
223  * lodirtybuf.  This may be necessary to push out excess dependencies or
224  * defragment the address space where a simple count of the number of dirty
225  * buffers is insufficient to characterize the demand for flushing them.
226  */
227 static int bd_speedupreq;
228 
229 /*
230  * This lock synchronizes access to bd_request.
231  */
232 static struct mtx bdlock;
233 
234 /*
235  * bogus page -- for I/O to/from partially complete buffers
236  * this is a temporary solution to the problem, but it is not
237  * really that bad.  it would be better to split the buffer
238  * for input in the case of buffers partially already in memory,
239  * but the code is intricate enough already.
240  */
241 vm_page_t bogus_page;
242 
243 /*
244  * Synchronization (sleep/wakeup) variable for active buffer space requests.
245  * Set when wait starts, cleared prior to wakeup().
246  * Used in runningbufwakeup() and waitrunningbufspace().
247  */
248 static int runningbufreq;
249 
250 /*
251  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
252  * waitrunningbufspace().
253  */
254 static struct mtx rbreqlock;
255 
256 /*
257  * Synchronization (sleep/wakeup) variable for buffer requests.
258  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
259  * by and/or.
260  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
261  * getnewbuf(), and getblk().
262  */
263 static int needsbuffer;
264 
265 /*
266  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
267  */
268 static struct mtx nblock;
269 
270 /*
271  * Definitions for the buffer free lists.
272  */
273 #define BUFFER_QUEUES	5	/* number of free buffer queues */
274 
275 #define QUEUE_NONE	0	/* on no queue */
276 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
277 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
278 #define QUEUE_EMPTYKVA	3	/* empty buffer headers w/KVA assignment */
279 #define QUEUE_EMPTY	4	/* empty buffer headers */
280 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
281 
282 /* Queues for free buffers with various properties */
283 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
284 
285 /* Lock for the bufqueues */
286 static struct mtx bqlock;
287 
288 /*
289  * Single global constant for BUF_WMESG, to avoid getting multiple references.
290  * buf_wmesg is referred from macros.
291  */
292 const char *buf_wmesg = BUF_WMESG;
293 
294 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
295 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
296 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
297 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
298 
299 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
300     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
301 static int
302 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
303 {
304 	long lvalue;
305 	int ivalue;
306 
307 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
308 		return (sysctl_handle_long(oidp, arg1, arg2, req));
309 	lvalue = *(long *)arg1;
310 	if (lvalue > INT_MAX)
311 		/* On overflow, still write out a long to trigger ENOMEM. */
312 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
313 	ivalue = lvalue;
314 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
315 }
316 #endif
317 
318 #ifdef DIRECTIO
319 extern void ffs_rawread_setup(void);
320 #endif /* DIRECTIO */
321 /*
322  *	numdirtywakeup:
323  *
324  *	If someone is blocked due to there being too many dirty buffers,
325  *	and numdirtybuffers is now reasonable, wake them up.
326  */
327 
328 static __inline void
329 numdirtywakeup(int level)
330 {
331 
332 	if (numdirtybuffers <= level) {
333 		mtx_lock(&nblock);
334 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
335 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
336 			wakeup(&needsbuffer);
337 		}
338 		mtx_unlock(&nblock);
339 	}
340 }
341 
342 /*
343  *	bufspacewakeup:
344  *
345  *	Called when buffer space is potentially available for recovery.
346  *	getnewbuf() will block on this flag when it is unable to free
347  *	sufficient buffer space.  Buffer space becomes recoverable when
348  *	bp's get placed back in the queues.
349  */
350 
351 static __inline void
352 bufspacewakeup(void)
353 {
354 
355 	/*
356 	 * If someone is waiting for BUF space, wake them up.  Even
357 	 * though we haven't freed the kva space yet, the waiting
358 	 * process will be able to now.
359 	 */
360 	mtx_lock(&nblock);
361 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
362 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
363 		wakeup(&needsbuffer);
364 	}
365 	mtx_unlock(&nblock);
366 }
367 
368 /*
369  * runningbufwakeup() - in-progress I/O accounting.
370  *
371  */
372 void
373 runningbufwakeup(struct buf *bp)
374 {
375 
376 	if (bp->b_runningbufspace) {
377 		atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
378 		bp->b_runningbufspace = 0;
379 		mtx_lock(&rbreqlock);
380 		if (runningbufreq && runningbufspace <= lorunningspace) {
381 			runningbufreq = 0;
382 			wakeup(&runningbufreq);
383 		}
384 		mtx_unlock(&rbreqlock);
385 	}
386 }
387 
388 /*
389  *	bufcountwakeup:
390  *
391  *	Called when a buffer has been added to one of the free queues to
392  *	account for the buffer and to wakeup anyone waiting for free buffers.
393  *	This typically occurs when large amounts of metadata are being handled
394  *	by the buffer cache ( else buffer space runs out first, usually ).
395  */
396 
397 static __inline void
398 bufcountwakeup(struct buf *bp)
399 {
400 	int old;
401 
402 	KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
403 	    ("buf %p already counted as free", bp));
404 	if (bp->b_bufobj != NULL)
405 		mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
406 	bp->b_vflags |= BV_INFREECNT;
407 	old = atomic_fetchadd_int(&numfreebuffers, 1);
408 	KASSERT(old >= 0 && old < nbuf,
409 	    ("numfreebuffers climbed to %d", old + 1));
410 	mtx_lock(&nblock);
411 	if (needsbuffer) {
412 		needsbuffer &= ~VFS_BIO_NEED_ANY;
413 		if (numfreebuffers >= hifreebuffers)
414 			needsbuffer &= ~VFS_BIO_NEED_FREE;
415 		wakeup(&needsbuffer);
416 	}
417 	mtx_unlock(&nblock);
418 }
419 
420 /*
421  *	waitrunningbufspace()
422  *
423  *	runningbufspace is a measure of the amount of I/O currently
424  *	running.  This routine is used in async-write situations to
425  *	prevent creating huge backups of pending writes to a device.
426  *	Only asynchronous writes are governed by this function.
427  *
428  *	Reads will adjust runningbufspace, but will not block based on it.
429  *	The read load has a side effect of reducing the allowed write load.
430  *
431  *	This does NOT turn an async write into a sync write.  It waits
432  *	for earlier writes to complete and generally returns before the
433  *	caller's write has reached the device.
434  */
435 void
436 waitrunningbufspace(void)
437 {
438 
439 	mtx_lock(&rbreqlock);
440 	while (runningbufspace > hirunningspace) {
441 		++runningbufreq;
442 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
443 	}
444 	mtx_unlock(&rbreqlock);
445 }
446 
447 
448 /*
449  *	vfs_buf_test_cache:
450  *
451  *	Called when a buffer is extended.  This function clears the B_CACHE
452  *	bit if the newly extended portion of the buffer does not contain
453  *	valid data.
454  */
455 static __inline
456 void
457 vfs_buf_test_cache(struct buf *bp,
458 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
459 		  vm_page_t m)
460 {
461 
462 	VM_OBJECT_ASSERT_WLOCKED(m->object);
463 	if (bp->b_flags & B_CACHE) {
464 		int base = (foff + off) & PAGE_MASK;
465 		if (vm_page_is_valid(m, base, size) == 0)
466 			bp->b_flags &= ~B_CACHE;
467 	}
468 }
469 
470 /* Wake up the buffer daemon if necessary */
471 static __inline
472 void
473 bd_wakeup(int dirtybuflevel)
474 {
475 
476 	mtx_lock(&bdlock);
477 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
478 		bd_request = 1;
479 		wakeup(&bd_request);
480 	}
481 	mtx_unlock(&bdlock);
482 }
483 
484 /*
485  * bd_speedup - speedup the buffer cache flushing code
486  */
487 
488 void
489 bd_speedup(void)
490 {
491 	int needwake;
492 
493 	mtx_lock(&bdlock);
494 	needwake = 0;
495 	if (bd_speedupreq == 0 || bd_request == 0)
496 		needwake = 1;
497 	bd_speedupreq = 1;
498 	bd_request = 1;
499 	if (needwake)
500 		wakeup(&bd_request);
501 	mtx_unlock(&bdlock);
502 }
503 
504 /*
505  * Calculating buffer cache scaling values and reserve space for buffer
506  * headers.  This is called during low level kernel initialization and
507  * may be called more then once.  We CANNOT write to the memory area
508  * being reserved at this time.
509  */
510 caddr_t
511 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
512 {
513 	int tuned_nbuf;
514 	long maxbuf;
515 
516 	/*
517 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
518 	 * PAGE_SIZE is >= 1K)
519 	 */
520 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
521 
522 	/*
523 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
524 	 * For the first 64MB of ram nominally allocate sufficient buffers to
525 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
526 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
527 	 * the buffer cache we limit the eventual kva reservation to
528 	 * maxbcache bytes.
529 	 *
530 	 * factor represents the 1/4 x ram conversion.
531 	 */
532 	if (nbuf == 0) {
533 		int factor = 4 * BKVASIZE / 1024;
534 
535 		nbuf = 50;
536 		if (physmem_est > 4096)
537 			nbuf += min((physmem_est - 4096) / factor,
538 			    65536 / factor);
539 		if (physmem_est > 65536)
540 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
541 
542 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
543 			nbuf = maxbcache / BKVASIZE;
544 		tuned_nbuf = 1;
545 	} else
546 		tuned_nbuf = 0;
547 
548 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
549 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
550 	if (nbuf > maxbuf) {
551 		if (!tuned_nbuf)
552 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
553 			    maxbuf);
554 		nbuf = maxbuf;
555 	}
556 
557 	/*
558 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
559 	 * We have no less then 16 and no more then 256.
560 	 */
561 	nswbuf = max(min(nbuf/4, 256), 16);
562 #ifdef NSWBUF_MIN
563 	if (nswbuf < NSWBUF_MIN)
564 		nswbuf = NSWBUF_MIN;
565 #endif
566 #ifdef DIRECTIO
567 	ffs_rawread_setup();
568 #endif
569 
570 	/*
571 	 * Reserve space for the buffer cache buffers
572 	 */
573 	swbuf = (void *)v;
574 	v = (caddr_t)(swbuf + nswbuf);
575 	buf = (void *)v;
576 	v = (caddr_t)(buf + nbuf);
577 
578 	return(v);
579 }
580 
581 /* Initialize the buffer subsystem.  Called before use of any buffers. */
582 void
583 bufinit(void)
584 {
585 	struct buf *bp;
586 	int i;
587 
588 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
589 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
590 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
591 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
592 
593 	/* next, make a null set of free lists */
594 	for (i = 0; i < BUFFER_QUEUES; i++)
595 		TAILQ_INIT(&bufqueues[i]);
596 
597 	/* finally, initialize each buffer header and stick on empty q */
598 	for (i = 0; i < nbuf; i++) {
599 		bp = &buf[i];
600 		bzero(bp, sizeof *bp);
601 		bp->b_flags = B_INVAL;	/* we're just an empty header */
602 		bp->b_rcred = NOCRED;
603 		bp->b_wcred = NOCRED;
604 		bp->b_qindex = QUEUE_EMPTY;
605 		bp->b_vflags = BV_INFREECNT;	/* buf is counted as free */
606 		bp->b_xflags = 0;
607 		LIST_INIT(&bp->b_dep);
608 		BUF_LOCKINIT(bp);
609 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
610 	}
611 
612 	/*
613 	 * maxbufspace is the absolute maximum amount of buffer space we are
614 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
615 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
616 	 * used by most other processes.  The differential is required to
617 	 * ensure that buf_daemon is able to run when other processes might
618 	 * be blocked waiting for buffer space.
619 	 *
620 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
621 	 * this may result in KVM fragmentation which is not handled optimally
622 	 * by the system.
623 	 */
624 	maxbufspace = (long)nbuf * BKVASIZE;
625 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
626 	lobufspace = hibufspace - MAXBSIZE;
627 
628 	/*
629 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
630 	 * arbitrarily and may need further tuning. It corresponds to
631 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
632 	 * which fits with many RAID controllers' tagged queuing limits.
633 	 * The lower 1 MiB limit is the historical upper limit for
634 	 * hirunningspace.
635 	 */
636 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
637 	    16 * 1024 * 1024), 1024 * 1024);
638 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
639 
640 /*
641  * Limit the amount of malloc memory since it is wired permanently into
642  * the kernel space.  Even though this is accounted for in the buffer
643  * allocation, we don't want the malloced region to grow uncontrolled.
644  * The malloc scheme improves memory utilization significantly on average
645  * (small) directories.
646  */
647 	maxbufmallocspace = hibufspace / 20;
648 
649 /*
650  * Reduce the chance of a deadlock occuring by limiting the number
651  * of delayed-write dirty buffers we allow to stack up.
652  */
653 	hidirtybuffers = nbuf / 4 + 20;
654 	dirtybufthresh = hidirtybuffers * 9 / 10;
655 	numdirtybuffers = 0;
656 /*
657  * To support extreme low-memory systems, make sure hidirtybuffers cannot
658  * eat up all available buffer space.  This occurs when our minimum cannot
659  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
660  * BKVASIZE'd buffers.
661  */
662 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
663 		hidirtybuffers >>= 1;
664 	}
665 	lodirtybuffers = hidirtybuffers / 2;
666 
667 /*
668  * Try to keep the number of free buffers in the specified range,
669  * and give special processes (e.g. like buf_daemon) access to an
670  * emergency reserve.
671  */
672 	lofreebuffers = nbuf / 18 + 5;
673 	hifreebuffers = 2 * lofreebuffers;
674 	numfreebuffers = nbuf;
675 
676 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
677 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
678 }
679 
680 /*
681  * bfreekva() - free the kva allocation for a buffer.
682  *
683  *	Since this call frees up buffer space, we call bufspacewakeup().
684  */
685 static void
686 bfreekva(struct buf *bp)
687 {
688 
689 	if (bp->b_kvasize) {
690 		atomic_add_int(&buffreekvacnt, 1);
691 		atomic_subtract_long(&bufspace, bp->b_kvasize);
692 		vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
693 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
694 		bp->b_kvasize = 0;
695 		bufspacewakeup();
696 	}
697 }
698 
699 /*
700  *	bremfree:
701  *
702  *	Mark the buffer for removal from the appropriate free list in brelse.
703  *
704  */
705 void
706 bremfree(struct buf *bp)
707 {
708 	int old;
709 
710 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
711 	KASSERT((bp->b_flags & B_REMFREE) == 0,
712 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
713 	KASSERT(bp->b_qindex != QUEUE_NONE,
714 	    ("bremfree: buffer %p not on a queue.", bp));
715 	BUF_ASSERT_HELD(bp);
716 
717 	bp->b_flags |= B_REMFREE;
718 	/* Fixup numfreebuffers count.  */
719 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
720 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
721 		    ("buf %p not counted in numfreebuffers", bp));
722 		if (bp->b_bufobj != NULL)
723 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
724 		bp->b_vflags &= ~BV_INFREECNT;
725 		old = atomic_fetchadd_int(&numfreebuffers, -1);
726 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
727 	}
728 }
729 
730 /*
731  *	bremfreef:
732  *
733  *	Force an immediate removal from a free list.  Used only in nfs when
734  *	it abuses the b_freelist pointer.
735  */
736 void
737 bremfreef(struct buf *bp)
738 {
739 	mtx_lock(&bqlock);
740 	bremfreel(bp);
741 	mtx_unlock(&bqlock);
742 }
743 
744 /*
745  *	bremfreel:
746  *
747  *	Removes a buffer from the free list, must be called with the
748  *	bqlock held.
749  */
750 static void
751 bremfreel(struct buf *bp)
752 {
753 	int old;
754 
755 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
756 	    bp, bp->b_vp, bp->b_flags);
757 	KASSERT(bp->b_qindex != QUEUE_NONE,
758 	    ("bremfreel: buffer %p not on a queue.", bp));
759 	BUF_ASSERT_HELD(bp);
760 	mtx_assert(&bqlock, MA_OWNED);
761 
762 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
763 	bp->b_qindex = QUEUE_NONE;
764 	/*
765 	 * If this was a delayed bremfree() we only need to remove the buffer
766 	 * from the queue and return the stats are already done.
767 	 */
768 	if (bp->b_flags & B_REMFREE) {
769 		bp->b_flags &= ~B_REMFREE;
770 		return;
771 	}
772 	/*
773 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
774 	 * delayed-write, the buffer was free and we must decrement
775 	 * numfreebuffers.
776 	 */
777 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
778 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
779 		    ("buf %p not counted in numfreebuffers", bp));
780 		if (bp->b_bufobj != NULL)
781 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
782 		bp->b_vflags &= ~BV_INFREECNT;
783 		old = atomic_fetchadd_int(&numfreebuffers, -1);
784 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
785 	}
786 }
787 
788 /*
789  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
790  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
791  * the buffer is valid and we do not have to do anything.
792  */
793 void
794 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
795     int cnt, struct ucred * cred)
796 {
797 	struct buf *rabp;
798 	int i;
799 
800 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
801 		if (inmem(vp, *rablkno))
802 			continue;
803 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
804 
805 		if ((rabp->b_flags & B_CACHE) == 0) {
806 			if (!TD_IS_IDLETHREAD(curthread))
807 				curthread->td_ru.ru_inblock++;
808 			rabp->b_flags |= B_ASYNC;
809 			rabp->b_flags &= ~B_INVAL;
810 			rabp->b_ioflags &= ~BIO_ERROR;
811 			rabp->b_iocmd = BIO_READ;
812 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
813 				rabp->b_rcred = crhold(cred);
814 			vfs_busy_pages(rabp, 0);
815 			BUF_KERNPROC(rabp);
816 			rabp->b_iooffset = dbtob(rabp->b_blkno);
817 			bstrategy(rabp);
818 		} else {
819 			brelse(rabp);
820 		}
821 	}
822 }
823 
824 /*
825  * Entry point for bread() and breadn() via #defines in sys/buf.h.
826  *
827  * Get a buffer with the specified data.  Look in the cache first.  We
828  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
829  * is set, the buffer is valid and we do not have to do anything, see
830  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
831  */
832 int
833 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
834     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
835 {
836 	struct buf *bp;
837 	int rv = 0, readwait = 0;
838 
839 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
840 	/*
841 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
842 	 */
843 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
844 	if (bp == NULL)
845 		return (EBUSY);
846 
847 	/* if not found in cache, do some I/O */
848 	if ((bp->b_flags & B_CACHE) == 0) {
849 		if (!TD_IS_IDLETHREAD(curthread))
850 			curthread->td_ru.ru_inblock++;
851 		bp->b_iocmd = BIO_READ;
852 		bp->b_flags &= ~B_INVAL;
853 		bp->b_ioflags &= ~BIO_ERROR;
854 		if (bp->b_rcred == NOCRED && cred != NOCRED)
855 			bp->b_rcred = crhold(cred);
856 		vfs_busy_pages(bp, 0);
857 		bp->b_iooffset = dbtob(bp->b_blkno);
858 		bstrategy(bp);
859 		++readwait;
860 	}
861 
862 	breada(vp, rablkno, rabsize, cnt, cred);
863 
864 	if (readwait) {
865 		rv = bufwait(bp);
866 	}
867 	return (rv);
868 }
869 
870 /*
871  * Write, release buffer on completion.  (Done by iodone
872  * if async).  Do not bother writing anything if the buffer
873  * is invalid.
874  *
875  * Note that we set B_CACHE here, indicating that buffer is
876  * fully valid and thus cacheable.  This is true even of NFS
877  * now so we set it generally.  This could be set either here
878  * or in biodone() since the I/O is synchronous.  We put it
879  * here.
880  */
881 int
882 bufwrite(struct buf *bp)
883 {
884 	int oldflags;
885 	struct vnode *vp;
886 	int vp_md;
887 
888 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
889 	if (bp->b_flags & B_INVAL) {
890 		brelse(bp);
891 		return (0);
892 	}
893 
894 	if (bp->b_flags & B_BARRIER)
895 		barrierwrites++;
896 
897 	oldflags = bp->b_flags;
898 
899 	BUF_ASSERT_HELD(bp);
900 
901 	if (bp->b_pin_count > 0)
902 		bunpin_wait(bp);
903 
904 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
905 	    ("FFS background buffer should not get here %p", bp));
906 
907 	vp = bp->b_vp;
908 	if (vp)
909 		vp_md = vp->v_vflag & VV_MD;
910 	else
911 		vp_md = 0;
912 
913 	/* Mark the buffer clean */
914 	bundirty(bp);
915 
916 	bp->b_flags &= ~B_DONE;
917 	bp->b_ioflags &= ~BIO_ERROR;
918 	bp->b_flags |= B_CACHE;
919 	bp->b_iocmd = BIO_WRITE;
920 
921 	bufobj_wref(bp->b_bufobj);
922 	vfs_busy_pages(bp, 1);
923 
924 	/*
925 	 * Normal bwrites pipeline writes
926 	 */
927 	bp->b_runningbufspace = bp->b_bufsize;
928 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
929 
930 	if (!TD_IS_IDLETHREAD(curthread))
931 		curthread->td_ru.ru_oublock++;
932 	if (oldflags & B_ASYNC)
933 		BUF_KERNPROC(bp);
934 	bp->b_iooffset = dbtob(bp->b_blkno);
935 	bstrategy(bp);
936 
937 	if ((oldflags & B_ASYNC) == 0) {
938 		int rtval = bufwait(bp);
939 		brelse(bp);
940 		return (rtval);
941 	} else {
942 		/*
943 		 * don't allow the async write to saturate the I/O
944 		 * system.  We will not deadlock here because
945 		 * we are blocking waiting for I/O that is already in-progress
946 		 * to complete. We do not block here if it is the update
947 		 * or syncer daemon trying to clean up as that can lead
948 		 * to deadlock.
949 		 */
950 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
951 			waitrunningbufspace();
952 	}
953 
954 	return (0);
955 }
956 
957 void
958 bufbdflush(struct bufobj *bo, struct buf *bp)
959 {
960 	struct buf *nbp;
961 
962 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
963 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
964 		altbufferflushes++;
965 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
966 		BO_LOCK(bo);
967 		/*
968 		 * Try to find a buffer to flush.
969 		 */
970 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
971 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
972 			    BUF_LOCK(nbp,
973 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
974 				continue;
975 			if (bp == nbp)
976 				panic("bdwrite: found ourselves");
977 			BO_UNLOCK(bo);
978 			/* Don't countdeps with the bo lock held. */
979 			if (buf_countdeps(nbp, 0)) {
980 				BO_LOCK(bo);
981 				BUF_UNLOCK(nbp);
982 				continue;
983 			}
984 			if (nbp->b_flags & B_CLUSTEROK) {
985 				vfs_bio_awrite(nbp);
986 			} else {
987 				bremfree(nbp);
988 				bawrite(nbp);
989 			}
990 			dirtybufferflushes++;
991 			break;
992 		}
993 		if (nbp == NULL)
994 			BO_UNLOCK(bo);
995 	}
996 }
997 
998 /*
999  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1000  * anything if the buffer is marked invalid.
1001  *
1002  * Note that since the buffer must be completely valid, we can safely
1003  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1004  * biodone() in order to prevent getblk from writing the buffer
1005  * out synchronously.
1006  */
1007 void
1008 bdwrite(struct buf *bp)
1009 {
1010 	struct thread *td = curthread;
1011 	struct vnode *vp;
1012 	struct bufobj *bo;
1013 
1014 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1015 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1016 	KASSERT((bp->b_flags & B_BARRIER) == 0,
1017 	    ("Barrier request in delayed write %p", bp));
1018 	BUF_ASSERT_HELD(bp);
1019 
1020 	if (bp->b_flags & B_INVAL) {
1021 		brelse(bp);
1022 		return;
1023 	}
1024 
1025 	/*
1026 	 * If we have too many dirty buffers, don't create any more.
1027 	 * If we are wildly over our limit, then force a complete
1028 	 * cleanup. Otherwise, just keep the situation from getting
1029 	 * out of control. Note that we have to avoid a recursive
1030 	 * disaster and not try to clean up after our own cleanup!
1031 	 */
1032 	vp = bp->b_vp;
1033 	bo = bp->b_bufobj;
1034 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1035 		td->td_pflags |= TDP_INBDFLUSH;
1036 		BO_BDFLUSH(bo, bp);
1037 		td->td_pflags &= ~TDP_INBDFLUSH;
1038 	} else
1039 		recursiveflushes++;
1040 
1041 	bdirty(bp);
1042 	/*
1043 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1044 	 * true even of NFS now.
1045 	 */
1046 	bp->b_flags |= B_CACHE;
1047 
1048 	/*
1049 	 * This bmap keeps the system from needing to do the bmap later,
1050 	 * perhaps when the system is attempting to do a sync.  Since it
1051 	 * is likely that the indirect block -- or whatever other datastructure
1052 	 * that the filesystem needs is still in memory now, it is a good
1053 	 * thing to do this.  Note also, that if the pageout daemon is
1054 	 * requesting a sync -- there might not be enough memory to do
1055 	 * the bmap then...  So, this is important to do.
1056 	 */
1057 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1058 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1059 	}
1060 
1061 	/*
1062 	 * Set the *dirty* buffer range based upon the VM system dirty
1063 	 * pages.
1064 	 *
1065 	 * Mark the buffer pages as clean.  We need to do this here to
1066 	 * satisfy the vnode_pager and the pageout daemon, so that it
1067 	 * thinks that the pages have been "cleaned".  Note that since
1068 	 * the pages are in a delayed write buffer -- the VFS layer
1069 	 * "will" see that the pages get written out on the next sync,
1070 	 * or perhaps the cluster will be completed.
1071 	 */
1072 	vfs_clean_pages_dirty_buf(bp);
1073 	bqrelse(bp);
1074 
1075 	/*
1076 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
1077 	 * buffers (midpoint between our recovery point and our stall
1078 	 * point).
1079 	 */
1080 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1081 
1082 	/*
1083 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1084 	 * due to the softdep code.
1085 	 */
1086 }
1087 
1088 /*
1089  *	bdirty:
1090  *
1091  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1092  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1093  *	itself to properly update it in the dirty/clean lists.  We mark it
1094  *	B_DONE to ensure that any asynchronization of the buffer properly
1095  *	clears B_DONE ( else a panic will occur later ).
1096  *
1097  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1098  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1099  *	should only be called if the buffer is known-good.
1100  *
1101  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1102  *	count.
1103  *
1104  *	The buffer must be on QUEUE_NONE.
1105  */
1106 void
1107 bdirty(struct buf *bp)
1108 {
1109 
1110 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1111 	    bp, bp->b_vp, bp->b_flags);
1112 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1113 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1114 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1115 	BUF_ASSERT_HELD(bp);
1116 	bp->b_flags &= ~(B_RELBUF);
1117 	bp->b_iocmd = BIO_WRITE;
1118 
1119 	if ((bp->b_flags & B_DELWRI) == 0) {
1120 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1121 		reassignbuf(bp);
1122 		atomic_add_int(&numdirtybuffers, 1);
1123 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1124 	}
1125 }
1126 
1127 /*
1128  *	bundirty:
1129  *
1130  *	Clear B_DELWRI for buffer.
1131  *
1132  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1133  *	count.
1134  *
1135  *	The buffer must be on QUEUE_NONE.
1136  */
1137 
1138 void
1139 bundirty(struct buf *bp)
1140 {
1141 
1142 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1143 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1144 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1145 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1146 	BUF_ASSERT_HELD(bp);
1147 
1148 	if (bp->b_flags & B_DELWRI) {
1149 		bp->b_flags &= ~B_DELWRI;
1150 		reassignbuf(bp);
1151 		atomic_subtract_int(&numdirtybuffers, 1);
1152 		numdirtywakeup(lodirtybuffers);
1153 	}
1154 	/*
1155 	 * Since it is now being written, we can clear its deferred write flag.
1156 	 */
1157 	bp->b_flags &= ~B_DEFERRED;
1158 }
1159 
1160 /*
1161  *	bawrite:
1162  *
1163  *	Asynchronous write.  Start output on a buffer, but do not wait for
1164  *	it to complete.  The buffer is released when the output completes.
1165  *
1166  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1167  *	B_INVAL buffers.  Not us.
1168  */
1169 void
1170 bawrite(struct buf *bp)
1171 {
1172 
1173 	bp->b_flags |= B_ASYNC;
1174 	(void) bwrite(bp);
1175 }
1176 
1177 /*
1178  *	babarrierwrite:
1179  *
1180  *	Asynchronous barrier write.  Start output on a buffer, but do not
1181  *	wait for it to complete.  Place a write barrier after this write so
1182  *	that this buffer and all buffers written before it are committed to
1183  *	the disk before any buffers written after this write are committed
1184  *	to the disk.  The buffer is released when the output completes.
1185  */
1186 void
1187 babarrierwrite(struct buf *bp)
1188 {
1189 
1190 	bp->b_flags |= B_ASYNC | B_BARRIER;
1191 	(void) bwrite(bp);
1192 }
1193 
1194 /*
1195  *	bbarrierwrite:
1196  *
1197  *	Synchronous barrier write.  Start output on a buffer and wait for
1198  *	it to complete.  Place a write barrier after this write so that
1199  *	this buffer and all buffers written before it are committed to
1200  *	the disk before any buffers written after this write are committed
1201  *	to the disk.  The buffer is released when the output completes.
1202  */
1203 int
1204 bbarrierwrite(struct buf *bp)
1205 {
1206 
1207 	bp->b_flags |= B_BARRIER;
1208 	return (bwrite(bp));
1209 }
1210 
1211 /*
1212  *	bwillwrite:
1213  *
1214  *	Called prior to the locking of any vnodes when we are expecting to
1215  *	write.  We do not want to starve the buffer cache with too many
1216  *	dirty buffers so we block here.  By blocking prior to the locking
1217  *	of any vnodes we attempt to avoid the situation where a locked vnode
1218  *	prevents the various system daemons from flushing related buffers.
1219  */
1220 
1221 void
1222 bwillwrite(void)
1223 {
1224 
1225 	if (numdirtybuffers >= hidirtybuffers) {
1226 		mtx_lock(&nblock);
1227 		while (numdirtybuffers >= hidirtybuffers) {
1228 			bd_wakeup(1);
1229 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1230 			msleep(&needsbuffer, &nblock,
1231 			    (PRIBIO + 4), "flswai", 0);
1232 		}
1233 		mtx_unlock(&nblock);
1234 	}
1235 }
1236 
1237 /*
1238  * Return true if we have too many dirty buffers.
1239  */
1240 int
1241 buf_dirty_count_severe(void)
1242 {
1243 
1244 	return(numdirtybuffers >= hidirtybuffers);
1245 }
1246 
1247 static __noinline int
1248 buf_vm_page_count_severe(void)
1249 {
1250 
1251 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1252 
1253 	return vm_page_count_severe();
1254 }
1255 
1256 /*
1257  *	brelse:
1258  *
1259  *	Release a busy buffer and, if requested, free its resources.  The
1260  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1261  *	to be accessed later as a cache entity or reused for other purposes.
1262  */
1263 void
1264 brelse(struct buf *bp)
1265 {
1266 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1267 	    bp, bp->b_vp, bp->b_flags);
1268 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1269 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1270 
1271 	if (BUF_LOCKRECURSED(bp)) {
1272 		/*
1273 		 * Do not process, in particular, do not handle the
1274 		 * B_INVAL/B_RELBUF and do not release to free list.
1275 		 */
1276 		BUF_UNLOCK(bp);
1277 		return;
1278 	}
1279 
1280 	if (bp->b_flags & B_MANAGED) {
1281 		bqrelse(bp);
1282 		return;
1283 	}
1284 
1285 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1286 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1287 		/*
1288 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1289 		 * pages from being scrapped.  If the error is anything
1290 		 * other than an I/O error (EIO), assume that retrying
1291 		 * is futile.
1292 		 */
1293 		bp->b_ioflags &= ~BIO_ERROR;
1294 		bdirty(bp);
1295 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1296 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1297 		/*
1298 		 * Either a failed I/O or we were asked to free or not
1299 		 * cache the buffer.
1300 		 */
1301 		bp->b_flags |= B_INVAL;
1302 		if (!LIST_EMPTY(&bp->b_dep))
1303 			buf_deallocate(bp);
1304 		if (bp->b_flags & B_DELWRI) {
1305 			atomic_subtract_int(&numdirtybuffers, 1);
1306 			numdirtywakeup(lodirtybuffers);
1307 		}
1308 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1309 		if ((bp->b_flags & B_VMIO) == 0) {
1310 			if (bp->b_bufsize)
1311 				allocbuf(bp, 0);
1312 			if (bp->b_vp)
1313 				brelvp(bp);
1314 		}
1315 	}
1316 
1317 	/*
1318 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1319 	 * is called with B_DELWRI set, the underlying pages may wind up
1320 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1321 	 * because pages associated with a B_DELWRI bp are marked clean.
1322 	 *
1323 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1324 	 * if B_DELWRI is set.
1325 	 *
1326 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1327 	 * on pages to return pages to the VM page queues.
1328 	 */
1329 	if (bp->b_flags & B_DELWRI)
1330 		bp->b_flags &= ~B_RELBUF;
1331 	else if (buf_vm_page_count_severe()) {
1332 		/*
1333 		 * The locking of the BO_LOCK is not necessary since
1334 		 * BKGRDINPROG cannot be set while we hold the buf
1335 		 * lock, it can only be cleared if it is already
1336 		 * pending.
1337 		 */
1338 		if (bp->b_vp) {
1339 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1340 				bp->b_flags |= B_RELBUF;
1341 		} else
1342 			bp->b_flags |= B_RELBUF;
1343 	}
1344 
1345 	/*
1346 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1347 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1348 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1349 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1350 	 *
1351 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1352 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1353 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1354 	 *
1355 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1356 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1357 	 * the commit state and we cannot afford to lose the buffer. If the
1358 	 * buffer has a background write in progress, we need to keep it
1359 	 * around to prevent it from being reconstituted and starting a second
1360 	 * background write.
1361 	 */
1362 	if ((bp->b_flags & B_VMIO)
1363 	    && !(bp->b_vp->v_mount != NULL &&
1364 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1365 		 !vn_isdisk(bp->b_vp, NULL) &&
1366 		 (bp->b_flags & B_DELWRI))
1367 	    ) {
1368 
1369 		int i, j, resid;
1370 		vm_page_t m;
1371 		off_t foff;
1372 		vm_pindex_t poff;
1373 		vm_object_t obj;
1374 
1375 		obj = bp->b_bufobj->bo_object;
1376 
1377 		/*
1378 		 * Get the base offset and length of the buffer.  Note that
1379 		 * in the VMIO case if the buffer block size is not
1380 		 * page-aligned then b_data pointer may not be page-aligned.
1381 		 * But our b_pages[] array *IS* page aligned.
1382 		 *
1383 		 * block sizes less then DEV_BSIZE (usually 512) are not
1384 		 * supported due to the page granularity bits (m->valid,
1385 		 * m->dirty, etc...).
1386 		 *
1387 		 * See man buf(9) for more information
1388 		 */
1389 		resid = bp->b_bufsize;
1390 		foff = bp->b_offset;
1391 		VM_OBJECT_WLOCK(obj);
1392 		for (i = 0; i < bp->b_npages; i++) {
1393 			int had_bogus = 0;
1394 
1395 			m = bp->b_pages[i];
1396 
1397 			/*
1398 			 * If we hit a bogus page, fixup *all* the bogus pages
1399 			 * now.
1400 			 */
1401 			if (m == bogus_page) {
1402 				poff = OFF_TO_IDX(bp->b_offset);
1403 				had_bogus = 1;
1404 
1405 				for (j = i; j < bp->b_npages; j++) {
1406 					vm_page_t mtmp;
1407 					mtmp = bp->b_pages[j];
1408 					if (mtmp == bogus_page) {
1409 						mtmp = vm_page_lookup(obj, poff + j);
1410 						if (!mtmp) {
1411 							panic("brelse: page missing\n");
1412 						}
1413 						bp->b_pages[j] = mtmp;
1414 					}
1415 				}
1416 
1417 				if ((bp->b_flags & B_INVAL) == 0) {
1418 					pmap_qenter(
1419 					    trunc_page((vm_offset_t)bp->b_data),
1420 					    bp->b_pages, bp->b_npages);
1421 				}
1422 				m = bp->b_pages[i];
1423 			}
1424 			if ((bp->b_flags & B_NOCACHE) ||
1425 			    (bp->b_ioflags & BIO_ERROR &&
1426 			     bp->b_iocmd == BIO_READ)) {
1427 				int poffset = foff & PAGE_MASK;
1428 				int presid = resid > (PAGE_SIZE - poffset) ?
1429 					(PAGE_SIZE - poffset) : resid;
1430 
1431 				KASSERT(presid >= 0, ("brelse: extra page"));
1432 				vm_page_set_invalid(m, poffset, presid);
1433 				if (had_bogus)
1434 					printf("avoided corruption bug in bogus_page/brelse code\n");
1435 			}
1436 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1437 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1438 		}
1439 		VM_OBJECT_WUNLOCK(obj);
1440 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1441 			vfs_vmio_release(bp);
1442 
1443 	} else if (bp->b_flags & B_VMIO) {
1444 
1445 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1446 			vfs_vmio_release(bp);
1447 		}
1448 
1449 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1450 		if (bp->b_bufsize != 0)
1451 			allocbuf(bp, 0);
1452 		if (bp->b_vp != NULL)
1453 			brelvp(bp);
1454 	}
1455 
1456 	/* enqueue */
1457 	mtx_lock(&bqlock);
1458 	/* Handle delayed bremfree() processing. */
1459 	if (bp->b_flags & B_REMFREE) {
1460 		struct bufobj *bo;
1461 
1462 		bo = bp->b_bufobj;
1463 		if (bo != NULL)
1464 			BO_LOCK(bo);
1465 		bremfreel(bp);
1466 		if (bo != NULL)
1467 			BO_UNLOCK(bo);
1468 	}
1469 	if (bp->b_qindex != QUEUE_NONE)
1470 		panic("brelse: free buffer onto another queue???");
1471 
1472 	/*
1473 	 * If the buffer has junk contents signal it and eventually
1474 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1475 	 * doesn't find it.
1476 	 */
1477 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1478 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1479 		bp->b_flags |= B_INVAL;
1480 	if (bp->b_flags & B_INVAL) {
1481 		if (bp->b_flags & B_DELWRI)
1482 			bundirty(bp);
1483 		if (bp->b_vp)
1484 			brelvp(bp);
1485 	}
1486 
1487 	/* buffers with no memory */
1488 	if (bp->b_bufsize == 0) {
1489 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1490 		if (bp->b_vflags & BV_BKGRDINPROG)
1491 			panic("losing buffer 1");
1492 		if (bp->b_kvasize) {
1493 			bp->b_qindex = QUEUE_EMPTYKVA;
1494 		} else {
1495 			bp->b_qindex = QUEUE_EMPTY;
1496 		}
1497 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1498 	/* buffers with junk contents */
1499 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1500 	    (bp->b_ioflags & BIO_ERROR)) {
1501 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1502 		if (bp->b_vflags & BV_BKGRDINPROG)
1503 			panic("losing buffer 2");
1504 		bp->b_qindex = QUEUE_CLEAN;
1505 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1506 	/* remaining buffers */
1507 	} else {
1508 		if (bp->b_flags & B_DELWRI)
1509 			bp->b_qindex = QUEUE_DIRTY;
1510 		else
1511 			bp->b_qindex = QUEUE_CLEAN;
1512 		if (bp->b_flags & B_AGE)
1513 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1514 		else
1515 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1516 	}
1517 	mtx_unlock(&bqlock);
1518 
1519 	/*
1520 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1521 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1522 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1523 	 * if B_INVAL is set ).
1524 	 */
1525 
1526 	if (!(bp->b_flags & B_DELWRI)) {
1527 		struct bufobj *bo;
1528 
1529 		bo = bp->b_bufobj;
1530 		if (bo != NULL)
1531 			BO_LOCK(bo);
1532 		bufcountwakeup(bp);
1533 		if (bo != NULL)
1534 			BO_UNLOCK(bo);
1535 	}
1536 
1537 	/*
1538 	 * Something we can maybe free or reuse
1539 	 */
1540 	if (bp->b_bufsize || bp->b_kvasize)
1541 		bufspacewakeup();
1542 
1543 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1544 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1545 		panic("brelse: not dirty");
1546 	/* unlock */
1547 	BUF_UNLOCK(bp);
1548 }
1549 
1550 /*
1551  * Release a buffer back to the appropriate queue but do not try to free
1552  * it.  The buffer is expected to be used again soon.
1553  *
1554  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1555  * biodone() to requeue an async I/O on completion.  It is also used when
1556  * known good buffers need to be requeued but we think we may need the data
1557  * again soon.
1558  *
1559  * XXX we should be able to leave the B_RELBUF hint set on completion.
1560  */
1561 void
1562 bqrelse(struct buf *bp)
1563 {
1564 	struct bufobj *bo;
1565 
1566 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1567 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1568 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1569 
1570 	if (BUF_LOCKRECURSED(bp)) {
1571 		/* do not release to free list */
1572 		BUF_UNLOCK(bp);
1573 		return;
1574 	}
1575 
1576 	bo = bp->b_bufobj;
1577 	if (bp->b_flags & B_MANAGED) {
1578 		if (bp->b_flags & B_REMFREE) {
1579 			mtx_lock(&bqlock);
1580 			if (bo != NULL)
1581 				BO_LOCK(bo);
1582 			bremfreel(bp);
1583 			if (bo != NULL)
1584 				BO_UNLOCK(bo);
1585 			mtx_unlock(&bqlock);
1586 		}
1587 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1588 		BUF_UNLOCK(bp);
1589 		return;
1590 	}
1591 
1592 	mtx_lock(&bqlock);
1593 	/* Handle delayed bremfree() processing. */
1594 	if (bp->b_flags & B_REMFREE) {
1595 		if (bo != NULL)
1596 			BO_LOCK(bo);
1597 		bremfreel(bp);
1598 		if (bo != NULL)
1599 			BO_UNLOCK(bo);
1600 	}
1601 	if (bp->b_qindex != QUEUE_NONE)
1602 		panic("bqrelse: free buffer onto another queue???");
1603 	/* buffers with stale but valid contents */
1604 	if (bp->b_flags & B_DELWRI) {
1605 		bp->b_qindex = QUEUE_DIRTY;
1606 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1607 	} else {
1608 		/*
1609 		 * The locking of the BO_LOCK for checking of the
1610 		 * BV_BKGRDINPROG is not necessary since the
1611 		 * BV_BKGRDINPROG cannot be set while we hold the buf
1612 		 * lock, it can only be cleared if it is already
1613 		 * pending.
1614 		 */
1615 		if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1616 			bp->b_qindex = QUEUE_CLEAN;
1617 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1618 			    b_freelist);
1619 		} else {
1620 			/*
1621 			 * We are too low on memory, we have to try to free
1622 			 * the buffer (most importantly: the wired pages
1623 			 * making up its backing store) *now*.
1624 			 */
1625 			mtx_unlock(&bqlock);
1626 			brelse(bp);
1627 			return;
1628 		}
1629 	}
1630 	mtx_unlock(&bqlock);
1631 
1632 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) {
1633 		if (bo != NULL)
1634 			BO_LOCK(bo);
1635 		bufcountwakeup(bp);
1636 		if (bo != NULL)
1637 			BO_UNLOCK(bo);
1638 	}
1639 
1640 	/*
1641 	 * Something we can maybe free or reuse.
1642 	 */
1643 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1644 		bufspacewakeup();
1645 
1646 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1647 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1648 		panic("bqrelse: not dirty");
1649 	/* unlock */
1650 	BUF_UNLOCK(bp);
1651 }
1652 
1653 /* Give pages used by the bp back to the VM system (where possible) */
1654 static void
1655 vfs_vmio_release(struct buf *bp)
1656 {
1657 	int i;
1658 	vm_page_t m;
1659 
1660 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1661 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
1662 	for (i = 0; i < bp->b_npages; i++) {
1663 		m = bp->b_pages[i];
1664 		bp->b_pages[i] = NULL;
1665 		/*
1666 		 * In order to keep page LRU ordering consistent, put
1667 		 * everything on the inactive queue.
1668 		 */
1669 		vm_page_lock(m);
1670 		vm_page_unwire(m, 0);
1671 		/*
1672 		 * We don't mess with busy pages, it is
1673 		 * the responsibility of the process that
1674 		 * busied the pages to deal with them.
1675 		 */
1676 		if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 &&
1677 		    m->wire_count == 0) {
1678 			/*
1679 			 * Might as well free the page if we can and it has
1680 			 * no valid data.  We also free the page if the
1681 			 * buffer was used for direct I/O
1682 			 */
1683 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1684 				vm_page_free(m);
1685 			} else if (bp->b_flags & B_DIRECT) {
1686 				vm_page_try_to_free(m);
1687 			} else if (buf_vm_page_count_severe()) {
1688 				vm_page_try_to_cache(m);
1689 			}
1690 		}
1691 		vm_page_unlock(m);
1692 	}
1693 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
1694 
1695 	if (bp->b_bufsize) {
1696 		bufspacewakeup();
1697 		bp->b_bufsize = 0;
1698 	}
1699 	bp->b_npages = 0;
1700 	bp->b_flags &= ~B_VMIO;
1701 	if (bp->b_vp)
1702 		brelvp(bp);
1703 }
1704 
1705 /*
1706  * Check to see if a block at a particular lbn is available for a clustered
1707  * write.
1708  */
1709 static int
1710 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1711 {
1712 	struct buf *bpa;
1713 	int match;
1714 
1715 	match = 0;
1716 
1717 	/* If the buf isn't in core skip it */
1718 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1719 		return (0);
1720 
1721 	/* If the buf is busy we don't want to wait for it */
1722 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1723 		return (0);
1724 
1725 	/* Only cluster with valid clusterable delayed write buffers */
1726 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1727 	    (B_DELWRI | B_CLUSTEROK))
1728 		goto done;
1729 
1730 	if (bpa->b_bufsize != size)
1731 		goto done;
1732 
1733 	/*
1734 	 * Check to see if it is in the expected place on disk and that the
1735 	 * block has been mapped.
1736 	 */
1737 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1738 		match = 1;
1739 done:
1740 	BUF_UNLOCK(bpa);
1741 	return (match);
1742 }
1743 
1744 /*
1745  *	vfs_bio_awrite:
1746  *
1747  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1748  *	This is much better then the old way of writing only one buffer at
1749  *	a time.  Note that we may not be presented with the buffers in the
1750  *	correct order, so we search for the cluster in both directions.
1751  */
1752 int
1753 vfs_bio_awrite(struct buf *bp)
1754 {
1755 	struct bufobj *bo;
1756 	int i;
1757 	int j;
1758 	daddr_t lblkno = bp->b_lblkno;
1759 	struct vnode *vp = bp->b_vp;
1760 	int ncl;
1761 	int nwritten;
1762 	int size;
1763 	int maxcl;
1764 
1765 	bo = &vp->v_bufobj;
1766 	/*
1767 	 * right now we support clustered writing only to regular files.  If
1768 	 * we find a clusterable block we could be in the middle of a cluster
1769 	 * rather then at the beginning.
1770 	 */
1771 	if ((vp->v_type == VREG) &&
1772 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1773 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1774 
1775 		size = vp->v_mount->mnt_stat.f_iosize;
1776 		maxcl = MAXPHYS / size;
1777 
1778 		BO_LOCK(bo);
1779 		for (i = 1; i < maxcl; i++)
1780 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1781 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1782 				break;
1783 
1784 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1785 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1786 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1787 				break;
1788 		BO_UNLOCK(bo);
1789 		--j;
1790 		ncl = i + j;
1791 		/*
1792 		 * this is a possible cluster write
1793 		 */
1794 		if (ncl != 1) {
1795 			BUF_UNLOCK(bp);
1796 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
1797 			    0);
1798 			return (nwritten);
1799 		}
1800 	}
1801 	bremfree(bp);
1802 	bp->b_flags |= B_ASYNC;
1803 	/*
1804 	 * default (old) behavior, writing out only one block
1805 	 *
1806 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1807 	 */
1808 	nwritten = bp->b_bufsize;
1809 	(void) bwrite(bp);
1810 
1811 	return (nwritten);
1812 }
1813 
1814 /*
1815  *	getnewbuf:
1816  *
1817  *	Find and initialize a new buffer header, freeing up existing buffers
1818  *	in the bufqueues as necessary.  The new buffer is returned locked.
1819  *
1820  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1821  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1822  *
1823  *	We block if:
1824  *		We have insufficient buffer headers
1825  *		We have insufficient buffer space
1826  *		buffer_map is too fragmented ( space reservation fails )
1827  *		If we have to flush dirty buffers ( but we try to avoid this )
1828  *
1829  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1830  *	Instead we ask the buf daemon to do it for us.  We attempt to
1831  *	avoid piecemeal wakeups of the pageout daemon.
1832  */
1833 
1834 static struct buf *
1835 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1836     int gbflags)
1837 {
1838 	struct thread *td;
1839 	struct buf *bp;
1840 	struct buf *nbp;
1841 	int defrag = 0;
1842 	int nqindex;
1843 	static int flushingbufs;
1844 
1845 	td = curthread;
1846 	/*
1847 	 * We can't afford to block since we might be holding a vnode lock,
1848 	 * which may prevent system daemons from running.  We deal with
1849 	 * low-memory situations by proactively returning memory and running
1850 	 * async I/O rather then sync I/O.
1851 	 */
1852 	atomic_add_int(&getnewbufcalls, 1);
1853 	atomic_subtract_int(&getnewbufrestarts, 1);
1854 restart:
1855 	atomic_add_int(&getnewbufrestarts, 1);
1856 
1857 	/*
1858 	 * Setup for scan.  If we do not have enough free buffers,
1859 	 * we setup a degenerate case that immediately fails.  Note
1860 	 * that if we are specially marked process, we are allowed to
1861 	 * dip into our reserves.
1862 	 *
1863 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1864 	 *
1865 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1866 	 * However, there are a number of cases (defragging, reusing, ...)
1867 	 * where we cannot backup.
1868 	 */
1869 	mtx_lock(&bqlock);
1870 	nqindex = QUEUE_EMPTYKVA;
1871 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1872 
1873 	if (nbp == NULL) {
1874 		/*
1875 		 * If no EMPTYKVA buffers and we are either
1876 		 * defragging or reusing, locate a CLEAN buffer
1877 		 * to free or reuse.  If bufspace useage is low
1878 		 * skip this step so we can allocate a new buffer.
1879 		 */
1880 		if (defrag || bufspace >= lobufspace) {
1881 			nqindex = QUEUE_CLEAN;
1882 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1883 		}
1884 
1885 		/*
1886 		 * If we could not find or were not allowed to reuse a
1887 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1888 		 * buffer.  We can only use an EMPTY buffer if allocating
1889 		 * its KVA would not otherwise run us out of buffer space.
1890 		 */
1891 		if (nbp == NULL && defrag == 0 &&
1892 		    bufspace + maxsize < hibufspace) {
1893 			nqindex = QUEUE_EMPTY;
1894 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1895 		}
1896 	}
1897 
1898 	/*
1899 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1900 	 * depending.
1901 	 */
1902 
1903 	while ((bp = nbp) != NULL) {
1904 		int qindex = nqindex;
1905 
1906 		/*
1907 		 * Calculate next bp ( we can only use it if we do not block
1908 		 * or do other fancy things ).
1909 		 */
1910 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1911 			switch(qindex) {
1912 			case QUEUE_EMPTY:
1913 				nqindex = QUEUE_EMPTYKVA;
1914 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1915 					break;
1916 				/* FALLTHROUGH */
1917 			case QUEUE_EMPTYKVA:
1918 				nqindex = QUEUE_CLEAN;
1919 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1920 					break;
1921 				/* FALLTHROUGH */
1922 			case QUEUE_CLEAN:
1923 				/*
1924 				 * nbp is NULL.
1925 				 */
1926 				break;
1927 			}
1928 		}
1929 		/*
1930 		 * If we are defragging then we need a buffer with
1931 		 * b_kvasize != 0.  XXX this situation should no longer
1932 		 * occur, if defrag is non-zero the buffer's b_kvasize
1933 		 * should also be non-zero at this point.  XXX
1934 		 */
1935 		if (defrag && bp->b_kvasize == 0) {
1936 			printf("Warning: defrag empty buffer %p\n", bp);
1937 			continue;
1938 		}
1939 
1940 		/*
1941 		 * Start freeing the bp.  This is somewhat involved.  nbp
1942 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1943 		 */
1944 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1945 			continue;
1946 		if (bp->b_vp) {
1947 			BO_LOCK(bp->b_bufobj);
1948 			if (bp->b_vflags & BV_BKGRDINPROG) {
1949 				BO_UNLOCK(bp->b_bufobj);
1950 				BUF_UNLOCK(bp);
1951 				continue;
1952 			}
1953 			BO_UNLOCK(bp->b_bufobj);
1954 		}
1955 		CTR6(KTR_BUF,
1956 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1957 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1958 		    bp->b_kvasize, bp->b_bufsize, qindex);
1959 
1960 		/*
1961 		 * Sanity Checks
1962 		 */
1963 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1964 
1965 		/*
1966 		 * Note: we no longer distinguish between VMIO and non-VMIO
1967 		 * buffers.
1968 		 */
1969 
1970 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1971 
1972 		if (bp->b_bufobj != NULL)
1973 			BO_LOCK(bp->b_bufobj);
1974 		bremfreel(bp);
1975 		if (bp->b_bufobj != NULL)
1976 			BO_UNLOCK(bp->b_bufobj);
1977 		mtx_unlock(&bqlock);
1978 
1979 		if (qindex == QUEUE_CLEAN) {
1980 			if (bp->b_flags & B_VMIO) {
1981 				bp->b_flags &= ~B_ASYNC;
1982 				vfs_vmio_release(bp);
1983 			}
1984 			if (bp->b_vp)
1985 				brelvp(bp);
1986 		}
1987 
1988 		/*
1989 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1990 		 * the scan from this point on.
1991 		 *
1992 		 * Get the rest of the buffer freed up.  b_kva* is still
1993 		 * valid after this operation.
1994 		 */
1995 
1996 		if (bp->b_rcred != NOCRED) {
1997 			crfree(bp->b_rcred);
1998 			bp->b_rcred = NOCRED;
1999 		}
2000 		if (bp->b_wcred != NOCRED) {
2001 			crfree(bp->b_wcred);
2002 			bp->b_wcred = NOCRED;
2003 		}
2004 		if (!LIST_EMPTY(&bp->b_dep))
2005 			buf_deallocate(bp);
2006 		if (bp->b_vflags & BV_BKGRDINPROG)
2007 			panic("losing buffer 3");
2008 		KASSERT(bp->b_vp == NULL,
2009 		    ("bp: %p still has vnode %p.  qindex: %d",
2010 		    bp, bp->b_vp, qindex));
2011 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2012 		   ("bp: %p still on a buffer list. xflags %X",
2013 		    bp, bp->b_xflags));
2014 
2015 		if (bp->b_bufsize)
2016 			allocbuf(bp, 0);
2017 
2018 		bp->b_flags = 0;
2019 		bp->b_ioflags = 0;
2020 		bp->b_xflags = 0;
2021 		KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
2022 		    ("buf %p still counted as free?", bp));
2023 		bp->b_vflags = 0;
2024 		bp->b_vp = NULL;
2025 		bp->b_blkno = bp->b_lblkno = 0;
2026 		bp->b_offset = NOOFFSET;
2027 		bp->b_iodone = 0;
2028 		bp->b_error = 0;
2029 		bp->b_resid = 0;
2030 		bp->b_bcount = 0;
2031 		bp->b_npages = 0;
2032 		bp->b_dirtyoff = bp->b_dirtyend = 0;
2033 		bp->b_bufobj = NULL;
2034 		bp->b_pin_count = 0;
2035 		bp->b_fsprivate1 = NULL;
2036 		bp->b_fsprivate2 = NULL;
2037 		bp->b_fsprivate3 = NULL;
2038 
2039 		LIST_INIT(&bp->b_dep);
2040 
2041 		/*
2042 		 * If we are defragging then free the buffer.
2043 		 */
2044 		if (defrag) {
2045 			bp->b_flags |= B_INVAL;
2046 			bfreekva(bp);
2047 			brelse(bp);
2048 			defrag = 0;
2049 			goto restart;
2050 		}
2051 
2052 		/*
2053 		 * Notify any waiters for the buffer lock about
2054 		 * identity change by freeing the buffer.
2055 		 */
2056 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2057 			bp->b_flags |= B_INVAL;
2058 			bfreekva(bp);
2059 			brelse(bp);
2060 			goto restart;
2061 		}
2062 
2063 		/*
2064 		 * If we are overcomitted then recover the buffer and its
2065 		 * KVM space.  This occurs in rare situations when multiple
2066 		 * processes are blocked in getnewbuf() or allocbuf().
2067 		 */
2068 		if (bufspace >= hibufspace)
2069 			flushingbufs = 1;
2070 		if (flushingbufs && bp->b_kvasize != 0) {
2071 			bp->b_flags |= B_INVAL;
2072 			bfreekva(bp);
2073 			brelse(bp);
2074 			goto restart;
2075 		}
2076 		if (bufspace < lobufspace)
2077 			flushingbufs = 0;
2078 		break;
2079 	}
2080 
2081 	/*
2082 	 * If we exhausted our list, sleep as appropriate.  We may have to
2083 	 * wakeup various daemons and write out some dirty buffers.
2084 	 *
2085 	 * Generally we are sleeping due to insufficient buffer space.
2086 	 */
2087 
2088 	if (bp == NULL) {
2089 		int flags, norunbuf;
2090 		char *waitmsg;
2091 		int fl;
2092 
2093 		if (defrag) {
2094 			flags = VFS_BIO_NEED_BUFSPACE;
2095 			waitmsg = "nbufkv";
2096 		} else if (bufspace >= hibufspace) {
2097 			waitmsg = "nbufbs";
2098 			flags = VFS_BIO_NEED_BUFSPACE;
2099 		} else {
2100 			waitmsg = "newbuf";
2101 			flags = VFS_BIO_NEED_ANY;
2102 		}
2103 		mtx_lock(&nblock);
2104 		needsbuffer |= flags;
2105 		mtx_unlock(&nblock);
2106 		mtx_unlock(&bqlock);
2107 
2108 		bd_speedup();	/* heeeelp */
2109 		if (gbflags & GB_NOWAIT_BD)
2110 			return (NULL);
2111 
2112 		mtx_lock(&nblock);
2113 		while (needsbuffer & flags) {
2114 			if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2115 				mtx_unlock(&nblock);
2116 				/*
2117 				 * getblk() is called with a vnode
2118 				 * locked, and some majority of the
2119 				 * dirty buffers may as well belong to
2120 				 * the vnode. Flushing the buffers
2121 				 * there would make a progress that
2122 				 * cannot be achieved by the
2123 				 * buf_daemon, that cannot lock the
2124 				 * vnode.
2125 				 */
2126 				norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2127 				    (td->td_pflags & TDP_NORUNNINGBUF);
2128 				/* play bufdaemon */
2129 				td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2130 				fl = buf_do_flush(vp);
2131 				td->td_pflags &= norunbuf;
2132 				mtx_lock(&nblock);
2133 				if (fl != 0)
2134 					continue;
2135 				if ((needsbuffer & flags) == 0)
2136 					break;
2137 			}
2138 			if (msleep(&needsbuffer, &nblock,
2139 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2140 				mtx_unlock(&nblock);
2141 				return (NULL);
2142 			}
2143 		}
2144 		mtx_unlock(&nblock);
2145 	} else {
2146 		/*
2147 		 * We finally have a valid bp.  We aren't quite out of the
2148 		 * woods, we still have to reserve kva space.  In order
2149 		 * to keep fragmentation sane we only allocate kva in
2150 		 * BKVASIZE chunks.
2151 		 */
2152 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2153 
2154 		if (maxsize != bp->b_kvasize) {
2155 			vm_offset_t addr = 0;
2156 			int rv;
2157 
2158 			bfreekva(bp);
2159 
2160 			vm_map_lock(buffer_map);
2161 			if (vm_map_findspace(buffer_map,
2162 			    vm_map_min(buffer_map), maxsize, &addr)) {
2163 				/*
2164 				 * Buffer map is too fragmented.
2165 				 * We must defragment the map.
2166 				 */
2167 				atomic_add_int(&bufdefragcnt, 1);
2168 				vm_map_unlock(buffer_map);
2169 				defrag = 1;
2170 				bp->b_flags |= B_INVAL;
2171 				brelse(bp);
2172 				goto restart;
2173 			}
2174 			rv = vm_map_insert(buffer_map, NULL, 0, addr,
2175 			    addr + maxsize, VM_PROT_ALL, VM_PROT_ALL,
2176 			    MAP_NOFAULT);
2177 			KASSERT(rv == KERN_SUCCESS,
2178 			    ("vm_map_insert(buffer_map) rv %d", rv));
2179 			vm_map_unlock(buffer_map);
2180 			bp->b_kvabase = (caddr_t)addr;
2181 			bp->b_kvasize = maxsize;
2182 			atomic_add_long(&bufspace, bp->b_kvasize);
2183 			atomic_add_int(&bufreusecnt, 1);
2184 		}
2185 		bp->b_saveaddr = bp->b_kvabase;
2186 		bp->b_data = bp->b_saveaddr;
2187 	}
2188 	return (bp);
2189 }
2190 
2191 /*
2192  *	buf_daemon:
2193  *
2194  *	buffer flushing daemon.  Buffers are normally flushed by the
2195  *	update daemon but if it cannot keep up this process starts to
2196  *	take the load in an attempt to prevent getnewbuf() from blocking.
2197  */
2198 
2199 static struct kproc_desc buf_kp = {
2200 	"bufdaemon",
2201 	buf_daemon,
2202 	&bufdaemonproc
2203 };
2204 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2205 
2206 static int
2207 buf_do_flush(struct vnode *vp)
2208 {
2209 	int flushed;
2210 
2211 	flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2212 	if (flushed == 0) {
2213 		/*
2214 		 * Could not find any buffers without rollback
2215 		 * dependencies, so just write the first one
2216 		 * in the hopes of eventually making progress.
2217 		 */
2218 		flushbufqueues(vp, QUEUE_DIRTY, 1);
2219 	}
2220 	return (flushed);
2221 }
2222 
2223 static void
2224 buf_daemon()
2225 {
2226 	int lodirtysave;
2227 
2228 	/*
2229 	 * This process needs to be suspended prior to shutdown sync.
2230 	 */
2231 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2232 	    SHUTDOWN_PRI_LAST);
2233 
2234 	/*
2235 	 * This process is allowed to take the buffer cache to the limit
2236 	 */
2237 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2238 	mtx_lock(&bdlock);
2239 	for (;;) {
2240 		bd_request = 0;
2241 		mtx_unlock(&bdlock);
2242 
2243 		kproc_suspend_check(bufdaemonproc);
2244 		lodirtysave = lodirtybuffers;
2245 		if (bd_speedupreq) {
2246 			lodirtybuffers = numdirtybuffers / 2;
2247 			bd_speedupreq = 0;
2248 		}
2249 		/*
2250 		 * Do the flush.  Limit the amount of in-transit I/O we
2251 		 * allow to build up, otherwise we would completely saturate
2252 		 * the I/O system.  Wakeup any waiting processes before we
2253 		 * normally would so they can run in parallel with our drain.
2254 		 */
2255 		while (numdirtybuffers > lodirtybuffers) {
2256 			if (buf_do_flush(NULL) == 0)
2257 				break;
2258 			kern_yield(PRI_USER);
2259 		}
2260 		lodirtybuffers = lodirtysave;
2261 
2262 		/*
2263 		 * Only clear bd_request if we have reached our low water
2264 		 * mark.  The buf_daemon normally waits 1 second and
2265 		 * then incrementally flushes any dirty buffers that have
2266 		 * built up, within reason.
2267 		 *
2268 		 * If we were unable to hit our low water mark and couldn't
2269 		 * find any flushable buffers, we sleep half a second.
2270 		 * Otherwise we loop immediately.
2271 		 */
2272 		mtx_lock(&bdlock);
2273 		if (numdirtybuffers <= lodirtybuffers) {
2274 			/*
2275 			 * We reached our low water mark, reset the
2276 			 * request and sleep until we are needed again.
2277 			 * The sleep is just so the suspend code works.
2278 			 */
2279 			bd_request = 0;
2280 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2281 		} else {
2282 			/*
2283 			 * We couldn't find any flushable dirty buffers but
2284 			 * still have too many dirty buffers, we
2285 			 * have to sleep and try again.  (rare)
2286 			 */
2287 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2288 		}
2289 	}
2290 }
2291 
2292 /*
2293  *	flushbufqueues:
2294  *
2295  *	Try to flush a buffer in the dirty queue.  We must be careful to
2296  *	free up B_INVAL buffers instead of write them, which NFS is
2297  *	particularly sensitive to.
2298  */
2299 static int flushwithdeps = 0;
2300 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2301     0, "Number of buffers flushed with dependecies that require rollbacks");
2302 
2303 static int
2304 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2305 {
2306 	struct buf *sentinel;
2307 	struct vnode *vp;
2308 	struct mount *mp;
2309 	struct buf *bp;
2310 	int hasdeps;
2311 	int flushed;
2312 	int target;
2313 
2314 	if (lvp == NULL) {
2315 		target = numdirtybuffers - lodirtybuffers;
2316 		if (flushdeps && target > 2)
2317 			target /= 2;
2318 	} else
2319 		target = flushbufqtarget;
2320 	flushed = 0;
2321 	bp = NULL;
2322 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2323 	sentinel->b_qindex = QUEUE_SENTINEL;
2324 	mtx_lock(&bqlock);
2325 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2326 	while (flushed != target) {
2327 		bp = TAILQ_NEXT(sentinel, b_freelist);
2328 		if (bp != NULL) {
2329 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2330 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2331 			    b_freelist);
2332 		} else
2333 			break;
2334 		/*
2335 		 * Skip sentinels inserted by other invocations of the
2336 		 * flushbufqueues(), taking care to not reorder them.
2337 		 */
2338 		if (bp->b_qindex == QUEUE_SENTINEL)
2339 			continue;
2340 		/*
2341 		 * Only flush the buffers that belong to the
2342 		 * vnode locked by the curthread.
2343 		 */
2344 		if (lvp != NULL && bp->b_vp != lvp)
2345 			continue;
2346 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2347 			continue;
2348 		if (bp->b_pin_count > 0) {
2349 			BUF_UNLOCK(bp);
2350 			continue;
2351 		}
2352 		BO_LOCK(bp->b_bufobj);
2353 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2354 		    (bp->b_flags & B_DELWRI) == 0) {
2355 			BO_UNLOCK(bp->b_bufobj);
2356 			BUF_UNLOCK(bp);
2357 			continue;
2358 		}
2359 		BO_UNLOCK(bp->b_bufobj);
2360 		if (bp->b_flags & B_INVAL) {
2361 			bremfreel(bp);
2362 			mtx_unlock(&bqlock);
2363 			brelse(bp);
2364 			flushed++;
2365 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2366 			mtx_lock(&bqlock);
2367 			continue;
2368 		}
2369 
2370 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2371 			if (flushdeps == 0) {
2372 				BUF_UNLOCK(bp);
2373 				continue;
2374 			}
2375 			hasdeps = 1;
2376 		} else
2377 			hasdeps = 0;
2378 		/*
2379 		 * We must hold the lock on a vnode before writing
2380 		 * one of its buffers. Otherwise we may confuse, or
2381 		 * in the case of a snapshot vnode, deadlock the
2382 		 * system.
2383 		 *
2384 		 * The lock order here is the reverse of the normal
2385 		 * of vnode followed by buf lock.  This is ok because
2386 		 * the NOWAIT will prevent deadlock.
2387 		 */
2388 		vp = bp->b_vp;
2389 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2390 			BUF_UNLOCK(bp);
2391 			continue;
2392 		}
2393 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2394 			mtx_unlock(&bqlock);
2395 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2396 			    bp, bp->b_vp, bp->b_flags);
2397 			if (curproc == bufdaemonproc)
2398 				vfs_bio_awrite(bp);
2399 			else {
2400 				bremfree(bp);
2401 				bwrite(bp);
2402 				notbufdflashes++;
2403 			}
2404 			vn_finished_write(mp);
2405 			VOP_UNLOCK(vp, 0);
2406 			flushwithdeps += hasdeps;
2407 			flushed++;
2408 
2409 			/*
2410 			 * Sleeping on runningbufspace while holding
2411 			 * vnode lock leads to deadlock.
2412 			 */
2413 			if (curproc == bufdaemonproc)
2414 				waitrunningbufspace();
2415 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2416 			mtx_lock(&bqlock);
2417 			continue;
2418 		}
2419 		vn_finished_write(mp);
2420 		BUF_UNLOCK(bp);
2421 	}
2422 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2423 	mtx_unlock(&bqlock);
2424 	free(sentinel, M_TEMP);
2425 	return (flushed);
2426 }
2427 
2428 /*
2429  * Check to see if a block is currently memory resident.
2430  */
2431 struct buf *
2432 incore(struct bufobj *bo, daddr_t blkno)
2433 {
2434 	struct buf *bp;
2435 
2436 	BO_LOCK(bo);
2437 	bp = gbincore(bo, blkno);
2438 	BO_UNLOCK(bo);
2439 	return (bp);
2440 }
2441 
2442 /*
2443  * Returns true if no I/O is needed to access the
2444  * associated VM object.  This is like incore except
2445  * it also hunts around in the VM system for the data.
2446  */
2447 
2448 static int
2449 inmem(struct vnode * vp, daddr_t blkno)
2450 {
2451 	vm_object_t obj;
2452 	vm_offset_t toff, tinc, size;
2453 	vm_page_t m;
2454 	vm_ooffset_t off;
2455 
2456 	ASSERT_VOP_LOCKED(vp, "inmem");
2457 
2458 	if (incore(&vp->v_bufobj, blkno))
2459 		return 1;
2460 	if (vp->v_mount == NULL)
2461 		return 0;
2462 	obj = vp->v_object;
2463 	if (obj == NULL)
2464 		return (0);
2465 
2466 	size = PAGE_SIZE;
2467 	if (size > vp->v_mount->mnt_stat.f_iosize)
2468 		size = vp->v_mount->mnt_stat.f_iosize;
2469 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2470 
2471 	VM_OBJECT_WLOCK(obj);
2472 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2473 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2474 		if (!m)
2475 			goto notinmem;
2476 		tinc = size;
2477 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2478 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2479 		if (vm_page_is_valid(m,
2480 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2481 			goto notinmem;
2482 	}
2483 	VM_OBJECT_WUNLOCK(obj);
2484 	return 1;
2485 
2486 notinmem:
2487 	VM_OBJECT_WUNLOCK(obj);
2488 	return (0);
2489 }
2490 
2491 /*
2492  * Set the dirty range for a buffer based on the status of the dirty
2493  * bits in the pages comprising the buffer.  The range is limited
2494  * to the size of the buffer.
2495  *
2496  * Tell the VM system that the pages associated with this buffer
2497  * are clean.  This is used for delayed writes where the data is
2498  * going to go to disk eventually without additional VM intevention.
2499  *
2500  * Note that while we only really need to clean through to b_bcount, we
2501  * just go ahead and clean through to b_bufsize.
2502  */
2503 static void
2504 vfs_clean_pages_dirty_buf(struct buf *bp)
2505 {
2506 	vm_ooffset_t foff, noff, eoff;
2507 	vm_page_t m;
2508 	int i;
2509 
2510 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2511 		return;
2512 
2513 	foff = bp->b_offset;
2514 	KASSERT(bp->b_offset != NOOFFSET,
2515 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2516 
2517 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2518 	vfs_drain_busy_pages(bp);
2519 	vfs_setdirty_locked_object(bp);
2520 	for (i = 0; i < bp->b_npages; i++) {
2521 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2522 		eoff = noff;
2523 		if (eoff > bp->b_offset + bp->b_bufsize)
2524 			eoff = bp->b_offset + bp->b_bufsize;
2525 		m = bp->b_pages[i];
2526 		vfs_page_set_validclean(bp, foff, m);
2527 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2528 		foff = noff;
2529 	}
2530 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2531 }
2532 
2533 static void
2534 vfs_setdirty_locked_object(struct buf *bp)
2535 {
2536 	vm_object_t object;
2537 	int i;
2538 
2539 	object = bp->b_bufobj->bo_object;
2540 	VM_OBJECT_ASSERT_WLOCKED(object);
2541 
2542 	/*
2543 	 * We qualify the scan for modified pages on whether the
2544 	 * object has been flushed yet.
2545 	 */
2546 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2547 		vm_offset_t boffset;
2548 		vm_offset_t eoffset;
2549 
2550 		/*
2551 		 * test the pages to see if they have been modified directly
2552 		 * by users through the VM system.
2553 		 */
2554 		for (i = 0; i < bp->b_npages; i++)
2555 			vm_page_test_dirty(bp->b_pages[i]);
2556 
2557 		/*
2558 		 * Calculate the encompassing dirty range, boffset and eoffset,
2559 		 * (eoffset - boffset) bytes.
2560 		 */
2561 
2562 		for (i = 0; i < bp->b_npages; i++) {
2563 			if (bp->b_pages[i]->dirty)
2564 				break;
2565 		}
2566 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2567 
2568 		for (i = bp->b_npages - 1; i >= 0; --i) {
2569 			if (bp->b_pages[i]->dirty) {
2570 				break;
2571 			}
2572 		}
2573 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2574 
2575 		/*
2576 		 * Fit it to the buffer.
2577 		 */
2578 
2579 		if (eoffset > bp->b_bcount)
2580 			eoffset = bp->b_bcount;
2581 
2582 		/*
2583 		 * If we have a good dirty range, merge with the existing
2584 		 * dirty range.
2585 		 */
2586 
2587 		if (boffset < eoffset) {
2588 			if (bp->b_dirtyoff > boffset)
2589 				bp->b_dirtyoff = boffset;
2590 			if (bp->b_dirtyend < eoffset)
2591 				bp->b_dirtyend = eoffset;
2592 		}
2593 	}
2594 }
2595 
2596 /*
2597  *	getblk:
2598  *
2599  *	Get a block given a specified block and offset into a file/device.
2600  *	The buffers B_DONE bit will be cleared on return, making it almost
2601  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2602  *	return.  The caller should clear B_INVAL prior to initiating a
2603  *	READ.
2604  *
2605  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2606  *	an existing buffer.
2607  *
2608  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2609  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2610  *	and then cleared based on the backing VM.  If the previous buffer is
2611  *	non-0-sized but invalid, B_CACHE will be cleared.
2612  *
2613  *	If getblk() must create a new buffer, the new buffer is returned with
2614  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2615  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2616  *	backing VM.
2617  *
2618  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2619  *	B_CACHE bit is clear.
2620  *
2621  *	What this means, basically, is that the caller should use B_CACHE to
2622  *	determine whether the buffer is fully valid or not and should clear
2623  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2624  *	the buffer by loading its data area with something, the caller needs
2625  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2626  *	the caller should set B_CACHE ( as an optimization ), else the caller
2627  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2628  *	a write attempt or if it was a successfull read.  If the caller
2629  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2630  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2631  */
2632 struct buf *
2633 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2634     int flags)
2635 {
2636 	struct buf *bp;
2637 	struct bufobj *bo;
2638 	int error;
2639 
2640 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2641 	ASSERT_VOP_LOCKED(vp, "getblk");
2642 	if (size > MAXBSIZE)
2643 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2644 
2645 	bo = &vp->v_bufobj;
2646 loop:
2647 	/*
2648 	 * Block if we are low on buffers.   Certain processes are allowed
2649 	 * to completely exhaust the buffer cache.
2650          *
2651          * If this check ever becomes a bottleneck it may be better to
2652          * move it into the else, when gbincore() fails.  At the moment
2653          * it isn't a problem.
2654          */
2655 	if (numfreebuffers == 0) {
2656 		if (TD_IS_IDLETHREAD(curthread))
2657 			return NULL;
2658 		mtx_lock(&nblock);
2659 		needsbuffer |= VFS_BIO_NEED_ANY;
2660 		mtx_unlock(&nblock);
2661 	}
2662 
2663 	BO_LOCK(bo);
2664 	bp = gbincore(bo, blkno);
2665 	if (bp != NULL) {
2666 		int lockflags;
2667 		/*
2668 		 * Buffer is in-core.  If the buffer is not busy nor managed,
2669 		 * it must be on a queue.
2670 		 */
2671 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2672 
2673 		if (flags & GB_LOCK_NOWAIT)
2674 			lockflags |= LK_NOWAIT;
2675 
2676 		error = BUF_TIMELOCK(bp, lockflags,
2677 		    BO_MTX(bo), "getblk", slpflag, slptimeo);
2678 
2679 		/*
2680 		 * If we slept and got the lock we have to restart in case
2681 		 * the buffer changed identities.
2682 		 */
2683 		if (error == ENOLCK)
2684 			goto loop;
2685 		/* We timed out or were interrupted. */
2686 		else if (error)
2687 			return (NULL);
2688 		/* If recursed, assume caller knows the rules. */
2689 		else if (BUF_LOCKRECURSED(bp))
2690 			goto end;
2691 
2692 		/*
2693 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2694 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2695 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2696 		 * backing VM cache.
2697 		 */
2698 		if (bp->b_flags & B_INVAL)
2699 			bp->b_flags &= ~B_CACHE;
2700 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2701 			bp->b_flags |= B_CACHE;
2702 		if (bp->b_flags & B_MANAGED)
2703 			MPASS(bp->b_qindex == QUEUE_NONE);
2704 		else {
2705 			BO_LOCK(bo);
2706 			bremfree(bp);
2707 			BO_UNLOCK(bo);
2708 		}
2709 
2710 		/*
2711 		 * check for size inconsistencies for non-VMIO case.
2712 		 */
2713 		if (bp->b_bcount != size) {
2714 			if ((bp->b_flags & B_VMIO) == 0 ||
2715 			    (size > bp->b_kvasize)) {
2716 				if (bp->b_flags & B_DELWRI) {
2717 					/*
2718 					 * If buffer is pinned and caller does
2719 					 * not want sleep  waiting for it to be
2720 					 * unpinned, bail out
2721 					 * */
2722 					if (bp->b_pin_count > 0) {
2723 						if (flags & GB_LOCK_NOWAIT) {
2724 							bqrelse(bp);
2725 							return (NULL);
2726 						} else {
2727 							bunpin_wait(bp);
2728 						}
2729 					}
2730 					bp->b_flags |= B_NOCACHE;
2731 					bwrite(bp);
2732 				} else {
2733 					if (LIST_EMPTY(&bp->b_dep)) {
2734 						bp->b_flags |= B_RELBUF;
2735 						brelse(bp);
2736 					} else {
2737 						bp->b_flags |= B_NOCACHE;
2738 						bwrite(bp);
2739 					}
2740 				}
2741 				goto loop;
2742 			}
2743 		}
2744 
2745 		/*
2746 		 * If the size is inconsistant in the VMIO case, we can resize
2747 		 * the buffer.  This might lead to B_CACHE getting set or
2748 		 * cleared.  If the size has not changed, B_CACHE remains
2749 		 * unchanged from its previous state.
2750 		 */
2751 
2752 		if (bp->b_bcount != size)
2753 			allocbuf(bp, size);
2754 
2755 		KASSERT(bp->b_offset != NOOFFSET,
2756 		    ("getblk: no buffer offset"));
2757 
2758 		/*
2759 		 * A buffer with B_DELWRI set and B_CACHE clear must
2760 		 * be committed before we can return the buffer in
2761 		 * order to prevent the caller from issuing a read
2762 		 * ( due to B_CACHE not being set ) and overwriting
2763 		 * it.
2764 		 *
2765 		 * Most callers, including NFS and FFS, need this to
2766 		 * operate properly either because they assume they
2767 		 * can issue a read if B_CACHE is not set, or because
2768 		 * ( for example ) an uncached B_DELWRI might loop due
2769 		 * to softupdates re-dirtying the buffer.  In the latter
2770 		 * case, B_CACHE is set after the first write completes,
2771 		 * preventing further loops.
2772 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2773 		 * above while extending the buffer, we cannot allow the
2774 		 * buffer to remain with B_CACHE set after the write
2775 		 * completes or it will represent a corrupt state.  To
2776 		 * deal with this we set B_NOCACHE to scrap the buffer
2777 		 * after the write.
2778 		 *
2779 		 * We might be able to do something fancy, like setting
2780 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2781 		 * so the below call doesn't set B_CACHE, but that gets real
2782 		 * confusing.  This is much easier.
2783 		 */
2784 
2785 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2786 			bp->b_flags |= B_NOCACHE;
2787 			bwrite(bp);
2788 			goto loop;
2789 		}
2790 		bp->b_flags &= ~B_DONE;
2791 	} else {
2792 		int bsize, maxsize, vmio;
2793 		off_t offset;
2794 
2795 		/*
2796 		 * Buffer is not in-core, create new buffer.  The buffer
2797 		 * returned by getnewbuf() is locked.  Note that the returned
2798 		 * buffer is also considered valid (not marked B_INVAL).
2799 		 */
2800 		BO_UNLOCK(bo);
2801 		/*
2802 		 * If the user does not want us to create the buffer, bail out
2803 		 * here.
2804 		 */
2805 		if (flags & GB_NOCREAT)
2806 			return NULL;
2807 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
2808 		offset = blkno * bsize;
2809 		vmio = vp->v_object != NULL;
2810 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2811 		maxsize = imax(maxsize, bsize);
2812 
2813 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2814 		if (bp == NULL) {
2815 			if (slpflag || slptimeo)
2816 				return NULL;
2817 			goto loop;
2818 		}
2819 
2820 		/*
2821 		 * This code is used to make sure that a buffer is not
2822 		 * created while the getnewbuf routine is blocked.
2823 		 * This can be a problem whether the vnode is locked or not.
2824 		 * If the buffer is created out from under us, we have to
2825 		 * throw away the one we just created.
2826 		 *
2827 		 * Note: this must occur before we associate the buffer
2828 		 * with the vp especially considering limitations in
2829 		 * the splay tree implementation when dealing with duplicate
2830 		 * lblkno's.
2831 		 */
2832 		BO_LOCK(bo);
2833 		if (gbincore(bo, blkno)) {
2834 			BO_UNLOCK(bo);
2835 			bp->b_flags |= B_INVAL;
2836 			brelse(bp);
2837 			goto loop;
2838 		}
2839 
2840 		/*
2841 		 * Insert the buffer into the hash, so that it can
2842 		 * be found by incore.
2843 		 */
2844 		bp->b_blkno = bp->b_lblkno = blkno;
2845 		bp->b_offset = offset;
2846 		bgetvp(vp, bp);
2847 		BO_UNLOCK(bo);
2848 
2849 		/*
2850 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2851 		 * buffer size starts out as 0, B_CACHE will be set by
2852 		 * allocbuf() for the VMIO case prior to it testing the
2853 		 * backing store for validity.
2854 		 */
2855 
2856 		if (vmio) {
2857 			bp->b_flags |= B_VMIO;
2858 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2859 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2860 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2861 		} else {
2862 			bp->b_flags &= ~B_VMIO;
2863 			KASSERT(bp->b_bufobj->bo_object == NULL,
2864 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2865 			    bp, bp->b_bufobj->bo_object));
2866 		}
2867 
2868 		allocbuf(bp, size);
2869 		bp->b_flags &= ~B_DONE;
2870 	}
2871 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2872 	BUF_ASSERT_HELD(bp);
2873 end:
2874 	KASSERT(bp->b_bufobj == bo,
2875 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2876 	return (bp);
2877 }
2878 
2879 /*
2880  * Get an empty, disassociated buffer of given size.  The buffer is initially
2881  * set to B_INVAL.
2882  */
2883 struct buf *
2884 geteblk(int size, int flags)
2885 {
2886 	struct buf *bp;
2887 	int maxsize;
2888 
2889 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2890 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2891 		if ((flags & GB_NOWAIT_BD) &&
2892 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
2893 			return (NULL);
2894 	}
2895 	allocbuf(bp, size);
2896 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2897 	BUF_ASSERT_HELD(bp);
2898 	return (bp);
2899 }
2900 
2901 
2902 /*
2903  * This code constitutes the buffer memory from either anonymous system
2904  * memory (in the case of non-VMIO operations) or from an associated
2905  * VM object (in the case of VMIO operations).  This code is able to
2906  * resize a buffer up or down.
2907  *
2908  * Note that this code is tricky, and has many complications to resolve
2909  * deadlock or inconsistant data situations.  Tread lightly!!!
2910  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2911  * the caller.  Calling this code willy nilly can result in the loss of data.
2912  *
2913  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2914  * B_CACHE for the non-VMIO case.
2915  */
2916 
2917 int
2918 allocbuf(struct buf *bp, int size)
2919 {
2920 	int newbsize, mbsize;
2921 	int i;
2922 
2923 	BUF_ASSERT_HELD(bp);
2924 
2925 	if (bp->b_kvasize < size)
2926 		panic("allocbuf: buffer too small");
2927 
2928 	if ((bp->b_flags & B_VMIO) == 0) {
2929 		caddr_t origbuf;
2930 		int origbufsize;
2931 		/*
2932 		 * Just get anonymous memory from the kernel.  Don't
2933 		 * mess with B_CACHE.
2934 		 */
2935 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2936 		if (bp->b_flags & B_MALLOC)
2937 			newbsize = mbsize;
2938 		else
2939 			newbsize = round_page(size);
2940 
2941 		if (newbsize < bp->b_bufsize) {
2942 			/*
2943 			 * malloced buffers are not shrunk
2944 			 */
2945 			if (bp->b_flags & B_MALLOC) {
2946 				if (newbsize) {
2947 					bp->b_bcount = size;
2948 				} else {
2949 					free(bp->b_data, M_BIOBUF);
2950 					if (bp->b_bufsize) {
2951 						atomic_subtract_long(
2952 						    &bufmallocspace,
2953 						    bp->b_bufsize);
2954 						bufspacewakeup();
2955 						bp->b_bufsize = 0;
2956 					}
2957 					bp->b_saveaddr = bp->b_kvabase;
2958 					bp->b_data = bp->b_saveaddr;
2959 					bp->b_bcount = 0;
2960 					bp->b_flags &= ~B_MALLOC;
2961 				}
2962 				return 1;
2963 			}
2964 			vm_hold_free_pages(bp, newbsize);
2965 		} else if (newbsize > bp->b_bufsize) {
2966 			/*
2967 			 * We only use malloced memory on the first allocation.
2968 			 * and revert to page-allocated memory when the buffer
2969 			 * grows.
2970 			 */
2971 			/*
2972 			 * There is a potential smp race here that could lead
2973 			 * to bufmallocspace slightly passing the max.  It
2974 			 * is probably extremely rare and not worth worrying
2975 			 * over.
2976 			 */
2977 			if ( (bufmallocspace < maxbufmallocspace) &&
2978 				(bp->b_bufsize == 0) &&
2979 				(mbsize <= PAGE_SIZE/2)) {
2980 
2981 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2982 				bp->b_bufsize = mbsize;
2983 				bp->b_bcount = size;
2984 				bp->b_flags |= B_MALLOC;
2985 				atomic_add_long(&bufmallocspace, mbsize);
2986 				return 1;
2987 			}
2988 			origbuf = NULL;
2989 			origbufsize = 0;
2990 			/*
2991 			 * If the buffer is growing on its other-than-first allocation,
2992 			 * then we revert to the page-allocation scheme.
2993 			 */
2994 			if (bp->b_flags & B_MALLOC) {
2995 				origbuf = bp->b_data;
2996 				origbufsize = bp->b_bufsize;
2997 				bp->b_data = bp->b_kvabase;
2998 				if (bp->b_bufsize) {
2999 					atomic_subtract_long(&bufmallocspace,
3000 					    bp->b_bufsize);
3001 					bufspacewakeup();
3002 					bp->b_bufsize = 0;
3003 				}
3004 				bp->b_flags &= ~B_MALLOC;
3005 				newbsize = round_page(newbsize);
3006 			}
3007 			vm_hold_load_pages(
3008 			    bp,
3009 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
3010 			    (vm_offset_t) bp->b_data + newbsize);
3011 			if (origbuf) {
3012 				bcopy(origbuf, bp->b_data, origbufsize);
3013 				free(origbuf, M_BIOBUF);
3014 			}
3015 		}
3016 	} else {
3017 		int desiredpages;
3018 
3019 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3020 		desiredpages = (size == 0) ? 0 :
3021 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3022 
3023 		if (bp->b_flags & B_MALLOC)
3024 			panic("allocbuf: VMIO buffer can't be malloced");
3025 		/*
3026 		 * Set B_CACHE initially if buffer is 0 length or will become
3027 		 * 0-length.
3028 		 */
3029 		if (size == 0 || bp->b_bufsize == 0)
3030 			bp->b_flags |= B_CACHE;
3031 
3032 		if (newbsize < bp->b_bufsize) {
3033 			/*
3034 			 * DEV_BSIZE aligned new buffer size is less then the
3035 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3036 			 * if we have to remove any pages.
3037 			 */
3038 			if (desiredpages < bp->b_npages) {
3039 				vm_page_t m;
3040 
3041 				pmap_qremove((vm_offset_t)trunc_page(
3042 				    (vm_offset_t)bp->b_data) +
3043 				    (desiredpages << PAGE_SHIFT),
3044 				    (bp->b_npages - desiredpages));
3045 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3046 				for (i = desiredpages; i < bp->b_npages; i++) {
3047 					/*
3048 					 * the page is not freed here -- it
3049 					 * is the responsibility of
3050 					 * vnode_pager_setsize
3051 					 */
3052 					m = bp->b_pages[i];
3053 					KASSERT(m != bogus_page,
3054 					    ("allocbuf: bogus page found"));
3055 					while (vm_page_sleep_if_busy(m, TRUE,
3056 					    "biodep"))
3057 						continue;
3058 
3059 					bp->b_pages[i] = NULL;
3060 					vm_page_lock(m);
3061 					vm_page_unwire(m, 0);
3062 					vm_page_unlock(m);
3063 				}
3064 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3065 				bp->b_npages = desiredpages;
3066 			}
3067 		} else if (size > bp->b_bcount) {
3068 			/*
3069 			 * We are growing the buffer, possibly in a
3070 			 * byte-granular fashion.
3071 			 */
3072 			vm_object_t obj;
3073 			vm_offset_t toff;
3074 			vm_offset_t tinc;
3075 
3076 			/*
3077 			 * Step 1, bring in the VM pages from the object,
3078 			 * allocating them if necessary.  We must clear
3079 			 * B_CACHE if these pages are not valid for the
3080 			 * range covered by the buffer.
3081 			 */
3082 
3083 			obj = bp->b_bufobj->bo_object;
3084 
3085 			VM_OBJECT_WLOCK(obj);
3086 			while (bp->b_npages < desiredpages) {
3087 				vm_page_t m;
3088 
3089 				/*
3090 				 * We must allocate system pages since blocking
3091 				 * here could interfere with paging I/O, no
3092 				 * matter which process we are.
3093 				 *
3094 				 * We can only test VPO_BUSY here.  Blocking on
3095 				 * m->busy might lead to a deadlock:
3096 				 *  vm_fault->getpages->cluster_read->allocbuf
3097 				 * Thus, we specify VM_ALLOC_IGN_SBUSY.
3098 				 */
3099 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3100 				    bp->b_npages, VM_ALLOC_NOBUSY |
3101 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3102 				    VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY |
3103 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3104 				if (m->valid == 0)
3105 					bp->b_flags &= ~B_CACHE;
3106 				bp->b_pages[bp->b_npages] = m;
3107 				++bp->b_npages;
3108 			}
3109 
3110 			/*
3111 			 * Step 2.  We've loaded the pages into the buffer,
3112 			 * we have to figure out if we can still have B_CACHE
3113 			 * set.  Note that B_CACHE is set according to the
3114 			 * byte-granular range ( bcount and size ), new the
3115 			 * aligned range ( newbsize ).
3116 			 *
3117 			 * The VM test is against m->valid, which is DEV_BSIZE
3118 			 * aligned.  Needless to say, the validity of the data
3119 			 * needs to also be DEV_BSIZE aligned.  Note that this
3120 			 * fails with NFS if the server or some other client
3121 			 * extends the file's EOF.  If our buffer is resized,
3122 			 * B_CACHE may remain set! XXX
3123 			 */
3124 
3125 			toff = bp->b_bcount;
3126 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3127 
3128 			while ((bp->b_flags & B_CACHE) && toff < size) {
3129 				vm_pindex_t pi;
3130 
3131 				if (tinc > (size - toff))
3132 					tinc = size - toff;
3133 
3134 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3135 				    PAGE_SHIFT;
3136 
3137 				vfs_buf_test_cache(
3138 				    bp,
3139 				    bp->b_offset,
3140 				    toff,
3141 				    tinc,
3142 				    bp->b_pages[pi]
3143 				);
3144 				toff += tinc;
3145 				tinc = PAGE_SIZE;
3146 			}
3147 			VM_OBJECT_WUNLOCK(obj);
3148 
3149 			/*
3150 			 * Step 3, fixup the KVM pmap.  Remember that
3151 			 * bp->b_data is relative to bp->b_offset, but
3152 			 * bp->b_offset may be offset into the first page.
3153 			 */
3154 
3155 			bp->b_data = (caddr_t)
3156 			    trunc_page((vm_offset_t)bp->b_data);
3157 			pmap_qenter(
3158 			    (vm_offset_t)bp->b_data,
3159 			    bp->b_pages,
3160 			    bp->b_npages
3161 			);
3162 
3163 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3164 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
3165 		}
3166 	}
3167 	if (newbsize < bp->b_bufsize)
3168 		bufspacewakeup();
3169 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3170 	bp->b_bcount = size;		/* requested buffer size	*/
3171 	return 1;
3172 }
3173 
3174 void
3175 biodone(struct bio *bp)
3176 {
3177 	struct mtx *mtxp;
3178 	void (*done)(struct bio *);
3179 
3180 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3181 	mtx_lock(mtxp);
3182 	bp->bio_flags |= BIO_DONE;
3183 	done = bp->bio_done;
3184 	if (done == NULL)
3185 		wakeup(bp);
3186 	mtx_unlock(mtxp);
3187 	if (done != NULL)
3188 		done(bp);
3189 }
3190 
3191 /*
3192  * Wait for a BIO to finish.
3193  *
3194  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3195  * case is not yet clear.
3196  */
3197 int
3198 biowait(struct bio *bp, const char *wchan)
3199 {
3200 	struct mtx *mtxp;
3201 
3202 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3203 	mtx_lock(mtxp);
3204 	while ((bp->bio_flags & BIO_DONE) == 0)
3205 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3206 	mtx_unlock(mtxp);
3207 	if (bp->bio_error != 0)
3208 		return (bp->bio_error);
3209 	if (!(bp->bio_flags & BIO_ERROR))
3210 		return (0);
3211 	return (EIO);
3212 }
3213 
3214 void
3215 biofinish(struct bio *bp, struct devstat *stat, int error)
3216 {
3217 
3218 	if (error) {
3219 		bp->bio_error = error;
3220 		bp->bio_flags |= BIO_ERROR;
3221 	}
3222 	if (stat != NULL)
3223 		devstat_end_transaction_bio(stat, bp);
3224 	biodone(bp);
3225 }
3226 
3227 /*
3228  *	bufwait:
3229  *
3230  *	Wait for buffer I/O completion, returning error status.  The buffer
3231  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3232  *	error and cleared.
3233  */
3234 int
3235 bufwait(struct buf *bp)
3236 {
3237 	if (bp->b_iocmd == BIO_READ)
3238 		bwait(bp, PRIBIO, "biord");
3239 	else
3240 		bwait(bp, PRIBIO, "biowr");
3241 	if (bp->b_flags & B_EINTR) {
3242 		bp->b_flags &= ~B_EINTR;
3243 		return (EINTR);
3244 	}
3245 	if (bp->b_ioflags & BIO_ERROR) {
3246 		return (bp->b_error ? bp->b_error : EIO);
3247 	} else {
3248 		return (0);
3249 	}
3250 }
3251 
3252  /*
3253   * Call back function from struct bio back up to struct buf.
3254   */
3255 static void
3256 bufdonebio(struct bio *bip)
3257 {
3258 	struct buf *bp;
3259 
3260 	bp = bip->bio_caller2;
3261 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3262 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3263 	bp->b_ioflags = bip->bio_flags;
3264 	bp->b_error = bip->bio_error;
3265 	if (bp->b_error)
3266 		bp->b_ioflags |= BIO_ERROR;
3267 	bufdone(bp);
3268 	g_destroy_bio(bip);
3269 }
3270 
3271 void
3272 dev_strategy(struct cdev *dev, struct buf *bp)
3273 {
3274 	struct cdevsw *csw;
3275 	struct bio *bip;
3276 	int ref;
3277 
3278 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3279 		panic("b_iocmd botch");
3280 	for (;;) {
3281 		bip = g_new_bio();
3282 		if (bip != NULL)
3283 			break;
3284 		/* Try again later */
3285 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3286 	}
3287 	bip->bio_cmd = bp->b_iocmd;
3288 	bip->bio_offset = bp->b_iooffset;
3289 	bip->bio_length = bp->b_bcount;
3290 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3291 	bip->bio_data = bp->b_data;
3292 	bip->bio_done = bufdonebio;
3293 	bip->bio_caller2 = bp;
3294 	bip->bio_dev = dev;
3295 	KASSERT(dev->si_refcount > 0,
3296 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3297 	    devtoname(dev)));
3298 	csw = dev_refthread(dev, &ref);
3299 	if (csw == NULL) {
3300 		g_destroy_bio(bip);
3301 		bp->b_error = ENXIO;
3302 		bp->b_ioflags = BIO_ERROR;
3303 		bufdone(bp);
3304 		return;
3305 	}
3306 	(*csw->d_strategy)(bip);
3307 	dev_relthread(dev, ref);
3308 }
3309 
3310 /*
3311  *	bufdone:
3312  *
3313  *	Finish I/O on a buffer, optionally calling a completion function.
3314  *	This is usually called from an interrupt so process blocking is
3315  *	not allowed.
3316  *
3317  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3318  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3319  *	assuming B_INVAL is clear.
3320  *
3321  *	For the VMIO case, we set B_CACHE if the op was a read and no
3322  *	read error occured, or if the op was a write.  B_CACHE is never
3323  *	set if the buffer is invalid or otherwise uncacheable.
3324  *
3325  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3326  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3327  *	in the biodone routine.
3328  */
3329 void
3330 bufdone(struct buf *bp)
3331 {
3332 	struct bufobj *dropobj;
3333 	void    (*biodone)(struct buf *);
3334 
3335 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3336 	dropobj = NULL;
3337 
3338 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3339 	BUF_ASSERT_HELD(bp);
3340 
3341 	runningbufwakeup(bp);
3342 	if (bp->b_iocmd == BIO_WRITE)
3343 		dropobj = bp->b_bufobj;
3344 	/* call optional completion function if requested */
3345 	if (bp->b_iodone != NULL) {
3346 		biodone = bp->b_iodone;
3347 		bp->b_iodone = NULL;
3348 		(*biodone) (bp);
3349 		if (dropobj)
3350 			bufobj_wdrop(dropobj);
3351 		return;
3352 	}
3353 
3354 	bufdone_finish(bp);
3355 
3356 	if (dropobj)
3357 		bufobj_wdrop(dropobj);
3358 }
3359 
3360 void
3361 bufdone_finish(struct buf *bp)
3362 {
3363 	BUF_ASSERT_HELD(bp);
3364 
3365 	if (!LIST_EMPTY(&bp->b_dep))
3366 		buf_complete(bp);
3367 
3368 	if (bp->b_flags & B_VMIO) {
3369 		vm_ooffset_t foff;
3370 		vm_page_t m;
3371 		vm_object_t obj;
3372 		struct vnode *vp;
3373 		int bogus, i, iosize;
3374 
3375 		obj = bp->b_bufobj->bo_object;
3376 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3377 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3378 		    obj->paging_in_progress, bp->b_npages));
3379 
3380 		vp = bp->b_vp;
3381 		KASSERT(vp->v_holdcnt > 0,
3382 		    ("biodone_finish: vnode %p has zero hold count", vp));
3383 		KASSERT(vp->v_object != NULL,
3384 		    ("biodone_finish: vnode %p has no vm_object", vp));
3385 
3386 		foff = bp->b_offset;
3387 		KASSERT(bp->b_offset != NOOFFSET,
3388 		    ("biodone_finish: bp %p has no buffer offset", bp));
3389 
3390 		/*
3391 		 * Set B_CACHE if the op was a normal read and no error
3392 		 * occured.  B_CACHE is set for writes in the b*write()
3393 		 * routines.
3394 		 */
3395 		iosize = bp->b_bcount - bp->b_resid;
3396 		if (bp->b_iocmd == BIO_READ &&
3397 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3398 		    !(bp->b_ioflags & BIO_ERROR)) {
3399 			bp->b_flags |= B_CACHE;
3400 		}
3401 		bogus = 0;
3402 		VM_OBJECT_WLOCK(obj);
3403 		for (i = 0; i < bp->b_npages; i++) {
3404 			int bogusflag = 0;
3405 			int resid;
3406 
3407 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3408 			if (resid > iosize)
3409 				resid = iosize;
3410 
3411 			/*
3412 			 * cleanup bogus pages, restoring the originals
3413 			 */
3414 			m = bp->b_pages[i];
3415 			if (m == bogus_page) {
3416 				bogus = bogusflag = 1;
3417 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3418 				if (m == NULL)
3419 					panic("biodone: page disappeared!");
3420 				bp->b_pages[i] = m;
3421 			}
3422 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3423 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3424 			    (intmax_t)foff, (uintmax_t)m->pindex));
3425 
3426 			/*
3427 			 * In the write case, the valid and clean bits are
3428 			 * already changed correctly ( see bdwrite() ), so we
3429 			 * only need to do this here in the read case.
3430 			 */
3431 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3432 				KASSERT((m->dirty & vm_page_bits(foff &
3433 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3434 				    " page %p has unexpected dirty bits", m));
3435 				vfs_page_set_valid(bp, foff, m);
3436 			}
3437 
3438 			vm_page_io_finish(m);
3439 			vm_object_pip_subtract(obj, 1);
3440 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3441 			iosize -= resid;
3442 		}
3443 		vm_object_pip_wakeupn(obj, 0);
3444 		VM_OBJECT_WUNLOCK(obj);
3445 		if (bogus)
3446 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3447 			    bp->b_pages, bp->b_npages);
3448 	}
3449 
3450 	/*
3451 	 * For asynchronous completions, release the buffer now. The brelse
3452 	 * will do a wakeup there if necessary - so no need to do a wakeup
3453 	 * here in the async case. The sync case always needs to do a wakeup.
3454 	 */
3455 
3456 	if (bp->b_flags & B_ASYNC) {
3457 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3458 			brelse(bp);
3459 		else
3460 			bqrelse(bp);
3461 	} else
3462 		bdone(bp);
3463 }
3464 
3465 /*
3466  * This routine is called in lieu of iodone in the case of
3467  * incomplete I/O.  This keeps the busy status for pages
3468  * consistant.
3469  */
3470 void
3471 vfs_unbusy_pages(struct buf *bp)
3472 {
3473 	int i;
3474 	vm_object_t obj;
3475 	vm_page_t m;
3476 
3477 	runningbufwakeup(bp);
3478 	if (!(bp->b_flags & B_VMIO))
3479 		return;
3480 
3481 	obj = bp->b_bufobj->bo_object;
3482 	VM_OBJECT_WLOCK(obj);
3483 	for (i = 0; i < bp->b_npages; i++) {
3484 		m = bp->b_pages[i];
3485 		if (m == bogus_page) {
3486 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3487 			if (!m)
3488 				panic("vfs_unbusy_pages: page missing\n");
3489 			bp->b_pages[i] = m;
3490 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3491 			    bp->b_pages, bp->b_npages);
3492 		}
3493 		vm_object_pip_subtract(obj, 1);
3494 		vm_page_io_finish(m);
3495 	}
3496 	vm_object_pip_wakeupn(obj, 0);
3497 	VM_OBJECT_WUNLOCK(obj);
3498 }
3499 
3500 /*
3501  * vfs_page_set_valid:
3502  *
3503  *	Set the valid bits in a page based on the supplied offset.   The
3504  *	range is restricted to the buffer's size.
3505  *
3506  *	This routine is typically called after a read completes.
3507  */
3508 static void
3509 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3510 {
3511 	vm_ooffset_t eoff;
3512 
3513 	/*
3514 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3515 	 * page boundary and eoff is not greater than the end of the buffer.
3516 	 * The end of the buffer, in this case, is our file EOF, not the
3517 	 * allocation size of the buffer.
3518 	 */
3519 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3520 	if (eoff > bp->b_offset + bp->b_bcount)
3521 		eoff = bp->b_offset + bp->b_bcount;
3522 
3523 	/*
3524 	 * Set valid range.  This is typically the entire buffer and thus the
3525 	 * entire page.
3526 	 */
3527 	if (eoff > off)
3528 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3529 }
3530 
3531 /*
3532  * vfs_page_set_validclean:
3533  *
3534  *	Set the valid bits and clear the dirty bits in a page based on the
3535  *	supplied offset.   The range is restricted to the buffer's size.
3536  */
3537 static void
3538 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3539 {
3540 	vm_ooffset_t soff, eoff;
3541 
3542 	/*
3543 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3544 	 * page boundry or cross the end of the buffer.  The end of the
3545 	 * buffer, in this case, is our file EOF, not the allocation size
3546 	 * of the buffer.
3547 	 */
3548 	soff = off;
3549 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3550 	if (eoff > bp->b_offset + bp->b_bcount)
3551 		eoff = bp->b_offset + bp->b_bcount;
3552 
3553 	/*
3554 	 * Set valid range.  This is typically the entire buffer and thus the
3555 	 * entire page.
3556 	 */
3557 	if (eoff > soff) {
3558 		vm_page_set_validclean(
3559 		    m,
3560 		   (vm_offset_t) (soff & PAGE_MASK),
3561 		   (vm_offset_t) (eoff - soff)
3562 		);
3563 	}
3564 }
3565 
3566 /*
3567  * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3568  * any page is busy, drain the flag.
3569  */
3570 static void
3571 vfs_drain_busy_pages(struct buf *bp)
3572 {
3573 	vm_page_t m;
3574 	int i, last_busied;
3575 
3576 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
3577 	last_busied = 0;
3578 	for (i = 0; i < bp->b_npages; i++) {
3579 		m = bp->b_pages[i];
3580 		if ((m->oflags & VPO_BUSY) != 0) {
3581 			for (; last_busied < i; last_busied++)
3582 				vm_page_busy(bp->b_pages[last_busied]);
3583 			while ((m->oflags & VPO_BUSY) != 0)
3584 				vm_page_sleep(m, "vbpage");
3585 		}
3586 	}
3587 	for (i = 0; i < last_busied; i++)
3588 		vm_page_wakeup(bp->b_pages[i]);
3589 }
3590 
3591 /*
3592  * This routine is called before a device strategy routine.
3593  * It is used to tell the VM system that paging I/O is in
3594  * progress, and treat the pages associated with the buffer
3595  * almost as being VPO_BUSY.  Also the object paging_in_progress
3596  * flag is handled to make sure that the object doesn't become
3597  * inconsistant.
3598  *
3599  * Since I/O has not been initiated yet, certain buffer flags
3600  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3601  * and should be ignored.
3602  */
3603 void
3604 vfs_busy_pages(struct buf *bp, int clear_modify)
3605 {
3606 	int i, bogus;
3607 	vm_object_t obj;
3608 	vm_ooffset_t foff;
3609 	vm_page_t m;
3610 
3611 	if (!(bp->b_flags & B_VMIO))
3612 		return;
3613 
3614 	obj = bp->b_bufobj->bo_object;
3615 	foff = bp->b_offset;
3616 	KASSERT(bp->b_offset != NOOFFSET,
3617 	    ("vfs_busy_pages: no buffer offset"));
3618 	VM_OBJECT_WLOCK(obj);
3619 	vfs_drain_busy_pages(bp);
3620 	if (bp->b_bufsize != 0)
3621 		vfs_setdirty_locked_object(bp);
3622 	bogus = 0;
3623 	for (i = 0; i < bp->b_npages; i++) {
3624 		m = bp->b_pages[i];
3625 
3626 		if ((bp->b_flags & B_CLUSTER) == 0) {
3627 			vm_object_pip_add(obj, 1);
3628 			vm_page_io_start(m);
3629 		}
3630 		/*
3631 		 * When readying a buffer for a read ( i.e
3632 		 * clear_modify == 0 ), it is important to do
3633 		 * bogus_page replacement for valid pages in
3634 		 * partially instantiated buffers.  Partially
3635 		 * instantiated buffers can, in turn, occur when
3636 		 * reconstituting a buffer from its VM backing store
3637 		 * base.  We only have to do this if B_CACHE is
3638 		 * clear ( which causes the I/O to occur in the
3639 		 * first place ).  The replacement prevents the read
3640 		 * I/O from overwriting potentially dirty VM-backed
3641 		 * pages.  XXX bogus page replacement is, uh, bogus.
3642 		 * It may not work properly with small-block devices.
3643 		 * We need to find a better way.
3644 		 */
3645 		if (clear_modify) {
3646 			pmap_remove_write(m);
3647 			vfs_page_set_validclean(bp, foff, m);
3648 		} else if (m->valid == VM_PAGE_BITS_ALL &&
3649 		    (bp->b_flags & B_CACHE) == 0) {
3650 			bp->b_pages[i] = bogus_page;
3651 			bogus++;
3652 		}
3653 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3654 	}
3655 	VM_OBJECT_WUNLOCK(obj);
3656 	if (bogus)
3657 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3658 		    bp->b_pages, bp->b_npages);
3659 }
3660 
3661 /*
3662  *	vfs_bio_set_valid:
3663  *
3664  *	Set the range within the buffer to valid.  The range is
3665  *	relative to the beginning of the buffer, b_offset.  Note that
3666  *	b_offset itself may be offset from the beginning of the first
3667  *	page.
3668  */
3669 void
3670 vfs_bio_set_valid(struct buf *bp, int base, int size)
3671 {
3672 	int i, n;
3673 	vm_page_t m;
3674 
3675 	if (!(bp->b_flags & B_VMIO))
3676 		return;
3677 
3678 	/*
3679 	 * Fixup base to be relative to beginning of first page.
3680 	 * Set initial n to be the maximum number of bytes in the
3681 	 * first page that can be validated.
3682 	 */
3683 	base += (bp->b_offset & PAGE_MASK);
3684 	n = PAGE_SIZE - (base & PAGE_MASK);
3685 
3686 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3687 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3688 		m = bp->b_pages[i];
3689 		if (n > size)
3690 			n = size;
3691 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
3692 		base += n;
3693 		size -= n;
3694 		n = PAGE_SIZE;
3695 	}
3696 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3697 }
3698 
3699 /*
3700  *	vfs_bio_clrbuf:
3701  *
3702  *	If the specified buffer is a non-VMIO buffer, clear the entire
3703  *	buffer.  If the specified buffer is a VMIO buffer, clear and
3704  *	validate only the previously invalid portions of the buffer.
3705  *	This routine essentially fakes an I/O, so we need to clear
3706  *	BIO_ERROR and B_INVAL.
3707  *
3708  *	Note that while we only theoretically need to clear through b_bcount,
3709  *	we go ahead and clear through b_bufsize.
3710  */
3711 void
3712 vfs_bio_clrbuf(struct buf *bp)
3713 {
3714 	int i, j, mask, sa, ea, slide;
3715 
3716 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3717 		clrbuf(bp);
3718 		return;
3719 	}
3720 	bp->b_flags &= ~B_INVAL;
3721 	bp->b_ioflags &= ~BIO_ERROR;
3722 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3723 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3724 	    (bp->b_offset & PAGE_MASK) == 0) {
3725 		if (bp->b_pages[0] == bogus_page)
3726 			goto unlock;
3727 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3728 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
3729 		if ((bp->b_pages[0]->valid & mask) == mask)
3730 			goto unlock;
3731 		if ((bp->b_pages[0]->valid & mask) == 0) {
3732 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
3733 			bp->b_pages[0]->valid |= mask;
3734 			goto unlock;
3735 		}
3736 	}
3737 	sa = bp->b_offset & PAGE_MASK;
3738 	slide = 0;
3739 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
3740 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
3741 		ea = slide & PAGE_MASK;
3742 		if (ea == 0)
3743 			ea = PAGE_SIZE;
3744 		if (bp->b_pages[i] == bogus_page)
3745 			continue;
3746 		j = sa / DEV_BSIZE;
3747 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3748 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
3749 		if ((bp->b_pages[i]->valid & mask) == mask)
3750 			continue;
3751 		if ((bp->b_pages[i]->valid & mask) == 0)
3752 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
3753 		else {
3754 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3755 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
3756 					pmap_zero_page_area(bp->b_pages[i],
3757 					    sa, DEV_BSIZE);
3758 				}
3759 			}
3760 		}
3761 		bp->b_pages[i]->valid |= mask;
3762 	}
3763 unlock:
3764 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3765 	bp->b_resid = 0;
3766 }
3767 
3768 /*
3769  * vm_hold_load_pages and vm_hold_free_pages get pages into
3770  * a buffers address space.  The pages are anonymous and are
3771  * not associated with a file object.
3772  */
3773 static void
3774 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3775 {
3776 	vm_offset_t pg;
3777 	vm_page_t p;
3778 	int index;
3779 
3780 	to = round_page(to);
3781 	from = round_page(from);
3782 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3783 
3784 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3785 tryagain:
3786 		/*
3787 		 * note: must allocate system pages since blocking here
3788 		 * could interfere with paging I/O, no matter which
3789 		 * process we are.
3790 		 */
3791 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
3792 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
3793 		if (p == NULL) {
3794 			VM_WAIT;
3795 			goto tryagain;
3796 		}
3797 		pmap_qenter(pg, &p, 1);
3798 		bp->b_pages[index] = p;
3799 	}
3800 	bp->b_npages = index;
3801 }
3802 
3803 /* Return pages associated with this buf to the vm system */
3804 static void
3805 vm_hold_free_pages(struct buf *bp, int newbsize)
3806 {
3807 	vm_offset_t from;
3808 	vm_page_t p;
3809 	int index, newnpages;
3810 
3811 	from = round_page((vm_offset_t)bp->b_data + newbsize);
3812 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3813 	if (bp->b_npages > newnpages)
3814 		pmap_qremove(from, bp->b_npages - newnpages);
3815 	for (index = newnpages; index < bp->b_npages; index++) {
3816 		p = bp->b_pages[index];
3817 		bp->b_pages[index] = NULL;
3818 		if (p->busy != 0)
3819 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3820 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
3821 		p->wire_count--;
3822 		vm_page_free(p);
3823 		atomic_subtract_int(&cnt.v_wire_count, 1);
3824 	}
3825 	bp->b_npages = newnpages;
3826 }
3827 
3828 /*
3829  * Map an IO request into kernel virtual address space.
3830  *
3831  * All requests are (re)mapped into kernel VA space.
3832  * Notice that we use b_bufsize for the size of the buffer
3833  * to be mapped.  b_bcount might be modified by the driver.
3834  *
3835  * Note that even if the caller determines that the address space should
3836  * be valid, a race or a smaller-file mapped into a larger space may
3837  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3838  * check the return value.
3839  */
3840 int
3841 vmapbuf(struct buf *bp)
3842 {
3843 	caddr_t kva;
3844 	vm_prot_t prot;
3845 	int pidx;
3846 
3847 	if (bp->b_bufsize < 0)
3848 		return (-1);
3849 	prot = VM_PROT_READ;
3850 	if (bp->b_iocmd == BIO_READ)
3851 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3852 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
3853 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
3854 	    btoc(MAXPHYS))) < 0)
3855 		return (-1);
3856 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3857 
3858 	kva = bp->b_saveaddr;
3859 	bp->b_npages = pidx;
3860 	bp->b_saveaddr = bp->b_data;
3861 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3862 	return(0);
3863 }
3864 
3865 /*
3866  * Free the io map PTEs associated with this IO operation.
3867  * We also invalidate the TLB entries and restore the original b_addr.
3868  */
3869 void
3870 vunmapbuf(struct buf *bp)
3871 {
3872 	int npages;
3873 
3874 	npages = bp->b_npages;
3875 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3876 	vm_page_unhold_pages(bp->b_pages, npages);
3877 
3878 	bp->b_data = bp->b_saveaddr;
3879 }
3880 
3881 void
3882 bdone(struct buf *bp)
3883 {
3884 	struct mtx *mtxp;
3885 
3886 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3887 	mtx_lock(mtxp);
3888 	bp->b_flags |= B_DONE;
3889 	wakeup(bp);
3890 	mtx_unlock(mtxp);
3891 }
3892 
3893 void
3894 bwait(struct buf *bp, u_char pri, const char *wchan)
3895 {
3896 	struct mtx *mtxp;
3897 
3898 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3899 	mtx_lock(mtxp);
3900 	while ((bp->b_flags & B_DONE) == 0)
3901 		msleep(bp, mtxp, pri, wchan, 0);
3902 	mtx_unlock(mtxp);
3903 }
3904 
3905 int
3906 bufsync(struct bufobj *bo, int waitfor)
3907 {
3908 
3909 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3910 }
3911 
3912 void
3913 bufstrategy(struct bufobj *bo, struct buf *bp)
3914 {
3915 	int i = 0;
3916 	struct vnode *vp;
3917 
3918 	vp = bp->b_vp;
3919 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3920 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3921 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3922 	i = VOP_STRATEGY(vp, bp);
3923 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3924 }
3925 
3926 void
3927 bufobj_wrefl(struct bufobj *bo)
3928 {
3929 
3930 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3931 	ASSERT_BO_LOCKED(bo);
3932 	bo->bo_numoutput++;
3933 }
3934 
3935 void
3936 bufobj_wref(struct bufobj *bo)
3937 {
3938 
3939 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3940 	BO_LOCK(bo);
3941 	bo->bo_numoutput++;
3942 	BO_UNLOCK(bo);
3943 }
3944 
3945 void
3946 bufobj_wdrop(struct bufobj *bo)
3947 {
3948 
3949 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3950 	BO_LOCK(bo);
3951 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3952 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3953 		bo->bo_flag &= ~BO_WWAIT;
3954 		wakeup(&bo->bo_numoutput);
3955 	}
3956 	BO_UNLOCK(bo);
3957 }
3958 
3959 int
3960 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3961 {
3962 	int error;
3963 
3964 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3965 	ASSERT_BO_LOCKED(bo);
3966 	error = 0;
3967 	while (bo->bo_numoutput) {
3968 		bo->bo_flag |= BO_WWAIT;
3969 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3970 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3971 		if (error)
3972 			break;
3973 	}
3974 	return (error);
3975 }
3976 
3977 void
3978 bpin(struct buf *bp)
3979 {
3980 	struct mtx *mtxp;
3981 
3982 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3983 	mtx_lock(mtxp);
3984 	bp->b_pin_count++;
3985 	mtx_unlock(mtxp);
3986 }
3987 
3988 void
3989 bunpin(struct buf *bp)
3990 {
3991 	struct mtx *mtxp;
3992 
3993 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3994 	mtx_lock(mtxp);
3995 	if (--bp->b_pin_count == 0)
3996 		wakeup(bp);
3997 	mtx_unlock(mtxp);
3998 }
3999 
4000 void
4001 bunpin_wait(struct buf *bp)
4002 {
4003 	struct mtx *mtxp;
4004 
4005 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4006 	mtx_lock(mtxp);
4007 	while (bp->b_pin_count > 0)
4008 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4009 	mtx_unlock(mtxp);
4010 }
4011 
4012 #include "opt_ddb.h"
4013 #ifdef DDB
4014 #include <ddb/ddb.h>
4015 
4016 /* DDB command to show buffer data */
4017 DB_SHOW_COMMAND(buffer, db_show_buffer)
4018 {
4019 	/* get args */
4020 	struct buf *bp = (struct buf *)addr;
4021 
4022 	if (!have_addr) {
4023 		db_printf("usage: show buffer <addr>\n");
4024 		return;
4025 	}
4026 
4027 	db_printf("buf at %p\n", bp);
4028 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4029 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4030 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4031 	db_printf(
4032 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4033 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4034 	    "b_dep = %p\n",
4035 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4036 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4037 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4038 	if (bp->b_npages) {
4039 		int i;
4040 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4041 		for (i = 0; i < bp->b_npages; i++) {
4042 			vm_page_t m;
4043 			m = bp->b_pages[i];
4044 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4045 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4046 			if ((i + 1) < bp->b_npages)
4047 				db_printf(",");
4048 		}
4049 		db_printf("\n");
4050 	}
4051 	db_printf(" ");
4052 	BUF_LOCKPRINTINFO(bp);
4053 }
4054 
4055 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4056 {
4057 	struct buf *bp;
4058 	int i;
4059 
4060 	for (i = 0; i < nbuf; i++) {
4061 		bp = &buf[i];
4062 		if (BUF_ISLOCKED(bp)) {
4063 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4064 			db_printf("\n");
4065 		}
4066 	}
4067 }
4068 
4069 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4070 {
4071 	struct vnode *vp;
4072 	struct buf *bp;
4073 
4074 	if (!have_addr) {
4075 		db_printf("usage: show vnodebufs <addr>\n");
4076 		return;
4077 	}
4078 	vp = (struct vnode *)addr;
4079 	db_printf("Clean buffers:\n");
4080 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4081 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4082 		db_printf("\n");
4083 	}
4084 	db_printf("Dirty buffers:\n");
4085 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4086 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4087 		db_printf("\n");
4088 	}
4089 }
4090 
4091 DB_COMMAND(countfreebufs, db_coundfreebufs)
4092 {
4093 	struct buf *bp;
4094 	int i, used = 0, nfree = 0;
4095 
4096 	if (have_addr) {
4097 		db_printf("usage: countfreebufs\n");
4098 		return;
4099 	}
4100 
4101 	for (i = 0; i < nbuf; i++) {
4102 		bp = &buf[i];
4103 		if ((bp->b_vflags & BV_INFREECNT) != 0)
4104 			nfree++;
4105 		else
4106 			used++;
4107 	}
4108 
4109 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4110 	    nfree + used);
4111 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4112 }
4113 #endif /* DDB */
4114