xref: /freebsd/sys/kern/vfs_bio.c (revision aa64588d28258aef88cc33b8043112e8856948d0)
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, vm_offset_t from,
99 		vm_offset_t to);
100 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
101 		vm_offset_t to);
102 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
103 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
104 		vm_page_t m);
105 static void vfs_drain_busy_pages(struct buf *bp);
106 static void vfs_clean_pages_dirty_buf(struct buf *bp);
107 static void vfs_setdirty_locked_object(struct buf *bp);
108 static void vfs_vmio_release(struct buf *bp);
109 static int vfs_bio_clcheck(struct vnode *vp, int size,
110 		daddr_t lblkno, daddr_t blkno);
111 static int buf_do_flush(struct vnode *vp);
112 static int flushbufqueues(struct vnode *, int, int);
113 static void buf_daemon(void);
114 static void bremfreel(struct buf *bp);
115 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
116     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
117 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
118 #endif
119 
120 int vmiodirenable = TRUE;
121 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
122     "Use the VM system for directory writes");
123 long runningbufspace;
124 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
125     "Amount of presently outstanding async buffer io");
126 static long bufspace;
127 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
128     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
129 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
130     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
131 #else
132 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
133     "Virtual memory used for buffers");
134 #endif
135 static long maxbufspace;
136 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
137     "Maximum allowed value of bufspace (including buf_daemon)");
138 static long bufmallocspace;
139 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
140     "Amount of malloced memory for buffers");
141 static long maxbufmallocspace;
142 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
143     "Maximum amount of malloced memory for buffers");
144 static long lobufspace;
145 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
146     "Minimum amount of buffers we want to have");
147 long hibufspace;
148 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
149     "Maximum allowed value of bufspace (excluding buf_daemon)");
150 static int bufreusecnt;
151 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
152     "Number of times we have reused a buffer");
153 static int buffreekvacnt;
154 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
155     "Number of times we have freed the KVA space from some buffer");
156 static int bufdefragcnt;
157 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
158     "Number of times we have had to repeat buffer allocation to defragment");
159 static long lorunningspace;
160 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
161     "Minimum preferred space used for in-progress I/O");
162 static long hirunningspace;
163 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
164     "Maximum amount of space to use for in-progress I/O");
165 int dirtybufferflushes;
166 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
167     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
168 int bdwriteskip;
169 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
170     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
171 int altbufferflushes;
172 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
173     0, "Number of fsync flushes to limit dirty buffers");
174 static int recursiveflushes;
175 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
176     0, "Number of flushes skipped due to being recursive");
177 static int numdirtybuffers;
178 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
179     "Number of buffers that are dirty (has unwritten changes) at the moment");
180 static int lodirtybuffers;
181 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
182     "How many buffers we want to have free before bufdaemon can sleep");
183 static int hidirtybuffers;
184 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
185     "When the number of dirty buffers is considered severe");
186 int dirtybufthresh;
187 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
188     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
189 static int numfreebuffers;
190 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
191     "Number of free buffers");
192 static int lofreebuffers;
193 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
194    "XXX Unused");
195 static int hifreebuffers;
196 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
197    "XXX Complicatedly unused");
198 static int getnewbufcalls;
199 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
200    "Number of calls to getnewbuf");
201 static int getnewbufrestarts;
202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
203     "Number of times getnewbuf has had to restart a buffer aquisition");
204 static int flushbufqtarget = 100;
205 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
206     "Amount of work to do in flushbufqueues when helping bufdaemon");
207 static long notbufdflashes;
208 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, &notbufdflashes, 0,
209     "Number of dirty buffer flushes done by the bufdaemon helpers");
210 
211 /*
212  * Wakeup point for bufdaemon, as well as indicator of whether it is already
213  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
214  * is idling.
215  */
216 static int bd_request;
217 
218 /*
219  * Request for the buf daemon to write more buffers than is indicated by
220  * lodirtybuf.  This may be necessary to push out excess dependencies or
221  * defragment the address space where a simple count of the number of dirty
222  * buffers is insufficient to characterize the demand for flushing them.
223  */
224 static int bd_speedupreq;
225 
226 /*
227  * This lock synchronizes access to bd_request.
228  */
229 static struct mtx bdlock;
230 
231 /*
232  * bogus page -- for I/O to/from partially complete buffers
233  * this is a temporary solution to the problem, but it is not
234  * really that bad.  it would be better to split the buffer
235  * for input in the case of buffers partially already in memory,
236  * but the code is intricate enough already.
237  */
238 vm_page_t bogus_page;
239 
240 /*
241  * Synchronization (sleep/wakeup) variable for active buffer space requests.
242  * Set when wait starts, cleared prior to wakeup().
243  * Used in runningbufwakeup() and waitrunningbufspace().
244  */
245 static int runningbufreq;
246 
247 /*
248  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
249  * waitrunningbufspace().
250  */
251 static struct mtx rbreqlock;
252 
253 /*
254  * Synchronization (sleep/wakeup) variable for buffer requests.
255  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
256  * by and/or.
257  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
258  * getnewbuf(), and getblk().
259  */
260 static int needsbuffer;
261 
262 /*
263  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
264  */
265 static struct mtx nblock;
266 
267 /*
268  * Definitions for the buffer free lists.
269  */
270 #define BUFFER_QUEUES	6	/* number of free buffer queues */
271 
272 #define QUEUE_NONE	0	/* on no queue */
273 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
274 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
275 #define QUEUE_DIRTY_GIANT 3	/* B_DELWRI buffers that need giant */
276 #define QUEUE_EMPTYKVA	4	/* empty buffer headers w/KVA assignment */
277 #define QUEUE_EMPTY	5	/* empty buffer headers */
278 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
279 
280 /* Queues for free buffers with various properties */
281 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
282 
283 /* Lock for the bufqueues */
284 static struct mtx bqlock;
285 
286 /*
287  * Single global constant for BUF_WMESG, to avoid getting multiple references.
288  * buf_wmesg is referred from macros.
289  */
290 const char *buf_wmesg = BUF_WMESG;
291 
292 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
293 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
294 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
295 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
296 
297 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
298     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
299 static int
300 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
301 {
302 	long lvalue;
303 	int ivalue;
304 
305 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
306 		return (sysctl_handle_long(oidp, arg1, arg2, req));
307 	lvalue = *(long *)arg1;
308 	if (lvalue > INT_MAX)
309 		/* On overflow, still write out a long to trigger ENOMEM. */
310 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
311 	ivalue = lvalue;
312 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
313 }
314 #endif
315 
316 #ifdef DIRECTIO
317 extern void ffs_rawread_setup(void);
318 #endif /* DIRECTIO */
319 /*
320  *	numdirtywakeup:
321  *
322  *	If someone is blocked due to there being too many dirty buffers,
323  *	and numdirtybuffers is now reasonable, wake them up.
324  */
325 
326 static __inline void
327 numdirtywakeup(int level)
328 {
329 
330 	if (numdirtybuffers <= level) {
331 		mtx_lock(&nblock);
332 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
333 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
334 			wakeup(&needsbuffer);
335 		}
336 		mtx_unlock(&nblock);
337 	}
338 }
339 
340 /*
341  *	bufspacewakeup:
342  *
343  *	Called when buffer space is potentially available for recovery.
344  *	getnewbuf() will block on this flag when it is unable to free
345  *	sufficient buffer space.  Buffer space becomes recoverable when
346  *	bp's get placed back in the queues.
347  */
348 
349 static __inline void
350 bufspacewakeup(void)
351 {
352 
353 	/*
354 	 * If someone is waiting for BUF space, wake them up.  Even
355 	 * though we haven't freed the kva space yet, the waiting
356 	 * process will be able to now.
357 	 */
358 	mtx_lock(&nblock);
359 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
360 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
361 		wakeup(&needsbuffer);
362 	}
363 	mtx_unlock(&nblock);
364 }
365 
366 /*
367  * runningbufwakeup() - in-progress I/O accounting.
368  *
369  */
370 void
371 runningbufwakeup(struct buf *bp)
372 {
373 
374 	if (bp->b_runningbufspace) {
375 		atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
376 		bp->b_runningbufspace = 0;
377 		mtx_lock(&rbreqlock);
378 		if (runningbufreq && runningbufspace <= lorunningspace) {
379 			runningbufreq = 0;
380 			wakeup(&runningbufreq);
381 		}
382 		mtx_unlock(&rbreqlock);
383 	}
384 }
385 
386 /*
387  *	bufcountwakeup:
388  *
389  *	Called when a buffer has been added to one of the free queues to
390  *	account for the buffer and to wakeup anyone waiting for free buffers.
391  *	This typically occurs when large amounts of metadata are being handled
392  *	by the buffer cache ( else buffer space runs out first, usually ).
393  */
394 
395 static __inline void
396 bufcountwakeup(void)
397 {
398 
399 	atomic_add_int(&numfreebuffers, 1);
400 	mtx_lock(&nblock);
401 	if (needsbuffer) {
402 		needsbuffer &= ~VFS_BIO_NEED_ANY;
403 		if (numfreebuffers >= hifreebuffers)
404 			needsbuffer &= ~VFS_BIO_NEED_FREE;
405 		wakeup(&needsbuffer);
406 	}
407 	mtx_unlock(&nblock);
408 }
409 
410 /*
411  *	waitrunningbufspace()
412  *
413  *	runningbufspace is a measure of the amount of I/O currently
414  *	running.  This routine is used in async-write situations to
415  *	prevent creating huge backups of pending writes to a device.
416  *	Only asynchronous writes are governed by this function.
417  *
418  *	Reads will adjust runningbufspace, but will not block based on it.
419  *	The read load has a side effect of reducing the allowed write load.
420  *
421  *	This does NOT turn an async write into a sync write.  It waits
422  *	for earlier writes to complete and generally returns before the
423  *	caller's write has reached the device.
424  */
425 void
426 waitrunningbufspace(void)
427 {
428 
429 	mtx_lock(&rbreqlock);
430 	while (runningbufspace > hirunningspace) {
431 		++runningbufreq;
432 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
433 	}
434 	mtx_unlock(&rbreqlock);
435 }
436 
437 
438 /*
439  *	vfs_buf_test_cache:
440  *
441  *	Called when a buffer is extended.  This function clears the B_CACHE
442  *	bit if the newly extended portion of the buffer does not contain
443  *	valid data.
444  */
445 static __inline
446 void
447 vfs_buf_test_cache(struct buf *bp,
448 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
449 		  vm_page_t m)
450 {
451 
452 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
453 	if (bp->b_flags & B_CACHE) {
454 		int base = (foff + off) & PAGE_MASK;
455 		if (vm_page_is_valid(m, base, size) == 0)
456 			bp->b_flags &= ~B_CACHE;
457 	}
458 }
459 
460 /* Wake up the buffer daemon if necessary */
461 static __inline
462 void
463 bd_wakeup(int dirtybuflevel)
464 {
465 
466 	mtx_lock(&bdlock);
467 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
468 		bd_request = 1;
469 		wakeup(&bd_request);
470 	}
471 	mtx_unlock(&bdlock);
472 }
473 
474 /*
475  * bd_speedup - speedup the buffer cache flushing code
476  */
477 
478 void
479 bd_speedup(void)
480 {
481 	int needwake;
482 
483 	mtx_lock(&bdlock);
484 	needwake = 0;
485 	if (bd_speedupreq == 0 || bd_request == 0)
486 		needwake = 1;
487 	bd_speedupreq = 1;
488 	bd_request = 1;
489 	if (needwake)
490 		wakeup(&bd_request);
491 	mtx_unlock(&bdlock);
492 }
493 
494 /*
495  * Calculating buffer cache scaling values and reserve space for buffer
496  * headers.  This is called during low level kernel initialization and
497  * may be called more then once.  We CANNOT write to the memory area
498  * being reserved at this time.
499  */
500 caddr_t
501 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
502 {
503 	int tuned_nbuf;
504 	long maxbuf;
505 
506 	/*
507 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
508 	 * PAGE_SIZE is >= 1K)
509 	 */
510 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
511 
512 	/*
513 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
514 	 * For the first 64MB of ram nominally allocate sufficient buffers to
515 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
516 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
517 	 * the buffer cache we limit the eventual kva reservation to
518 	 * maxbcache bytes.
519 	 *
520 	 * factor represents the 1/4 x ram conversion.
521 	 */
522 	if (nbuf == 0) {
523 		int factor = 4 * BKVASIZE / 1024;
524 
525 		nbuf = 50;
526 		if (physmem_est > 4096)
527 			nbuf += min((physmem_est - 4096) / factor,
528 			    65536 / factor);
529 		if (physmem_est > 65536)
530 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
531 
532 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
533 			nbuf = maxbcache / BKVASIZE;
534 		tuned_nbuf = 1;
535 	} else
536 		tuned_nbuf = 0;
537 
538 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
539 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
540 	if (nbuf > maxbuf) {
541 		if (!tuned_nbuf)
542 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
543 			    maxbuf);
544 		nbuf = maxbuf;
545 	}
546 
547 	/*
548 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
549 	 * We have no less then 16 and no more then 256.
550 	 */
551 	nswbuf = max(min(nbuf/4, 256), 16);
552 #ifdef NSWBUF_MIN
553 	if (nswbuf < NSWBUF_MIN)
554 		nswbuf = NSWBUF_MIN;
555 #endif
556 #ifdef DIRECTIO
557 	ffs_rawread_setup();
558 #endif
559 
560 	/*
561 	 * Reserve space for the buffer cache buffers
562 	 */
563 	swbuf = (void *)v;
564 	v = (caddr_t)(swbuf + nswbuf);
565 	buf = (void *)v;
566 	v = (caddr_t)(buf + nbuf);
567 
568 	return(v);
569 }
570 
571 /* Initialize the buffer subsystem.  Called before use of any buffers. */
572 void
573 bufinit(void)
574 {
575 	struct buf *bp;
576 	int i;
577 
578 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
579 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
580 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
581 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
582 
583 	/* next, make a null set of free lists */
584 	for (i = 0; i < BUFFER_QUEUES; i++)
585 		TAILQ_INIT(&bufqueues[i]);
586 
587 	/* finally, initialize each buffer header and stick on empty q */
588 	for (i = 0; i < nbuf; i++) {
589 		bp = &buf[i];
590 		bzero(bp, sizeof *bp);
591 		bp->b_flags = B_INVAL;	/* we're just an empty header */
592 		bp->b_rcred = NOCRED;
593 		bp->b_wcred = NOCRED;
594 		bp->b_qindex = QUEUE_EMPTY;
595 		bp->b_vflags = 0;
596 		bp->b_xflags = 0;
597 		LIST_INIT(&bp->b_dep);
598 		BUF_LOCKINIT(bp);
599 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
600 	}
601 
602 	/*
603 	 * maxbufspace is the absolute maximum amount of buffer space we are
604 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
605 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
606 	 * used by most other processes.  The differential is required to
607 	 * ensure that buf_daemon is able to run when other processes might
608 	 * be blocked waiting for buffer space.
609 	 *
610 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
611 	 * this may result in KVM fragmentation which is not handled optimally
612 	 * by the system.
613 	 */
614 	maxbufspace = (long)nbuf * BKVASIZE;
615 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
616 	lobufspace = hibufspace - MAXBSIZE;
617 
618 	lorunningspace = 512 * 1024;
619 	hirunningspace = 1024 * 1024;
620 
621 /*
622  * Limit the amount of malloc memory since it is wired permanently into
623  * the kernel space.  Even though this is accounted for in the buffer
624  * allocation, we don't want the malloced region to grow uncontrolled.
625  * The malloc scheme improves memory utilization significantly on average
626  * (small) directories.
627  */
628 	maxbufmallocspace = hibufspace / 20;
629 
630 /*
631  * Reduce the chance of a deadlock occuring by limiting the number
632  * of delayed-write dirty buffers we allow to stack up.
633  */
634 	hidirtybuffers = nbuf / 4 + 20;
635 	dirtybufthresh = hidirtybuffers * 9 / 10;
636 	numdirtybuffers = 0;
637 /*
638  * To support extreme low-memory systems, make sure hidirtybuffers cannot
639  * eat up all available buffer space.  This occurs when our minimum cannot
640  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
641  * BKVASIZE'd (8K) buffers.
642  */
643 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
644 		hidirtybuffers >>= 1;
645 	}
646 	lodirtybuffers = hidirtybuffers / 2;
647 
648 /*
649  * Try to keep the number of free buffers in the specified range,
650  * and give special processes (e.g. like buf_daemon) access to an
651  * emergency reserve.
652  */
653 	lofreebuffers = nbuf / 18 + 5;
654 	hifreebuffers = 2 * lofreebuffers;
655 	numfreebuffers = nbuf;
656 
657 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
658 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
659 }
660 
661 /*
662  * bfreekva() - free the kva allocation for a buffer.
663  *
664  *	Since this call frees up buffer space, we call bufspacewakeup().
665  */
666 static void
667 bfreekva(struct buf *bp)
668 {
669 
670 	if (bp->b_kvasize) {
671 		atomic_add_int(&buffreekvacnt, 1);
672 		atomic_subtract_long(&bufspace, bp->b_kvasize);
673 		vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
674 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
675 		bp->b_kvasize = 0;
676 		bufspacewakeup();
677 	}
678 }
679 
680 /*
681  *	bremfree:
682  *
683  *	Mark the buffer for removal from the appropriate free list in brelse.
684  *
685  */
686 void
687 bremfree(struct buf *bp)
688 {
689 
690 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
691 	KASSERT((bp->b_flags & B_REMFREE) == 0,
692 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
693 	KASSERT(bp->b_qindex != QUEUE_NONE,
694 	    ("bremfree: buffer %p not on a queue.", bp));
695 	BUF_ASSERT_HELD(bp);
696 
697 	bp->b_flags |= B_REMFREE;
698 	/* Fixup numfreebuffers count.  */
699 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
700 		atomic_subtract_int(&numfreebuffers, 1);
701 }
702 
703 /*
704  *	bremfreef:
705  *
706  *	Force an immediate removal from a free list.  Used only in nfs when
707  *	it abuses the b_freelist pointer.
708  */
709 void
710 bremfreef(struct buf *bp)
711 {
712 	mtx_lock(&bqlock);
713 	bremfreel(bp);
714 	mtx_unlock(&bqlock);
715 }
716 
717 /*
718  *	bremfreel:
719  *
720  *	Removes a buffer from the free list, must be called with the
721  *	bqlock held.
722  */
723 static void
724 bremfreel(struct buf *bp)
725 {
726 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
727 	    bp, bp->b_vp, bp->b_flags);
728 	KASSERT(bp->b_qindex != QUEUE_NONE,
729 	    ("bremfreel: buffer %p not on a queue.", bp));
730 	BUF_ASSERT_HELD(bp);
731 	mtx_assert(&bqlock, MA_OWNED);
732 
733 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
734 	bp->b_qindex = QUEUE_NONE;
735 	/*
736 	 * If this was a delayed bremfree() we only need to remove the buffer
737 	 * from the queue and return the stats are already done.
738 	 */
739 	if (bp->b_flags & B_REMFREE) {
740 		bp->b_flags &= ~B_REMFREE;
741 		return;
742 	}
743 	/*
744 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
745 	 * delayed-write, the buffer was free and we must decrement
746 	 * numfreebuffers.
747 	 */
748 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
749 		atomic_subtract_int(&numfreebuffers, 1);
750 }
751 
752 
753 /*
754  * Get a buffer with the specified data.  Look in the cache first.  We
755  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
756  * is set, the buffer is valid and we do not have to do anything ( see
757  * getblk() ).  This is really just a special case of breadn().
758  */
759 int
760 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
761     struct buf **bpp)
762 {
763 
764 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
765 }
766 
767 /*
768  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
769  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
770  * the buffer is valid and we do not have to do anything.
771  */
772 void
773 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
774     int cnt, struct ucred * cred)
775 {
776 	struct buf *rabp;
777 	int i;
778 
779 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
780 		if (inmem(vp, *rablkno))
781 			continue;
782 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
783 
784 		if ((rabp->b_flags & B_CACHE) == 0) {
785 			if (!TD_IS_IDLETHREAD(curthread))
786 				curthread->td_ru.ru_inblock++;
787 			rabp->b_flags |= B_ASYNC;
788 			rabp->b_flags &= ~B_INVAL;
789 			rabp->b_ioflags &= ~BIO_ERROR;
790 			rabp->b_iocmd = BIO_READ;
791 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
792 				rabp->b_rcred = crhold(cred);
793 			vfs_busy_pages(rabp, 0);
794 			BUF_KERNPROC(rabp);
795 			rabp->b_iooffset = dbtob(rabp->b_blkno);
796 			bstrategy(rabp);
797 		} else {
798 			brelse(rabp);
799 		}
800 	}
801 }
802 
803 /*
804  * Operates like bread, but also starts asynchronous I/O on
805  * read-ahead blocks.
806  */
807 int
808 breadn(struct vnode * vp, daddr_t blkno, int size,
809     daddr_t * rablkno, int *rabsize,
810     int cnt, struct ucred * cred, struct buf **bpp)
811 {
812 	struct buf *bp;
813 	int rv = 0, readwait = 0;
814 
815 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
816 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
817 
818 	/* if not found in cache, do some I/O */
819 	if ((bp->b_flags & B_CACHE) == 0) {
820 		if (!TD_IS_IDLETHREAD(curthread))
821 			curthread->td_ru.ru_inblock++;
822 		bp->b_iocmd = BIO_READ;
823 		bp->b_flags &= ~B_INVAL;
824 		bp->b_ioflags &= ~BIO_ERROR;
825 		if (bp->b_rcred == NOCRED && cred != NOCRED)
826 			bp->b_rcred = crhold(cred);
827 		vfs_busy_pages(bp, 0);
828 		bp->b_iooffset = dbtob(bp->b_blkno);
829 		bstrategy(bp);
830 		++readwait;
831 	}
832 
833 	breada(vp, rablkno, rabsize, cnt, cred);
834 
835 	if (readwait) {
836 		rv = bufwait(bp);
837 	}
838 	return (rv);
839 }
840 
841 /*
842  * Write, release buffer on completion.  (Done by iodone
843  * if async).  Do not bother writing anything if the buffer
844  * is invalid.
845  *
846  * Note that we set B_CACHE here, indicating that buffer is
847  * fully valid and thus cacheable.  This is true even of NFS
848  * now so we set it generally.  This could be set either here
849  * or in biodone() since the I/O is synchronous.  We put it
850  * here.
851  */
852 int
853 bufwrite(struct buf *bp)
854 {
855 	int oldflags;
856 	struct vnode *vp;
857 	int vp_md;
858 
859 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
860 	if (bp->b_flags & B_INVAL) {
861 		brelse(bp);
862 		return (0);
863 	}
864 
865 	oldflags = bp->b_flags;
866 
867 	BUF_ASSERT_HELD(bp);
868 
869 	if (bp->b_pin_count > 0)
870 		bunpin_wait(bp);
871 
872 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
873 	    ("FFS background buffer should not get here %p", bp));
874 
875 	vp = bp->b_vp;
876 	if (vp)
877 		vp_md = vp->v_vflag & VV_MD;
878 	else
879 		vp_md = 0;
880 
881 	/* Mark the buffer clean */
882 	bundirty(bp);
883 
884 	bp->b_flags &= ~B_DONE;
885 	bp->b_ioflags &= ~BIO_ERROR;
886 	bp->b_flags |= B_CACHE;
887 	bp->b_iocmd = BIO_WRITE;
888 
889 	bufobj_wref(bp->b_bufobj);
890 	vfs_busy_pages(bp, 1);
891 
892 	/*
893 	 * Normal bwrites pipeline writes
894 	 */
895 	bp->b_runningbufspace = bp->b_bufsize;
896 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
897 
898 	if (!TD_IS_IDLETHREAD(curthread))
899 		curthread->td_ru.ru_oublock++;
900 	if (oldflags & B_ASYNC)
901 		BUF_KERNPROC(bp);
902 	bp->b_iooffset = dbtob(bp->b_blkno);
903 	bstrategy(bp);
904 
905 	if ((oldflags & B_ASYNC) == 0) {
906 		int rtval = bufwait(bp);
907 		brelse(bp);
908 		return (rtval);
909 	} else {
910 		/*
911 		 * don't allow the async write to saturate the I/O
912 		 * system.  We will not deadlock here because
913 		 * we are blocking waiting for I/O that is already in-progress
914 		 * to complete. We do not block here if it is the update
915 		 * or syncer daemon trying to clean up as that can lead
916 		 * to deadlock.
917 		 */
918 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
919 			waitrunningbufspace();
920 	}
921 
922 	return (0);
923 }
924 
925 void
926 bufbdflush(struct bufobj *bo, struct buf *bp)
927 {
928 	struct buf *nbp;
929 
930 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
931 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
932 		altbufferflushes++;
933 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
934 		BO_LOCK(bo);
935 		/*
936 		 * Try to find a buffer to flush.
937 		 */
938 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
939 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
940 			    BUF_LOCK(nbp,
941 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
942 				continue;
943 			if (bp == nbp)
944 				panic("bdwrite: found ourselves");
945 			BO_UNLOCK(bo);
946 			/* Don't countdeps with the bo lock held. */
947 			if (buf_countdeps(nbp, 0)) {
948 				BO_LOCK(bo);
949 				BUF_UNLOCK(nbp);
950 				continue;
951 			}
952 			if (nbp->b_flags & B_CLUSTEROK) {
953 				vfs_bio_awrite(nbp);
954 			} else {
955 				bremfree(nbp);
956 				bawrite(nbp);
957 			}
958 			dirtybufferflushes++;
959 			break;
960 		}
961 		if (nbp == NULL)
962 			BO_UNLOCK(bo);
963 	}
964 }
965 
966 /*
967  * Delayed write. (Buffer is marked dirty).  Do not bother writing
968  * anything if the buffer is marked invalid.
969  *
970  * Note that since the buffer must be completely valid, we can safely
971  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
972  * biodone() in order to prevent getblk from writing the buffer
973  * out synchronously.
974  */
975 void
976 bdwrite(struct buf *bp)
977 {
978 	struct thread *td = curthread;
979 	struct vnode *vp;
980 	struct bufobj *bo;
981 
982 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
983 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
984 	BUF_ASSERT_HELD(bp);
985 
986 	if (bp->b_flags & B_INVAL) {
987 		brelse(bp);
988 		return;
989 	}
990 
991 	/*
992 	 * If we have too many dirty buffers, don't create any more.
993 	 * If we are wildly over our limit, then force a complete
994 	 * cleanup. Otherwise, just keep the situation from getting
995 	 * out of control. Note that we have to avoid a recursive
996 	 * disaster and not try to clean up after our own cleanup!
997 	 */
998 	vp = bp->b_vp;
999 	bo = bp->b_bufobj;
1000 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1001 		td->td_pflags |= TDP_INBDFLUSH;
1002 		BO_BDFLUSH(bo, bp);
1003 		td->td_pflags &= ~TDP_INBDFLUSH;
1004 	} else
1005 		recursiveflushes++;
1006 
1007 	bdirty(bp);
1008 	/*
1009 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1010 	 * true even of NFS now.
1011 	 */
1012 	bp->b_flags |= B_CACHE;
1013 
1014 	/*
1015 	 * This bmap keeps the system from needing to do the bmap later,
1016 	 * perhaps when the system is attempting to do a sync.  Since it
1017 	 * is likely that the indirect block -- or whatever other datastructure
1018 	 * that the filesystem needs is still in memory now, it is a good
1019 	 * thing to do this.  Note also, that if the pageout daemon is
1020 	 * requesting a sync -- there might not be enough memory to do
1021 	 * the bmap then...  So, this is important to do.
1022 	 */
1023 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1024 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1025 	}
1026 
1027 	/*
1028 	 * Set the *dirty* buffer range based upon the VM system dirty
1029 	 * pages.
1030 	 *
1031 	 * Mark the buffer pages as clean.  We need to do this here to
1032 	 * satisfy the vnode_pager and the pageout daemon, so that it
1033 	 * thinks that the pages have been "cleaned".  Note that since
1034 	 * the pages are in a delayed write buffer -- the VFS layer
1035 	 * "will" see that the pages get written out on the next sync,
1036 	 * or perhaps the cluster will be completed.
1037 	 */
1038 	vfs_clean_pages_dirty_buf(bp);
1039 	bqrelse(bp);
1040 
1041 	/*
1042 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
1043 	 * buffers (midpoint between our recovery point and our stall
1044 	 * point).
1045 	 */
1046 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1047 
1048 	/*
1049 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1050 	 * due to the softdep code.
1051 	 */
1052 }
1053 
1054 /*
1055  *	bdirty:
1056  *
1057  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1058  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1059  *	itself to properly update it in the dirty/clean lists.  We mark it
1060  *	B_DONE to ensure that any asynchronization of the buffer properly
1061  *	clears B_DONE ( else a panic will occur later ).
1062  *
1063  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1064  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1065  *	should only be called if the buffer is known-good.
1066  *
1067  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1068  *	count.
1069  *
1070  *	The buffer must be on QUEUE_NONE.
1071  */
1072 void
1073 bdirty(struct buf *bp)
1074 {
1075 
1076 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1077 	    bp, bp->b_vp, bp->b_flags);
1078 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1079 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1080 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1081 	BUF_ASSERT_HELD(bp);
1082 	bp->b_flags &= ~(B_RELBUF);
1083 	bp->b_iocmd = BIO_WRITE;
1084 
1085 	if ((bp->b_flags & B_DELWRI) == 0) {
1086 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1087 		reassignbuf(bp);
1088 		atomic_add_int(&numdirtybuffers, 1);
1089 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1090 	}
1091 }
1092 
1093 /*
1094  *	bundirty:
1095  *
1096  *	Clear B_DELWRI for buffer.
1097  *
1098  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1099  *	count.
1100  *
1101  *	The buffer must be on QUEUE_NONE.
1102  */
1103 
1104 void
1105 bundirty(struct buf *bp)
1106 {
1107 
1108 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1109 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1110 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1111 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1112 	BUF_ASSERT_HELD(bp);
1113 
1114 	if (bp->b_flags & B_DELWRI) {
1115 		bp->b_flags &= ~B_DELWRI;
1116 		reassignbuf(bp);
1117 		atomic_subtract_int(&numdirtybuffers, 1);
1118 		numdirtywakeup(lodirtybuffers);
1119 	}
1120 	/*
1121 	 * Since it is now being written, we can clear its deferred write flag.
1122 	 */
1123 	bp->b_flags &= ~B_DEFERRED;
1124 }
1125 
1126 /*
1127  *	bawrite:
1128  *
1129  *	Asynchronous write.  Start output on a buffer, but do not wait for
1130  *	it to complete.  The buffer is released when the output completes.
1131  *
1132  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1133  *	B_INVAL buffers.  Not us.
1134  */
1135 void
1136 bawrite(struct buf *bp)
1137 {
1138 
1139 	bp->b_flags |= B_ASYNC;
1140 	(void) bwrite(bp);
1141 }
1142 
1143 /*
1144  *	bwillwrite:
1145  *
1146  *	Called prior to the locking of any vnodes when we are expecting to
1147  *	write.  We do not want to starve the buffer cache with too many
1148  *	dirty buffers so we block here.  By blocking prior to the locking
1149  *	of any vnodes we attempt to avoid the situation where a locked vnode
1150  *	prevents the various system daemons from flushing related buffers.
1151  */
1152 
1153 void
1154 bwillwrite(void)
1155 {
1156 
1157 	if (numdirtybuffers >= hidirtybuffers) {
1158 		mtx_lock(&nblock);
1159 		while (numdirtybuffers >= hidirtybuffers) {
1160 			bd_wakeup(1);
1161 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1162 			msleep(&needsbuffer, &nblock,
1163 			    (PRIBIO + 4), "flswai", 0);
1164 		}
1165 		mtx_unlock(&nblock);
1166 	}
1167 }
1168 
1169 /*
1170  * Return true if we have too many dirty buffers.
1171  */
1172 int
1173 buf_dirty_count_severe(void)
1174 {
1175 
1176 	return(numdirtybuffers >= hidirtybuffers);
1177 }
1178 
1179 static __noinline int
1180 buf_vm_page_count_severe(void)
1181 {
1182 
1183 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1184 
1185 	return vm_page_count_severe();
1186 }
1187 
1188 /*
1189  *	brelse:
1190  *
1191  *	Release a busy buffer and, if requested, free its resources.  The
1192  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1193  *	to be accessed later as a cache entity or reused for other purposes.
1194  */
1195 void
1196 brelse(struct buf *bp)
1197 {
1198 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1199 	    bp, bp->b_vp, bp->b_flags);
1200 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1201 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1202 
1203 	if (bp->b_flags & B_MANAGED) {
1204 		bqrelse(bp);
1205 		return;
1206 	}
1207 
1208 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1209 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1210 		/*
1211 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1212 		 * pages from being scrapped.  If the error is anything
1213 		 * other than an I/O error (EIO), assume that retrying
1214 		 * is futile.
1215 		 */
1216 		bp->b_ioflags &= ~BIO_ERROR;
1217 		bdirty(bp);
1218 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1219 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1220 		/*
1221 		 * Either a failed I/O or we were asked to free or not
1222 		 * cache the buffer.
1223 		 */
1224 		bp->b_flags |= B_INVAL;
1225 		if (!LIST_EMPTY(&bp->b_dep))
1226 			buf_deallocate(bp);
1227 		if (bp->b_flags & B_DELWRI) {
1228 			atomic_subtract_int(&numdirtybuffers, 1);
1229 			numdirtywakeup(lodirtybuffers);
1230 		}
1231 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1232 		if ((bp->b_flags & B_VMIO) == 0) {
1233 			if (bp->b_bufsize)
1234 				allocbuf(bp, 0);
1235 			if (bp->b_vp)
1236 				brelvp(bp);
1237 		}
1238 	}
1239 
1240 	/*
1241 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1242 	 * is called with B_DELWRI set, the underlying pages may wind up
1243 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1244 	 * because pages associated with a B_DELWRI bp are marked clean.
1245 	 *
1246 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1247 	 * if B_DELWRI is set.
1248 	 *
1249 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1250 	 * on pages to return pages to the VM page queues.
1251 	 */
1252 	if (bp->b_flags & B_DELWRI)
1253 		bp->b_flags &= ~B_RELBUF;
1254 	else if (buf_vm_page_count_severe()) {
1255 		/*
1256 		 * The locking of the BO_LOCK is not necessary since
1257 		 * BKGRDINPROG cannot be set while we hold the buf
1258 		 * lock, it can only be cleared if it is already
1259 		 * pending.
1260 		 */
1261 		if (bp->b_vp) {
1262 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1263 				bp->b_flags |= B_RELBUF;
1264 		} else
1265 			bp->b_flags |= B_RELBUF;
1266 	}
1267 
1268 	/*
1269 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1270 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1271 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1272 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1273 	 *
1274 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1275 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1276 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1277 	 *
1278 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1279 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1280 	 * the commit state and we cannot afford to lose the buffer. If the
1281 	 * buffer has a background write in progress, we need to keep it
1282 	 * around to prevent it from being reconstituted and starting a second
1283 	 * background write.
1284 	 */
1285 	if ((bp->b_flags & B_VMIO)
1286 	    && !(bp->b_vp->v_mount != NULL &&
1287 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1288 		 !vn_isdisk(bp->b_vp, NULL) &&
1289 		 (bp->b_flags & B_DELWRI))
1290 	    ) {
1291 
1292 		int i, j, resid;
1293 		vm_page_t m;
1294 		off_t foff;
1295 		vm_pindex_t poff;
1296 		vm_object_t obj;
1297 
1298 		obj = bp->b_bufobj->bo_object;
1299 
1300 		/*
1301 		 * Get the base offset and length of the buffer.  Note that
1302 		 * in the VMIO case if the buffer block size is not
1303 		 * page-aligned then b_data pointer may not be page-aligned.
1304 		 * But our b_pages[] array *IS* page aligned.
1305 		 *
1306 		 * block sizes less then DEV_BSIZE (usually 512) are not
1307 		 * supported due to the page granularity bits (m->valid,
1308 		 * m->dirty, etc...).
1309 		 *
1310 		 * See man buf(9) for more information
1311 		 */
1312 		resid = bp->b_bufsize;
1313 		foff = bp->b_offset;
1314 		VM_OBJECT_LOCK(obj);
1315 		for (i = 0; i < bp->b_npages; i++) {
1316 			int had_bogus = 0;
1317 
1318 			m = bp->b_pages[i];
1319 
1320 			/*
1321 			 * If we hit a bogus page, fixup *all* the bogus pages
1322 			 * now.
1323 			 */
1324 			if (m == bogus_page) {
1325 				poff = OFF_TO_IDX(bp->b_offset);
1326 				had_bogus = 1;
1327 
1328 				for (j = i; j < bp->b_npages; j++) {
1329 					vm_page_t mtmp;
1330 					mtmp = bp->b_pages[j];
1331 					if (mtmp == bogus_page) {
1332 						mtmp = vm_page_lookup(obj, poff + j);
1333 						if (!mtmp) {
1334 							panic("brelse: page missing\n");
1335 						}
1336 						bp->b_pages[j] = mtmp;
1337 					}
1338 				}
1339 
1340 				if ((bp->b_flags & B_INVAL) == 0) {
1341 					pmap_qenter(
1342 					    trunc_page((vm_offset_t)bp->b_data),
1343 					    bp->b_pages, bp->b_npages);
1344 				}
1345 				m = bp->b_pages[i];
1346 			}
1347 			if ((bp->b_flags & B_NOCACHE) ||
1348 			    (bp->b_ioflags & BIO_ERROR &&
1349 			     bp->b_iocmd == BIO_READ)) {
1350 				int poffset = foff & PAGE_MASK;
1351 				int presid = resid > (PAGE_SIZE - poffset) ?
1352 					(PAGE_SIZE - poffset) : resid;
1353 
1354 				KASSERT(presid >= 0, ("brelse: extra page"));
1355 				vm_page_set_invalid(m, poffset, presid);
1356 				if (had_bogus)
1357 					printf("avoided corruption bug in bogus_page/brelse code\n");
1358 			}
1359 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1360 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1361 		}
1362 		VM_OBJECT_UNLOCK(obj);
1363 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1364 			vfs_vmio_release(bp);
1365 
1366 	} else if (bp->b_flags & B_VMIO) {
1367 
1368 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1369 			vfs_vmio_release(bp);
1370 		}
1371 
1372 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1373 		if (bp->b_bufsize != 0)
1374 			allocbuf(bp, 0);
1375 		if (bp->b_vp != NULL)
1376 			brelvp(bp);
1377 	}
1378 
1379 	if (BUF_LOCKRECURSED(bp)) {
1380 		/* do not release to free list */
1381 		BUF_UNLOCK(bp);
1382 		return;
1383 	}
1384 
1385 	/* enqueue */
1386 	mtx_lock(&bqlock);
1387 	/* Handle delayed bremfree() processing. */
1388 	if (bp->b_flags & B_REMFREE)
1389 		bremfreel(bp);
1390 	if (bp->b_qindex != QUEUE_NONE)
1391 		panic("brelse: free buffer onto another queue???");
1392 
1393 	/*
1394 	 * If the buffer has junk contents signal it and eventually
1395 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1396 	 * doesn't find it.
1397 	 */
1398 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1399 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1400 		bp->b_flags |= B_INVAL;
1401 	if (bp->b_flags & B_INVAL) {
1402 		if (bp->b_flags & B_DELWRI)
1403 			bundirty(bp);
1404 		if (bp->b_vp)
1405 			brelvp(bp);
1406 	}
1407 
1408 	/* buffers with no memory */
1409 	if (bp->b_bufsize == 0) {
1410 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1411 		if (bp->b_vflags & BV_BKGRDINPROG)
1412 			panic("losing buffer 1");
1413 		if (bp->b_kvasize) {
1414 			bp->b_qindex = QUEUE_EMPTYKVA;
1415 		} else {
1416 			bp->b_qindex = QUEUE_EMPTY;
1417 		}
1418 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1419 	/* buffers with junk contents */
1420 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1421 	    (bp->b_ioflags & BIO_ERROR)) {
1422 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1423 		if (bp->b_vflags & BV_BKGRDINPROG)
1424 			panic("losing buffer 2");
1425 		bp->b_qindex = QUEUE_CLEAN;
1426 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1427 	/* remaining buffers */
1428 	} else {
1429 		if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1430 		    (B_DELWRI|B_NEEDSGIANT))
1431 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1432 		else if (bp->b_flags & B_DELWRI)
1433 			bp->b_qindex = QUEUE_DIRTY;
1434 		else
1435 			bp->b_qindex = QUEUE_CLEAN;
1436 		if (bp->b_flags & B_AGE)
1437 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1438 		else
1439 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1440 	}
1441 	mtx_unlock(&bqlock);
1442 
1443 	/*
1444 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1445 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1446 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1447 	 * if B_INVAL is set ).
1448 	 */
1449 
1450 	if (!(bp->b_flags & B_DELWRI))
1451 		bufcountwakeup();
1452 
1453 	/*
1454 	 * Something we can maybe free or reuse
1455 	 */
1456 	if (bp->b_bufsize || bp->b_kvasize)
1457 		bufspacewakeup();
1458 
1459 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1460 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1461 		panic("brelse: not dirty");
1462 	/* unlock */
1463 	BUF_UNLOCK(bp);
1464 }
1465 
1466 /*
1467  * Release a buffer back to the appropriate queue but do not try to free
1468  * it.  The buffer is expected to be used again soon.
1469  *
1470  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1471  * biodone() to requeue an async I/O on completion.  It is also used when
1472  * known good buffers need to be requeued but we think we may need the data
1473  * again soon.
1474  *
1475  * XXX we should be able to leave the B_RELBUF hint set on completion.
1476  */
1477 void
1478 bqrelse(struct buf *bp)
1479 {
1480 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1481 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1482 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1483 
1484 	if (BUF_LOCKRECURSED(bp)) {
1485 		/* do not release to free list */
1486 		BUF_UNLOCK(bp);
1487 		return;
1488 	}
1489 
1490 	if (bp->b_flags & B_MANAGED) {
1491 		if (bp->b_flags & B_REMFREE) {
1492 			mtx_lock(&bqlock);
1493 			bremfreel(bp);
1494 			mtx_unlock(&bqlock);
1495 		}
1496 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1497 		BUF_UNLOCK(bp);
1498 		return;
1499 	}
1500 
1501 	mtx_lock(&bqlock);
1502 	/* Handle delayed bremfree() processing. */
1503 	if (bp->b_flags & B_REMFREE)
1504 		bremfreel(bp);
1505 	if (bp->b_qindex != QUEUE_NONE)
1506 		panic("bqrelse: free buffer onto another queue???");
1507 	/* buffers with stale but valid contents */
1508 	if (bp->b_flags & B_DELWRI) {
1509 		if (bp->b_flags & B_NEEDSGIANT)
1510 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1511 		else
1512 			bp->b_qindex = QUEUE_DIRTY;
1513 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1514 	} else {
1515 		/*
1516 		 * The locking of the BO_LOCK for checking of the
1517 		 * BV_BKGRDINPROG is not necessary since the
1518 		 * BV_BKGRDINPROG cannot be set while we hold the buf
1519 		 * lock, it can only be cleared if it is already
1520 		 * pending.
1521 		 */
1522 		if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1523 			bp->b_qindex = QUEUE_CLEAN;
1524 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1525 			    b_freelist);
1526 		} else {
1527 			/*
1528 			 * We are too low on memory, we have to try to free
1529 			 * the buffer (most importantly: the wired pages
1530 			 * making up its backing store) *now*.
1531 			 */
1532 			mtx_unlock(&bqlock);
1533 			brelse(bp);
1534 			return;
1535 		}
1536 	}
1537 	mtx_unlock(&bqlock);
1538 
1539 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1540 		bufcountwakeup();
1541 
1542 	/*
1543 	 * Something we can maybe free or reuse.
1544 	 */
1545 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1546 		bufspacewakeup();
1547 
1548 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1549 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1550 		panic("bqrelse: not dirty");
1551 	/* unlock */
1552 	BUF_UNLOCK(bp);
1553 }
1554 
1555 /* Give pages used by the bp back to the VM system (where possible) */
1556 static void
1557 vfs_vmio_release(struct buf *bp)
1558 {
1559 	int i;
1560 	vm_page_t m;
1561 
1562 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1563 	for (i = 0; i < bp->b_npages; i++) {
1564 		m = bp->b_pages[i];
1565 		bp->b_pages[i] = NULL;
1566 		/*
1567 		 * In order to keep page LRU ordering consistent, put
1568 		 * everything on the inactive queue.
1569 		 */
1570 		vm_page_lock(m);
1571 		vm_page_unwire(m, 0);
1572 		/*
1573 		 * We don't mess with busy pages, it is
1574 		 * the responsibility of the process that
1575 		 * busied the pages to deal with them.
1576 		 */
1577 		if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 &&
1578 		    m->wire_count == 0) {
1579 			/*
1580 			 * Might as well free the page if we can and it has
1581 			 * no valid data.  We also free the page if the
1582 			 * buffer was used for direct I/O
1583 			 */
1584 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1585 			    m->hold_count == 0) {
1586 				vm_page_free(m);
1587 			} else if (bp->b_flags & B_DIRECT) {
1588 				vm_page_try_to_free(m);
1589 			} else if (buf_vm_page_count_severe()) {
1590 				vm_page_try_to_cache(m);
1591 			}
1592 		}
1593 		vm_page_unlock(m);
1594 	}
1595 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1596 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1597 
1598 	if (bp->b_bufsize) {
1599 		bufspacewakeup();
1600 		bp->b_bufsize = 0;
1601 	}
1602 	bp->b_npages = 0;
1603 	bp->b_flags &= ~B_VMIO;
1604 	if (bp->b_vp)
1605 		brelvp(bp);
1606 }
1607 
1608 /*
1609  * Check to see if a block at a particular lbn is available for a clustered
1610  * write.
1611  */
1612 static int
1613 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1614 {
1615 	struct buf *bpa;
1616 	int match;
1617 
1618 	match = 0;
1619 
1620 	/* If the buf isn't in core skip it */
1621 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1622 		return (0);
1623 
1624 	/* If the buf is busy we don't want to wait for it */
1625 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1626 		return (0);
1627 
1628 	/* Only cluster with valid clusterable delayed write buffers */
1629 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1630 	    (B_DELWRI | B_CLUSTEROK))
1631 		goto done;
1632 
1633 	if (bpa->b_bufsize != size)
1634 		goto done;
1635 
1636 	/*
1637 	 * Check to see if it is in the expected place on disk and that the
1638 	 * block has been mapped.
1639 	 */
1640 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1641 		match = 1;
1642 done:
1643 	BUF_UNLOCK(bpa);
1644 	return (match);
1645 }
1646 
1647 /*
1648  *	vfs_bio_awrite:
1649  *
1650  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1651  *	This is much better then the old way of writing only one buffer at
1652  *	a time.  Note that we may not be presented with the buffers in the
1653  *	correct order, so we search for the cluster in both directions.
1654  */
1655 int
1656 vfs_bio_awrite(struct buf *bp)
1657 {
1658 	struct bufobj *bo;
1659 	int i;
1660 	int j;
1661 	daddr_t lblkno = bp->b_lblkno;
1662 	struct vnode *vp = bp->b_vp;
1663 	int ncl;
1664 	int nwritten;
1665 	int size;
1666 	int maxcl;
1667 
1668 	bo = &vp->v_bufobj;
1669 	/*
1670 	 * right now we support clustered writing only to regular files.  If
1671 	 * we find a clusterable block we could be in the middle of a cluster
1672 	 * rather then at the beginning.
1673 	 */
1674 	if ((vp->v_type == VREG) &&
1675 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1676 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1677 
1678 		size = vp->v_mount->mnt_stat.f_iosize;
1679 		maxcl = MAXPHYS / size;
1680 
1681 		BO_LOCK(bo);
1682 		for (i = 1; i < maxcl; i++)
1683 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1684 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1685 				break;
1686 
1687 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1688 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1689 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1690 				break;
1691 		BO_UNLOCK(bo);
1692 		--j;
1693 		ncl = i + j;
1694 		/*
1695 		 * this is a possible cluster write
1696 		 */
1697 		if (ncl != 1) {
1698 			BUF_UNLOCK(bp);
1699 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1700 			return nwritten;
1701 		}
1702 	}
1703 	bremfree(bp);
1704 	bp->b_flags |= B_ASYNC;
1705 	/*
1706 	 * default (old) behavior, writing out only one block
1707 	 *
1708 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1709 	 */
1710 	nwritten = bp->b_bufsize;
1711 	(void) bwrite(bp);
1712 
1713 	return nwritten;
1714 }
1715 
1716 /*
1717  *	getnewbuf:
1718  *
1719  *	Find and initialize a new buffer header, freeing up existing buffers
1720  *	in the bufqueues as necessary.  The new buffer is returned locked.
1721  *
1722  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1723  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1724  *
1725  *	We block if:
1726  *		We have insufficient buffer headers
1727  *		We have insufficient buffer space
1728  *		buffer_map is too fragmented ( space reservation fails )
1729  *		If we have to flush dirty buffers ( but we try to avoid this )
1730  *
1731  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1732  *	Instead we ask the buf daemon to do it for us.  We attempt to
1733  *	avoid piecemeal wakeups of the pageout daemon.
1734  */
1735 
1736 static struct buf *
1737 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1738     int gbflags)
1739 {
1740 	struct thread *td;
1741 	struct buf *bp;
1742 	struct buf *nbp;
1743 	int defrag = 0;
1744 	int nqindex;
1745 	static int flushingbufs;
1746 
1747 	td = curthread;
1748 	/*
1749 	 * We can't afford to block since we might be holding a vnode lock,
1750 	 * which may prevent system daemons from running.  We deal with
1751 	 * low-memory situations by proactively returning memory and running
1752 	 * async I/O rather then sync I/O.
1753 	 */
1754 	atomic_add_int(&getnewbufcalls, 1);
1755 	atomic_subtract_int(&getnewbufrestarts, 1);
1756 restart:
1757 	atomic_add_int(&getnewbufrestarts, 1);
1758 
1759 	/*
1760 	 * Setup for scan.  If we do not have enough free buffers,
1761 	 * we setup a degenerate case that immediately fails.  Note
1762 	 * that if we are specially marked process, we are allowed to
1763 	 * dip into our reserves.
1764 	 *
1765 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1766 	 *
1767 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1768 	 * However, there are a number of cases (defragging, reusing, ...)
1769 	 * where we cannot backup.
1770 	 */
1771 	mtx_lock(&bqlock);
1772 	nqindex = QUEUE_EMPTYKVA;
1773 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1774 
1775 	if (nbp == NULL) {
1776 		/*
1777 		 * If no EMPTYKVA buffers and we are either
1778 		 * defragging or reusing, locate a CLEAN buffer
1779 		 * to free or reuse.  If bufspace useage is low
1780 		 * skip this step so we can allocate a new buffer.
1781 		 */
1782 		if (defrag || bufspace >= lobufspace) {
1783 			nqindex = QUEUE_CLEAN;
1784 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1785 		}
1786 
1787 		/*
1788 		 * If we could not find or were not allowed to reuse a
1789 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1790 		 * buffer.  We can only use an EMPTY buffer if allocating
1791 		 * its KVA would not otherwise run us out of buffer space.
1792 		 */
1793 		if (nbp == NULL && defrag == 0 &&
1794 		    bufspace + maxsize < hibufspace) {
1795 			nqindex = QUEUE_EMPTY;
1796 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1797 		}
1798 	}
1799 
1800 	/*
1801 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1802 	 * depending.
1803 	 */
1804 
1805 	while ((bp = nbp) != NULL) {
1806 		int qindex = nqindex;
1807 
1808 		/*
1809 		 * Calculate next bp ( we can only use it if we do not block
1810 		 * or do other fancy things ).
1811 		 */
1812 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1813 			switch(qindex) {
1814 			case QUEUE_EMPTY:
1815 				nqindex = QUEUE_EMPTYKVA;
1816 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1817 					break;
1818 				/* FALLTHROUGH */
1819 			case QUEUE_EMPTYKVA:
1820 				nqindex = QUEUE_CLEAN;
1821 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1822 					break;
1823 				/* FALLTHROUGH */
1824 			case QUEUE_CLEAN:
1825 				/*
1826 				 * nbp is NULL.
1827 				 */
1828 				break;
1829 			}
1830 		}
1831 		/*
1832 		 * If we are defragging then we need a buffer with
1833 		 * b_kvasize != 0.  XXX this situation should no longer
1834 		 * occur, if defrag is non-zero the buffer's b_kvasize
1835 		 * should also be non-zero at this point.  XXX
1836 		 */
1837 		if (defrag && bp->b_kvasize == 0) {
1838 			printf("Warning: defrag empty buffer %p\n", bp);
1839 			continue;
1840 		}
1841 
1842 		/*
1843 		 * Start freeing the bp.  This is somewhat involved.  nbp
1844 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1845 		 */
1846 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1847 			continue;
1848 		if (bp->b_vp) {
1849 			BO_LOCK(bp->b_bufobj);
1850 			if (bp->b_vflags & BV_BKGRDINPROG) {
1851 				BO_UNLOCK(bp->b_bufobj);
1852 				BUF_UNLOCK(bp);
1853 				continue;
1854 			}
1855 			BO_UNLOCK(bp->b_bufobj);
1856 		}
1857 		CTR6(KTR_BUF,
1858 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1859 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1860 		    bp->b_kvasize, bp->b_bufsize, qindex);
1861 
1862 		/*
1863 		 * Sanity Checks
1864 		 */
1865 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1866 
1867 		/*
1868 		 * Note: we no longer distinguish between VMIO and non-VMIO
1869 		 * buffers.
1870 		 */
1871 
1872 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1873 
1874 		bremfreel(bp);
1875 		mtx_unlock(&bqlock);
1876 
1877 		if (qindex == QUEUE_CLEAN) {
1878 			if (bp->b_flags & B_VMIO) {
1879 				bp->b_flags &= ~B_ASYNC;
1880 				vfs_vmio_release(bp);
1881 			}
1882 			if (bp->b_vp)
1883 				brelvp(bp);
1884 		}
1885 
1886 		/*
1887 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1888 		 * the scan from this point on.
1889 		 *
1890 		 * Get the rest of the buffer freed up.  b_kva* is still
1891 		 * valid after this operation.
1892 		 */
1893 
1894 		if (bp->b_rcred != NOCRED) {
1895 			crfree(bp->b_rcred);
1896 			bp->b_rcred = NOCRED;
1897 		}
1898 		if (bp->b_wcred != NOCRED) {
1899 			crfree(bp->b_wcred);
1900 			bp->b_wcred = NOCRED;
1901 		}
1902 		if (!LIST_EMPTY(&bp->b_dep))
1903 			buf_deallocate(bp);
1904 		if (bp->b_vflags & BV_BKGRDINPROG)
1905 			panic("losing buffer 3");
1906 		KASSERT(bp->b_vp == NULL,
1907 		    ("bp: %p still has vnode %p.  qindex: %d",
1908 		    bp, bp->b_vp, qindex));
1909 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1910 		   ("bp: %p still on a buffer list. xflags %X",
1911 		    bp, bp->b_xflags));
1912 
1913 		if (bp->b_bufsize)
1914 			allocbuf(bp, 0);
1915 
1916 		bp->b_flags = 0;
1917 		bp->b_ioflags = 0;
1918 		bp->b_xflags = 0;
1919 		bp->b_vflags = 0;
1920 		bp->b_vp = NULL;
1921 		bp->b_blkno = bp->b_lblkno = 0;
1922 		bp->b_offset = NOOFFSET;
1923 		bp->b_iodone = 0;
1924 		bp->b_error = 0;
1925 		bp->b_resid = 0;
1926 		bp->b_bcount = 0;
1927 		bp->b_npages = 0;
1928 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1929 		bp->b_bufobj = NULL;
1930 		bp->b_pin_count = 0;
1931 		bp->b_fsprivate1 = NULL;
1932 		bp->b_fsprivate2 = NULL;
1933 		bp->b_fsprivate3 = NULL;
1934 
1935 		LIST_INIT(&bp->b_dep);
1936 
1937 		/*
1938 		 * If we are defragging then free the buffer.
1939 		 */
1940 		if (defrag) {
1941 			bp->b_flags |= B_INVAL;
1942 			bfreekva(bp);
1943 			brelse(bp);
1944 			defrag = 0;
1945 			goto restart;
1946 		}
1947 
1948 		/*
1949 		 * Notify any waiters for the buffer lock about
1950 		 * identity change by freeing the buffer.
1951 		 */
1952 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
1953 			bp->b_flags |= B_INVAL;
1954 			bfreekva(bp);
1955 			brelse(bp);
1956 			goto restart;
1957 		}
1958 
1959 		/*
1960 		 * If we are overcomitted then recover the buffer and its
1961 		 * KVM space.  This occurs in rare situations when multiple
1962 		 * processes are blocked in getnewbuf() or allocbuf().
1963 		 */
1964 		if (bufspace >= hibufspace)
1965 			flushingbufs = 1;
1966 		if (flushingbufs && bp->b_kvasize != 0) {
1967 			bp->b_flags |= B_INVAL;
1968 			bfreekva(bp);
1969 			brelse(bp);
1970 			goto restart;
1971 		}
1972 		if (bufspace < lobufspace)
1973 			flushingbufs = 0;
1974 		break;
1975 	}
1976 
1977 	/*
1978 	 * If we exhausted our list, sleep as appropriate.  We may have to
1979 	 * wakeup various daemons and write out some dirty buffers.
1980 	 *
1981 	 * Generally we are sleeping due to insufficient buffer space.
1982 	 */
1983 
1984 	if (bp == NULL) {
1985 		int flags, norunbuf;
1986 		char *waitmsg;
1987 		int fl;
1988 
1989 		if (defrag) {
1990 			flags = VFS_BIO_NEED_BUFSPACE;
1991 			waitmsg = "nbufkv";
1992 		} else if (bufspace >= hibufspace) {
1993 			waitmsg = "nbufbs";
1994 			flags = VFS_BIO_NEED_BUFSPACE;
1995 		} else {
1996 			waitmsg = "newbuf";
1997 			flags = VFS_BIO_NEED_ANY;
1998 		}
1999 		mtx_lock(&nblock);
2000 		needsbuffer |= flags;
2001 		mtx_unlock(&nblock);
2002 		mtx_unlock(&bqlock);
2003 
2004 		bd_speedup();	/* heeeelp */
2005 		if (gbflags & GB_NOWAIT_BD)
2006 			return (NULL);
2007 
2008 		mtx_lock(&nblock);
2009 		while (needsbuffer & flags) {
2010 			if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2011 				mtx_unlock(&nblock);
2012 				/*
2013 				 * getblk() is called with a vnode
2014 				 * locked, and some majority of the
2015 				 * dirty buffers may as well belong to
2016 				 * the vnode. Flushing the buffers
2017 				 * there would make a progress that
2018 				 * cannot be achieved by the
2019 				 * buf_daemon, that cannot lock the
2020 				 * vnode.
2021 				 */
2022 				norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2023 				    (td->td_pflags & TDP_NORUNNINGBUF);
2024 				/* play bufdaemon */
2025 				td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2026 				fl = buf_do_flush(vp);
2027 				td->td_pflags &= norunbuf;
2028 				mtx_lock(&nblock);
2029 				if (fl != 0)
2030 					continue;
2031 				if ((needsbuffer & flags) == 0)
2032 					break;
2033 			}
2034 			if (msleep(&needsbuffer, &nblock,
2035 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2036 				mtx_unlock(&nblock);
2037 				return (NULL);
2038 			}
2039 		}
2040 		mtx_unlock(&nblock);
2041 	} else {
2042 		/*
2043 		 * We finally have a valid bp.  We aren't quite out of the
2044 		 * woods, we still have to reserve kva space.  In order
2045 		 * to keep fragmentation sane we only allocate kva in
2046 		 * BKVASIZE chunks.
2047 		 */
2048 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2049 
2050 		if (maxsize != bp->b_kvasize) {
2051 			vm_offset_t addr = 0;
2052 
2053 			bfreekva(bp);
2054 
2055 			vm_map_lock(buffer_map);
2056 			if (vm_map_findspace(buffer_map,
2057 				vm_map_min(buffer_map), maxsize, &addr)) {
2058 				/*
2059 				 * Uh oh.  Buffer map is to fragmented.  We
2060 				 * must defragment the map.
2061 				 */
2062 				atomic_add_int(&bufdefragcnt, 1);
2063 				vm_map_unlock(buffer_map);
2064 				defrag = 1;
2065 				bp->b_flags |= B_INVAL;
2066 				brelse(bp);
2067 				goto restart;
2068 			}
2069 			if (addr) {
2070 				vm_map_insert(buffer_map, NULL, 0,
2071 					addr, addr + maxsize,
2072 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2073 
2074 				bp->b_kvabase = (caddr_t) addr;
2075 				bp->b_kvasize = maxsize;
2076 				atomic_add_long(&bufspace, bp->b_kvasize);
2077 				atomic_add_int(&bufreusecnt, 1);
2078 			}
2079 			vm_map_unlock(buffer_map);
2080 		}
2081 		bp->b_saveaddr = bp->b_kvabase;
2082 		bp->b_data = bp->b_saveaddr;
2083 	}
2084 	return(bp);
2085 }
2086 
2087 /*
2088  *	buf_daemon:
2089  *
2090  *	buffer flushing daemon.  Buffers are normally flushed by the
2091  *	update daemon but if it cannot keep up this process starts to
2092  *	take the load in an attempt to prevent getnewbuf() from blocking.
2093  */
2094 
2095 static struct kproc_desc buf_kp = {
2096 	"bufdaemon",
2097 	buf_daemon,
2098 	&bufdaemonproc
2099 };
2100 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2101 
2102 static int
2103 buf_do_flush(struct vnode *vp)
2104 {
2105 	int flushed;
2106 
2107 	flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2108 	/* The list empty check here is slightly racy */
2109 	if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2110 		mtx_lock(&Giant);
2111 		flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0);
2112 		mtx_unlock(&Giant);
2113 	}
2114 	if (flushed == 0) {
2115 		/*
2116 		 * Could not find any buffers without rollback
2117 		 * dependencies, so just write the first one
2118 		 * in the hopes of eventually making progress.
2119 		 */
2120 		flushbufqueues(vp, QUEUE_DIRTY, 1);
2121 		if (!TAILQ_EMPTY(
2122 			    &bufqueues[QUEUE_DIRTY_GIANT])) {
2123 			mtx_lock(&Giant);
2124 			flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1);
2125 			mtx_unlock(&Giant);
2126 		}
2127 	}
2128 	return (flushed);
2129 }
2130 
2131 static void
2132 buf_daemon()
2133 {
2134 	int lodirtysave;
2135 
2136 	/*
2137 	 * This process needs to be suspended prior to shutdown sync.
2138 	 */
2139 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2140 	    SHUTDOWN_PRI_LAST);
2141 
2142 	/*
2143 	 * This process is allowed to take the buffer cache to the limit
2144 	 */
2145 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2146 	mtx_lock(&bdlock);
2147 	for (;;) {
2148 		bd_request = 0;
2149 		mtx_unlock(&bdlock);
2150 
2151 		kproc_suspend_check(bufdaemonproc);
2152 		lodirtysave = lodirtybuffers;
2153 		if (bd_speedupreq) {
2154 			lodirtybuffers = numdirtybuffers / 2;
2155 			bd_speedupreq = 0;
2156 		}
2157 		/*
2158 		 * Do the flush.  Limit the amount of in-transit I/O we
2159 		 * allow to build up, otherwise we would completely saturate
2160 		 * the I/O system.  Wakeup any waiting processes before we
2161 		 * normally would so they can run in parallel with our drain.
2162 		 */
2163 		while (numdirtybuffers > lodirtybuffers) {
2164 			if (buf_do_flush(NULL) == 0)
2165 				break;
2166 			uio_yield();
2167 		}
2168 		lodirtybuffers = lodirtysave;
2169 
2170 		/*
2171 		 * Only clear bd_request if we have reached our low water
2172 		 * mark.  The buf_daemon normally waits 1 second and
2173 		 * then incrementally flushes any dirty buffers that have
2174 		 * built up, within reason.
2175 		 *
2176 		 * If we were unable to hit our low water mark and couldn't
2177 		 * find any flushable buffers, we sleep half a second.
2178 		 * Otherwise we loop immediately.
2179 		 */
2180 		mtx_lock(&bdlock);
2181 		if (numdirtybuffers <= lodirtybuffers) {
2182 			/*
2183 			 * We reached our low water mark, reset the
2184 			 * request and sleep until we are needed again.
2185 			 * The sleep is just so the suspend code works.
2186 			 */
2187 			bd_request = 0;
2188 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2189 		} else {
2190 			/*
2191 			 * We couldn't find any flushable dirty buffers but
2192 			 * still have too many dirty buffers, we
2193 			 * have to sleep and try again.  (rare)
2194 			 */
2195 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2196 		}
2197 	}
2198 }
2199 
2200 /*
2201  *	flushbufqueues:
2202  *
2203  *	Try to flush a buffer in the dirty queue.  We must be careful to
2204  *	free up B_INVAL buffers instead of write them, which NFS is
2205  *	particularly sensitive to.
2206  */
2207 static int flushwithdeps = 0;
2208 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2209     0, "Number of buffers flushed with dependecies that require rollbacks");
2210 
2211 static int
2212 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2213 {
2214 	struct buf *sentinel;
2215 	struct vnode *vp;
2216 	struct mount *mp;
2217 	struct buf *bp;
2218 	int hasdeps;
2219 	int flushed;
2220 	int target;
2221 
2222 	if (lvp == NULL) {
2223 		target = numdirtybuffers - lodirtybuffers;
2224 		if (flushdeps && target > 2)
2225 			target /= 2;
2226 	} else
2227 		target = flushbufqtarget;
2228 	flushed = 0;
2229 	bp = NULL;
2230 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2231 	sentinel->b_qindex = QUEUE_SENTINEL;
2232 	mtx_lock(&bqlock);
2233 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2234 	while (flushed != target) {
2235 		bp = TAILQ_NEXT(sentinel, b_freelist);
2236 		if (bp != NULL) {
2237 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2238 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2239 			    b_freelist);
2240 		} else
2241 			break;
2242 		/*
2243 		 * Skip sentinels inserted by other invocations of the
2244 		 * flushbufqueues(), taking care to not reorder them.
2245 		 */
2246 		if (bp->b_qindex == QUEUE_SENTINEL)
2247 			continue;
2248 		/*
2249 		 * Only flush the buffers that belong to the
2250 		 * vnode locked by the curthread.
2251 		 */
2252 		if (lvp != NULL && bp->b_vp != lvp)
2253 			continue;
2254 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2255 			continue;
2256 		if (bp->b_pin_count > 0) {
2257 			BUF_UNLOCK(bp);
2258 			continue;
2259 		}
2260 		BO_LOCK(bp->b_bufobj);
2261 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2262 		    (bp->b_flags & B_DELWRI) == 0) {
2263 			BO_UNLOCK(bp->b_bufobj);
2264 			BUF_UNLOCK(bp);
2265 			continue;
2266 		}
2267 		BO_UNLOCK(bp->b_bufobj);
2268 		if (bp->b_flags & B_INVAL) {
2269 			bremfreel(bp);
2270 			mtx_unlock(&bqlock);
2271 			brelse(bp);
2272 			flushed++;
2273 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2274 			mtx_lock(&bqlock);
2275 			continue;
2276 		}
2277 
2278 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2279 			if (flushdeps == 0) {
2280 				BUF_UNLOCK(bp);
2281 				continue;
2282 			}
2283 			hasdeps = 1;
2284 		} else
2285 			hasdeps = 0;
2286 		/*
2287 		 * We must hold the lock on a vnode before writing
2288 		 * one of its buffers. Otherwise we may confuse, or
2289 		 * in the case of a snapshot vnode, deadlock the
2290 		 * system.
2291 		 *
2292 		 * The lock order here is the reverse of the normal
2293 		 * of vnode followed by buf lock.  This is ok because
2294 		 * the NOWAIT will prevent deadlock.
2295 		 */
2296 		vp = bp->b_vp;
2297 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2298 			BUF_UNLOCK(bp);
2299 			continue;
2300 		}
2301 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2302 			mtx_unlock(&bqlock);
2303 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2304 			    bp, bp->b_vp, bp->b_flags);
2305 			if (curproc == bufdaemonproc)
2306 				vfs_bio_awrite(bp);
2307 			else {
2308 				bremfree(bp);
2309 				bwrite(bp);
2310 				notbufdflashes++;
2311 			}
2312 			vn_finished_write(mp);
2313 			VOP_UNLOCK(vp, 0);
2314 			flushwithdeps += hasdeps;
2315 			flushed++;
2316 
2317 			/*
2318 			 * Sleeping on runningbufspace while holding
2319 			 * vnode lock leads to deadlock.
2320 			 */
2321 			if (curproc == bufdaemonproc)
2322 				waitrunningbufspace();
2323 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2324 			mtx_lock(&bqlock);
2325 			continue;
2326 		}
2327 		vn_finished_write(mp);
2328 		BUF_UNLOCK(bp);
2329 	}
2330 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2331 	mtx_unlock(&bqlock);
2332 	free(sentinel, M_TEMP);
2333 	return (flushed);
2334 }
2335 
2336 /*
2337  * Check to see if a block is currently memory resident.
2338  */
2339 struct buf *
2340 incore(struct bufobj *bo, daddr_t blkno)
2341 {
2342 	struct buf *bp;
2343 
2344 	BO_LOCK(bo);
2345 	bp = gbincore(bo, blkno);
2346 	BO_UNLOCK(bo);
2347 	return (bp);
2348 }
2349 
2350 /*
2351  * Returns true if no I/O is needed to access the
2352  * associated VM object.  This is like incore except
2353  * it also hunts around in the VM system for the data.
2354  */
2355 
2356 static int
2357 inmem(struct vnode * vp, daddr_t blkno)
2358 {
2359 	vm_object_t obj;
2360 	vm_offset_t toff, tinc, size;
2361 	vm_page_t m;
2362 	vm_ooffset_t off;
2363 
2364 	ASSERT_VOP_LOCKED(vp, "inmem");
2365 
2366 	if (incore(&vp->v_bufobj, blkno))
2367 		return 1;
2368 	if (vp->v_mount == NULL)
2369 		return 0;
2370 	obj = vp->v_object;
2371 	if (obj == NULL)
2372 		return (0);
2373 
2374 	size = PAGE_SIZE;
2375 	if (size > vp->v_mount->mnt_stat.f_iosize)
2376 		size = vp->v_mount->mnt_stat.f_iosize;
2377 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2378 
2379 	VM_OBJECT_LOCK(obj);
2380 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2381 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2382 		if (!m)
2383 			goto notinmem;
2384 		tinc = size;
2385 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2386 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2387 		if (vm_page_is_valid(m,
2388 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2389 			goto notinmem;
2390 	}
2391 	VM_OBJECT_UNLOCK(obj);
2392 	return 1;
2393 
2394 notinmem:
2395 	VM_OBJECT_UNLOCK(obj);
2396 	return (0);
2397 }
2398 
2399 /*
2400  * Set the dirty range for a buffer based on the status of the dirty
2401  * bits in the pages comprising the buffer.  The range is limited
2402  * to the size of the buffer.
2403  *
2404  * Tell the VM system that the pages associated with this buffer
2405  * are clean.  This is used for delayed writes where the data is
2406  * going to go to disk eventually without additional VM intevention.
2407  *
2408  * Note that while we only really need to clean through to b_bcount, we
2409  * just go ahead and clean through to b_bufsize.
2410  */
2411 static void
2412 vfs_clean_pages_dirty_buf(struct buf *bp)
2413 {
2414 	vm_ooffset_t foff, noff, eoff;
2415 	vm_page_t m;
2416 	int i;
2417 
2418 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2419 		return;
2420 
2421 	foff = bp->b_offset;
2422 	KASSERT(bp->b_offset != NOOFFSET,
2423 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2424 
2425 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2426 	vfs_drain_busy_pages(bp);
2427 	vfs_setdirty_locked_object(bp);
2428 	for (i = 0; i < bp->b_npages; i++) {
2429 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2430 		eoff = noff;
2431 		if (eoff > bp->b_offset + bp->b_bufsize)
2432 			eoff = bp->b_offset + bp->b_bufsize;
2433 		m = bp->b_pages[i];
2434 		vfs_page_set_validclean(bp, foff, m);
2435 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2436 		foff = noff;
2437 	}
2438 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2439 }
2440 
2441 static void
2442 vfs_setdirty_locked_object(struct buf *bp)
2443 {
2444 	vm_object_t object;
2445 	int i;
2446 
2447 	object = bp->b_bufobj->bo_object;
2448 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2449 
2450 	/*
2451 	 * We qualify the scan for modified pages on whether the
2452 	 * object has been flushed yet.
2453 	 */
2454 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2455 		vm_offset_t boffset;
2456 		vm_offset_t eoffset;
2457 
2458 		/*
2459 		 * test the pages to see if they have been modified directly
2460 		 * by users through the VM system.
2461 		 */
2462 		for (i = 0; i < bp->b_npages; i++)
2463 			vm_page_test_dirty(bp->b_pages[i]);
2464 
2465 		/*
2466 		 * Calculate the encompassing dirty range, boffset and eoffset,
2467 		 * (eoffset - boffset) bytes.
2468 		 */
2469 
2470 		for (i = 0; i < bp->b_npages; i++) {
2471 			if (bp->b_pages[i]->dirty)
2472 				break;
2473 		}
2474 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2475 
2476 		for (i = bp->b_npages - 1; i >= 0; --i) {
2477 			if (bp->b_pages[i]->dirty) {
2478 				break;
2479 			}
2480 		}
2481 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2482 
2483 		/*
2484 		 * Fit it to the buffer.
2485 		 */
2486 
2487 		if (eoffset > bp->b_bcount)
2488 			eoffset = bp->b_bcount;
2489 
2490 		/*
2491 		 * If we have a good dirty range, merge with the existing
2492 		 * dirty range.
2493 		 */
2494 
2495 		if (boffset < eoffset) {
2496 			if (bp->b_dirtyoff > boffset)
2497 				bp->b_dirtyoff = boffset;
2498 			if (bp->b_dirtyend < eoffset)
2499 				bp->b_dirtyend = eoffset;
2500 		}
2501 	}
2502 }
2503 
2504 /*
2505  *	getblk:
2506  *
2507  *	Get a block given a specified block and offset into a file/device.
2508  *	The buffers B_DONE bit will be cleared on return, making it almost
2509  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2510  *	return.  The caller should clear B_INVAL prior to initiating a
2511  *	READ.
2512  *
2513  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2514  *	an existing buffer.
2515  *
2516  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2517  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2518  *	and then cleared based on the backing VM.  If the previous buffer is
2519  *	non-0-sized but invalid, B_CACHE will be cleared.
2520  *
2521  *	If getblk() must create a new buffer, the new buffer is returned with
2522  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2523  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2524  *	backing VM.
2525  *
2526  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2527  *	B_CACHE bit is clear.
2528  *
2529  *	What this means, basically, is that the caller should use B_CACHE to
2530  *	determine whether the buffer is fully valid or not and should clear
2531  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2532  *	the buffer by loading its data area with something, the caller needs
2533  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2534  *	the caller should set B_CACHE ( as an optimization ), else the caller
2535  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2536  *	a write attempt or if it was a successfull read.  If the caller
2537  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2538  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2539  */
2540 struct buf *
2541 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2542     int flags)
2543 {
2544 	struct buf *bp;
2545 	struct bufobj *bo;
2546 	int error;
2547 
2548 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2549 	ASSERT_VOP_LOCKED(vp, "getblk");
2550 	if (size > MAXBSIZE)
2551 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2552 
2553 	bo = &vp->v_bufobj;
2554 loop:
2555 	/*
2556 	 * Block if we are low on buffers.   Certain processes are allowed
2557 	 * to completely exhaust the buffer cache.
2558          *
2559          * If this check ever becomes a bottleneck it may be better to
2560          * move it into the else, when gbincore() fails.  At the moment
2561          * it isn't a problem.
2562 	 *
2563 	 * XXX remove if 0 sections (clean this up after its proven)
2564          */
2565 	if (numfreebuffers == 0) {
2566 		if (TD_IS_IDLETHREAD(curthread))
2567 			return NULL;
2568 		mtx_lock(&nblock);
2569 		needsbuffer |= VFS_BIO_NEED_ANY;
2570 		mtx_unlock(&nblock);
2571 	}
2572 
2573 	BO_LOCK(bo);
2574 	bp = gbincore(bo, blkno);
2575 	if (bp != NULL) {
2576 		int lockflags;
2577 		/*
2578 		 * Buffer is in-core.  If the buffer is not busy, it must
2579 		 * be on a queue.
2580 		 */
2581 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2582 
2583 		if (flags & GB_LOCK_NOWAIT)
2584 			lockflags |= LK_NOWAIT;
2585 
2586 		error = BUF_TIMELOCK(bp, lockflags,
2587 		    BO_MTX(bo), "getblk", slpflag, slptimeo);
2588 
2589 		/*
2590 		 * If we slept and got the lock we have to restart in case
2591 		 * the buffer changed identities.
2592 		 */
2593 		if (error == ENOLCK)
2594 			goto loop;
2595 		/* We timed out or were interrupted. */
2596 		else if (error)
2597 			return (NULL);
2598 
2599 		/*
2600 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2601 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2602 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2603 		 * backing VM cache.
2604 		 */
2605 		if (bp->b_flags & B_INVAL)
2606 			bp->b_flags &= ~B_CACHE;
2607 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2608 			bp->b_flags |= B_CACHE;
2609 		bremfree(bp);
2610 
2611 		/*
2612 		 * check for size inconsistancies for non-VMIO case.
2613 		 */
2614 
2615 		if (bp->b_bcount != size) {
2616 			if ((bp->b_flags & B_VMIO) == 0 ||
2617 			    (size > bp->b_kvasize)) {
2618 				if (bp->b_flags & B_DELWRI) {
2619 					/*
2620 					 * If buffer is pinned and caller does
2621 					 * not want sleep  waiting for it to be
2622 					 * unpinned, bail out
2623 					 * */
2624 					if (bp->b_pin_count > 0) {
2625 						if (flags & GB_LOCK_NOWAIT) {
2626 							bqrelse(bp);
2627 							return (NULL);
2628 						} else {
2629 							bunpin_wait(bp);
2630 						}
2631 					}
2632 					bp->b_flags |= B_NOCACHE;
2633 					bwrite(bp);
2634 				} else {
2635 					if (LIST_EMPTY(&bp->b_dep)) {
2636 						bp->b_flags |= B_RELBUF;
2637 						brelse(bp);
2638 					} else {
2639 						bp->b_flags |= B_NOCACHE;
2640 						bwrite(bp);
2641 					}
2642 				}
2643 				goto loop;
2644 			}
2645 		}
2646 
2647 		/*
2648 		 * If the size is inconsistant in the VMIO case, we can resize
2649 		 * the buffer.  This might lead to B_CACHE getting set or
2650 		 * cleared.  If the size has not changed, B_CACHE remains
2651 		 * unchanged from its previous state.
2652 		 */
2653 
2654 		if (bp->b_bcount != size)
2655 			allocbuf(bp, size);
2656 
2657 		KASSERT(bp->b_offset != NOOFFSET,
2658 		    ("getblk: no buffer offset"));
2659 
2660 		/*
2661 		 * A buffer with B_DELWRI set and B_CACHE clear must
2662 		 * be committed before we can return the buffer in
2663 		 * order to prevent the caller from issuing a read
2664 		 * ( due to B_CACHE not being set ) and overwriting
2665 		 * it.
2666 		 *
2667 		 * Most callers, including NFS and FFS, need this to
2668 		 * operate properly either because they assume they
2669 		 * can issue a read if B_CACHE is not set, or because
2670 		 * ( for example ) an uncached B_DELWRI might loop due
2671 		 * to softupdates re-dirtying the buffer.  In the latter
2672 		 * case, B_CACHE is set after the first write completes,
2673 		 * preventing further loops.
2674 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2675 		 * above while extending the buffer, we cannot allow the
2676 		 * buffer to remain with B_CACHE set after the write
2677 		 * completes or it will represent a corrupt state.  To
2678 		 * deal with this we set B_NOCACHE to scrap the buffer
2679 		 * after the write.
2680 		 *
2681 		 * We might be able to do something fancy, like setting
2682 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2683 		 * so the below call doesn't set B_CACHE, but that gets real
2684 		 * confusing.  This is much easier.
2685 		 */
2686 
2687 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2688 			bp->b_flags |= B_NOCACHE;
2689 			bwrite(bp);
2690 			goto loop;
2691 		}
2692 		bp->b_flags &= ~B_DONE;
2693 	} else {
2694 		int bsize, maxsize, vmio;
2695 		off_t offset;
2696 
2697 		/*
2698 		 * Buffer is not in-core, create new buffer.  The buffer
2699 		 * returned by getnewbuf() is locked.  Note that the returned
2700 		 * buffer is also considered valid (not marked B_INVAL).
2701 		 */
2702 		BO_UNLOCK(bo);
2703 		/*
2704 		 * If the user does not want us to create the buffer, bail out
2705 		 * here.
2706 		 */
2707 		if (flags & GB_NOCREAT)
2708 			return NULL;
2709 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
2710 		offset = blkno * bsize;
2711 		vmio = vp->v_object != NULL;
2712 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2713 		maxsize = imax(maxsize, bsize);
2714 
2715 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2716 		if (bp == NULL) {
2717 			if (slpflag || slptimeo)
2718 				return NULL;
2719 			goto loop;
2720 		}
2721 
2722 		/*
2723 		 * This code is used to make sure that a buffer is not
2724 		 * created while the getnewbuf routine is blocked.
2725 		 * This can be a problem whether the vnode is locked or not.
2726 		 * If the buffer is created out from under us, we have to
2727 		 * throw away the one we just created.
2728 		 *
2729 		 * Note: this must occur before we associate the buffer
2730 		 * with the vp especially considering limitations in
2731 		 * the splay tree implementation when dealing with duplicate
2732 		 * lblkno's.
2733 		 */
2734 		BO_LOCK(bo);
2735 		if (gbincore(bo, blkno)) {
2736 			BO_UNLOCK(bo);
2737 			bp->b_flags |= B_INVAL;
2738 			brelse(bp);
2739 			goto loop;
2740 		}
2741 
2742 		/*
2743 		 * Insert the buffer into the hash, so that it can
2744 		 * be found by incore.
2745 		 */
2746 		bp->b_blkno = bp->b_lblkno = blkno;
2747 		bp->b_offset = offset;
2748 		bgetvp(vp, bp);
2749 		BO_UNLOCK(bo);
2750 
2751 		/*
2752 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2753 		 * buffer size starts out as 0, B_CACHE will be set by
2754 		 * allocbuf() for the VMIO case prior to it testing the
2755 		 * backing store for validity.
2756 		 */
2757 
2758 		if (vmio) {
2759 			bp->b_flags |= B_VMIO;
2760 #if defined(VFS_BIO_DEBUG)
2761 			if (vn_canvmio(vp) != TRUE)
2762 				printf("getblk: VMIO on vnode type %d\n",
2763 					vp->v_type);
2764 #endif
2765 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2766 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2767 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2768 		} else {
2769 			bp->b_flags &= ~B_VMIO;
2770 			KASSERT(bp->b_bufobj->bo_object == NULL,
2771 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2772 			    bp, bp->b_bufobj->bo_object));
2773 		}
2774 
2775 		allocbuf(bp, size);
2776 		bp->b_flags &= ~B_DONE;
2777 	}
2778 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2779 	BUF_ASSERT_HELD(bp);
2780 	KASSERT(bp->b_bufobj == bo,
2781 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2782 	return (bp);
2783 }
2784 
2785 /*
2786  * Get an empty, disassociated buffer of given size.  The buffer is initially
2787  * set to B_INVAL.
2788  */
2789 struct buf *
2790 geteblk(int size, int flags)
2791 {
2792 	struct buf *bp;
2793 	int maxsize;
2794 
2795 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2796 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2797 		if ((flags & GB_NOWAIT_BD) &&
2798 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
2799 			return (NULL);
2800 	}
2801 	allocbuf(bp, size);
2802 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2803 	BUF_ASSERT_HELD(bp);
2804 	return (bp);
2805 }
2806 
2807 
2808 /*
2809  * This code constitutes the buffer memory from either anonymous system
2810  * memory (in the case of non-VMIO operations) or from an associated
2811  * VM object (in the case of VMIO operations).  This code is able to
2812  * resize a buffer up or down.
2813  *
2814  * Note that this code is tricky, and has many complications to resolve
2815  * deadlock or inconsistant data situations.  Tread lightly!!!
2816  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2817  * the caller.  Calling this code willy nilly can result in the loss of data.
2818  *
2819  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2820  * B_CACHE for the non-VMIO case.
2821  */
2822 
2823 int
2824 allocbuf(struct buf *bp, int size)
2825 {
2826 	int newbsize, mbsize;
2827 	int i;
2828 
2829 	BUF_ASSERT_HELD(bp);
2830 
2831 	if (bp->b_kvasize < size)
2832 		panic("allocbuf: buffer too small");
2833 
2834 	if ((bp->b_flags & B_VMIO) == 0) {
2835 		caddr_t origbuf;
2836 		int origbufsize;
2837 		/*
2838 		 * Just get anonymous memory from the kernel.  Don't
2839 		 * mess with B_CACHE.
2840 		 */
2841 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2842 		if (bp->b_flags & B_MALLOC)
2843 			newbsize = mbsize;
2844 		else
2845 			newbsize = round_page(size);
2846 
2847 		if (newbsize < bp->b_bufsize) {
2848 			/*
2849 			 * malloced buffers are not shrunk
2850 			 */
2851 			if (bp->b_flags & B_MALLOC) {
2852 				if (newbsize) {
2853 					bp->b_bcount = size;
2854 				} else {
2855 					free(bp->b_data, M_BIOBUF);
2856 					if (bp->b_bufsize) {
2857 						atomic_subtract_long(
2858 						    &bufmallocspace,
2859 						    bp->b_bufsize);
2860 						bufspacewakeup();
2861 						bp->b_bufsize = 0;
2862 					}
2863 					bp->b_saveaddr = bp->b_kvabase;
2864 					bp->b_data = bp->b_saveaddr;
2865 					bp->b_bcount = 0;
2866 					bp->b_flags &= ~B_MALLOC;
2867 				}
2868 				return 1;
2869 			}
2870 			vm_hold_free_pages(
2871 			    bp,
2872 			    (vm_offset_t) bp->b_data + newbsize,
2873 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2874 		} else if (newbsize > bp->b_bufsize) {
2875 			/*
2876 			 * We only use malloced memory on the first allocation.
2877 			 * and revert to page-allocated memory when the buffer
2878 			 * grows.
2879 			 */
2880 			/*
2881 			 * There is a potential smp race here that could lead
2882 			 * to bufmallocspace slightly passing the max.  It
2883 			 * is probably extremely rare and not worth worrying
2884 			 * over.
2885 			 */
2886 			if ( (bufmallocspace < maxbufmallocspace) &&
2887 				(bp->b_bufsize == 0) &&
2888 				(mbsize <= PAGE_SIZE/2)) {
2889 
2890 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2891 				bp->b_bufsize = mbsize;
2892 				bp->b_bcount = size;
2893 				bp->b_flags |= B_MALLOC;
2894 				atomic_add_long(&bufmallocspace, mbsize);
2895 				return 1;
2896 			}
2897 			origbuf = NULL;
2898 			origbufsize = 0;
2899 			/*
2900 			 * If the buffer is growing on its other-than-first allocation,
2901 			 * then we revert to the page-allocation scheme.
2902 			 */
2903 			if (bp->b_flags & B_MALLOC) {
2904 				origbuf = bp->b_data;
2905 				origbufsize = bp->b_bufsize;
2906 				bp->b_data = bp->b_kvabase;
2907 				if (bp->b_bufsize) {
2908 					atomic_subtract_long(&bufmallocspace,
2909 					    bp->b_bufsize);
2910 					bufspacewakeup();
2911 					bp->b_bufsize = 0;
2912 				}
2913 				bp->b_flags &= ~B_MALLOC;
2914 				newbsize = round_page(newbsize);
2915 			}
2916 			vm_hold_load_pages(
2917 			    bp,
2918 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2919 			    (vm_offset_t) bp->b_data + newbsize);
2920 			if (origbuf) {
2921 				bcopy(origbuf, bp->b_data, origbufsize);
2922 				free(origbuf, M_BIOBUF);
2923 			}
2924 		}
2925 	} else {
2926 		int desiredpages;
2927 
2928 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2929 		desiredpages = (size == 0) ? 0 :
2930 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2931 
2932 		if (bp->b_flags & B_MALLOC)
2933 			panic("allocbuf: VMIO buffer can't be malloced");
2934 		/*
2935 		 * Set B_CACHE initially if buffer is 0 length or will become
2936 		 * 0-length.
2937 		 */
2938 		if (size == 0 || bp->b_bufsize == 0)
2939 			bp->b_flags |= B_CACHE;
2940 
2941 		if (newbsize < bp->b_bufsize) {
2942 			/*
2943 			 * DEV_BSIZE aligned new buffer size is less then the
2944 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2945 			 * if we have to remove any pages.
2946 			 */
2947 			if (desiredpages < bp->b_npages) {
2948 				vm_page_t m;
2949 
2950 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2951 				for (i = desiredpages; i < bp->b_npages; i++) {
2952 					/*
2953 					 * the page is not freed here -- it
2954 					 * is the responsibility of
2955 					 * vnode_pager_setsize
2956 					 */
2957 					m = bp->b_pages[i];
2958 					KASSERT(m != bogus_page,
2959 					    ("allocbuf: bogus page found"));
2960 					while (vm_page_sleep_if_busy(m, TRUE,
2961 					    "biodep"))
2962 						continue;
2963 
2964 					bp->b_pages[i] = NULL;
2965 					vm_page_lock(m);
2966 					vm_page_unwire(m, 0);
2967 					vm_page_unlock(m);
2968 				}
2969 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2970 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2971 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2972 				bp->b_npages = desiredpages;
2973 			}
2974 		} else if (size > bp->b_bcount) {
2975 			/*
2976 			 * We are growing the buffer, possibly in a
2977 			 * byte-granular fashion.
2978 			 */
2979 			vm_object_t obj;
2980 			vm_offset_t toff;
2981 			vm_offset_t tinc;
2982 
2983 			/*
2984 			 * Step 1, bring in the VM pages from the object,
2985 			 * allocating them if necessary.  We must clear
2986 			 * B_CACHE if these pages are not valid for the
2987 			 * range covered by the buffer.
2988 			 */
2989 
2990 			obj = bp->b_bufobj->bo_object;
2991 
2992 			VM_OBJECT_LOCK(obj);
2993 			while (bp->b_npages < desiredpages) {
2994 				vm_page_t m;
2995 				vm_pindex_t pi;
2996 
2997 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2998 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2999 					/*
3000 					 * note: must allocate system pages
3001 					 * since blocking here could intefere
3002 					 * with paging I/O, no matter which
3003 					 * process we are.
3004 					 */
3005 					m = vm_page_alloc(obj, pi,
3006 					    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
3007 					    VM_ALLOC_WIRED);
3008 					if (m == NULL) {
3009 						atomic_add_int(&vm_pageout_deficit,
3010 						    desiredpages - bp->b_npages);
3011 						VM_OBJECT_UNLOCK(obj);
3012 						VM_WAIT;
3013 						VM_OBJECT_LOCK(obj);
3014 					} else {
3015 						if (m->valid == 0)
3016 							bp->b_flags &= ~B_CACHE;
3017 						bp->b_pages[bp->b_npages] = m;
3018 						++bp->b_npages;
3019 					}
3020 					continue;
3021 				}
3022 
3023 				/*
3024 				 * We found a page.  If we have to sleep on it,
3025 				 * retry because it might have gotten freed out
3026 				 * from under us.
3027 				 *
3028 				 * We can only test VPO_BUSY here.  Blocking on
3029 				 * m->busy might lead to a deadlock:
3030 				 *
3031 				 *  vm_fault->getpages->cluster_read->allocbuf
3032 				 *
3033 				 */
3034 				if ((m->oflags & VPO_BUSY) != 0) {
3035 					/*
3036 					 * Reference the page before unlocking
3037 					 * and sleeping so that the page daemon
3038 					 * is less likely to reclaim it.
3039 					 */
3040 					vm_page_lock_queues();
3041 					vm_page_flag_set(m, PG_REFERENCED);
3042 					vm_page_sleep(m, "pgtblk");
3043 					continue;
3044 				}
3045 
3046 				/*
3047 				 * We have a good page.
3048 				 */
3049 				vm_page_lock(m);
3050 				vm_page_wire(m);
3051 				vm_page_unlock(m);
3052 				bp->b_pages[bp->b_npages] = m;
3053 				++bp->b_npages;
3054 			}
3055 
3056 			/*
3057 			 * Step 2.  We've loaded the pages into the buffer,
3058 			 * we have to figure out if we can still have B_CACHE
3059 			 * set.  Note that B_CACHE is set according to the
3060 			 * byte-granular range ( bcount and size ), new the
3061 			 * aligned range ( newbsize ).
3062 			 *
3063 			 * The VM test is against m->valid, which is DEV_BSIZE
3064 			 * aligned.  Needless to say, the validity of the data
3065 			 * needs to also be DEV_BSIZE aligned.  Note that this
3066 			 * fails with NFS if the server or some other client
3067 			 * extends the file's EOF.  If our buffer is resized,
3068 			 * B_CACHE may remain set! XXX
3069 			 */
3070 
3071 			toff = bp->b_bcount;
3072 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3073 
3074 			while ((bp->b_flags & B_CACHE) && toff < size) {
3075 				vm_pindex_t pi;
3076 
3077 				if (tinc > (size - toff))
3078 					tinc = size - toff;
3079 
3080 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3081 				    PAGE_SHIFT;
3082 
3083 				vfs_buf_test_cache(
3084 				    bp,
3085 				    bp->b_offset,
3086 				    toff,
3087 				    tinc,
3088 				    bp->b_pages[pi]
3089 				);
3090 				toff += tinc;
3091 				tinc = PAGE_SIZE;
3092 			}
3093 			VM_OBJECT_UNLOCK(obj);
3094 
3095 			/*
3096 			 * Step 3, fixup the KVM pmap.  Remember that
3097 			 * bp->b_data is relative to bp->b_offset, but
3098 			 * bp->b_offset may be offset into the first page.
3099 			 */
3100 
3101 			bp->b_data = (caddr_t)
3102 			    trunc_page((vm_offset_t)bp->b_data);
3103 			pmap_qenter(
3104 			    (vm_offset_t)bp->b_data,
3105 			    bp->b_pages,
3106 			    bp->b_npages
3107 			);
3108 
3109 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3110 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
3111 		}
3112 	}
3113 	if (newbsize < bp->b_bufsize)
3114 		bufspacewakeup();
3115 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3116 	bp->b_bcount = size;		/* requested buffer size	*/
3117 	return 1;
3118 }
3119 
3120 void
3121 biodone(struct bio *bp)
3122 {
3123 	struct mtx *mtxp;
3124 	void (*done)(struct bio *);
3125 
3126 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3127 	mtx_lock(mtxp);
3128 	bp->bio_flags |= BIO_DONE;
3129 	done = bp->bio_done;
3130 	if (done == NULL)
3131 		wakeup(bp);
3132 	mtx_unlock(mtxp);
3133 	if (done != NULL)
3134 		done(bp);
3135 }
3136 
3137 /*
3138  * Wait for a BIO to finish.
3139  *
3140  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3141  * case is not yet clear.
3142  */
3143 int
3144 biowait(struct bio *bp, const char *wchan)
3145 {
3146 	struct mtx *mtxp;
3147 
3148 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3149 	mtx_lock(mtxp);
3150 	while ((bp->bio_flags & BIO_DONE) == 0)
3151 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3152 	mtx_unlock(mtxp);
3153 	if (bp->bio_error != 0)
3154 		return (bp->bio_error);
3155 	if (!(bp->bio_flags & BIO_ERROR))
3156 		return (0);
3157 	return (EIO);
3158 }
3159 
3160 void
3161 biofinish(struct bio *bp, struct devstat *stat, int error)
3162 {
3163 
3164 	if (error) {
3165 		bp->bio_error = error;
3166 		bp->bio_flags |= BIO_ERROR;
3167 	}
3168 	if (stat != NULL)
3169 		devstat_end_transaction_bio(stat, bp);
3170 	biodone(bp);
3171 }
3172 
3173 /*
3174  *	bufwait:
3175  *
3176  *	Wait for buffer I/O completion, returning error status.  The buffer
3177  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3178  *	error and cleared.
3179  */
3180 int
3181 bufwait(struct buf *bp)
3182 {
3183 	if (bp->b_iocmd == BIO_READ)
3184 		bwait(bp, PRIBIO, "biord");
3185 	else
3186 		bwait(bp, PRIBIO, "biowr");
3187 	if (bp->b_flags & B_EINTR) {
3188 		bp->b_flags &= ~B_EINTR;
3189 		return (EINTR);
3190 	}
3191 	if (bp->b_ioflags & BIO_ERROR) {
3192 		return (bp->b_error ? bp->b_error : EIO);
3193 	} else {
3194 		return (0);
3195 	}
3196 }
3197 
3198  /*
3199   * Call back function from struct bio back up to struct buf.
3200   */
3201 static void
3202 bufdonebio(struct bio *bip)
3203 {
3204 	struct buf *bp;
3205 
3206 	bp = bip->bio_caller2;
3207 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3208 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3209 	bp->b_ioflags = bip->bio_flags;
3210 	bp->b_error = bip->bio_error;
3211 	if (bp->b_error)
3212 		bp->b_ioflags |= BIO_ERROR;
3213 	bufdone(bp);
3214 	g_destroy_bio(bip);
3215 }
3216 
3217 void
3218 dev_strategy(struct cdev *dev, struct buf *bp)
3219 {
3220 	struct cdevsw *csw;
3221 	struct bio *bip;
3222 
3223 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3224 		panic("b_iocmd botch");
3225 	for (;;) {
3226 		bip = g_new_bio();
3227 		if (bip != NULL)
3228 			break;
3229 		/* Try again later */
3230 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3231 	}
3232 	bip->bio_cmd = bp->b_iocmd;
3233 	bip->bio_offset = bp->b_iooffset;
3234 	bip->bio_length = bp->b_bcount;
3235 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3236 	bip->bio_data = bp->b_data;
3237 	bip->bio_done = bufdonebio;
3238 	bip->bio_caller2 = bp;
3239 	bip->bio_dev = dev;
3240 	KASSERT(dev->si_refcount > 0,
3241 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3242 	    devtoname(dev)));
3243 	csw = dev_refthread(dev);
3244 	if (csw == NULL) {
3245 		g_destroy_bio(bip);
3246 		bp->b_error = ENXIO;
3247 		bp->b_ioflags = BIO_ERROR;
3248 		bufdone(bp);
3249 		return;
3250 	}
3251 	(*csw->d_strategy)(bip);
3252 	dev_relthread(dev);
3253 }
3254 
3255 /*
3256  *	bufdone:
3257  *
3258  *	Finish I/O on a buffer, optionally calling a completion function.
3259  *	This is usually called from an interrupt so process blocking is
3260  *	not allowed.
3261  *
3262  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3263  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3264  *	assuming B_INVAL is clear.
3265  *
3266  *	For the VMIO case, we set B_CACHE if the op was a read and no
3267  *	read error occured, or if the op was a write.  B_CACHE is never
3268  *	set if the buffer is invalid or otherwise uncacheable.
3269  *
3270  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3271  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3272  *	in the biodone routine.
3273  */
3274 void
3275 bufdone(struct buf *bp)
3276 {
3277 	struct bufobj *dropobj;
3278 	void    (*biodone)(struct buf *);
3279 
3280 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3281 	dropobj = NULL;
3282 
3283 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3284 	BUF_ASSERT_HELD(bp);
3285 
3286 	runningbufwakeup(bp);
3287 	if (bp->b_iocmd == BIO_WRITE)
3288 		dropobj = bp->b_bufobj;
3289 	/* call optional completion function if requested */
3290 	if (bp->b_iodone != NULL) {
3291 		biodone = bp->b_iodone;
3292 		bp->b_iodone = NULL;
3293 		(*biodone) (bp);
3294 		if (dropobj)
3295 			bufobj_wdrop(dropobj);
3296 		return;
3297 	}
3298 
3299 	bufdone_finish(bp);
3300 
3301 	if (dropobj)
3302 		bufobj_wdrop(dropobj);
3303 }
3304 
3305 void
3306 bufdone_finish(struct buf *bp)
3307 {
3308 	BUF_ASSERT_HELD(bp);
3309 
3310 	if (!LIST_EMPTY(&bp->b_dep))
3311 		buf_complete(bp);
3312 
3313 	if (bp->b_flags & B_VMIO) {
3314 		int i;
3315 		vm_ooffset_t foff;
3316 		vm_page_t m;
3317 		vm_object_t obj;
3318 		int iosize;
3319 		struct vnode *vp = bp->b_vp;
3320 
3321 		obj = bp->b_bufobj->bo_object;
3322 
3323 #if defined(VFS_BIO_DEBUG)
3324 		mp_fixme("usecount and vflag accessed without locks.");
3325 		if (vp->v_usecount == 0) {
3326 			panic("biodone: zero vnode ref count");
3327 		}
3328 
3329 		KASSERT(vp->v_object != NULL,
3330 			("biodone: vnode %p has no vm_object", vp));
3331 #endif
3332 
3333 		foff = bp->b_offset;
3334 		KASSERT(bp->b_offset != NOOFFSET,
3335 		    ("biodone: no buffer offset"));
3336 
3337 		VM_OBJECT_LOCK(obj);
3338 #if defined(VFS_BIO_DEBUG)
3339 		if (obj->paging_in_progress < bp->b_npages) {
3340 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3341 			    obj->paging_in_progress, bp->b_npages);
3342 		}
3343 #endif
3344 
3345 		/*
3346 		 * Set B_CACHE if the op was a normal read and no error
3347 		 * occured.  B_CACHE is set for writes in the b*write()
3348 		 * routines.
3349 		 */
3350 		iosize = bp->b_bcount - bp->b_resid;
3351 		if (bp->b_iocmd == BIO_READ &&
3352 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3353 		    !(bp->b_ioflags & BIO_ERROR)) {
3354 			bp->b_flags |= B_CACHE;
3355 		}
3356 		for (i = 0; i < bp->b_npages; i++) {
3357 			int bogusflag = 0;
3358 			int resid;
3359 
3360 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3361 			if (resid > iosize)
3362 				resid = iosize;
3363 
3364 			/*
3365 			 * cleanup bogus pages, restoring the originals
3366 			 */
3367 			m = bp->b_pages[i];
3368 			if (m == bogus_page) {
3369 				bogusflag = 1;
3370 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3371 				if (m == NULL)
3372 					panic("biodone: page disappeared!");
3373 				bp->b_pages[i] = m;
3374 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3375 				    bp->b_pages, bp->b_npages);
3376 			}
3377 #if defined(VFS_BIO_DEBUG)
3378 			if (OFF_TO_IDX(foff) != m->pindex) {
3379 				printf(
3380 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3381 				    (intmax_t)foff, (uintmax_t)m->pindex);
3382 			}
3383 #endif
3384 
3385 			/*
3386 			 * In the write case, the valid and clean bits are
3387 			 * already changed correctly ( see bdwrite() ), so we
3388 			 * only need to do this here in the read case.
3389 			 */
3390 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3391 				KASSERT((m->dirty & vm_page_bits(foff &
3392 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3393 				    " page %p has unexpected dirty bits", m));
3394 				vfs_page_set_valid(bp, foff, m);
3395 			}
3396 
3397 			/*
3398 			 * when debugging new filesystems or buffer I/O methods, this
3399 			 * is the most common error that pops up.  if you see this, you
3400 			 * have not set the page busy flag correctly!!!
3401 			 */
3402 			if (m->busy == 0) {
3403 				printf("biodone: page busy < 0, "
3404 				    "pindex: %d, foff: 0x(%x,%x), "
3405 				    "resid: %d, index: %d\n",
3406 				    (int) m->pindex, (int)(foff >> 32),
3407 						(int) foff & 0xffffffff, resid, i);
3408 				if (!vn_isdisk(vp, NULL))
3409 					printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3410 					    (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3411 					    (intmax_t) bp->b_lblkno,
3412 					    bp->b_flags, bp->b_npages);
3413 				else
3414 					printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3415 					    (intmax_t) bp->b_lblkno,
3416 					    bp->b_flags, bp->b_npages);
3417 				printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3418 				    (u_long)m->valid, (u_long)m->dirty,
3419 				    m->wire_count);
3420 				panic("biodone: page busy < 0\n");
3421 			}
3422 			vm_page_io_finish(m);
3423 			vm_object_pip_subtract(obj, 1);
3424 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3425 			iosize -= resid;
3426 		}
3427 		vm_object_pip_wakeupn(obj, 0);
3428 		VM_OBJECT_UNLOCK(obj);
3429 	}
3430 
3431 	/*
3432 	 * For asynchronous completions, release the buffer now. The brelse
3433 	 * will do a wakeup there if necessary - so no need to do a wakeup
3434 	 * here in the async case. The sync case always needs to do a wakeup.
3435 	 */
3436 
3437 	if (bp->b_flags & B_ASYNC) {
3438 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3439 			brelse(bp);
3440 		else
3441 			bqrelse(bp);
3442 	} else
3443 		bdone(bp);
3444 }
3445 
3446 /*
3447  * This routine is called in lieu of iodone in the case of
3448  * incomplete I/O.  This keeps the busy status for pages
3449  * consistant.
3450  */
3451 void
3452 vfs_unbusy_pages(struct buf *bp)
3453 {
3454 	int i;
3455 	vm_object_t obj;
3456 	vm_page_t m;
3457 
3458 	runningbufwakeup(bp);
3459 	if (!(bp->b_flags & B_VMIO))
3460 		return;
3461 
3462 	obj = bp->b_bufobj->bo_object;
3463 	VM_OBJECT_LOCK(obj);
3464 	for (i = 0; i < bp->b_npages; i++) {
3465 		m = bp->b_pages[i];
3466 		if (m == bogus_page) {
3467 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3468 			if (!m)
3469 				panic("vfs_unbusy_pages: page missing\n");
3470 			bp->b_pages[i] = m;
3471 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3472 			    bp->b_pages, bp->b_npages);
3473 		}
3474 		vm_object_pip_subtract(obj, 1);
3475 		vm_page_io_finish(m);
3476 	}
3477 	vm_object_pip_wakeupn(obj, 0);
3478 	VM_OBJECT_UNLOCK(obj);
3479 }
3480 
3481 /*
3482  * vfs_page_set_valid:
3483  *
3484  *	Set the valid bits in a page based on the supplied offset.   The
3485  *	range is restricted to the buffer's size.
3486  *
3487  *	This routine is typically called after a read completes.
3488  */
3489 static void
3490 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3491 {
3492 	vm_ooffset_t eoff;
3493 
3494 	/*
3495 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3496 	 * page boundary and eoff is not greater than the end of the buffer.
3497 	 * The end of the buffer, in this case, is our file EOF, not the
3498 	 * allocation size of the buffer.
3499 	 */
3500 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3501 	if (eoff > bp->b_offset + bp->b_bcount)
3502 		eoff = bp->b_offset + bp->b_bcount;
3503 
3504 	/*
3505 	 * Set valid range.  This is typically the entire buffer and thus the
3506 	 * entire page.
3507 	 */
3508 	if (eoff > off)
3509 		vm_page_set_valid(m, off & PAGE_MASK, eoff - off);
3510 }
3511 
3512 /*
3513  * vfs_page_set_validclean:
3514  *
3515  *	Set the valid bits and clear the dirty bits in a page based on the
3516  *	supplied offset.   The range is restricted to the buffer's size.
3517  */
3518 static void
3519 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3520 {
3521 	vm_ooffset_t soff, eoff;
3522 
3523 	/*
3524 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3525 	 * page boundry or cross the end of the buffer.  The end of the
3526 	 * buffer, in this case, is our file EOF, not the allocation size
3527 	 * of the buffer.
3528 	 */
3529 	soff = off;
3530 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3531 	if (eoff > bp->b_offset + bp->b_bcount)
3532 		eoff = bp->b_offset + bp->b_bcount;
3533 
3534 	/*
3535 	 * Set valid range.  This is typically the entire buffer and thus the
3536 	 * entire page.
3537 	 */
3538 	if (eoff > soff) {
3539 		vm_page_set_validclean(
3540 		    m,
3541 		   (vm_offset_t) (soff & PAGE_MASK),
3542 		   (vm_offset_t) (eoff - soff)
3543 		);
3544 	}
3545 }
3546 
3547 /*
3548  * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3549  * any page is busy, drain the flag.
3550  */
3551 static void
3552 vfs_drain_busy_pages(struct buf *bp)
3553 {
3554 	vm_page_t m;
3555 	int i, last_busied;
3556 
3557 	VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED);
3558 	last_busied = 0;
3559 	for (i = 0; i < bp->b_npages; i++) {
3560 		m = bp->b_pages[i];
3561 		if ((m->oflags & VPO_BUSY) != 0) {
3562 			for (; last_busied < i; last_busied++)
3563 				vm_page_busy(bp->b_pages[last_busied]);
3564 			while ((m->oflags & VPO_BUSY) != 0)
3565 				vm_page_sleep(m, "vbpage");
3566 		}
3567 	}
3568 	for (i = 0; i < last_busied; i++)
3569 		vm_page_wakeup(bp->b_pages[i]);
3570 }
3571 
3572 /*
3573  * This routine is called before a device strategy routine.
3574  * It is used to tell the VM system that paging I/O is in
3575  * progress, and treat the pages associated with the buffer
3576  * almost as being VPO_BUSY.  Also the object paging_in_progress
3577  * flag is handled to make sure that the object doesn't become
3578  * inconsistant.
3579  *
3580  * Since I/O has not been initiated yet, certain buffer flags
3581  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3582  * and should be ignored.
3583  */
3584 void
3585 vfs_busy_pages(struct buf *bp, int clear_modify)
3586 {
3587 	int i, bogus;
3588 	vm_object_t obj;
3589 	vm_ooffset_t foff;
3590 	vm_page_t m;
3591 
3592 	if (!(bp->b_flags & B_VMIO))
3593 		return;
3594 
3595 	obj = bp->b_bufobj->bo_object;
3596 	foff = bp->b_offset;
3597 	KASSERT(bp->b_offset != NOOFFSET,
3598 	    ("vfs_busy_pages: no buffer offset"));
3599 	VM_OBJECT_LOCK(obj);
3600 	vfs_drain_busy_pages(bp);
3601 	if (bp->b_bufsize != 0)
3602 		vfs_setdirty_locked_object(bp);
3603 	bogus = 0;
3604 	for (i = 0; i < bp->b_npages; i++) {
3605 		m = bp->b_pages[i];
3606 
3607 		if ((bp->b_flags & B_CLUSTER) == 0) {
3608 			vm_object_pip_add(obj, 1);
3609 			vm_page_io_start(m);
3610 		}
3611 		/*
3612 		 * When readying a buffer for a read ( i.e
3613 		 * clear_modify == 0 ), it is important to do
3614 		 * bogus_page replacement for valid pages in
3615 		 * partially instantiated buffers.  Partially
3616 		 * instantiated buffers can, in turn, occur when
3617 		 * reconstituting a buffer from its VM backing store
3618 		 * base.  We only have to do this if B_CACHE is
3619 		 * clear ( which causes the I/O to occur in the
3620 		 * first place ).  The replacement prevents the read
3621 		 * I/O from overwriting potentially dirty VM-backed
3622 		 * pages.  XXX bogus page replacement is, uh, bogus.
3623 		 * It may not work properly with small-block devices.
3624 		 * We need to find a better way.
3625 		 */
3626 		if (clear_modify) {
3627 			pmap_remove_write(m);
3628 			vfs_page_set_validclean(bp, foff, m);
3629 		} else if (m->valid == VM_PAGE_BITS_ALL &&
3630 		    (bp->b_flags & B_CACHE) == 0) {
3631 			bp->b_pages[i] = bogus_page;
3632 			bogus++;
3633 		}
3634 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3635 	}
3636 	VM_OBJECT_UNLOCK(obj);
3637 	if (bogus)
3638 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3639 		    bp->b_pages, bp->b_npages);
3640 }
3641 
3642 /*
3643  *	vfs_bio_set_valid:
3644  *
3645  *	Set the range within the buffer to valid.  The range is
3646  *	relative to the beginning of the buffer, b_offset.  Note that
3647  *	b_offset itself may be offset from the beginning of the first
3648  *	page.
3649  */
3650 void
3651 vfs_bio_set_valid(struct buf *bp, int base, int size)
3652 {
3653 	int i, n;
3654 	vm_page_t m;
3655 
3656 	if (!(bp->b_flags & B_VMIO))
3657 		return;
3658 
3659 	/*
3660 	 * Fixup base to be relative to beginning of first page.
3661 	 * Set initial n to be the maximum number of bytes in the
3662 	 * first page that can be validated.
3663 	 */
3664 	base += (bp->b_offset & PAGE_MASK);
3665 	n = PAGE_SIZE - (base & PAGE_MASK);
3666 
3667 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3668 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3669 		m = bp->b_pages[i];
3670 		if (n > size)
3671 			n = size;
3672 		vm_page_set_valid(m, base & PAGE_MASK, n);
3673 		base += n;
3674 		size -= n;
3675 		n = PAGE_SIZE;
3676 	}
3677 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3678 }
3679 
3680 /*
3681  *	vfs_bio_clrbuf:
3682  *
3683  *	If the specified buffer is a non-VMIO buffer, clear the entire
3684  *	buffer.  If the specified buffer is a VMIO buffer, clear and
3685  *	validate only the previously invalid portions of the buffer.
3686  *	This routine essentially fakes an I/O, so we need to clear
3687  *	BIO_ERROR and B_INVAL.
3688  *
3689  *	Note that while we only theoretically need to clear through b_bcount,
3690  *	we go ahead and clear through b_bufsize.
3691  */
3692 void
3693 vfs_bio_clrbuf(struct buf *bp)
3694 {
3695 	int i, j, mask;
3696 	caddr_t sa, ea;
3697 
3698 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3699 		clrbuf(bp);
3700 		return;
3701 	}
3702 	bp->b_flags &= ~B_INVAL;
3703 	bp->b_ioflags &= ~BIO_ERROR;
3704 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3705 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3706 	    (bp->b_offset & PAGE_MASK) == 0) {
3707 		if (bp->b_pages[0] == bogus_page)
3708 			goto unlock;
3709 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3710 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3711 		if ((bp->b_pages[0]->valid & mask) == mask)
3712 			goto unlock;
3713 		if ((bp->b_pages[0]->valid & mask) == 0) {
3714 			bzero(bp->b_data, bp->b_bufsize);
3715 			bp->b_pages[0]->valid |= mask;
3716 			goto unlock;
3717 		}
3718 	}
3719 	ea = sa = bp->b_data;
3720 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3721 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3722 		ea = (caddr_t)(vm_offset_t)ulmin(
3723 		    (u_long)(vm_offset_t)ea,
3724 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3725 		if (bp->b_pages[i] == bogus_page)
3726 			continue;
3727 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3728 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3729 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3730 		if ((bp->b_pages[i]->valid & mask) == mask)
3731 			continue;
3732 		if ((bp->b_pages[i]->valid & mask) == 0)
3733 			bzero(sa, ea - sa);
3734 		else {
3735 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3736 				if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3737 					bzero(sa, DEV_BSIZE);
3738 			}
3739 		}
3740 		bp->b_pages[i]->valid |= mask;
3741 	}
3742 unlock:
3743 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3744 	bp->b_resid = 0;
3745 }
3746 
3747 /*
3748  * vm_hold_load_pages and vm_hold_free_pages get pages into
3749  * a buffers address space.  The pages are anonymous and are
3750  * not associated with a file object.
3751  */
3752 static void
3753 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3754 {
3755 	vm_offset_t pg;
3756 	vm_page_t p;
3757 	int index;
3758 
3759 	to = round_page(to);
3760 	from = round_page(from);
3761 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3762 
3763 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3764 tryagain:
3765 		/*
3766 		 * note: must allocate system pages since blocking here
3767 		 * could interfere with paging I/O, no matter which
3768 		 * process we are.
3769 		 */
3770 		p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ |
3771 		    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3772 		if (!p) {
3773 			atomic_add_int(&vm_pageout_deficit,
3774 			    (to - pg) >> PAGE_SHIFT);
3775 			VM_WAIT;
3776 			goto tryagain;
3777 		}
3778 		pmap_qenter(pg, &p, 1);
3779 		bp->b_pages[index] = p;
3780 	}
3781 	bp->b_npages = index;
3782 }
3783 
3784 /* Return pages associated with this buf to the vm system */
3785 static void
3786 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3787 {
3788 	vm_offset_t pg;
3789 	vm_page_t p;
3790 	int index, newnpages;
3791 
3792 	from = round_page(from);
3793 	to = round_page(to);
3794 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3795 
3796 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3797 		p = bp->b_pages[index];
3798 		if (p && (index < bp->b_npages)) {
3799 			if (p->busy) {
3800 				printf(
3801 			    "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3802 				    (intmax_t)bp->b_blkno,
3803 				    (intmax_t)bp->b_lblkno);
3804 			}
3805 			bp->b_pages[index] = NULL;
3806 			pmap_qremove(pg, 1);
3807 			p->wire_count--;
3808 			vm_page_free(p);
3809 			atomic_subtract_int(&cnt.v_wire_count, 1);
3810 		}
3811 	}
3812 	bp->b_npages = newnpages;
3813 }
3814 
3815 /*
3816  * Map an IO request into kernel virtual address space.
3817  *
3818  * All requests are (re)mapped into kernel VA space.
3819  * Notice that we use b_bufsize for the size of the buffer
3820  * to be mapped.  b_bcount might be modified by the driver.
3821  *
3822  * Note that even if the caller determines that the address space should
3823  * be valid, a race or a smaller-file mapped into a larger space may
3824  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3825  * check the return value.
3826  */
3827 int
3828 vmapbuf(struct buf *bp)
3829 {
3830 	caddr_t addr, kva;
3831 	vm_prot_t prot;
3832 	int pidx, i;
3833 	struct vm_page *m;
3834 	struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3835 
3836 	if (bp->b_bufsize < 0)
3837 		return (-1);
3838 	prot = VM_PROT_READ;
3839 	if (bp->b_iocmd == BIO_READ)
3840 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3841 	for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3842 	     addr < bp->b_data + bp->b_bufsize;
3843 	     addr += PAGE_SIZE, pidx++) {
3844 		/*
3845 		 * Do the vm_fault if needed; do the copy-on-write thing
3846 		 * when reading stuff off device into memory.
3847 		 *
3848 		 * NOTE! Must use pmap_extract() because addr may be in
3849 		 * the userland address space, and kextract is only guarenteed
3850 		 * to work for the kernland address space (see: sparc64 port).
3851 		 */
3852 retry:
3853 		if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3854 		    prot) < 0) {
3855 			for (i = 0; i < pidx; ++i) {
3856 				vm_page_lock(bp->b_pages[i]);
3857 				vm_page_unhold(bp->b_pages[i]);
3858 				vm_page_unlock(bp->b_pages[i]);
3859 				bp->b_pages[i] = NULL;
3860 			}
3861 			return(-1);
3862 		}
3863 		m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3864 		if (m == NULL)
3865 			goto retry;
3866 		bp->b_pages[pidx] = m;
3867 	}
3868 	if (pidx > btoc(MAXPHYS))
3869 		panic("vmapbuf: mapped more than MAXPHYS");
3870 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3871 
3872 	kva = bp->b_saveaddr;
3873 	bp->b_npages = pidx;
3874 	bp->b_saveaddr = bp->b_data;
3875 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3876 	return(0);
3877 }
3878 
3879 /*
3880  * Free the io map PTEs associated with this IO operation.
3881  * We also invalidate the TLB entries and restore the original b_addr.
3882  */
3883 void
3884 vunmapbuf(struct buf *bp)
3885 {
3886 	int pidx;
3887 	int npages;
3888 
3889 	npages = bp->b_npages;
3890 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3891 	for (pidx = 0; pidx < npages; pidx++) {
3892 		vm_page_lock(bp->b_pages[pidx]);
3893 		vm_page_unhold(bp->b_pages[pidx]);
3894 		vm_page_unlock(bp->b_pages[pidx]);
3895 	}
3896 
3897 	bp->b_data = bp->b_saveaddr;
3898 }
3899 
3900 void
3901 bdone(struct buf *bp)
3902 {
3903 	struct mtx *mtxp;
3904 
3905 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3906 	mtx_lock(mtxp);
3907 	bp->b_flags |= B_DONE;
3908 	wakeup(bp);
3909 	mtx_unlock(mtxp);
3910 }
3911 
3912 void
3913 bwait(struct buf *bp, u_char pri, const char *wchan)
3914 {
3915 	struct mtx *mtxp;
3916 
3917 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3918 	mtx_lock(mtxp);
3919 	while ((bp->b_flags & B_DONE) == 0)
3920 		msleep(bp, mtxp, pri, wchan, 0);
3921 	mtx_unlock(mtxp);
3922 }
3923 
3924 int
3925 bufsync(struct bufobj *bo, int waitfor)
3926 {
3927 
3928 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3929 }
3930 
3931 void
3932 bufstrategy(struct bufobj *bo, struct buf *bp)
3933 {
3934 	int i = 0;
3935 	struct vnode *vp;
3936 
3937 	vp = bp->b_vp;
3938 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3939 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3940 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3941 	i = VOP_STRATEGY(vp, bp);
3942 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3943 }
3944 
3945 void
3946 bufobj_wrefl(struct bufobj *bo)
3947 {
3948 
3949 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3950 	ASSERT_BO_LOCKED(bo);
3951 	bo->bo_numoutput++;
3952 }
3953 
3954 void
3955 bufobj_wref(struct bufobj *bo)
3956 {
3957 
3958 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3959 	BO_LOCK(bo);
3960 	bo->bo_numoutput++;
3961 	BO_UNLOCK(bo);
3962 }
3963 
3964 void
3965 bufobj_wdrop(struct bufobj *bo)
3966 {
3967 
3968 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3969 	BO_LOCK(bo);
3970 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3971 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3972 		bo->bo_flag &= ~BO_WWAIT;
3973 		wakeup(&bo->bo_numoutput);
3974 	}
3975 	BO_UNLOCK(bo);
3976 }
3977 
3978 int
3979 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3980 {
3981 	int error;
3982 
3983 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3984 	ASSERT_BO_LOCKED(bo);
3985 	error = 0;
3986 	while (bo->bo_numoutput) {
3987 		bo->bo_flag |= BO_WWAIT;
3988 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3989 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3990 		if (error)
3991 			break;
3992 	}
3993 	return (error);
3994 }
3995 
3996 void
3997 bpin(struct buf *bp)
3998 {
3999 	struct mtx *mtxp;
4000 
4001 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4002 	mtx_lock(mtxp);
4003 	bp->b_pin_count++;
4004 	mtx_unlock(mtxp);
4005 }
4006 
4007 void
4008 bunpin(struct buf *bp)
4009 {
4010 	struct mtx *mtxp;
4011 
4012 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4013 	mtx_lock(mtxp);
4014 	if (--bp->b_pin_count == 0)
4015 		wakeup(bp);
4016 	mtx_unlock(mtxp);
4017 }
4018 
4019 void
4020 bunpin_wait(struct buf *bp)
4021 {
4022 	struct mtx *mtxp;
4023 
4024 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4025 	mtx_lock(mtxp);
4026 	while (bp->b_pin_count > 0)
4027 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4028 	mtx_unlock(mtxp);
4029 }
4030 
4031 #include "opt_ddb.h"
4032 #ifdef DDB
4033 #include <ddb/ddb.h>
4034 
4035 /* DDB command to show buffer data */
4036 DB_SHOW_COMMAND(buffer, db_show_buffer)
4037 {
4038 	/* get args */
4039 	struct buf *bp = (struct buf *)addr;
4040 
4041 	if (!have_addr) {
4042 		db_printf("usage: show buffer <addr>\n");
4043 		return;
4044 	}
4045 
4046 	db_printf("buf at %p\n", bp);
4047 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4048 	db_printf(
4049 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4050 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n",
4051 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4052 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4053 	    bp->b_dep.lh_first);
4054 	if (bp->b_npages) {
4055 		int i;
4056 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4057 		for (i = 0; i < bp->b_npages; i++) {
4058 			vm_page_t m;
4059 			m = bp->b_pages[i];
4060 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4061 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4062 			if ((i + 1) < bp->b_npages)
4063 				db_printf(",");
4064 		}
4065 		db_printf("\n");
4066 	}
4067 	db_printf(" ");
4068 	lockmgr_printinfo(&bp->b_lock);
4069 }
4070 
4071 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4072 {
4073 	struct buf *bp;
4074 	int i;
4075 
4076 	for (i = 0; i < nbuf; i++) {
4077 		bp = &buf[i];
4078 		if (BUF_ISLOCKED(bp)) {
4079 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4080 			db_printf("\n");
4081 		}
4082 	}
4083 }
4084 
4085 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4086 {
4087 	struct vnode *vp;
4088 	struct buf *bp;
4089 
4090 	if (!have_addr) {
4091 		db_printf("usage: show vnodebufs <addr>\n");
4092 		return;
4093 	}
4094 	vp = (struct vnode *)addr;
4095 	db_printf("Clean buffers:\n");
4096 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4097 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4098 		db_printf("\n");
4099 	}
4100 	db_printf("Dirty buffers:\n");
4101 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4102 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4103 		db_printf("\n");
4104 	}
4105 }
4106 #endif /* DDB */
4107