xref: /freebsd/sys/kern/vfs_bio.c (revision a10cee30c94cf5944826d2a495e9cdf339dfbcc8)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * this file contains a new buffer I/O scheme implementing a coherent
30  * VM object and buffer cache scheme.  Pains have been taken to make
31  * sure that the performance degradation associated with schemes such
32  * as this is not realized.
33  *
34  * Author:  John S. Dyson
35  * Significant help during the development and debugging phases
36  * had been provided by David Greenman, also of the FreeBSD core team.
37  *
38  * see man buf(9) for more info.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bio.h>
47 #include <sys/conf.h>
48 #include <sys/buf.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
51 #include <sys/fail.h>
52 #include <sys/limits.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mount.h>
56 #include <sys/mutex.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
59 #include <sys/proc.h>
60 #include <sys/resourcevar.h>
61 #include <sys/sysctl.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
64 #include <geom/geom.h>
65 #include <vm/vm.h>
66 #include <vm/vm_param.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_pageout.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_extern.h>
72 #include <vm/vm_map.h>
73 #include "opt_compat.h"
74 #include "opt_directio.h"
75 #include "opt_swap.h"
76 
77 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
78 
79 struct	bio_ops bioops;		/* I/O operation notification */
80 
81 struct	buf_ops buf_ops_bio = {
82 	.bop_name	=	"buf_ops_bio",
83 	.bop_write	=	bufwrite,
84 	.bop_strategy	=	bufstrategy,
85 	.bop_sync	=	bufsync,
86 	.bop_bdflush	=	bufbdflush,
87 };
88 
89 /*
90  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
91  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
92  */
93 struct buf *buf;		/* buffer header pool */
94 
95 static struct proc *bufdaemonproc;
96 
97 static int inmem(struct vnode *vp, daddr_t blkno);
98 static void vm_hold_free_pages(struct buf *bp, int newbsize);
99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
100 		vm_offset_t to);
101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
103 		vm_page_t m);
104 static void vfs_drain_busy_pages(struct buf *bp);
105 static void vfs_clean_pages_dirty_buf(struct buf *bp);
106 static void vfs_setdirty_locked_object(struct buf *bp);
107 static void vfs_vmio_release(struct buf *bp);
108 static int vfs_bio_clcheck(struct vnode *vp, int size,
109 		daddr_t lblkno, daddr_t blkno);
110 static int buf_do_flush(struct vnode *vp);
111 static int flushbufqueues(struct vnode *, int, int);
112 static void buf_daemon(void);
113 static void bremfreel(struct buf *bp);
114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
115     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
117 #endif
118 
119 int vmiodirenable = TRUE;
120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
121     "Use the VM system for directory writes");
122 long runningbufspace;
123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
124     "Amount of presently outstanding async buffer io");
125 static long bufspace;
126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
127     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
129     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
130 #else
131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
132     "Virtual memory used for buffers");
133 #endif
134 static long maxbufspace;
135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
136     "Maximum allowed value of bufspace (including buf_daemon)");
137 static long bufmallocspace;
138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
139     "Amount of malloced memory for buffers");
140 static long maxbufmallocspace;
141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
142     "Maximum amount of malloced memory for buffers");
143 static long lobufspace;
144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
145     "Minimum amount of buffers we want to have");
146 long hibufspace;
147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
148     "Maximum allowed value of bufspace (excluding buf_daemon)");
149 static int bufreusecnt;
150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
151     "Number of times we have reused a buffer");
152 static int buffreekvacnt;
153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
154     "Number of times we have freed the KVA space from some buffer");
155 static int bufdefragcnt;
156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
157     "Number of times we have had to repeat buffer allocation to defragment");
158 static long lorunningspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
160     "Minimum preferred space used for in-progress I/O");
161 static long hirunningspace;
162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
163     "Maximum amount of space to use for in-progress I/O");
164 int dirtybufferflushes;
165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
166     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
167 int bdwriteskip;
168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
169     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
170 int altbufferflushes;
171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
172     0, "Number of fsync flushes to limit dirty buffers");
173 static int recursiveflushes;
174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
175     0, "Number of flushes skipped due to being recursive");
176 static int numdirtybuffers;
177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
178     "Number of buffers that are dirty (has unwritten changes) at the moment");
179 static int lodirtybuffers;
180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
181     "How many buffers we want to have free before bufdaemon can sleep");
182 static int hidirtybuffers;
183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
184     "When the number of dirty buffers is considered severe");
185 int dirtybufthresh;
186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
187     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
188 static int numfreebuffers;
189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
190     "Number of free buffers");
191 static int lofreebuffers;
192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
193    "XXX Unused");
194 static int hifreebuffers;
195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
196    "XXX Complicatedly unused");
197 static int getnewbufcalls;
198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
199    "Number of calls to getnewbuf");
200 static int getnewbufrestarts;
201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
202     "Number of times getnewbuf has had to restart a buffer aquisition");
203 static int flushbufqtarget = 100;
204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
205     "Amount of work to do in flushbufqueues when helping bufdaemon");
206 static long notbufdflashes;
207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, &notbufdflashes, 0,
208     "Number of dirty buffer flushes done by the bufdaemon helpers");
209 
210 /*
211  * Wakeup point for bufdaemon, as well as indicator of whether it is already
212  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
213  * is idling.
214  */
215 static int bd_request;
216 
217 /*
218  * Request for the buf daemon to write more buffers than is indicated by
219  * lodirtybuf.  This may be necessary to push out excess dependencies or
220  * defragment the address space where a simple count of the number of dirty
221  * buffers is insufficient to characterize the demand for flushing them.
222  */
223 static int bd_speedupreq;
224 
225 /*
226  * This lock synchronizes access to bd_request.
227  */
228 static struct mtx bdlock;
229 
230 /*
231  * bogus page -- for I/O to/from partially complete buffers
232  * this is a temporary solution to the problem, but it is not
233  * really that bad.  it would be better to split the buffer
234  * for input in the case of buffers partially already in memory,
235  * but the code is intricate enough already.
236  */
237 vm_page_t bogus_page;
238 
239 /*
240  * Synchronization (sleep/wakeup) variable for active buffer space requests.
241  * Set when wait starts, cleared prior to wakeup().
242  * Used in runningbufwakeup() and waitrunningbufspace().
243  */
244 static int runningbufreq;
245 
246 /*
247  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
248  * waitrunningbufspace().
249  */
250 static struct mtx rbreqlock;
251 
252 /*
253  * Synchronization (sleep/wakeup) variable for buffer requests.
254  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
255  * by and/or.
256  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
257  * getnewbuf(), and getblk().
258  */
259 static int needsbuffer;
260 
261 /*
262  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
263  */
264 static struct mtx nblock;
265 
266 /*
267  * Definitions for the buffer free lists.
268  */
269 #define BUFFER_QUEUES	6	/* number of free buffer queues */
270 
271 #define QUEUE_NONE	0	/* on no queue */
272 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
273 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
274 #define QUEUE_DIRTY_GIANT 3	/* B_DELWRI buffers that need giant */
275 #define QUEUE_EMPTYKVA	4	/* empty buffer headers w/KVA assignment */
276 #define QUEUE_EMPTY	5	/* empty buffer headers */
277 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
278 
279 /* Queues for free buffers with various properties */
280 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
281 
282 /* Lock for the bufqueues */
283 static struct mtx bqlock;
284 
285 /*
286  * Single global constant for BUF_WMESG, to avoid getting multiple references.
287  * buf_wmesg is referred from macros.
288  */
289 const char *buf_wmesg = BUF_WMESG;
290 
291 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
292 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
293 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
294 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
295 
296 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
297     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
298 static int
299 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
300 {
301 	long lvalue;
302 	int ivalue;
303 
304 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
305 		return (sysctl_handle_long(oidp, arg1, arg2, req));
306 	lvalue = *(long *)arg1;
307 	if (lvalue > INT_MAX)
308 		/* On overflow, still write out a long to trigger ENOMEM. */
309 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
310 	ivalue = lvalue;
311 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
312 }
313 #endif
314 
315 #ifdef DIRECTIO
316 extern void ffs_rawread_setup(void);
317 #endif /* DIRECTIO */
318 /*
319  *	numdirtywakeup:
320  *
321  *	If someone is blocked due to there being too many dirty buffers,
322  *	and numdirtybuffers is now reasonable, wake them up.
323  */
324 
325 static __inline void
326 numdirtywakeup(int level)
327 {
328 
329 	if (numdirtybuffers <= level) {
330 		mtx_lock(&nblock);
331 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
332 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
333 			wakeup(&needsbuffer);
334 		}
335 		mtx_unlock(&nblock);
336 	}
337 }
338 
339 /*
340  *	bufspacewakeup:
341  *
342  *	Called when buffer space is potentially available for recovery.
343  *	getnewbuf() will block on this flag when it is unable to free
344  *	sufficient buffer space.  Buffer space becomes recoverable when
345  *	bp's get placed back in the queues.
346  */
347 
348 static __inline void
349 bufspacewakeup(void)
350 {
351 
352 	/*
353 	 * If someone is waiting for BUF space, wake them up.  Even
354 	 * though we haven't freed the kva space yet, the waiting
355 	 * process will be able to now.
356 	 */
357 	mtx_lock(&nblock);
358 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
359 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
360 		wakeup(&needsbuffer);
361 	}
362 	mtx_unlock(&nblock);
363 }
364 
365 /*
366  * runningbufwakeup() - in-progress I/O accounting.
367  *
368  */
369 void
370 runningbufwakeup(struct buf *bp)
371 {
372 
373 	if (bp->b_runningbufspace) {
374 		atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
375 		bp->b_runningbufspace = 0;
376 		mtx_lock(&rbreqlock);
377 		if (runningbufreq && runningbufspace <= lorunningspace) {
378 			runningbufreq = 0;
379 			wakeup(&runningbufreq);
380 		}
381 		mtx_unlock(&rbreqlock);
382 	}
383 }
384 
385 /*
386  *	bufcountwakeup:
387  *
388  *	Called when a buffer has been added to one of the free queues to
389  *	account for the buffer and to wakeup anyone waiting for free buffers.
390  *	This typically occurs when large amounts of metadata are being handled
391  *	by the buffer cache ( else buffer space runs out first, usually ).
392  */
393 
394 static __inline void
395 bufcountwakeup(struct buf *bp)
396 {
397 	int old;
398 
399 	KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
400 	    ("buf %p already counted as free", bp));
401 	if (bp->b_bufobj != NULL)
402 		mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
403 	bp->b_vflags |= BV_INFREECNT;
404 	old = atomic_fetchadd_int(&numfreebuffers, 1);
405 	KASSERT(old >= 0 && old < nbuf,
406 	    ("numfreebuffers climbed to %d", old + 1));
407 	mtx_lock(&nblock);
408 	if (needsbuffer) {
409 		needsbuffer &= ~VFS_BIO_NEED_ANY;
410 		if (numfreebuffers >= hifreebuffers)
411 			needsbuffer &= ~VFS_BIO_NEED_FREE;
412 		wakeup(&needsbuffer);
413 	}
414 	mtx_unlock(&nblock);
415 }
416 
417 /*
418  *	waitrunningbufspace()
419  *
420  *	runningbufspace is a measure of the amount of I/O currently
421  *	running.  This routine is used in async-write situations to
422  *	prevent creating huge backups of pending writes to a device.
423  *	Only asynchronous writes are governed by this function.
424  *
425  *	Reads will adjust runningbufspace, but will not block based on it.
426  *	The read load has a side effect of reducing the allowed write load.
427  *
428  *	This does NOT turn an async write into a sync write.  It waits
429  *	for earlier writes to complete and generally returns before the
430  *	caller's write has reached the device.
431  */
432 void
433 waitrunningbufspace(void)
434 {
435 
436 	mtx_lock(&rbreqlock);
437 	while (runningbufspace > hirunningspace) {
438 		++runningbufreq;
439 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
440 	}
441 	mtx_unlock(&rbreqlock);
442 }
443 
444 
445 /*
446  *	vfs_buf_test_cache:
447  *
448  *	Called when a buffer is extended.  This function clears the B_CACHE
449  *	bit if the newly extended portion of the buffer does not contain
450  *	valid data.
451  */
452 static __inline
453 void
454 vfs_buf_test_cache(struct buf *bp,
455 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
456 		  vm_page_t m)
457 {
458 
459 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
460 	if (bp->b_flags & B_CACHE) {
461 		int base = (foff + off) & PAGE_MASK;
462 		if (vm_page_is_valid(m, base, size) == 0)
463 			bp->b_flags &= ~B_CACHE;
464 	}
465 }
466 
467 /* Wake up the buffer daemon if necessary */
468 static __inline
469 void
470 bd_wakeup(int dirtybuflevel)
471 {
472 
473 	mtx_lock(&bdlock);
474 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
475 		bd_request = 1;
476 		wakeup(&bd_request);
477 	}
478 	mtx_unlock(&bdlock);
479 }
480 
481 /*
482  * bd_speedup - speedup the buffer cache flushing code
483  */
484 
485 void
486 bd_speedup(void)
487 {
488 	int needwake;
489 
490 	mtx_lock(&bdlock);
491 	needwake = 0;
492 	if (bd_speedupreq == 0 || bd_request == 0)
493 		needwake = 1;
494 	bd_speedupreq = 1;
495 	bd_request = 1;
496 	if (needwake)
497 		wakeup(&bd_request);
498 	mtx_unlock(&bdlock);
499 }
500 
501 /*
502  * Calculating buffer cache scaling values and reserve space for buffer
503  * headers.  This is called during low level kernel initialization and
504  * may be called more then once.  We CANNOT write to the memory area
505  * being reserved at this time.
506  */
507 caddr_t
508 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
509 {
510 	int tuned_nbuf;
511 	long maxbuf;
512 
513 	/*
514 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
515 	 * PAGE_SIZE is >= 1K)
516 	 */
517 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
518 
519 	/*
520 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
521 	 * For the first 64MB of ram nominally allocate sufficient buffers to
522 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
523 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
524 	 * the buffer cache we limit the eventual kva reservation to
525 	 * maxbcache bytes.
526 	 *
527 	 * factor represents the 1/4 x ram conversion.
528 	 */
529 	if (nbuf == 0) {
530 		int factor = 4 * BKVASIZE / 1024;
531 
532 		nbuf = 50;
533 		if (physmem_est > 4096)
534 			nbuf += min((physmem_est - 4096) / factor,
535 			    65536 / factor);
536 		if (physmem_est > 65536)
537 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
538 
539 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
540 			nbuf = maxbcache / BKVASIZE;
541 		tuned_nbuf = 1;
542 	} else
543 		tuned_nbuf = 0;
544 
545 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
546 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
547 	if (nbuf > maxbuf) {
548 		if (!tuned_nbuf)
549 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
550 			    maxbuf);
551 		nbuf = maxbuf;
552 	}
553 
554 	/*
555 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
556 	 * We have no less then 16 and no more then 256.
557 	 */
558 	nswbuf = max(min(nbuf/4, 256), 16);
559 #ifdef NSWBUF_MIN
560 	if (nswbuf < NSWBUF_MIN)
561 		nswbuf = NSWBUF_MIN;
562 #endif
563 #ifdef DIRECTIO
564 	ffs_rawread_setup();
565 #endif
566 
567 	/*
568 	 * Reserve space for the buffer cache buffers
569 	 */
570 	swbuf = (void *)v;
571 	v = (caddr_t)(swbuf + nswbuf);
572 	buf = (void *)v;
573 	v = (caddr_t)(buf + nbuf);
574 
575 	return(v);
576 }
577 
578 /* Initialize the buffer subsystem.  Called before use of any buffers. */
579 void
580 bufinit(void)
581 {
582 	struct buf *bp;
583 	int i;
584 
585 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
586 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
587 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
588 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
589 
590 	/* next, make a null set of free lists */
591 	for (i = 0; i < BUFFER_QUEUES; i++)
592 		TAILQ_INIT(&bufqueues[i]);
593 
594 	/* finally, initialize each buffer header and stick on empty q */
595 	for (i = 0; i < nbuf; i++) {
596 		bp = &buf[i];
597 		bzero(bp, sizeof *bp);
598 		bp->b_flags = B_INVAL;	/* we're just an empty header */
599 		bp->b_rcred = NOCRED;
600 		bp->b_wcred = NOCRED;
601 		bp->b_qindex = QUEUE_EMPTY;
602 		bp->b_vflags = BV_INFREECNT;	/* buf is counted as free */
603 		bp->b_xflags = 0;
604 		LIST_INIT(&bp->b_dep);
605 		BUF_LOCKINIT(bp);
606 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
607 	}
608 
609 	/*
610 	 * maxbufspace is the absolute maximum amount of buffer space we are
611 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
612 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
613 	 * used by most other processes.  The differential is required to
614 	 * ensure that buf_daemon is able to run when other processes might
615 	 * be blocked waiting for buffer space.
616 	 *
617 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
618 	 * this may result in KVM fragmentation which is not handled optimally
619 	 * by the system.
620 	 */
621 	maxbufspace = (long)nbuf * BKVASIZE;
622 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
623 	lobufspace = hibufspace - MAXBSIZE;
624 
625 	/*
626 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
627 	 * arbitrarily and may need further tuning. It corresponds to
628 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
629 	 * which fits with many RAID controllers' tagged queuing limits.
630 	 * The lower 1 MiB limit is the historical upper limit for
631 	 * hirunningspace.
632 	 */
633 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
634 	    16 * 1024 * 1024), 1024 * 1024);
635 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
636 
637 /*
638  * Limit the amount of malloc memory since it is wired permanently into
639  * the kernel space.  Even though this is accounted for in the buffer
640  * allocation, we don't want the malloced region to grow uncontrolled.
641  * The malloc scheme improves memory utilization significantly on average
642  * (small) directories.
643  */
644 	maxbufmallocspace = hibufspace / 20;
645 
646 /*
647  * Reduce the chance of a deadlock occuring by limiting the number
648  * of delayed-write dirty buffers we allow to stack up.
649  */
650 	hidirtybuffers = nbuf / 4 + 20;
651 	dirtybufthresh = hidirtybuffers * 9 / 10;
652 	numdirtybuffers = 0;
653 /*
654  * To support extreme low-memory systems, make sure hidirtybuffers cannot
655  * eat up all available buffer space.  This occurs when our minimum cannot
656  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
657  * BKVASIZE'd buffers.
658  */
659 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
660 		hidirtybuffers >>= 1;
661 	}
662 	lodirtybuffers = hidirtybuffers / 2;
663 
664 /*
665  * Try to keep the number of free buffers in the specified range,
666  * and give special processes (e.g. like buf_daemon) access to an
667  * emergency reserve.
668  */
669 	lofreebuffers = nbuf / 18 + 5;
670 	hifreebuffers = 2 * lofreebuffers;
671 	numfreebuffers = nbuf;
672 
673 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
674 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
675 }
676 
677 /*
678  * bfreekva() - free the kva allocation for a buffer.
679  *
680  *	Since this call frees up buffer space, we call bufspacewakeup().
681  */
682 static void
683 bfreekva(struct buf *bp)
684 {
685 
686 	if (bp->b_kvasize) {
687 		atomic_add_int(&buffreekvacnt, 1);
688 		atomic_subtract_long(&bufspace, bp->b_kvasize);
689 		vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
690 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
691 		bp->b_kvasize = 0;
692 		bufspacewakeup();
693 	}
694 }
695 
696 /*
697  *	bremfree:
698  *
699  *	Mark the buffer for removal from the appropriate free list in brelse.
700  *
701  */
702 void
703 bremfree(struct buf *bp)
704 {
705 	int old;
706 
707 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
708 	KASSERT((bp->b_flags & B_REMFREE) == 0,
709 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
710 	KASSERT(bp->b_qindex != QUEUE_NONE,
711 	    ("bremfree: buffer %p not on a queue.", bp));
712 	BUF_ASSERT_HELD(bp);
713 
714 	bp->b_flags |= B_REMFREE;
715 	/* Fixup numfreebuffers count.  */
716 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
717 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
718 		    ("buf %p not counted in numfreebuffers", bp));
719 		if (bp->b_bufobj != NULL)
720 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
721 		bp->b_vflags &= ~BV_INFREECNT;
722 		old = atomic_fetchadd_int(&numfreebuffers, -1);
723 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
724 	}
725 }
726 
727 /*
728  *	bremfreef:
729  *
730  *	Force an immediate removal from a free list.  Used only in nfs when
731  *	it abuses the b_freelist pointer.
732  */
733 void
734 bremfreef(struct buf *bp)
735 {
736 	mtx_lock(&bqlock);
737 	bremfreel(bp);
738 	mtx_unlock(&bqlock);
739 }
740 
741 /*
742  *	bremfreel:
743  *
744  *	Removes a buffer from the free list, must be called with the
745  *	bqlock held.
746  */
747 static void
748 bremfreel(struct buf *bp)
749 {
750 	int old;
751 
752 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
753 	    bp, bp->b_vp, bp->b_flags);
754 	KASSERT(bp->b_qindex != QUEUE_NONE,
755 	    ("bremfreel: buffer %p not on a queue.", bp));
756 	BUF_ASSERT_HELD(bp);
757 	mtx_assert(&bqlock, MA_OWNED);
758 
759 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
760 	bp->b_qindex = QUEUE_NONE;
761 	/*
762 	 * If this was a delayed bremfree() we only need to remove the buffer
763 	 * from the queue and return the stats are already done.
764 	 */
765 	if (bp->b_flags & B_REMFREE) {
766 		bp->b_flags &= ~B_REMFREE;
767 		return;
768 	}
769 	/*
770 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
771 	 * delayed-write, the buffer was free and we must decrement
772 	 * numfreebuffers.
773 	 */
774 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
775 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
776 		    ("buf %p not counted in numfreebuffers", bp));
777 		if (bp->b_bufobj != NULL)
778 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
779 		bp->b_vflags &= ~BV_INFREECNT;
780 		old = atomic_fetchadd_int(&numfreebuffers, -1);
781 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
782 	}
783 }
784 
785 /*
786  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
787  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
788  * the buffer is valid and we do not have to do anything.
789  */
790 void
791 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
792     int cnt, struct ucred * cred)
793 {
794 	struct buf *rabp;
795 	int i;
796 
797 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
798 		if (inmem(vp, *rablkno))
799 			continue;
800 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
801 
802 		if ((rabp->b_flags & B_CACHE) == 0) {
803 			if (!TD_IS_IDLETHREAD(curthread))
804 				curthread->td_ru.ru_inblock++;
805 			rabp->b_flags |= B_ASYNC;
806 			rabp->b_flags &= ~B_INVAL;
807 			rabp->b_ioflags &= ~BIO_ERROR;
808 			rabp->b_iocmd = BIO_READ;
809 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
810 				rabp->b_rcred = crhold(cred);
811 			vfs_busy_pages(rabp, 0);
812 			BUF_KERNPROC(rabp);
813 			rabp->b_iooffset = dbtob(rabp->b_blkno);
814 			bstrategy(rabp);
815 		} else {
816 			brelse(rabp);
817 		}
818 	}
819 }
820 
821 /*
822  * Entry point for bread() and breadn() via #defines in sys/buf.h.
823  *
824  * Get a buffer with the specified data.  Look in the cache first.  We
825  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
826  * is set, the buffer is valid and we do not have to do anything, see
827  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
828  */
829 int
830 breadn_flags(struct vnode * vp, daddr_t blkno, int size,
831     daddr_t * rablkno, int *rabsize, int cnt,
832     struct ucred * cred, int flags, struct buf **bpp)
833 {
834 	struct buf *bp;
835 	int rv = 0, readwait = 0;
836 
837 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
838 	/*
839 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
840 	 */
841 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
842 	if (bp == NULL)
843 		return (EBUSY);
844 
845 	/* if not found in cache, do some I/O */
846 	if ((bp->b_flags & B_CACHE) == 0) {
847 		if (!TD_IS_IDLETHREAD(curthread))
848 			curthread->td_ru.ru_inblock++;
849 		bp->b_iocmd = BIO_READ;
850 		bp->b_flags &= ~B_INVAL;
851 		bp->b_ioflags &= ~BIO_ERROR;
852 		if (bp->b_rcred == NOCRED && cred != NOCRED)
853 			bp->b_rcred = crhold(cred);
854 		vfs_busy_pages(bp, 0);
855 		bp->b_iooffset = dbtob(bp->b_blkno);
856 		bstrategy(bp);
857 		++readwait;
858 	}
859 
860 	breada(vp, rablkno, rabsize, cnt, cred);
861 
862 	if (readwait) {
863 		rv = bufwait(bp);
864 	}
865 	return (rv);
866 }
867 
868 /*
869  * Write, release buffer on completion.  (Done by iodone
870  * if async).  Do not bother writing anything if the buffer
871  * is invalid.
872  *
873  * Note that we set B_CACHE here, indicating that buffer is
874  * fully valid and thus cacheable.  This is true even of NFS
875  * now so we set it generally.  This could be set either here
876  * or in biodone() since the I/O is synchronous.  We put it
877  * here.
878  */
879 int
880 bufwrite(struct buf *bp)
881 {
882 	int oldflags;
883 	struct vnode *vp;
884 	int vp_md;
885 
886 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
887 	if (bp->b_flags & B_INVAL) {
888 		brelse(bp);
889 		return (0);
890 	}
891 
892 	oldflags = bp->b_flags;
893 
894 	BUF_ASSERT_HELD(bp);
895 
896 	if (bp->b_pin_count > 0)
897 		bunpin_wait(bp);
898 
899 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
900 	    ("FFS background buffer should not get here %p", bp));
901 
902 	vp = bp->b_vp;
903 	if (vp)
904 		vp_md = vp->v_vflag & VV_MD;
905 	else
906 		vp_md = 0;
907 
908 	/* Mark the buffer clean */
909 	bundirty(bp);
910 
911 	bp->b_flags &= ~B_DONE;
912 	bp->b_ioflags &= ~BIO_ERROR;
913 	bp->b_flags |= B_CACHE;
914 	bp->b_iocmd = BIO_WRITE;
915 
916 	bufobj_wref(bp->b_bufobj);
917 	vfs_busy_pages(bp, 1);
918 
919 	/*
920 	 * Normal bwrites pipeline writes
921 	 */
922 	bp->b_runningbufspace = bp->b_bufsize;
923 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
924 
925 	if (!TD_IS_IDLETHREAD(curthread))
926 		curthread->td_ru.ru_oublock++;
927 	if (oldflags & B_ASYNC)
928 		BUF_KERNPROC(bp);
929 	bp->b_iooffset = dbtob(bp->b_blkno);
930 	bstrategy(bp);
931 
932 	if ((oldflags & B_ASYNC) == 0) {
933 		int rtval = bufwait(bp);
934 		brelse(bp);
935 		return (rtval);
936 	} else {
937 		/*
938 		 * don't allow the async write to saturate the I/O
939 		 * system.  We will not deadlock here because
940 		 * we are blocking waiting for I/O that is already in-progress
941 		 * to complete. We do not block here if it is the update
942 		 * or syncer daemon trying to clean up as that can lead
943 		 * to deadlock.
944 		 */
945 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
946 			waitrunningbufspace();
947 	}
948 
949 	return (0);
950 }
951 
952 void
953 bufbdflush(struct bufobj *bo, struct buf *bp)
954 {
955 	struct buf *nbp;
956 
957 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
958 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
959 		altbufferflushes++;
960 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
961 		BO_LOCK(bo);
962 		/*
963 		 * Try to find a buffer to flush.
964 		 */
965 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
966 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
967 			    BUF_LOCK(nbp,
968 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
969 				continue;
970 			if (bp == nbp)
971 				panic("bdwrite: found ourselves");
972 			BO_UNLOCK(bo);
973 			/* Don't countdeps with the bo lock held. */
974 			if (buf_countdeps(nbp, 0)) {
975 				BO_LOCK(bo);
976 				BUF_UNLOCK(nbp);
977 				continue;
978 			}
979 			if (nbp->b_flags & B_CLUSTEROK) {
980 				vfs_bio_awrite(nbp);
981 			} else {
982 				bremfree(nbp);
983 				bawrite(nbp);
984 			}
985 			dirtybufferflushes++;
986 			break;
987 		}
988 		if (nbp == NULL)
989 			BO_UNLOCK(bo);
990 	}
991 }
992 
993 /*
994  * Delayed write. (Buffer is marked dirty).  Do not bother writing
995  * anything if the buffer is marked invalid.
996  *
997  * Note that since the buffer must be completely valid, we can safely
998  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
999  * biodone() in order to prevent getblk from writing the buffer
1000  * out synchronously.
1001  */
1002 void
1003 bdwrite(struct buf *bp)
1004 {
1005 	struct thread *td = curthread;
1006 	struct vnode *vp;
1007 	struct bufobj *bo;
1008 
1009 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1010 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1011 	BUF_ASSERT_HELD(bp);
1012 
1013 	if (bp->b_flags & B_INVAL) {
1014 		brelse(bp);
1015 		return;
1016 	}
1017 
1018 	/*
1019 	 * If we have too many dirty buffers, don't create any more.
1020 	 * If we are wildly over our limit, then force a complete
1021 	 * cleanup. Otherwise, just keep the situation from getting
1022 	 * out of control. Note that we have to avoid a recursive
1023 	 * disaster and not try to clean up after our own cleanup!
1024 	 */
1025 	vp = bp->b_vp;
1026 	bo = bp->b_bufobj;
1027 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1028 		td->td_pflags |= TDP_INBDFLUSH;
1029 		BO_BDFLUSH(bo, bp);
1030 		td->td_pflags &= ~TDP_INBDFLUSH;
1031 	} else
1032 		recursiveflushes++;
1033 
1034 	bdirty(bp);
1035 	/*
1036 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1037 	 * true even of NFS now.
1038 	 */
1039 	bp->b_flags |= B_CACHE;
1040 
1041 	/*
1042 	 * This bmap keeps the system from needing to do the bmap later,
1043 	 * perhaps when the system is attempting to do a sync.  Since it
1044 	 * is likely that the indirect block -- or whatever other datastructure
1045 	 * that the filesystem needs is still in memory now, it is a good
1046 	 * thing to do this.  Note also, that if the pageout daemon is
1047 	 * requesting a sync -- there might not be enough memory to do
1048 	 * the bmap then...  So, this is important to do.
1049 	 */
1050 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1051 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1052 	}
1053 
1054 	/*
1055 	 * Set the *dirty* buffer range based upon the VM system dirty
1056 	 * pages.
1057 	 *
1058 	 * Mark the buffer pages as clean.  We need to do this here to
1059 	 * satisfy the vnode_pager and the pageout daemon, so that it
1060 	 * thinks that the pages have been "cleaned".  Note that since
1061 	 * the pages are in a delayed write buffer -- the VFS layer
1062 	 * "will" see that the pages get written out on the next sync,
1063 	 * or perhaps the cluster will be completed.
1064 	 */
1065 	vfs_clean_pages_dirty_buf(bp);
1066 	bqrelse(bp);
1067 
1068 	/*
1069 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
1070 	 * buffers (midpoint between our recovery point and our stall
1071 	 * point).
1072 	 */
1073 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1074 
1075 	/*
1076 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1077 	 * due to the softdep code.
1078 	 */
1079 }
1080 
1081 /*
1082  *	bdirty:
1083  *
1084  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1085  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1086  *	itself to properly update it in the dirty/clean lists.  We mark it
1087  *	B_DONE to ensure that any asynchronization of the buffer properly
1088  *	clears B_DONE ( else a panic will occur later ).
1089  *
1090  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1091  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1092  *	should only be called if the buffer is known-good.
1093  *
1094  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1095  *	count.
1096  *
1097  *	The buffer must be on QUEUE_NONE.
1098  */
1099 void
1100 bdirty(struct buf *bp)
1101 {
1102 
1103 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1104 	    bp, bp->b_vp, bp->b_flags);
1105 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1106 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1107 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1108 	BUF_ASSERT_HELD(bp);
1109 	bp->b_flags &= ~(B_RELBUF);
1110 	bp->b_iocmd = BIO_WRITE;
1111 
1112 	if ((bp->b_flags & B_DELWRI) == 0) {
1113 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1114 		reassignbuf(bp);
1115 		atomic_add_int(&numdirtybuffers, 1);
1116 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1117 	}
1118 }
1119 
1120 /*
1121  *	bundirty:
1122  *
1123  *	Clear B_DELWRI for buffer.
1124  *
1125  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1126  *	count.
1127  *
1128  *	The buffer must be on QUEUE_NONE.
1129  */
1130 
1131 void
1132 bundirty(struct buf *bp)
1133 {
1134 
1135 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1136 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1137 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1138 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1139 	BUF_ASSERT_HELD(bp);
1140 
1141 	if (bp->b_flags & B_DELWRI) {
1142 		bp->b_flags &= ~B_DELWRI;
1143 		reassignbuf(bp);
1144 		atomic_subtract_int(&numdirtybuffers, 1);
1145 		numdirtywakeup(lodirtybuffers);
1146 	}
1147 	/*
1148 	 * Since it is now being written, we can clear its deferred write flag.
1149 	 */
1150 	bp->b_flags &= ~B_DEFERRED;
1151 }
1152 
1153 /*
1154  *	bawrite:
1155  *
1156  *	Asynchronous write.  Start output on a buffer, but do not wait for
1157  *	it to complete.  The buffer is released when the output completes.
1158  *
1159  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1160  *	B_INVAL buffers.  Not us.
1161  */
1162 void
1163 bawrite(struct buf *bp)
1164 {
1165 
1166 	bp->b_flags |= B_ASYNC;
1167 	(void) bwrite(bp);
1168 }
1169 
1170 /*
1171  *	bwillwrite:
1172  *
1173  *	Called prior to the locking of any vnodes when we are expecting to
1174  *	write.  We do not want to starve the buffer cache with too many
1175  *	dirty buffers so we block here.  By blocking prior to the locking
1176  *	of any vnodes we attempt to avoid the situation where a locked vnode
1177  *	prevents the various system daemons from flushing related buffers.
1178  */
1179 
1180 void
1181 bwillwrite(void)
1182 {
1183 
1184 	if (numdirtybuffers >= hidirtybuffers) {
1185 		mtx_lock(&nblock);
1186 		while (numdirtybuffers >= hidirtybuffers) {
1187 			bd_wakeup(1);
1188 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1189 			msleep(&needsbuffer, &nblock,
1190 			    (PRIBIO + 4), "flswai", 0);
1191 		}
1192 		mtx_unlock(&nblock);
1193 	}
1194 }
1195 
1196 /*
1197  * Return true if we have too many dirty buffers.
1198  */
1199 int
1200 buf_dirty_count_severe(void)
1201 {
1202 
1203 	return(numdirtybuffers >= hidirtybuffers);
1204 }
1205 
1206 static __noinline int
1207 buf_vm_page_count_severe(void)
1208 {
1209 
1210 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1211 
1212 	return vm_page_count_severe();
1213 }
1214 
1215 /*
1216  *	brelse:
1217  *
1218  *	Release a busy buffer and, if requested, free its resources.  The
1219  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1220  *	to be accessed later as a cache entity or reused for other purposes.
1221  */
1222 void
1223 brelse(struct buf *bp)
1224 {
1225 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1226 	    bp, bp->b_vp, bp->b_flags);
1227 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1228 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1229 
1230 	if (bp->b_flags & B_MANAGED) {
1231 		bqrelse(bp);
1232 		return;
1233 	}
1234 
1235 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1236 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1237 		/*
1238 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1239 		 * pages from being scrapped.  If the error is anything
1240 		 * other than an I/O error (EIO), assume that retrying
1241 		 * is futile.
1242 		 */
1243 		bp->b_ioflags &= ~BIO_ERROR;
1244 		bdirty(bp);
1245 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1246 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1247 		/*
1248 		 * Either a failed I/O or we were asked to free or not
1249 		 * cache the buffer.
1250 		 */
1251 		bp->b_flags |= B_INVAL;
1252 		if (!LIST_EMPTY(&bp->b_dep))
1253 			buf_deallocate(bp);
1254 		if (bp->b_flags & B_DELWRI) {
1255 			atomic_subtract_int(&numdirtybuffers, 1);
1256 			numdirtywakeup(lodirtybuffers);
1257 		}
1258 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1259 		if ((bp->b_flags & B_VMIO) == 0) {
1260 			if (bp->b_bufsize)
1261 				allocbuf(bp, 0);
1262 			if (bp->b_vp)
1263 				brelvp(bp);
1264 		}
1265 	}
1266 
1267 	/*
1268 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1269 	 * is called with B_DELWRI set, the underlying pages may wind up
1270 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1271 	 * because pages associated with a B_DELWRI bp are marked clean.
1272 	 *
1273 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1274 	 * if B_DELWRI is set.
1275 	 *
1276 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1277 	 * on pages to return pages to the VM page queues.
1278 	 */
1279 	if (bp->b_flags & B_DELWRI)
1280 		bp->b_flags &= ~B_RELBUF;
1281 	else if (buf_vm_page_count_severe()) {
1282 		/*
1283 		 * The locking of the BO_LOCK is not necessary since
1284 		 * BKGRDINPROG cannot be set while we hold the buf
1285 		 * lock, it can only be cleared if it is already
1286 		 * pending.
1287 		 */
1288 		if (bp->b_vp) {
1289 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1290 				bp->b_flags |= B_RELBUF;
1291 		} else
1292 			bp->b_flags |= B_RELBUF;
1293 	}
1294 
1295 	/*
1296 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1297 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1298 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1299 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1300 	 *
1301 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1302 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1303 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1304 	 *
1305 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1306 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1307 	 * the commit state and we cannot afford to lose the buffer. If the
1308 	 * buffer has a background write in progress, we need to keep it
1309 	 * around to prevent it from being reconstituted and starting a second
1310 	 * background write.
1311 	 */
1312 	if ((bp->b_flags & B_VMIO)
1313 	    && !(bp->b_vp->v_mount != NULL &&
1314 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1315 		 !vn_isdisk(bp->b_vp, NULL) &&
1316 		 (bp->b_flags & B_DELWRI))
1317 	    ) {
1318 
1319 		int i, j, resid;
1320 		vm_page_t m;
1321 		off_t foff;
1322 		vm_pindex_t poff;
1323 		vm_object_t obj;
1324 
1325 		obj = bp->b_bufobj->bo_object;
1326 
1327 		/*
1328 		 * Get the base offset and length of the buffer.  Note that
1329 		 * in the VMIO case if the buffer block size is not
1330 		 * page-aligned then b_data pointer may not be page-aligned.
1331 		 * But our b_pages[] array *IS* page aligned.
1332 		 *
1333 		 * block sizes less then DEV_BSIZE (usually 512) are not
1334 		 * supported due to the page granularity bits (m->valid,
1335 		 * m->dirty, etc...).
1336 		 *
1337 		 * See man buf(9) for more information
1338 		 */
1339 		resid = bp->b_bufsize;
1340 		foff = bp->b_offset;
1341 		VM_OBJECT_LOCK(obj);
1342 		for (i = 0; i < bp->b_npages; i++) {
1343 			int had_bogus = 0;
1344 
1345 			m = bp->b_pages[i];
1346 
1347 			/*
1348 			 * If we hit a bogus page, fixup *all* the bogus pages
1349 			 * now.
1350 			 */
1351 			if (m == bogus_page) {
1352 				poff = OFF_TO_IDX(bp->b_offset);
1353 				had_bogus = 1;
1354 
1355 				for (j = i; j < bp->b_npages; j++) {
1356 					vm_page_t mtmp;
1357 					mtmp = bp->b_pages[j];
1358 					if (mtmp == bogus_page) {
1359 						mtmp = vm_page_lookup(obj, poff + j);
1360 						if (!mtmp) {
1361 							panic("brelse: page missing\n");
1362 						}
1363 						bp->b_pages[j] = mtmp;
1364 					}
1365 				}
1366 
1367 				if ((bp->b_flags & B_INVAL) == 0) {
1368 					pmap_qenter(
1369 					    trunc_page((vm_offset_t)bp->b_data),
1370 					    bp->b_pages, bp->b_npages);
1371 				}
1372 				m = bp->b_pages[i];
1373 			}
1374 			if ((bp->b_flags & B_NOCACHE) ||
1375 			    (bp->b_ioflags & BIO_ERROR &&
1376 			     bp->b_iocmd == BIO_READ)) {
1377 				int poffset = foff & PAGE_MASK;
1378 				int presid = resid > (PAGE_SIZE - poffset) ?
1379 					(PAGE_SIZE - poffset) : resid;
1380 
1381 				KASSERT(presid >= 0, ("brelse: extra page"));
1382 				vm_page_set_invalid(m, poffset, presid);
1383 				if (had_bogus)
1384 					printf("avoided corruption bug in bogus_page/brelse code\n");
1385 			}
1386 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1387 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1388 		}
1389 		VM_OBJECT_UNLOCK(obj);
1390 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1391 			vfs_vmio_release(bp);
1392 
1393 	} else if (bp->b_flags & B_VMIO) {
1394 
1395 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1396 			vfs_vmio_release(bp);
1397 		}
1398 
1399 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1400 		if (bp->b_bufsize != 0)
1401 			allocbuf(bp, 0);
1402 		if (bp->b_vp != NULL)
1403 			brelvp(bp);
1404 	}
1405 
1406 	if (BUF_LOCKRECURSED(bp)) {
1407 		/* do not release to free list */
1408 		BUF_UNLOCK(bp);
1409 		return;
1410 	}
1411 
1412 	/* enqueue */
1413 	mtx_lock(&bqlock);
1414 	/* Handle delayed bremfree() processing. */
1415 	if (bp->b_flags & B_REMFREE) {
1416 		struct bufobj *bo;
1417 
1418 		bo = bp->b_bufobj;
1419 		if (bo != NULL)
1420 			BO_LOCK(bo);
1421 		bremfreel(bp);
1422 		if (bo != NULL)
1423 			BO_UNLOCK(bo);
1424 	}
1425 	if (bp->b_qindex != QUEUE_NONE)
1426 		panic("brelse: free buffer onto another queue???");
1427 
1428 	/*
1429 	 * If the buffer has junk contents signal it and eventually
1430 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1431 	 * doesn't find it.
1432 	 */
1433 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1434 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1435 		bp->b_flags |= B_INVAL;
1436 	if (bp->b_flags & B_INVAL) {
1437 		if (bp->b_flags & B_DELWRI)
1438 			bundirty(bp);
1439 		if (bp->b_vp)
1440 			brelvp(bp);
1441 	}
1442 
1443 	/* buffers with no memory */
1444 	if (bp->b_bufsize == 0) {
1445 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1446 		if (bp->b_vflags & BV_BKGRDINPROG)
1447 			panic("losing buffer 1");
1448 		if (bp->b_kvasize) {
1449 			bp->b_qindex = QUEUE_EMPTYKVA;
1450 		} else {
1451 			bp->b_qindex = QUEUE_EMPTY;
1452 		}
1453 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1454 	/* buffers with junk contents */
1455 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1456 	    (bp->b_ioflags & BIO_ERROR)) {
1457 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1458 		if (bp->b_vflags & BV_BKGRDINPROG)
1459 			panic("losing buffer 2");
1460 		bp->b_qindex = QUEUE_CLEAN;
1461 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1462 	/* remaining buffers */
1463 	} else {
1464 		if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1465 		    (B_DELWRI|B_NEEDSGIANT))
1466 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1467 		else if (bp->b_flags & B_DELWRI)
1468 			bp->b_qindex = QUEUE_DIRTY;
1469 		else
1470 			bp->b_qindex = QUEUE_CLEAN;
1471 		if (bp->b_flags & B_AGE)
1472 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1473 		else
1474 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1475 	}
1476 	mtx_unlock(&bqlock);
1477 
1478 	/*
1479 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1480 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1481 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1482 	 * if B_INVAL is set ).
1483 	 */
1484 
1485 	if (!(bp->b_flags & B_DELWRI)) {
1486 		struct bufobj *bo;
1487 
1488 		bo = bp->b_bufobj;
1489 		if (bo != NULL)
1490 			BO_LOCK(bo);
1491 		bufcountwakeup(bp);
1492 		if (bo != NULL)
1493 			BO_UNLOCK(bo);
1494 	}
1495 
1496 	/*
1497 	 * Something we can maybe free or reuse
1498 	 */
1499 	if (bp->b_bufsize || bp->b_kvasize)
1500 		bufspacewakeup();
1501 
1502 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1503 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1504 		panic("brelse: not dirty");
1505 	/* unlock */
1506 	BUF_UNLOCK(bp);
1507 }
1508 
1509 /*
1510  * Release a buffer back to the appropriate queue but do not try to free
1511  * it.  The buffer is expected to be used again soon.
1512  *
1513  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1514  * biodone() to requeue an async I/O on completion.  It is also used when
1515  * known good buffers need to be requeued but we think we may need the data
1516  * again soon.
1517  *
1518  * XXX we should be able to leave the B_RELBUF hint set on completion.
1519  */
1520 void
1521 bqrelse(struct buf *bp)
1522 {
1523 	struct bufobj *bo;
1524 
1525 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1526 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1527 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1528 
1529 	if (BUF_LOCKRECURSED(bp)) {
1530 		/* do not release to free list */
1531 		BUF_UNLOCK(bp);
1532 		return;
1533 	}
1534 
1535 	bo = bp->b_bufobj;
1536 	if (bp->b_flags & B_MANAGED) {
1537 		if (bp->b_flags & B_REMFREE) {
1538 			mtx_lock(&bqlock);
1539 			if (bo != NULL)
1540 				BO_LOCK(bo);
1541 			bremfreel(bp);
1542 			if (bo != NULL)
1543 				BO_UNLOCK(bo);
1544 			mtx_unlock(&bqlock);
1545 		}
1546 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1547 		BUF_UNLOCK(bp);
1548 		return;
1549 	}
1550 
1551 	mtx_lock(&bqlock);
1552 	/* Handle delayed bremfree() processing. */
1553 	if (bp->b_flags & B_REMFREE) {
1554 		if (bo != NULL)
1555 			BO_LOCK(bo);
1556 		bremfreel(bp);
1557 		if (bo != NULL)
1558 			BO_UNLOCK(bo);
1559 	}
1560 	if (bp->b_qindex != QUEUE_NONE)
1561 		panic("bqrelse: free buffer onto another queue???");
1562 	/* buffers with stale but valid contents */
1563 	if (bp->b_flags & B_DELWRI) {
1564 		if (bp->b_flags & B_NEEDSGIANT)
1565 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1566 		else
1567 			bp->b_qindex = QUEUE_DIRTY;
1568 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1569 	} else {
1570 		/*
1571 		 * The locking of the BO_LOCK for checking of the
1572 		 * BV_BKGRDINPROG is not necessary since the
1573 		 * BV_BKGRDINPROG cannot be set while we hold the buf
1574 		 * lock, it can only be cleared if it is already
1575 		 * pending.
1576 		 */
1577 		if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1578 			bp->b_qindex = QUEUE_CLEAN;
1579 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1580 			    b_freelist);
1581 		} else {
1582 			/*
1583 			 * We are too low on memory, we have to try to free
1584 			 * the buffer (most importantly: the wired pages
1585 			 * making up its backing store) *now*.
1586 			 */
1587 			mtx_unlock(&bqlock);
1588 			brelse(bp);
1589 			return;
1590 		}
1591 	}
1592 	mtx_unlock(&bqlock);
1593 
1594 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) {
1595 		if (bo != NULL)
1596 			BO_LOCK(bo);
1597 		bufcountwakeup(bp);
1598 		if (bo != NULL)
1599 			BO_UNLOCK(bo);
1600 	}
1601 
1602 	/*
1603 	 * Something we can maybe free or reuse.
1604 	 */
1605 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1606 		bufspacewakeup();
1607 
1608 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1609 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1610 		panic("bqrelse: not dirty");
1611 	/* unlock */
1612 	BUF_UNLOCK(bp);
1613 }
1614 
1615 /* Give pages used by the bp back to the VM system (where possible) */
1616 static void
1617 vfs_vmio_release(struct buf *bp)
1618 {
1619 	int i;
1620 	vm_page_t m;
1621 
1622 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1623 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1624 	for (i = 0; i < bp->b_npages; i++) {
1625 		m = bp->b_pages[i];
1626 		bp->b_pages[i] = NULL;
1627 		/*
1628 		 * In order to keep page LRU ordering consistent, put
1629 		 * everything on the inactive queue.
1630 		 */
1631 		vm_page_lock(m);
1632 		vm_page_unwire(m, 0);
1633 		/*
1634 		 * We don't mess with busy pages, it is
1635 		 * the responsibility of the process that
1636 		 * busied the pages to deal with them.
1637 		 */
1638 		if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 &&
1639 		    m->wire_count == 0) {
1640 			/*
1641 			 * Might as well free the page if we can and it has
1642 			 * no valid data.  We also free the page if the
1643 			 * buffer was used for direct I/O
1644 			 */
1645 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1646 				vm_page_free(m);
1647 			} else if (bp->b_flags & B_DIRECT) {
1648 				vm_page_try_to_free(m);
1649 			} else if (buf_vm_page_count_severe()) {
1650 				vm_page_try_to_cache(m);
1651 			}
1652 		}
1653 		vm_page_unlock(m);
1654 	}
1655 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1656 
1657 	if (bp->b_bufsize) {
1658 		bufspacewakeup();
1659 		bp->b_bufsize = 0;
1660 	}
1661 	bp->b_npages = 0;
1662 	bp->b_flags &= ~B_VMIO;
1663 	if (bp->b_vp)
1664 		brelvp(bp);
1665 }
1666 
1667 /*
1668  * Check to see if a block at a particular lbn is available for a clustered
1669  * write.
1670  */
1671 static int
1672 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1673 {
1674 	struct buf *bpa;
1675 	int match;
1676 
1677 	match = 0;
1678 
1679 	/* If the buf isn't in core skip it */
1680 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1681 		return (0);
1682 
1683 	/* If the buf is busy we don't want to wait for it */
1684 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1685 		return (0);
1686 
1687 	/* Only cluster with valid clusterable delayed write buffers */
1688 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1689 	    (B_DELWRI | B_CLUSTEROK))
1690 		goto done;
1691 
1692 	if (bpa->b_bufsize != size)
1693 		goto done;
1694 
1695 	/*
1696 	 * Check to see if it is in the expected place on disk and that the
1697 	 * block has been mapped.
1698 	 */
1699 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1700 		match = 1;
1701 done:
1702 	BUF_UNLOCK(bpa);
1703 	return (match);
1704 }
1705 
1706 /*
1707  *	vfs_bio_awrite:
1708  *
1709  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1710  *	This is much better then the old way of writing only one buffer at
1711  *	a time.  Note that we may not be presented with the buffers in the
1712  *	correct order, so we search for the cluster in both directions.
1713  */
1714 int
1715 vfs_bio_awrite(struct buf *bp)
1716 {
1717 	struct bufobj *bo;
1718 	int i;
1719 	int j;
1720 	daddr_t lblkno = bp->b_lblkno;
1721 	struct vnode *vp = bp->b_vp;
1722 	int ncl;
1723 	int nwritten;
1724 	int size;
1725 	int maxcl;
1726 
1727 	bo = &vp->v_bufobj;
1728 	/*
1729 	 * right now we support clustered writing only to regular files.  If
1730 	 * we find a clusterable block we could be in the middle of a cluster
1731 	 * rather then at the beginning.
1732 	 */
1733 	if ((vp->v_type == VREG) &&
1734 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1735 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1736 
1737 		size = vp->v_mount->mnt_stat.f_iosize;
1738 		maxcl = MAXPHYS / size;
1739 
1740 		BO_LOCK(bo);
1741 		for (i = 1; i < maxcl; i++)
1742 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1743 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1744 				break;
1745 
1746 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1747 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1748 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1749 				break;
1750 		BO_UNLOCK(bo);
1751 		--j;
1752 		ncl = i + j;
1753 		/*
1754 		 * this is a possible cluster write
1755 		 */
1756 		if (ncl != 1) {
1757 			BUF_UNLOCK(bp);
1758 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1759 			return nwritten;
1760 		}
1761 	}
1762 	bremfree(bp);
1763 	bp->b_flags |= B_ASYNC;
1764 	/*
1765 	 * default (old) behavior, writing out only one block
1766 	 *
1767 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1768 	 */
1769 	nwritten = bp->b_bufsize;
1770 	(void) bwrite(bp);
1771 
1772 	return nwritten;
1773 }
1774 
1775 /*
1776  *	getnewbuf:
1777  *
1778  *	Find and initialize a new buffer header, freeing up existing buffers
1779  *	in the bufqueues as necessary.  The new buffer is returned locked.
1780  *
1781  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1782  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1783  *
1784  *	We block if:
1785  *		We have insufficient buffer headers
1786  *		We have insufficient buffer space
1787  *		buffer_map is too fragmented ( space reservation fails )
1788  *		If we have to flush dirty buffers ( but we try to avoid this )
1789  *
1790  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1791  *	Instead we ask the buf daemon to do it for us.  We attempt to
1792  *	avoid piecemeal wakeups of the pageout daemon.
1793  */
1794 
1795 static struct buf *
1796 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1797     int gbflags)
1798 {
1799 	struct thread *td;
1800 	struct buf *bp;
1801 	struct buf *nbp;
1802 	int defrag = 0;
1803 	int nqindex;
1804 	static int flushingbufs;
1805 
1806 	td = curthread;
1807 	/*
1808 	 * We can't afford to block since we might be holding a vnode lock,
1809 	 * which may prevent system daemons from running.  We deal with
1810 	 * low-memory situations by proactively returning memory and running
1811 	 * async I/O rather then sync I/O.
1812 	 */
1813 	atomic_add_int(&getnewbufcalls, 1);
1814 	atomic_subtract_int(&getnewbufrestarts, 1);
1815 restart:
1816 	atomic_add_int(&getnewbufrestarts, 1);
1817 
1818 	/*
1819 	 * Setup for scan.  If we do not have enough free buffers,
1820 	 * we setup a degenerate case that immediately fails.  Note
1821 	 * that if we are specially marked process, we are allowed to
1822 	 * dip into our reserves.
1823 	 *
1824 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1825 	 *
1826 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1827 	 * However, there are a number of cases (defragging, reusing, ...)
1828 	 * where we cannot backup.
1829 	 */
1830 	mtx_lock(&bqlock);
1831 	nqindex = QUEUE_EMPTYKVA;
1832 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1833 
1834 	if (nbp == NULL) {
1835 		/*
1836 		 * If no EMPTYKVA buffers and we are either
1837 		 * defragging or reusing, locate a CLEAN buffer
1838 		 * to free or reuse.  If bufspace useage is low
1839 		 * skip this step so we can allocate a new buffer.
1840 		 */
1841 		if (defrag || bufspace >= lobufspace) {
1842 			nqindex = QUEUE_CLEAN;
1843 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1844 		}
1845 
1846 		/*
1847 		 * If we could not find or were not allowed to reuse a
1848 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1849 		 * buffer.  We can only use an EMPTY buffer if allocating
1850 		 * its KVA would not otherwise run us out of buffer space.
1851 		 */
1852 		if (nbp == NULL && defrag == 0 &&
1853 		    bufspace + maxsize < hibufspace) {
1854 			nqindex = QUEUE_EMPTY;
1855 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1856 		}
1857 	}
1858 
1859 	/*
1860 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1861 	 * depending.
1862 	 */
1863 
1864 	while ((bp = nbp) != NULL) {
1865 		int qindex = nqindex;
1866 
1867 		/*
1868 		 * Calculate next bp ( we can only use it if we do not block
1869 		 * or do other fancy things ).
1870 		 */
1871 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1872 			switch(qindex) {
1873 			case QUEUE_EMPTY:
1874 				nqindex = QUEUE_EMPTYKVA;
1875 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1876 					break;
1877 				/* FALLTHROUGH */
1878 			case QUEUE_EMPTYKVA:
1879 				nqindex = QUEUE_CLEAN;
1880 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1881 					break;
1882 				/* FALLTHROUGH */
1883 			case QUEUE_CLEAN:
1884 				/*
1885 				 * nbp is NULL.
1886 				 */
1887 				break;
1888 			}
1889 		}
1890 		/*
1891 		 * If we are defragging then we need a buffer with
1892 		 * b_kvasize != 0.  XXX this situation should no longer
1893 		 * occur, if defrag is non-zero the buffer's b_kvasize
1894 		 * should also be non-zero at this point.  XXX
1895 		 */
1896 		if (defrag && bp->b_kvasize == 0) {
1897 			printf("Warning: defrag empty buffer %p\n", bp);
1898 			continue;
1899 		}
1900 
1901 		/*
1902 		 * Start freeing the bp.  This is somewhat involved.  nbp
1903 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1904 		 */
1905 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1906 			continue;
1907 		if (bp->b_vp) {
1908 			BO_LOCK(bp->b_bufobj);
1909 			if (bp->b_vflags & BV_BKGRDINPROG) {
1910 				BO_UNLOCK(bp->b_bufobj);
1911 				BUF_UNLOCK(bp);
1912 				continue;
1913 			}
1914 			BO_UNLOCK(bp->b_bufobj);
1915 		}
1916 		CTR6(KTR_BUF,
1917 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1918 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1919 		    bp->b_kvasize, bp->b_bufsize, qindex);
1920 
1921 		/*
1922 		 * Sanity Checks
1923 		 */
1924 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1925 
1926 		/*
1927 		 * Note: we no longer distinguish between VMIO and non-VMIO
1928 		 * buffers.
1929 		 */
1930 
1931 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1932 
1933 		if (bp->b_bufobj != NULL)
1934 			BO_LOCK(bp->b_bufobj);
1935 		bremfreel(bp);
1936 		if (bp->b_bufobj != NULL)
1937 			BO_UNLOCK(bp->b_bufobj);
1938 		mtx_unlock(&bqlock);
1939 
1940 		if (qindex == QUEUE_CLEAN) {
1941 			if (bp->b_flags & B_VMIO) {
1942 				bp->b_flags &= ~B_ASYNC;
1943 				vfs_vmio_release(bp);
1944 			}
1945 			if (bp->b_vp)
1946 				brelvp(bp);
1947 		}
1948 
1949 		/*
1950 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1951 		 * the scan from this point on.
1952 		 *
1953 		 * Get the rest of the buffer freed up.  b_kva* is still
1954 		 * valid after this operation.
1955 		 */
1956 
1957 		if (bp->b_rcred != NOCRED) {
1958 			crfree(bp->b_rcred);
1959 			bp->b_rcred = NOCRED;
1960 		}
1961 		if (bp->b_wcred != NOCRED) {
1962 			crfree(bp->b_wcred);
1963 			bp->b_wcred = NOCRED;
1964 		}
1965 		if (!LIST_EMPTY(&bp->b_dep))
1966 			buf_deallocate(bp);
1967 		if (bp->b_vflags & BV_BKGRDINPROG)
1968 			panic("losing buffer 3");
1969 		KASSERT(bp->b_vp == NULL,
1970 		    ("bp: %p still has vnode %p.  qindex: %d",
1971 		    bp, bp->b_vp, qindex));
1972 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1973 		   ("bp: %p still on a buffer list. xflags %X",
1974 		    bp, bp->b_xflags));
1975 
1976 		if (bp->b_bufsize)
1977 			allocbuf(bp, 0);
1978 
1979 		bp->b_flags = 0;
1980 		bp->b_ioflags = 0;
1981 		bp->b_xflags = 0;
1982 		KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
1983 		    ("buf %p still counted as free?", bp));
1984 		bp->b_vflags = 0;
1985 		bp->b_vp = NULL;
1986 		bp->b_blkno = bp->b_lblkno = 0;
1987 		bp->b_offset = NOOFFSET;
1988 		bp->b_iodone = 0;
1989 		bp->b_error = 0;
1990 		bp->b_resid = 0;
1991 		bp->b_bcount = 0;
1992 		bp->b_npages = 0;
1993 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1994 		bp->b_bufobj = NULL;
1995 		bp->b_pin_count = 0;
1996 		bp->b_fsprivate1 = NULL;
1997 		bp->b_fsprivate2 = NULL;
1998 		bp->b_fsprivate3 = NULL;
1999 
2000 		LIST_INIT(&bp->b_dep);
2001 
2002 		/*
2003 		 * If we are defragging then free the buffer.
2004 		 */
2005 		if (defrag) {
2006 			bp->b_flags |= B_INVAL;
2007 			bfreekva(bp);
2008 			brelse(bp);
2009 			defrag = 0;
2010 			goto restart;
2011 		}
2012 
2013 		/*
2014 		 * Notify any waiters for the buffer lock about
2015 		 * identity change by freeing the buffer.
2016 		 */
2017 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2018 			bp->b_flags |= B_INVAL;
2019 			bfreekva(bp);
2020 			brelse(bp);
2021 			goto restart;
2022 		}
2023 
2024 		/*
2025 		 * If we are overcomitted then recover the buffer and its
2026 		 * KVM space.  This occurs in rare situations when multiple
2027 		 * processes are blocked in getnewbuf() or allocbuf().
2028 		 */
2029 		if (bufspace >= hibufspace)
2030 			flushingbufs = 1;
2031 		if (flushingbufs && bp->b_kvasize != 0) {
2032 			bp->b_flags |= B_INVAL;
2033 			bfreekva(bp);
2034 			brelse(bp);
2035 			goto restart;
2036 		}
2037 		if (bufspace < lobufspace)
2038 			flushingbufs = 0;
2039 		break;
2040 	}
2041 
2042 	/*
2043 	 * If we exhausted our list, sleep as appropriate.  We may have to
2044 	 * wakeup various daemons and write out some dirty buffers.
2045 	 *
2046 	 * Generally we are sleeping due to insufficient buffer space.
2047 	 */
2048 
2049 	if (bp == NULL) {
2050 		int flags, norunbuf;
2051 		char *waitmsg;
2052 		int fl;
2053 
2054 		if (defrag) {
2055 			flags = VFS_BIO_NEED_BUFSPACE;
2056 			waitmsg = "nbufkv";
2057 		} else if (bufspace >= hibufspace) {
2058 			waitmsg = "nbufbs";
2059 			flags = VFS_BIO_NEED_BUFSPACE;
2060 		} else {
2061 			waitmsg = "newbuf";
2062 			flags = VFS_BIO_NEED_ANY;
2063 		}
2064 		mtx_lock(&nblock);
2065 		needsbuffer |= flags;
2066 		mtx_unlock(&nblock);
2067 		mtx_unlock(&bqlock);
2068 
2069 		bd_speedup();	/* heeeelp */
2070 		if (gbflags & GB_NOWAIT_BD)
2071 			return (NULL);
2072 
2073 		mtx_lock(&nblock);
2074 		while (needsbuffer & flags) {
2075 			if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2076 				mtx_unlock(&nblock);
2077 				/*
2078 				 * getblk() is called with a vnode
2079 				 * locked, and some majority of the
2080 				 * dirty buffers may as well belong to
2081 				 * the vnode. Flushing the buffers
2082 				 * there would make a progress that
2083 				 * cannot be achieved by the
2084 				 * buf_daemon, that cannot lock the
2085 				 * vnode.
2086 				 */
2087 				norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2088 				    (td->td_pflags & TDP_NORUNNINGBUF);
2089 				/* play bufdaemon */
2090 				td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2091 				fl = buf_do_flush(vp);
2092 				td->td_pflags &= norunbuf;
2093 				mtx_lock(&nblock);
2094 				if (fl != 0)
2095 					continue;
2096 				if ((needsbuffer & flags) == 0)
2097 					break;
2098 			}
2099 			if (msleep(&needsbuffer, &nblock,
2100 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2101 				mtx_unlock(&nblock);
2102 				return (NULL);
2103 			}
2104 		}
2105 		mtx_unlock(&nblock);
2106 	} else {
2107 		/*
2108 		 * We finally have a valid bp.  We aren't quite out of the
2109 		 * woods, we still have to reserve kva space.  In order
2110 		 * to keep fragmentation sane we only allocate kva in
2111 		 * BKVASIZE chunks.
2112 		 */
2113 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2114 
2115 		if (maxsize != bp->b_kvasize) {
2116 			vm_offset_t addr = 0;
2117 
2118 			bfreekva(bp);
2119 
2120 			vm_map_lock(buffer_map);
2121 			if (vm_map_findspace(buffer_map,
2122 				vm_map_min(buffer_map), maxsize, &addr)) {
2123 				/*
2124 				 * Uh oh.  Buffer map is to fragmented.  We
2125 				 * must defragment the map.
2126 				 */
2127 				atomic_add_int(&bufdefragcnt, 1);
2128 				vm_map_unlock(buffer_map);
2129 				defrag = 1;
2130 				bp->b_flags |= B_INVAL;
2131 				brelse(bp);
2132 				goto restart;
2133 			}
2134 			if (addr) {
2135 				vm_map_insert(buffer_map, NULL, 0,
2136 					addr, addr + maxsize,
2137 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2138 
2139 				bp->b_kvabase = (caddr_t) addr;
2140 				bp->b_kvasize = maxsize;
2141 				atomic_add_long(&bufspace, bp->b_kvasize);
2142 				atomic_add_int(&bufreusecnt, 1);
2143 			}
2144 			vm_map_unlock(buffer_map);
2145 		}
2146 		bp->b_saveaddr = bp->b_kvabase;
2147 		bp->b_data = bp->b_saveaddr;
2148 	}
2149 	return(bp);
2150 }
2151 
2152 /*
2153  *	buf_daemon:
2154  *
2155  *	buffer flushing daemon.  Buffers are normally flushed by the
2156  *	update daemon but if it cannot keep up this process starts to
2157  *	take the load in an attempt to prevent getnewbuf() from blocking.
2158  */
2159 
2160 static struct kproc_desc buf_kp = {
2161 	"bufdaemon",
2162 	buf_daemon,
2163 	&bufdaemonproc
2164 };
2165 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2166 
2167 static int
2168 buf_do_flush(struct vnode *vp)
2169 {
2170 	int flushed;
2171 
2172 	flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2173 	/* The list empty check here is slightly racy */
2174 	if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2175 		mtx_lock(&Giant);
2176 		flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0);
2177 		mtx_unlock(&Giant);
2178 	}
2179 	if (flushed == 0) {
2180 		/*
2181 		 * Could not find any buffers without rollback
2182 		 * dependencies, so just write the first one
2183 		 * in the hopes of eventually making progress.
2184 		 */
2185 		flushbufqueues(vp, QUEUE_DIRTY, 1);
2186 		if (!TAILQ_EMPTY(
2187 			    &bufqueues[QUEUE_DIRTY_GIANT])) {
2188 			mtx_lock(&Giant);
2189 			flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1);
2190 			mtx_unlock(&Giant);
2191 		}
2192 	}
2193 	return (flushed);
2194 }
2195 
2196 static void
2197 buf_daemon()
2198 {
2199 	int lodirtysave;
2200 
2201 	/*
2202 	 * This process needs to be suspended prior to shutdown sync.
2203 	 */
2204 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2205 	    SHUTDOWN_PRI_LAST);
2206 
2207 	/*
2208 	 * This process is allowed to take the buffer cache to the limit
2209 	 */
2210 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2211 	mtx_lock(&bdlock);
2212 	for (;;) {
2213 		bd_request = 0;
2214 		mtx_unlock(&bdlock);
2215 
2216 		kproc_suspend_check(bufdaemonproc);
2217 		lodirtysave = lodirtybuffers;
2218 		if (bd_speedupreq) {
2219 			lodirtybuffers = numdirtybuffers / 2;
2220 			bd_speedupreq = 0;
2221 		}
2222 		/*
2223 		 * Do the flush.  Limit the amount of in-transit I/O we
2224 		 * allow to build up, otherwise we would completely saturate
2225 		 * the I/O system.  Wakeup any waiting processes before we
2226 		 * normally would so they can run in parallel with our drain.
2227 		 */
2228 		while (numdirtybuffers > lodirtybuffers) {
2229 			if (buf_do_flush(NULL) == 0)
2230 				break;
2231 			kern_yield(PRI_UNCHANGED);
2232 		}
2233 		lodirtybuffers = lodirtysave;
2234 
2235 		/*
2236 		 * Only clear bd_request if we have reached our low water
2237 		 * mark.  The buf_daemon normally waits 1 second and
2238 		 * then incrementally flushes any dirty buffers that have
2239 		 * built up, within reason.
2240 		 *
2241 		 * If we were unable to hit our low water mark and couldn't
2242 		 * find any flushable buffers, we sleep half a second.
2243 		 * Otherwise we loop immediately.
2244 		 */
2245 		mtx_lock(&bdlock);
2246 		if (numdirtybuffers <= lodirtybuffers) {
2247 			/*
2248 			 * We reached our low water mark, reset the
2249 			 * request and sleep until we are needed again.
2250 			 * The sleep is just so the suspend code works.
2251 			 */
2252 			bd_request = 0;
2253 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2254 		} else {
2255 			/*
2256 			 * We couldn't find any flushable dirty buffers but
2257 			 * still have too many dirty buffers, we
2258 			 * have to sleep and try again.  (rare)
2259 			 */
2260 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2261 		}
2262 	}
2263 }
2264 
2265 /*
2266  *	flushbufqueues:
2267  *
2268  *	Try to flush a buffer in the dirty queue.  We must be careful to
2269  *	free up B_INVAL buffers instead of write them, which NFS is
2270  *	particularly sensitive to.
2271  */
2272 static int flushwithdeps = 0;
2273 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2274     0, "Number of buffers flushed with dependecies that require rollbacks");
2275 
2276 static int
2277 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2278 {
2279 	struct buf *sentinel;
2280 	struct vnode *vp;
2281 	struct mount *mp;
2282 	struct buf *bp;
2283 	int hasdeps;
2284 	int flushed;
2285 	int target;
2286 
2287 	if (lvp == NULL) {
2288 		target = numdirtybuffers - lodirtybuffers;
2289 		if (flushdeps && target > 2)
2290 			target /= 2;
2291 	} else
2292 		target = flushbufqtarget;
2293 	flushed = 0;
2294 	bp = NULL;
2295 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2296 	sentinel->b_qindex = QUEUE_SENTINEL;
2297 	mtx_lock(&bqlock);
2298 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2299 	while (flushed != target) {
2300 		bp = TAILQ_NEXT(sentinel, b_freelist);
2301 		if (bp != NULL) {
2302 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2303 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2304 			    b_freelist);
2305 		} else
2306 			break;
2307 		/*
2308 		 * Skip sentinels inserted by other invocations of the
2309 		 * flushbufqueues(), taking care to not reorder them.
2310 		 */
2311 		if (bp->b_qindex == QUEUE_SENTINEL)
2312 			continue;
2313 		/*
2314 		 * Only flush the buffers that belong to the
2315 		 * vnode locked by the curthread.
2316 		 */
2317 		if (lvp != NULL && bp->b_vp != lvp)
2318 			continue;
2319 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2320 			continue;
2321 		if (bp->b_pin_count > 0) {
2322 			BUF_UNLOCK(bp);
2323 			continue;
2324 		}
2325 		BO_LOCK(bp->b_bufobj);
2326 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2327 		    (bp->b_flags & B_DELWRI) == 0) {
2328 			BO_UNLOCK(bp->b_bufobj);
2329 			BUF_UNLOCK(bp);
2330 			continue;
2331 		}
2332 		BO_UNLOCK(bp->b_bufobj);
2333 		if (bp->b_flags & B_INVAL) {
2334 			bremfreel(bp);
2335 			mtx_unlock(&bqlock);
2336 			brelse(bp);
2337 			flushed++;
2338 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2339 			mtx_lock(&bqlock);
2340 			continue;
2341 		}
2342 
2343 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2344 			if (flushdeps == 0) {
2345 				BUF_UNLOCK(bp);
2346 				continue;
2347 			}
2348 			hasdeps = 1;
2349 		} else
2350 			hasdeps = 0;
2351 		/*
2352 		 * We must hold the lock on a vnode before writing
2353 		 * one of its buffers. Otherwise we may confuse, or
2354 		 * in the case of a snapshot vnode, deadlock the
2355 		 * system.
2356 		 *
2357 		 * The lock order here is the reverse of the normal
2358 		 * of vnode followed by buf lock.  This is ok because
2359 		 * the NOWAIT will prevent deadlock.
2360 		 */
2361 		vp = bp->b_vp;
2362 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2363 			BUF_UNLOCK(bp);
2364 			continue;
2365 		}
2366 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2367 			mtx_unlock(&bqlock);
2368 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2369 			    bp, bp->b_vp, bp->b_flags);
2370 			if (curproc == bufdaemonproc)
2371 				vfs_bio_awrite(bp);
2372 			else {
2373 				bremfree(bp);
2374 				bwrite(bp);
2375 				notbufdflashes++;
2376 			}
2377 			vn_finished_write(mp);
2378 			VOP_UNLOCK(vp, 0);
2379 			flushwithdeps += hasdeps;
2380 			flushed++;
2381 
2382 			/*
2383 			 * Sleeping on runningbufspace while holding
2384 			 * vnode lock leads to deadlock.
2385 			 */
2386 			if (curproc == bufdaemonproc)
2387 				waitrunningbufspace();
2388 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2389 			mtx_lock(&bqlock);
2390 			continue;
2391 		}
2392 		vn_finished_write(mp);
2393 		BUF_UNLOCK(bp);
2394 	}
2395 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2396 	mtx_unlock(&bqlock);
2397 	free(sentinel, M_TEMP);
2398 	return (flushed);
2399 }
2400 
2401 /*
2402  * Check to see if a block is currently memory resident.
2403  */
2404 struct buf *
2405 incore(struct bufobj *bo, daddr_t blkno)
2406 {
2407 	struct buf *bp;
2408 
2409 	BO_LOCK(bo);
2410 	bp = gbincore(bo, blkno);
2411 	BO_UNLOCK(bo);
2412 	return (bp);
2413 }
2414 
2415 /*
2416  * Returns true if no I/O is needed to access the
2417  * associated VM object.  This is like incore except
2418  * it also hunts around in the VM system for the data.
2419  */
2420 
2421 static int
2422 inmem(struct vnode * vp, daddr_t blkno)
2423 {
2424 	vm_object_t obj;
2425 	vm_offset_t toff, tinc, size;
2426 	vm_page_t m;
2427 	vm_ooffset_t off;
2428 
2429 	ASSERT_VOP_LOCKED(vp, "inmem");
2430 
2431 	if (incore(&vp->v_bufobj, blkno))
2432 		return 1;
2433 	if (vp->v_mount == NULL)
2434 		return 0;
2435 	obj = vp->v_object;
2436 	if (obj == NULL)
2437 		return (0);
2438 
2439 	size = PAGE_SIZE;
2440 	if (size > vp->v_mount->mnt_stat.f_iosize)
2441 		size = vp->v_mount->mnt_stat.f_iosize;
2442 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2443 
2444 	VM_OBJECT_LOCK(obj);
2445 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2446 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2447 		if (!m)
2448 			goto notinmem;
2449 		tinc = size;
2450 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2451 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2452 		if (vm_page_is_valid(m,
2453 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2454 			goto notinmem;
2455 	}
2456 	VM_OBJECT_UNLOCK(obj);
2457 	return 1;
2458 
2459 notinmem:
2460 	VM_OBJECT_UNLOCK(obj);
2461 	return (0);
2462 }
2463 
2464 /*
2465  * Set the dirty range for a buffer based on the status of the dirty
2466  * bits in the pages comprising the buffer.  The range is limited
2467  * to the size of the buffer.
2468  *
2469  * Tell the VM system that the pages associated with this buffer
2470  * are clean.  This is used for delayed writes where the data is
2471  * going to go to disk eventually without additional VM intevention.
2472  *
2473  * Note that while we only really need to clean through to b_bcount, we
2474  * just go ahead and clean through to b_bufsize.
2475  */
2476 static void
2477 vfs_clean_pages_dirty_buf(struct buf *bp)
2478 {
2479 	vm_ooffset_t foff, noff, eoff;
2480 	vm_page_t m;
2481 	int i;
2482 
2483 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2484 		return;
2485 
2486 	foff = bp->b_offset;
2487 	KASSERT(bp->b_offset != NOOFFSET,
2488 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2489 
2490 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2491 	vfs_drain_busy_pages(bp);
2492 	vfs_setdirty_locked_object(bp);
2493 	for (i = 0; i < bp->b_npages; i++) {
2494 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2495 		eoff = noff;
2496 		if (eoff > bp->b_offset + bp->b_bufsize)
2497 			eoff = bp->b_offset + bp->b_bufsize;
2498 		m = bp->b_pages[i];
2499 		vfs_page_set_validclean(bp, foff, m);
2500 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2501 		foff = noff;
2502 	}
2503 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2504 }
2505 
2506 static void
2507 vfs_setdirty_locked_object(struct buf *bp)
2508 {
2509 	vm_object_t object;
2510 	int i;
2511 
2512 	object = bp->b_bufobj->bo_object;
2513 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2514 
2515 	/*
2516 	 * We qualify the scan for modified pages on whether the
2517 	 * object has been flushed yet.
2518 	 */
2519 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2520 		vm_offset_t boffset;
2521 		vm_offset_t eoffset;
2522 
2523 		/*
2524 		 * test the pages to see if they have been modified directly
2525 		 * by users through the VM system.
2526 		 */
2527 		for (i = 0; i < bp->b_npages; i++)
2528 			vm_page_test_dirty(bp->b_pages[i]);
2529 
2530 		/*
2531 		 * Calculate the encompassing dirty range, boffset and eoffset,
2532 		 * (eoffset - boffset) bytes.
2533 		 */
2534 
2535 		for (i = 0; i < bp->b_npages; i++) {
2536 			if (bp->b_pages[i]->dirty)
2537 				break;
2538 		}
2539 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2540 
2541 		for (i = bp->b_npages - 1; i >= 0; --i) {
2542 			if (bp->b_pages[i]->dirty) {
2543 				break;
2544 			}
2545 		}
2546 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2547 
2548 		/*
2549 		 * Fit it to the buffer.
2550 		 */
2551 
2552 		if (eoffset > bp->b_bcount)
2553 			eoffset = bp->b_bcount;
2554 
2555 		/*
2556 		 * If we have a good dirty range, merge with the existing
2557 		 * dirty range.
2558 		 */
2559 
2560 		if (boffset < eoffset) {
2561 			if (bp->b_dirtyoff > boffset)
2562 				bp->b_dirtyoff = boffset;
2563 			if (bp->b_dirtyend < eoffset)
2564 				bp->b_dirtyend = eoffset;
2565 		}
2566 	}
2567 }
2568 
2569 /*
2570  *	getblk:
2571  *
2572  *	Get a block given a specified block and offset into a file/device.
2573  *	The buffers B_DONE bit will be cleared on return, making it almost
2574  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2575  *	return.  The caller should clear B_INVAL prior to initiating a
2576  *	READ.
2577  *
2578  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2579  *	an existing buffer.
2580  *
2581  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2582  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2583  *	and then cleared based on the backing VM.  If the previous buffer is
2584  *	non-0-sized but invalid, B_CACHE will be cleared.
2585  *
2586  *	If getblk() must create a new buffer, the new buffer is returned with
2587  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2588  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2589  *	backing VM.
2590  *
2591  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2592  *	B_CACHE bit is clear.
2593  *
2594  *	What this means, basically, is that the caller should use B_CACHE to
2595  *	determine whether the buffer is fully valid or not and should clear
2596  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2597  *	the buffer by loading its data area with something, the caller needs
2598  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2599  *	the caller should set B_CACHE ( as an optimization ), else the caller
2600  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2601  *	a write attempt or if it was a successfull read.  If the caller
2602  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2603  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2604  */
2605 struct buf *
2606 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2607     int flags)
2608 {
2609 	struct buf *bp;
2610 	struct bufobj *bo;
2611 	int error;
2612 
2613 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2614 	ASSERT_VOP_LOCKED(vp, "getblk");
2615 	if (size > MAXBSIZE)
2616 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2617 
2618 	bo = &vp->v_bufobj;
2619 loop:
2620 	/*
2621 	 * Block if we are low on buffers.   Certain processes are allowed
2622 	 * to completely exhaust the buffer cache.
2623          *
2624          * If this check ever becomes a bottleneck it may be better to
2625          * move it into the else, when gbincore() fails.  At the moment
2626          * it isn't a problem.
2627 	 *
2628 	 * XXX remove if 0 sections (clean this up after its proven)
2629          */
2630 	if (numfreebuffers == 0) {
2631 		if (TD_IS_IDLETHREAD(curthread))
2632 			return NULL;
2633 		mtx_lock(&nblock);
2634 		needsbuffer |= VFS_BIO_NEED_ANY;
2635 		mtx_unlock(&nblock);
2636 	}
2637 
2638 	BO_LOCK(bo);
2639 	bp = gbincore(bo, blkno);
2640 	if (bp != NULL) {
2641 		int lockflags;
2642 		/*
2643 		 * Buffer is in-core.  If the buffer is not busy nor managed,
2644 		 * it must be on a queue.
2645 		 */
2646 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2647 
2648 		if (flags & GB_LOCK_NOWAIT)
2649 			lockflags |= LK_NOWAIT;
2650 
2651 		error = BUF_TIMELOCK(bp, lockflags,
2652 		    BO_MTX(bo), "getblk", slpflag, slptimeo);
2653 
2654 		/*
2655 		 * If we slept and got the lock we have to restart in case
2656 		 * the buffer changed identities.
2657 		 */
2658 		if (error == ENOLCK)
2659 			goto loop;
2660 		/* We timed out or were interrupted. */
2661 		else if (error)
2662 			return (NULL);
2663 
2664 		/*
2665 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2666 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2667 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2668 		 * backing VM cache.
2669 		 */
2670 		if (bp->b_flags & B_INVAL)
2671 			bp->b_flags &= ~B_CACHE;
2672 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2673 			bp->b_flags |= B_CACHE;
2674 		if (bp->b_flags & B_MANAGED)
2675 			MPASS(bp->b_qindex == QUEUE_NONE);
2676 		else {
2677 			BO_LOCK(bo);
2678 			bremfree(bp);
2679 			BO_UNLOCK(bo);
2680 		}
2681 
2682 		/*
2683 		 * check for size inconsistancies for non-VMIO case.
2684 		 */
2685 
2686 		if (bp->b_bcount != size) {
2687 			if ((bp->b_flags & B_VMIO) == 0 ||
2688 			    (size > bp->b_kvasize)) {
2689 				if (bp->b_flags & B_DELWRI) {
2690 					/*
2691 					 * If buffer is pinned and caller does
2692 					 * not want sleep  waiting for it to be
2693 					 * unpinned, bail out
2694 					 * */
2695 					if (bp->b_pin_count > 0) {
2696 						if (flags & GB_LOCK_NOWAIT) {
2697 							bqrelse(bp);
2698 							return (NULL);
2699 						} else {
2700 							bunpin_wait(bp);
2701 						}
2702 					}
2703 					bp->b_flags |= B_NOCACHE;
2704 					bwrite(bp);
2705 				} else {
2706 					if (LIST_EMPTY(&bp->b_dep)) {
2707 						bp->b_flags |= B_RELBUF;
2708 						brelse(bp);
2709 					} else {
2710 						bp->b_flags |= B_NOCACHE;
2711 						bwrite(bp);
2712 					}
2713 				}
2714 				goto loop;
2715 			}
2716 		}
2717 
2718 		/*
2719 		 * If the size is inconsistant in the VMIO case, we can resize
2720 		 * the buffer.  This might lead to B_CACHE getting set or
2721 		 * cleared.  If the size has not changed, B_CACHE remains
2722 		 * unchanged from its previous state.
2723 		 */
2724 
2725 		if (bp->b_bcount != size)
2726 			allocbuf(bp, size);
2727 
2728 		KASSERT(bp->b_offset != NOOFFSET,
2729 		    ("getblk: no buffer offset"));
2730 
2731 		/*
2732 		 * A buffer with B_DELWRI set and B_CACHE clear must
2733 		 * be committed before we can return the buffer in
2734 		 * order to prevent the caller from issuing a read
2735 		 * ( due to B_CACHE not being set ) and overwriting
2736 		 * it.
2737 		 *
2738 		 * Most callers, including NFS and FFS, need this to
2739 		 * operate properly either because they assume they
2740 		 * can issue a read if B_CACHE is not set, or because
2741 		 * ( for example ) an uncached B_DELWRI might loop due
2742 		 * to softupdates re-dirtying the buffer.  In the latter
2743 		 * case, B_CACHE is set after the first write completes,
2744 		 * preventing further loops.
2745 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2746 		 * above while extending the buffer, we cannot allow the
2747 		 * buffer to remain with B_CACHE set after the write
2748 		 * completes or it will represent a corrupt state.  To
2749 		 * deal with this we set B_NOCACHE to scrap the buffer
2750 		 * after the write.
2751 		 *
2752 		 * We might be able to do something fancy, like setting
2753 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2754 		 * so the below call doesn't set B_CACHE, but that gets real
2755 		 * confusing.  This is much easier.
2756 		 */
2757 
2758 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2759 			bp->b_flags |= B_NOCACHE;
2760 			bwrite(bp);
2761 			goto loop;
2762 		}
2763 		bp->b_flags &= ~B_DONE;
2764 	} else {
2765 		int bsize, maxsize, vmio;
2766 		off_t offset;
2767 
2768 		/*
2769 		 * Buffer is not in-core, create new buffer.  The buffer
2770 		 * returned by getnewbuf() is locked.  Note that the returned
2771 		 * buffer is also considered valid (not marked B_INVAL).
2772 		 */
2773 		BO_UNLOCK(bo);
2774 		/*
2775 		 * If the user does not want us to create the buffer, bail out
2776 		 * here.
2777 		 */
2778 		if (flags & GB_NOCREAT)
2779 			return NULL;
2780 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
2781 		offset = blkno * bsize;
2782 		vmio = vp->v_object != NULL;
2783 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2784 		maxsize = imax(maxsize, bsize);
2785 
2786 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2787 		if (bp == NULL) {
2788 			if (slpflag || slptimeo)
2789 				return NULL;
2790 			goto loop;
2791 		}
2792 
2793 		/*
2794 		 * This code is used to make sure that a buffer is not
2795 		 * created while the getnewbuf routine is blocked.
2796 		 * This can be a problem whether the vnode is locked or not.
2797 		 * If the buffer is created out from under us, we have to
2798 		 * throw away the one we just created.
2799 		 *
2800 		 * Note: this must occur before we associate the buffer
2801 		 * with the vp especially considering limitations in
2802 		 * the splay tree implementation when dealing with duplicate
2803 		 * lblkno's.
2804 		 */
2805 		BO_LOCK(bo);
2806 		if (gbincore(bo, blkno)) {
2807 			BO_UNLOCK(bo);
2808 			bp->b_flags |= B_INVAL;
2809 			brelse(bp);
2810 			goto loop;
2811 		}
2812 
2813 		/*
2814 		 * Insert the buffer into the hash, so that it can
2815 		 * be found by incore.
2816 		 */
2817 		bp->b_blkno = bp->b_lblkno = blkno;
2818 		bp->b_offset = offset;
2819 		bgetvp(vp, bp);
2820 		BO_UNLOCK(bo);
2821 
2822 		/*
2823 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2824 		 * buffer size starts out as 0, B_CACHE will be set by
2825 		 * allocbuf() for the VMIO case prior to it testing the
2826 		 * backing store for validity.
2827 		 */
2828 
2829 		if (vmio) {
2830 			bp->b_flags |= B_VMIO;
2831 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2832 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2833 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2834 		} else {
2835 			bp->b_flags &= ~B_VMIO;
2836 			KASSERT(bp->b_bufobj->bo_object == NULL,
2837 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2838 			    bp, bp->b_bufobj->bo_object));
2839 		}
2840 
2841 		allocbuf(bp, size);
2842 		bp->b_flags &= ~B_DONE;
2843 	}
2844 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2845 	BUF_ASSERT_HELD(bp);
2846 	KASSERT(bp->b_bufobj == bo,
2847 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2848 	return (bp);
2849 }
2850 
2851 /*
2852  * Get an empty, disassociated buffer of given size.  The buffer is initially
2853  * set to B_INVAL.
2854  */
2855 struct buf *
2856 geteblk(int size, int flags)
2857 {
2858 	struct buf *bp;
2859 	int maxsize;
2860 
2861 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2862 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2863 		if ((flags & GB_NOWAIT_BD) &&
2864 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
2865 			return (NULL);
2866 	}
2867 	allocbuf(bp, size);
2868 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2869 	BUF_ASSERT_HELD(bp);
2870 	return (bp);
2871 }
2872 
2873 
2874 /*
2875  * This code constitutes the buffer memory from either anonymous system
2876  * memory (in the case of non-VMIO operations) or from an associated
2877  * VM object (in the case of VMIO operations).  This code is able to
2878  * resize a buffer up or down.
2879  *
2880  * Note that this code is tricky, and has many complications to resolve
2881  * deadlock or inconsistant data situations.  Tread lightly!!!
2882  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2883  * the caller.  Calling this code willy nilly can result in the loss of data.
2884  *
2885  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2886  * B_CACHE for the non-VMIO case.
2887  */
2888 
2889 int
2890 allocbuf(struct buf *bp, int size)
2891 {
2892 	int newbsize, mbsize;
2893 	int i;
2894 
2895 	BUF_ASSERT_HELD(bp);
2896 
2897 	if (bp->b_kvasize < size)
2898 		panic("allocbuf: buffer too small");
2899 
2900 	if ((bp->b_flags & B_VMIO) == 0) {
2901 		caddr_t origbuf;
2902 		int origbufsize;
2903 		/*
2904 		 * Just get anonymous memory from the kernel.  Don't
2905 		 * mess with B_CACHE.
2906 		 */
2907 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2908 		if (bp->b_flags & B_MALLOC)
2909 			newbsize = mbsize;
2910 		else
2911 			newbsize = round_page(size);
2912 
2913 		if (newbsize < bp->b_bufsize) {
2914 			/*
2915 			 * malloced buffers are not shrunk
2916 			 */
2917 			if (bp->b_flags & B_MALLOC) {
2918 				if (newbsize) {
2919 					bp->b_bcount = size;
2920 				} else {
2921 					free(bp->b_data, M_BIOBUF);
2922 					if (bp->b_bufsize) {
2923 						atomic_subtract_long(
2924 						    &bufmallocspace,
2925 						    bp->b_bufsize);
2926 						bufspacewakeup();
2927 						bp->b_bufsize = 0;
2928 					}
2929 					bp->b_saveaddr = bp->b_kvabase;
2930 					bp->b_data = bp->b_saveaddr;
2931 					bp->b_bcount = 0;
2932 					bp->b_flags &= ~B_MALLOC;
2933 				}
2934 				return 1;
2935 			}
2936 			vm_hold_free_pages(bp, newbsize);
2937 		} else if (newbsize > bp->b_bufsize) {
2938 			/*
2939 			 * We only use malloced memory on the first allocation.
2940 			 * and revert to page-allocated memory when the buffer
2941 			 * grows.
2942 			 */
2943 			/*
2944 			 * There is a potential smp race here that could lead
2945 			 * to bufmallocspace slightly passing the max.  It
2946 			 * is probably extremely rare and not worth worrying
2947 			 * over.
2948 			 */
2949 			if ( (bufmallocspace < maxbufmallocspace) &&
2950 				(bp->b_bufsize == 0) &&
2951 				(mbsize <= PAGE_SIZE/2)) {
2952 
2953 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2954 				bp->b_bufsize = mbsize;
2955 				bp->b_bcount = size;
2956 				bp->b_flags |= B_MALLOC;
2957 				atomic_add_long(&bufmallocspace, mbsize);
2958 				return 1;
2959 			}
2960 			origbuf = NULL;
2961 			origbufsize = 0;
2962 			/*
2963 			 * If the buffer is growing on its other-than-first allocation,
2964 			 * then we revert to the page-allocation scheme.
2965 			 */
2966 			if (bp->b_flags & B_MALLOC) {
2967 				origbuf = bp->b_data;
2968 				origbufsize = bp->b_bufsize;
2969 				bp->b_data = bp->b_kvabase;
2970 				if (bp->b_bufsize) {
2971 					atomic_subtract_long(&bufmallocspace,
2972 					    bp->b_bufsize);
2973 					bufspacewakeup();
2974 					bp->b_bufsize = 0;
2975 				}
2976 				bp->b_flags &= ~B_MALLOC;
2977 				newbsize = round_page(newbsize);
2978 			}
2979 			vm_hold_load_pages(
2980 			    bp,
2981 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2982 			    (vm_offset_t) bp->b_data + newbsize);
2983 			if (origbuf) {
2984 				bcopy(origbuf, bp->b_data, origbufsize);
2985 				free(origbuf, M_BIOBUF);
2986 			}
2987 		}
2988 	} else {
2989 		int desiredpages;
2990 
2991 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2992 		desiredpages = (size == 0) ? 0 :
2993 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2994 
2995 		if (bp->b_flags & B_MALLOC)
2996 			panic("allocbuf: VMIO buffer can't be malloced");
2997 		/*
2998 		 * Set B_CACHE initially if buffer is 0 length or will become
2999 		 * 0-length.
3000 		 */
3001 		if (size == 0 || bp->b_bufsize == 0)
3002 			bp->b_flags |= B_CACHE;
3003 
3004 		if (newbsize < bp->b_bufsize) {
3005 			/*
3006 			 * DEV_BSIZE aligned new buffer size is less then the
3007 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3008 			 * if we have to remove any pages.
3009 			 */
3010 			if (desiredpages < bp->b_npages) {
3011 				vm_page_t m;
3012 
3013 				pmap_qremove((vm_offset_t)trunc_page(
3014 				    (vm_offset_t)bp->b_data) +
3015 				    (desiredpages << PAGE_SHIFT),
3016 				    (bp->b_npages - desiredpages));
3017 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3018 				for (i = desiredpages; i < bp->b_npages; i++) {
3019 					/*
3020 					 * the page is not freed here -- it
3021 					 * is the responsibility of
3022 					 * vnode_pager_setsize
3023 					 */
3024 					m = bp->b_pages[i];
3025 					KASSERT(m != bogus_page,
3026 					    ("allocbuf: bogus page found"));
3027 					while (vm_page_sleep_if_busy(m, TRUE,
3028 					    "biodep"))
3029 						continue;
3030 
3031 					bp->b_pages[i] = NULL;
3032 					vm_page_lock(m);
3033 					vm_page_unwire(m, 0);
3034 					vm_page_unlock(m);
3035 				}
3036 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3037 				bp->b_npages = desiredpages;
3038 			}
3039 		} else if (size > bp->b_bcount) {
3040 			/*
3041 			 * We are growing the buffer, possibly in a
3042 			 * byte-granular fashion.
3043 			 */
3044 			vm_object_t obj;
3045 			vm_offset_t toff;
3046 			vm_offset_t tinc;
3047 
3048 			/*
3049 			 * Step 1, bring in the VM pages from the object,
3050 			 * allocating them if necessary.  We must clear
3051 			 * B_CACHE if these pages are not valid for the
3052 			 * range covered by the buffer.
3053 			 */
3054 
3055 			obj = bp->b_bufobj->bo_object;
3056 
3057 			VM_OBJECT_LOCK(obj);
3058 			while (bp->b_npages < desiredpages) {
3059 				vm_page_t m;
3060 
3061 				/*
3062 				 * We must allocate system pages since blocking
3063 				 * here could interfere with paging I/O, no
3064 				 * matter which process we are.
3065 				 *
3066 				 * We can only test VPO_BUSY here.  Blocking on
3067 				 * m->busy might lead to a deadlock:
3068 				 *  vm_fault->getpages->cluster_read->allocbuf
3069 				 * Thus, we specify VM_ALLOC_IGN_SBUSY.
3070 				 */
3071 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3072 				    bp->b_npages, VM_ALLOC_NOBUSY |
3073 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3074 				    VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY |
3075 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3076 				if (m->valid == 0)
3077 					bp->b_flags &= ~B_CACHE;
3078 				bp->b_pages[bp->b_npages] = m;
3079 				++bp->b_npages;
3080 			}
3081 
3082 			/*
3083 			 * Step 2.  We've loaded the pages into the buffer,
3084 			 * we have to figure out if we can still have B_CACHE
3085 			 * set.  Note that B_CACHE is set according to the
3086 			 * byte-granular range ( bcount and size ), new the
3087 			 * aligned range ( newbsize ).
3088 			 *
3089 			 * The VM test is against m->valid, which is DEV_BSIZE
3090 			 * aligned.  Needless to say, the validity of the data
3091 			 * needs to also be DEV_BSIZE aligned.  Note that this
3092 			 * fails with NFS if the server or some other client
3093 			 * extends the file's EOF.  If our buffer is resized,
3094 			 * B_CACHE may remain set! XXX
3095 			 */
3096 
3097 			toff = bp->b_bcount;
3098 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3099 
3100 			while ((bp->b_flags & B_CACHE) && toff < size) {
3101 				vm_pindex_t pi;
3102 
3103 				if (tinc > (size - toff))
3104 					tinc = size - toff;
3105 
3106 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3107 				    PAGE_SHIFT;
3108 
3109 				vfs_buf_test_cache(
3110 				    bp,
3111 				    bp->b_offset,
3112 				    toff,
3113 				    tinc,
3114 				    bp->b_pages[pi]
3115 				);
3116 				toff += tinc;
3117 				tinc = PAGE_SIZE;
3118 			}
3119 			VM_OBJECT_UNLOCK(obj);
3120 
3121 			/*
3122 			 * Step 3, fixup the KVM pmap.  Remember that
3123 			 * bp->b_data is relative to bp->b_offset, but
3124 			 * bp->b_offset may be offset into the first page.
3125 			 */
3126 
3127 			bp->b_data = (caddr_t)
3128 			    trunc_page((vm_offset_t)bp->b_data);
3129 			pmap_qenter(
3130 			    (vm_offset_t)bp->b_data,
3131 			    bp->b_pages,
3132 			    bp->b_npages
3133 			);
3134 
3135 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3136 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
3137 		}
3138 	}
3139 	if (newbsize < bp->b_bufsize)
3140 		bufspacewakeup();
3141 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3142 	bp->b_bcount = size;		/* requested buffer size	*/
3143 	return 1;
3144 }
3145 
3146 void
3147 biodone(struct bio *bp)
3148 {
3149 	struct mtx *mtxp;
3150 	void (*done)(struct bio *);
3151 
3152 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3153 	mtx_lock(mtxp);
3154 	bp->bio_flags |= BIO_DONE;
3155 	done = bp->bio_done;
3156 	if (done == NULL)
3157 		wakeup(bp);
3158 	mtx_unlock(mtxp);
3159 	if (done != NULL)
3160 		done(bp);
3161 }
3162 
3163 /*
3164  * Wait for a BIO to finish.
3165  *
3166  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3167  * case is not yet clear.
3168  */
3169 int
3170 biowait(struct bio *bp, const char *wchan)
3171 {
3172 	struct mtx *mtxp;
3173 
3174 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3175 	mtx_lock(mtxp);
3176 	while ((bp->bio_flags & BIO_DONE) == 0)
3177 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3178 	mtx_unlock(mtxp);
3179 	if (bp->bio_error != 0)
3180 		return (bp->bio_error);
3181 	if (!(bp->bio_flags & BIO_ERROR))
3182 		return (0);
3183 	return (EIO);
3184 }
3185 
3186 void
3187 biofinish(struct bio *bp, struct devstat *stat, int error)
3188 {
3189 
3190 	if (error) {
3191 		bp->bio_error = error;
3192 		bp->bio_flags |= BIO_ERROR;
3193 	}
3194 	if (stat != NULL)
3195 		devstat_end_transaction_bio(stat, bp);
3196 	biodone(bp);
3197 }
3198 
3199 /*
3200  *	bufwait:
3201  *
3202  *	Wait for buffer I/O completion, returning error status.  The buffer
3203  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3204  *	error and cleared.
3205  */
3206 int
3207 bufwait(struct buf *bp)
3208 {
3209 	if (bp->b_iocmd == BIO_READ)
3210 		bwait(bp, PRIBIO, "biord");
3211 	else
3212 		bwait(bp, PRIBIO, "biowr");
3213 	if (bp->b_flags & B_EINTR) {
3214 		bp->b_flags &= ~B_EINTR;
3215 		return (EINTR);
3216 	}
3217 	if (bp->b_ioflags & BIO_ERROR) {
3218 		return (bp->b_error ? bp->b_error : EIO);
3219 	} else {
3220 		return (0);
3221 	}
3222 }
3223 
3224  /*
3225   * Call back function from struct bio back up to struct buf.
3226   */
3227 static void
3228 bufdonebio(struct bio *bip)
3229 {
3230 	struct buf *bp;
3231 
3232 	bp = bip->bio_caller2;
3233 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3234 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3235 	bp->b_ioflags = bip->bio_flags;
3236 	bp->b_error = bip->bio_error;
3237 	if (bp->b_error)
3238 		bp->b_ioflags |= BIO_ERROR;
3239 	bufdone(bp);
3240 	g_destroy_bio(bip);
3241 }
3242 
3243 void
3244 dev_strategy(struct cdev *dev, struct buf *bp)
3245 {
3246 	struct cdevsw *csw;
3247 	struct bio *bip;
3248 	int ref;
3249 
3250 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3251 		panic("b_iocmd botch");
3252 	for (;;) {
3253 		bip = g_new_bio();
3254 		if (bip != NULL)
3255 			break;
3256 		/* Try again later */
3257 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3258 	}
3259 	bip->bio_cmd = bp->b_iocmd;
3260 	bip->bio_offset = bp->b_iooffset;
3261 	bip->bio_length = bp->b_bcount;
3262 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3263 	bip->bio_data = bp->b_data;
3264 	bip->bio_done = bufdonebio;
3265 	bip->bio_caller2 = bp;
3266 	bip->bio_dev = dev;
3267 	KASSERT(dev->si_refcount > 0,
3268 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3269 	    devtoname(dev)));
3270 	csw = dev_refthread(dev, &ref);
3271 	if (csw == NULL) {
3272 		g_destroy_bio(bip);
3273 		bp->b_error = ENXIO;
3274 		bp->b_ioflags = BIO_ERROR;
3275 		bufdone(bp);
3276 		return;
3277 	}
3278 	(*csw->d_strategy)(bip);
3279 	dev_relthread(dev, ref);
3280 }
3281 
3282 /*
3283  *	bufdone:
3284  *
3285  *	Finish I/O on a buffer, optionally calling a completion function.
3286  *	This is usually called from an interrupt so process blocking is
3287  *	not allowed.
3288  *
3289  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3290  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3291  *	assuming B_INVAL is clear.
3292  *
3293  *	For the VMIO case, we set B_CACHE if the op was a read and no
3294  *	read error occured, or if the op was a write.  B_CACHE is never
3295  *	set if the buffer is invalid or otherwise uncacheable.
3296  *
3297  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3298  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3299  *	in the biodone routine.
3300  */
3301 void
3302 bufdone(struct buf *bp)
3303 {
3304 	struct bufobj *dropobj;
3305 	void    (*biodone)(struct buf *);
3306 
3307 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3308 	dropobj = NULL;
3309 
3310 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3311 	BUF_ASSERT_HELD(bp);
3312 
3313 	runningbufwakeup(bp);
3314 	if (bp->b_iocmd == BIO_WRITE)
3315 		dropobj = bp->b_bufobj;
3316 	/* call optional completion function if requested */
3317 	if (bp->b_iodone != NULL) {
3318 		biodone = bp->b_iodone;
3319 		bp->b_iodone = NULL;
3320 		(*biodone) (bp);
3321 		if (dropobj)
3322 			bufobj_wdrop(dropobj);
3323 		return;
3324 	}
3325 
3326 	bufdone_finish(bp);
3327 
3328 	if (dropobj)
3329 		bufobj_wdrop(dropobj);
3330 }
3331 
3332 void
3333 bufdone_finish(struct buf *bp)
3334 {
3335 	BUF_ASSERT_HELD(bp);
3336 
3337 	if (!LIST_EMPTY(&bp->b_dep))
3338 		buf_complete(bp);
3339 
3340 	if (bp->b_flags & B_VMIO) {
3341 		vm_ooffset_t foff;
3342 		vm_page_t m;
3343 		vm_object_t obj;
3344 		struct vnode *vp;
3345 		int bogus, i, iosize;
3346 
3347 		obj = bp->b_bufobj->bo_object;
3348 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3349 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3350 		    obj->paging_in_progress, bp->b_npages));
3351 
3352 		vp = bp->b_vp;
3353 		KASSERT(vp->v_holdcnt > 0,
3354 		    ("biodone_finish: vnode %p has zero hold count", vp));
3355 		KASSERT(vp->v_object != NULL,
3356 		    ("biodone_finish: vnode %p has no vm_object", vp));
3357 
3358 		foff = bp->b_offset;
3359 		KASSERT(bp->b_offset != NOOFFSET,
3360 		    ("biodone_finish: bp %p has no buffer offset", bp));
3361 
3362 		/*
3363 		 * Set B_CACHE if the op was a normal read and no error
3364 		 * occured.  B_CACHE is set for writes in the b*write()
3365 		 * routines.
3366 		 */
3367 		iosize = bp->b_bcount - bp->b_resid;
3368 		if (bp->b_iocmd == BIO_READ &&
3369 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3370 		    !(bp->b_ioflags & BIO_ERROR)) {
3371 			bp->b_flags |= B_CACHE;
3372 		}
3373 		bogus = 0;
3374 		VM_OBJECT_LOCK(obj);
3375 		for (i = 0; i < bp->b_npages; i++) {
3376 			int bogusflag = 0;
3377 			int resid;
3378 
3379 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3380 			if (resid > iosize)
3381 				resid = iosize;
3382 
3383 			/*
3384 			 * cleanup bogus pages, restoring the originals
3385 			 */
3386 			m = bp->b_pages[i];
3387 			if (m == bogus_page) {
3388 				bogus = bogusflag = 1;
3389 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3390 				if (m == NULL)
3391 					panic("biodone: page disappeared!");
3392 				bp->b_pages[i] = m;
3393 			}
3394 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3395 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3396 			    (intmax_t)foff, (uintmax_t)m->pindex));
3397 
3398 			/*
3399 			 * In the write case, the valid and clean bits are
3400 			 * already changed correctly ( see bdwrite() ), so we
3401 			 * only need to do this here in the read case.
3402 			 */
3403 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3404 				KASSERT((m->dirty & vm_page_bits(foff &
3405 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3406 				    " page %p has unexpected dirty bits", m));
3407 				vfs_page_set_valid(bp, foff, m);
3408 			}
3409 
3410 			vm_page_io_finish(m);
3411 			vm_object_pip_subtract(obj, 1);
3412 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3413 			iosize -= resid;
3414 		}
3415 		vm_object_pip_wakeupn(obj, 0);
3416 		VM_OBJECT_UNLOCK(obj);
3417 		if (bogus)
3418 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3419 			    bp->b_pages, bp->b_npages);
3420 	}
3421 
3422 	/*
3423 	 * For asynchronous completions, release the buffer now. The brelse
3424 	 * will do a wakeup there if necessary - so no need to do a wakeup
3425 	 * here in the async case. The sync case always needs to do a wakeup.
3426 	 */
3427 
3428 	if (bp->b_flags & B_ASYNC) {
3429 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3430 			brelse(bp);
3431 		else
3432 			bqrelse(bp);
3433 	} else
3434 		bdone(bp);
3435 }
3436 
3437 /*
3438  * This routine is called in lieu of iodone in the case of
3439  * incomplete I/O.  This keeps the busy status for pages
3440  * consistant.
3441  */
3442 void
3443 vfs_unbusy_pages(struct buf *bp)
3444 {
3445 	int i;
3446 	vm_object_t obj;
3447 	vm_page_t m;
3448 
3449 	runningbufwakeup(bp);
3450 	if (!(bp->b_flags & B_VMIO))
3451 		return;
3452 
3453 	obj = bp->b_bufobj->bo_object;
3454 	VM_OBJECT_LOCK(obj);
3455 	for (i = 0; i < bp->b_npages; i++) {
3456 		m = bp->b_pages[i];
3457 		if (m == bogus_page) {
3458 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3459 			if (!m)
3460 				panic("vfs_unbusy_pages: page missing\n");
3461 			bp->b_pages[i] = m;
3462 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3463 			    bp->b_pages, bp->b_npages);
3464 		}
3465 		vm_object_pip_subtract(obj, 1);
3466 		vm_page_io_finish(m);
3467 	}
3468 	vm_object_pip_wakeupn(obj, 0);
3469 	VM_OBJECT_UNLOCK(obj);
3470 }
3471 
3472 /*
3473  * vfs_page_set_valid:
3474  *
3475  *	Set the valid bits in a page based on the supplied offset.   The
3476  *	range is restricted to the buffer's size.
3477  *
3478  *	This routine is typically called after a read completes.
3479  */
3480 static void
3481 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3482 {
3483 	vm_ooffset_t eoff;
3484 
3485 	/*
3486 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3487 	 * page boundary and eoff is not greater than the end of the buffer.
3488 	 * The end of the buffer, in this case, is our file EOF, not the
3489 	 * allocation size of the buffer.
3490 	 */
3491 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3492 	if (eoff > bp->b_offset + bp->b_bcount)
3493 		eoff = bp->b_offset + bp->b_bcount;
3494 
3495 	/*
3496 	 * Set valid range.  This is typically the entire buffer and thus the
3497 	 * entire page.
3498 	 */
3499 	if (eoff > off)
3500 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3501 }
3502 
3503 /*
3504  * vfs_page_set_validclean:
3505  *
3506  *	Set the valid bits and clear the dirty bits in a page based on the
3507  *	supplied offset.   The range is restricted to the buffer's size.
3508  */
3509 static void
3510 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3511 {
3512 	vm_ooffset_t soff, eoff;
3513 
3514 	/*
3515 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3516 	 * page boundry or cross the end of the buffer.  The end of the
3517 	 * buffer, in this case, is our file EOF, not the allocation size
3518 	 * of the buffer.
3519 	 */
3520 	soff = off;
3521 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3522 	if (eoff > bp->b_offset + bp->b_bcount)
3523 		eoff = bp->b_offset + bp->b_bcount;
3524 
3525 	/*
3526 	 * Set valid range.  This is typically the entire buffer and thus the
3527 	 * entire page.
3528 	 */
3529 	if (eoff > soff) {
3530 		vm_page_set_validclean(
3531 		    m,
3532 		   (vm_offset_t) (soff & PAGE_MASK),
3533 		   (vm_offset_t) (eoff - soff)
3534 		);
3535 	}
3536 }
3537 
3538 /*
3539  * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3540  * any page is busy, drain the flag.
3541  */
3542 static void
3543 vfs_drain_busy_pages(struct buf *bp)
3544 {
3545 	vm_page_t m;
3546 	int i, last_busied;
3547 
3548 	VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED);
3549 	last_busied = 0;
3550 	for (i = 0; i < bp->b_npages; i++) {
3551 		m = bp->b_pages[i];
3552 		if ((m->oflags & VPO_BUSY) != 0) {
3553 			for (; last_busied < i; last_busied++)
3554 				vm_page_busy(bp->b_pages[last_busied]);
3555 			while ((m->oflags & VPO_BUSY) != 0)
3556 				vm_page_sleep(m, "vbpage");
3557 		}
3558 	}
3559 	for (i = 0; i < last_busied; i++)
3560 		vm_page_wakeup(bp->b_pages[i]);
3561 }
3562 
3563 /*
3564  * This routine is called before a device strategy routine.
3565  * It is used to tell the VM system that paging I/O is in
3566  * progress, and treat the pages associated with the buffer
3567  * almost as being VPO_BUSY.  Also the object paging_in_progress
3568  * flag is handled to make sure that the object doesn't become
3569  * inconsistant.
3570  *
3571  * Since I/O has not been initiated yet, certain buffer flags
3572  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3573  * and should be ignored.
3574  */
3575 void
3576 vfs_busy_pages(struct buf *bp, int clear_modify)
3577 {
3578 	int i, bogus;
3579 	vm_object_t obj;
3580 	vm_ooffset_t foff;
3581 	vm_page_t m;
3582 
3583 	if (!(bp->b_flags & B_VMIO))
3584 		return;
3585 
3586 	obj = bp->b_bufobj->bo_object;
3587 	foff = bp->b_offset;
3588 	KASSERT(bp->b_offset != NOOFFSET,
3589 	    ("vfs_busy_pages: no buffer offset"));
3590 	VM_OBJECT_LOCK(obj);
3591 	vfs_drain_busy_pages(bp);
3592 	if (bp->b_bufsize != 0)
3593 		vfs_setdirty_locked_object(bp);
3594 	bogus = 0;
3595 	for (i = 0; i < bp->b_npages; i++) {
3596 		m = bp->b_pages[i];
3597 
3598 		if ((bp->b_flags & B_CLUSTER) == 0) {
3599 			vm_object_pip_add(obj, 1);
3600 			vm_page_io_start(m);
3601 		}
3602 		/*
3603 		 * When readying a buffer for a read ( i.e
3604 		 * clear_modify == 0 ), it is important to do
3605 		 * bogus_page replacement for valid pages in
3606 		 * partially instantiated buffers.  Partially
3607 		 * instantiated buffers can, in turn, occur when
3608 		 * reconstituting a buffer from its VM backing store
3609 		 * base.  We only have to do this if B_CACHE is
3610 		 * clear ( which causes the I/O to occur in the
3611 		 * first place ).  The replacement prevents the read
3612 		 * I/O from overwriting potentially dirty VM-backed
3613 		 * pages.  XXX bogus page replacement is, uh, bogus.
3614 		 * It may not work properly with small-block devices.
3615 		 * We need to find a better way.
3616 		 */
3617 		if (clear_modify) {
3618 			pmap_remove_write(m);
3619 			vfs_page_set_validclean(bp, foff, m);
3620 		} else if (m->valid == VM_PAGE_BITS_ALL &&
3621 		    (bp->b_flags & B_CACHE) == 0) {
3622 			bp->b_pages[i] = bogus_page;
3623 			bogus++;
3624 		}
3625 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3626 	}
3627 	VM_OBJECT_UNLOCK(obj);
3628 	if (bogus)
3629 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3630 		    bp->b_pages, bp->b_npages);
3631 }
3632 
3633 /*
3634  *	vfs_bio_set_valid:
3635  *
3636  *	Set the range within the buffer to valid.  The range is
3637  *	relative to the beginning of the buffer, b_offset.  Note that
3638  *	b_offset itself may be offset from the beginning of the first
3639  *	page.
3640  */
3641 void
3642 vfs_bio_set_valid(struct buf *bp, int base, int size)
3643 {
3644 	int i, n;
3645 	vm_page_t m;
3646 
3647 	if (!(bp->b_flags & B_VMIO))
3648 		return;
3649 
3650 	/*
3651 	 * Fixup base to be relative to beginning of first page.
3652 	 * Set initial n to be the maximum number of bytes in the
3653 	 * first page that can be validated.
3654 	 */
3655 	base += (bp->b_offset & PAGE_MASK);
3656 	n = PAGE_SIZE - (base & PAGE_MASK);
3657 
3658 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3659 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3660 		m = bp->b_pages[i];
3661 		if (n > size)
3662 			n = size;
3663 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
3664 		base += n;
3665 		size -= n;
3666 		n = PAGE_SIZE;
3667 	}
3668 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3669 }
3670 
3671 /*
3672  *	vfs_bio_clrbuf:
3673  *
3674  *	If the specified buffer is a non-VMIO buffer, clear the entire
3675  *	buffer.  If the specified buffer is a VMIO buffer, clear and
3676  *	validate only the previously invalid portions of the buffer.
3677  *	This routine essentially fakes an I/O, so we need to clear
3678  *	BIO_ERROR and B_INVAL.
3679  *
3680  *	Note that while we only theoretically need to clear through b_bcount,
3681  *	we go ahead and clear through b_bufsize.
3682  */
3683 void
3684 vfs_bio_clrbuf(struct buf *bp)
3685 {
3686 	int i, j, mask;
3687 	caddr_t sa, ea;
3688 
3689 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3690 		clrbuf(bp);
3691 		return;
3692 	}
3693 	bp->b_flags &= ~B_INVAL;
3694 	bp->b_ioflags &= ~BIO_ERROR;
3695 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3696 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3697 	    (bp->b_offset & PAGE_MASK) == 0) {
3698 		if (bp->b_pages[0] == bogus_page)
3699 			goto unlock;
3700 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3701 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3702 		if ((bp->b_pages[0]->valid & mask) == mask)
3703 			goto unlock;
3704 		if ((bp->b_pages[0]->valid & mask) == 0) {
3705 			bzero(bp->b_data, bp->b_bufsize);
3706 			bp->b_pages[0]->valid |= mask;
3707 			goto unlock;
3708 		}
3709 	}
3710 	ea = sa = bp->b_data;
3711 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3712 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3713 		ea = (caddr_t)(vm_offset_t)ulmin(
3714 		    (u_long)(vm_offset_t)ea,
3715 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3716 		if (bp->b_pages[i] == bogus_page)
3717 			continue;
3718 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3719 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3720 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3721 		if ((bp->b_pages[i]->valid & mask) == mask)
3722 			continue;
3723 		if ((bp->b_pages[i]->valid & mask) == 0)
3724 			bzero(sa, ea - sa);
3725 		else {
3726 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3727 				if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3728 					bzero(sa, DEV_BSIZE);
3729 			}
3730 		}
3731 		bp->b_pages[i]->valid |= mask;
3732 	}
3733 unlock:
3734 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3735 	bp->b_resid = 0;
3736 }
3737 
3738 /*
3739  * vm_hold_load_pages and vm_hold_free_pages get pages into
3740  * a buffers address space.  The pages are anonymous and are
3741  * not associated with a file object.
3742  */
3743 static void
3744 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3745 {
3746 	vm_offset_t pg;
3747 	vm_page_t p;
3748 	int index;
3749 
3750 	to = round_page(to);
3751 	from = round_page(from);
3752 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3753 
3754 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3755 tryagain:
3756 		/*
3757 		 * note: must allocate system pages since blocking here
3758 		 * could interfere with paging I/O, no matter which
3759 		 * process we are.
3760 		 */
3761 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
3762 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
3763 		if (p == NULL) {
3764 			VM_WAIT;
3765 			goto tryagain;
3766 		}
3767 		pmap_qenter(pg, &p, 1);
3768 		bp->b_pages[index] = p;
3769 	}
3770 	bp->b_npages = index;
3771 }
3772 
3773 /* Return pages associated with this buf to the vm system */
3774 static void
3775 vm_hold_free_pages(struct buf *bp, int newbsize)
3776 {
3777 	vm_offset_t from;
3778 	vm_page_t p;
3779 	int index, newnpages;
3780 
3781 	from = round_page((vm_offset_t)bp->b_data + newbsize);
3782 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3783 	if (bp->b_npages > newnpages)
3784 		pmap_qremove(from, bp->b_npages - newnpages);
3785 	for (index = newnpages; index < bp->b_npages; index++) {
3786 		p = bp->b_pages[index];
3787 		bp->b_pages[index] = NULL;
3788 		if (p->busy != 0)
3789 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3790 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
3791 		p->wire_count--;
3792 		vm_page_free(p);
3793 		atomic_subtract_int(&cnt.v_wire_count, 1);
3794 	}
3795 	bp->b_npages = newnpages;
3796 }
3797 
3798 /*
3799  * Map an IO request into kernel virtual address space.
3800  *
3801  * All requests are (re)mapped into kernel VA space.
3802  * Notice that we use b_bufsize for the size of the buffer
3803  * to be mapped.  b_bcount might be modified by the driver.
3804  *
3805  * Note that even if the caller determines that the address space should
3806  * be valid, a race or a smaller-file mapped into a larger space may
3807  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3808  * check the return value.
3809  */
3810 int
3811 vmapbuf(struct buf *bp)
3812 {
3813 	caddr_t kva;
3814 	vm_prot_t prot;
3815 	int pidx;
3816 
3817 	if (bp->b_bufsize < 0)
3818 		return (-1);
3819 	prot = VM_PROT_READ;
3820 	if (bp->b_iocmd == BIO_READ)
3821 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3822 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
3823 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
3824 	    btoc(MAXPHYS))) < 0)
3825 		return (-1);
3826 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3827 
3828 	kva = bp->b_saveaddr;
3829 	bp->b_npages = pidx;
3830 	bp->b_saveaddr = bp->b_data;
3831 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3832 	return(0);
3833 }
3834 
3835 /*
3836  * Free the io map PTEs associated with this IO operation.
3837  * We also invalidate the TLB entries and restore the original b_addr.
3838  */
3839 void
3840 vunmapbuf(struct buf *bp)
3841 {
3842 	int npages;
3843 
3844 	npages = bp->b_npages;
3845 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3846 	vm_page_unhold_pages(bp->b_pages, npages);
3847 
3848 	bp->b_data = bp->b_saveaddr;
3849 }
3850 
3851 void
3852 bdone(struct buf *bp)
3853 {
3854 	struct mtx *mtxp;
3855 
3856 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3857 	mtx_lock(mtxp);
3858 	bp->b_flags |= B_DONE;
3859 	wakeup(bp);
3860 	mtx_unlock(mtxp);
3861 }
3862 
3863 void
3864 bwait(struct buf *bp, u_char pri, const char *wchan)
3865 {
3866 	struct mtx *mtxp;
3867 
3868 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3869 	mtx_lock(mtxp);
3870 	while ((bp->b_flags & B_DONE) == 0)
3871 		msleep(bp, mtxp, pri, wchan, 0);
3872 	mtx_unlock(mtxp);
3873 }
3874 
3875 int
3876 bufsync(struct bufobj *bo, int waitfor)
3877 {
3878 
3879 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3880 }
3881 
3882 void
3883 bufstrategy(struct bufobj *bo, struct buf *bp)
3884 {
3885 	int i = 0;
3886 	struct vnode *vp;
3887 
3888 	vp = bp->b_vp;
3889 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3890 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3891 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3892 	i = VOP_STRATEGY(vp, bp);
3893 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3894 }
3895 
3896 void
3897 bufobj_wrefl(struct bufobj *bo)
3898 {
3899 
3900 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3901 	ASSERT_BO_LOCKED(bo);
3902 	bo->bo_numoutput++;
3903 }
3904 
3905 void
3906 bufobj_wref(struct bufobj *bo)
3907 {
3908 
3909 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3910 	BO_LOCK(bo);
3911 	bo->bo_numoutput++;
3912 	BO_UNLOCK(bo);
3913 }
3914 
3915 void
3916 bufobj_wdrop(struct bufobj *bo)
3917 {
3918 
3919 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3920 	BO_LOCK(bo);
3921 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3922 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3923 		bo->bo_flag &= ~BO_WWAIT;
3924 		wakeup(&bo->bo_numoutput);
3925 	}
3926 	BO_UNLOCK(bo);
3927 }
3928 
3929 int
3930 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3931 {
3932 	int error;
3933 
3934 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3935 	ASSERT_BO_LOCKED(bo);
3936 	error = 0;
3937 	while (bo->bo_numoutput) {
3938 		bo->bo_flag |= BO_WWAIT;
3939 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3940 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3941 		if (error)
3942 			break;
3943 	}
3944 	return (error);
3945 }
3946 
3947 void
3948 bpin(struct buf *bp)
3949 {
3950 	struct mtx *mtxp;
3951 
3952 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3953 	mtx_lock(mtxp);
3954 	bp->b_pin_count++;
3955 	mtx_unlock(mtxp);
3956 }
3957 
3958 void
3959 bunpin(struct buf *bp)
3960 {
3961 	struct mtx *mtxp;
3962 
3963 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3964 	mtx_lock(mtxp);
3965 	if (--bp->b_pin_count == 0)
3966 		wakeup(bp);
3967 	mtx_unlock(mtxp);
3968 }
3969 
3970 void
3971 bunpin_wait(struct buf *bp)
3972 {
3973 	struct mtx *mtxp;
3974 
3975 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3976 	mtx_lock(mtxp);
3977 	while (bp->b_pin_count > 0)
3978 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
3979 	mtx_unlock(mtxp);
3980 }
3981 
3982 #include "opt_ddb.h"
3983 #ifdef DDB
3984 #include <ddb/ddb.h>
3985 
3986 /* DDB command to show buffer data */
3987 DB_SHOW_COMMAND(buffer, db_show_buffer)
3988 {
3989 	/* get args */
3990 	struct buf *bp = (struct buf *)addr;
3991 
3992 	if (!have_addr) {
3993 		db_printf("usage: show buffer <addr>\n");
3994 		return;
3995 	}
3996 
3997 	db_printf("buf at %p\n", bp);
3998 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
3999 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4000 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4001 	db_printf(
4002 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4003 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4004 	    "b_dep = %p\n",
4005 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4006 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4007 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4008 	if (bp->b_npages) {
4009 		int i;
4010 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4011 		for (i = 0; i < bp->b_npages; i++) {
4012 			vm_page_t m;
4013 			m = bp->b_pages[i];
4014 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4015 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4016 			if ((i + 1) < bp->b_npages)
4017 				db_printf(",");
4018 		}
4019 		db_printf("\n");
4020 	}
4021 	db_printf(" ");
4022 	BUF_LOCKPRINTINFO(bp);
4023 }
4024 
4025 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4026 {
4027 	struct buf *bp;
4028 	int i;
4029 
4030 	for (i = 0; i < nbuf; i++) {
4031 		bp = &buf[i];
4032 		if (BUF_ISLOCKED(bp)) {
4033 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4034 			db_printf("\n");
4035 		}
4036 	}
4037 }
4038 
4039 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4040 {
4041 	struct vnode *vp;
4042 	struct buf *bp;
4043 
4044 	if (!have_addr) {
4045 		db_printf("usage: show vnodebufs <addr>\n");
4046 		return;
4047 	}
4048 	vp = (struct vnode *)addr;
4049 	db_printf("Clean buffers:\n");
4050 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4051 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4052 		db_printf("\n");
4053 	}
4054 	db_printf("Dirty buffers:\n");
4055 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4056 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4057 		db_printf("\n");
4058 	}
4059 }
4060 
4061 DB_COMMAND(countfreebufs, db_coundfreebufs)
4062 {
4063 	struct buf *bp;
4064 	int i, used = 0, nfree = 0;
4065 
4066 	if (have_addr) {
4067 		db_printf("usage: countfreebufs\n");
4068 		return;
4069 	}
4070 
4071 	for (i = 0; i < nbuf; i++) {
4072 		bp = &buf[i];
4073 		if ((bp->b_vflags & BV_INFREECNT) != 0)
4074 			nfree++;
4075 		else
4076 			used++;
4077 	}
4078 
4079 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4080 	    nfree + used);
4081 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4082 }
4083 #endif /* DDB */
4084