xref: /freebsd/sys/kern/vfs_bio.c (revision 23f282aa31e9b6fceacd449020e936e98d6f2298)
1 /*
2  * Copyright (c) 1994,1997 John S. Dyson
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice immediately at the beginning of the file, without modification,
10  *    this list of conditions, and the following disclaimer.
11  * 2. Absolutely no warranty of function or purpose is made by the author
12  *		John S. Dyson.
13  *
14  * $FreeBSD$
15  */
16 
17 /*
18  * this file contains a new buffer I/O scheme implementing a coherent
19  * VM object and buffer cache scheme.  Pains have been taken to make
20  * sure that the performance degradation associated with schemes such
21  * as this is not realized.
22  *
23  * Author:  John S. Dyson
24  * Significant help during the development and debugging phases
25  * had been provided by David Greenman, also of the FreeBSD core team.
26  *
27  * see man buf(9) for more info.
28  */
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/buf.h>
33 #include <sys/eventhandler.h>
34 #include <sys/lock.h>
35 #include <sys/malloc.h>
36 #include <sys/mount.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
39 #include <sys/proc.h>
40 #include <sys/reboot.h>
41 #include <sys/resourcevar.h>
42 #include <sys/sysctl.h>
43 #include <sys/vmmeter.h>
44 #include <sys/vnode.h>
45 #include <vm/vm.h>
46 #include <vm/vm_param.h>
47 #include <vm/vm_kern.h>
48 #include <vm/vm_pageout.h>
49 #include <vm/vm_page.h>
50 #include <vm/vm_object.h>
51 #include <vm/vm_extern.h>
52 #include <vm/vm_map.h>
53 
54 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
55 
56 struct	bio_ops bioops;		/* I/O operation notification */
57 
58 struct buf *buf;		/* buffer header pool */
59 struct swqueue bswlist;
60 struct simplelock buftimelock;	/* Interlock on setting prio and timo */
61 
62 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
63 		vm_offset_t to);
64 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
65 		vm_offset_t to);
66 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
67 			       int pageno, vm_page_t m);
68 static void vfs_clean_pages(struct buf * bp);
69 static void vfs_setdirty(struct buf *bp);
70 static void vfs_vmio_release(struct buf *bp);
71 static void vfs_backgroundwritedone(struct buf *bp);
72 static int flushbufqueues(void);
73 
74 static int bd_request;
75 
76 static void buf_daemon __P((void));
77 /*
78  * bogus page -- for I/O to/from partially complete buffers
79  * this is a temporary solution to the problem, but it is not
80  * really that bad.  it would be better to split the buffer
81  * for input in the case of buffers partially already in memory,
82  * but the code is intricate enough already.
83  */
84 vm_page_t bogus_page;
85 int runningbufspace;
86 int vmiodirenable = FALSE;
87 static vm_offset_t bogus_offset;
88 
89 static int bufspace, maxbufspace,
90 	bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
91 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
92 static int maxbdrun;
93 static int needsbuffer;
94 static int numdirtybuffers, hidirtybuffers;
95 static int numfreebuffers, lofreebuffers, hifreebuffers;
96 static int getnewbufcalls;
97 static int getnewbufrestarts;
98 
99 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD,
100 	&numdirtybuffers, 0, "");
101 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW,
102 	&hidirtybuffers, 0, "");
103 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD,
104 	&numfreebuffers, 0, "");
105 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW,
106 	&lofreebuffers, 0, "");
107 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW,
108 	&hifreebuffers, 0, "");
109 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD,
110 	&runningbufspace, 0, "");
111 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD,
112 	&maxbufspace, 0, "");
113 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD,
114 	&hibufspace, 0, "");
115 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD,
116 	&lobufspace, 0, "");
117 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD,
118 	&bufspace, 0, "");
119 SYSCTL_INT(_vfs, OID_AUTO, maxbdrun, CTLFLAG_RW,
120 	&maxbdrun, 0, "");
121 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW,
122 	&maxbufmallocspace, 0, "");
123 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD,
124 	&bufmallocspace, 0, "");
125 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW,
126 	&getnewbufcalls, 0, "");
127 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW,
128 	&getnewbufrestarts, 0, "");
129 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW,
130 	&vmiodirenable, 0, "");
131 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW,
132 	&bufdefragcnt, 0, "");
133 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW,
134 	&buffreekvacnt, 0, "");
135 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW,
136 	&bufreusecnt, 0, "");
137 
138 static int bufhashmask;
139 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
140 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } };
141 char *buf_wmesg = BUF_WMESG;
142 
143 extern int vm_swap_size;
144 
145 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
146 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
147 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
148 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
149 
150 /*
151  * Buffer hash table code.  Note that the logical block scans linearly, which
152  * gives us some L1 cache locality.
153  */
154 
155 static __inline
156 struct bufhashhdr *
157 bufhash(struct vnode *vnp, daddr_t bn)
158 {
159 	return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]);
160 }
161 
162 /*
163  *	numdirtywakeup:
164  *
165  *	If someone is blocked due to there being too many dirty buffers,
166  *	and numdirtybuffers is now reasonable, wake them up.
167  */
168 
169 static __inline void
170 numdirtywakeup(void)
171 {
172 	if (numdirtybuffers < hidirtybuffers) {
173 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
174 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
175 			wakeup(&needsbuffer);
176 		}
177 	}
178 }
179 
180 /*
181  *	bufspacewakeup:
182  *
183  *	Called when buffer space is potentially available for recovery.
184  *	getnewbuf() will block on this flag when it is unable to free
185  *	sufficient buffer space.  Buffer space becomes recoverable when
186  *	bp's get placed back in the queues.
187  */
188 
189 static __inline void
190 bufspacewakeup(void)
191 {
192 	/*
193 	 * If someone is waiting for BUF space, wake them up.  Even
194 	 * though we haven't freed the kva space yet, the waiting
195 	 * process will be able to now.
196 	 */
197 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
198 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
199 		wakeup(&needsbuffer);
200 	}
201 }
202 
203 /*
204  *	bufcountwakeup:
205  *
206  *	Called when a buffer has been added to one of the free queues to
207  *	account for the buffer and to wakeup anyone waiting for free buffers.
208  *	This typically occurs when large amounts of metadata are being handled
209  *	by the buffer cache ( else buffer space runs out first, usually ).
210  */
211 
212 static __inline void
213 bufcountwakeup(void)
214 {
215 	++numfreebuffers;
216 	if (needsbuffer) {
217 		needsbuffer &= ~VFS_BIO_NEED_ANY;
218 		if (numfreebuffers >= hifreebuffers)
219 			needsbuffer &= ~VFS_BIO_NEED_FREE;
220 		wakeup(&needsbuffer);
221 	}
222 }
223 
224 /*
225  *	vfs_buf_test_cache:
226  *
227  *	Called when a buffer is extended.  This function clears the B_CACHE
228  *	bit if the newly extended portion of the buffer does not contain
229  *	valid data.
230  */
231 static __inline__
232 void
233 vfs_buf_test_cache(struct buf *bp,
234 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
235 		  vm_page_t m)
236 {
237 	if (bp->b_flags & B_CACHE) {
238 		int base = (foff + off) & PAGE_MASK;
239 		if (vm_page_is_valid(m, base, size) == 0)
240 			bp->b_flags &= ~B_CACHE;
241 	}
242 }
243 
244 static __inline__
245 void
246 bd_wakeup(int dirtybuflevel)
247 {
248 	if (numdirtybuffers >= dirtybuflevel && bd_request == 0) {
249 		bd_request = 1;
250 		wakeup(&bd_request);
251 	}
252 }
253 
254 /*
255  * bd_speedup - speedup the buffer cache flushing code
256  */
257 
258 static __inline__
259 void
260 bd_speedup(void)
261 {
262 	bd_wakeup(1);
263 }
264 
265 /*
266  * Initialize buffer headers and related structures.
267  */
268 
269 caddr_t
270 bufhashinit(caddr_t vaddr)
271 {
272 	/* first, make a null hash table */
273 	for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1)
274 		;
275 	bufhashtbl = (void *)vaddr;
276 	vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask;
277 	--bufhashmask;
278 	return(vaddr);
279 }
280 
281 void
282 bufinit(void)
283 {
284 	struct buf *bp;
285 	int i;
286 
287 	TAILQ_INIT(&bswlist);
288 	LIST_INIT(&invalhash);
289 	simple_lock_init(&buftimelock);
290 
291 	for (i = 0; i <= bufhashmask; i++)
292 		LIST_INIT(&bufhashtbl[i]);
293 
294 	/* next, make a null set of free lists */
295 	for (i = 0; i < BUFFER_QUEUES; i++)
296 		TAILQ_INIT(&bufqueues[i]);
297 
298 	/* finally, initialize each buffer header and stick on empty q */
299 	for (i = 0; i < nbuf; i++) {
300 		bp = &buf[i];
301 		bzero(bp, sizeof *bp);
302 		bp->b_flags = B_INVAL;	/* we're just an empty header */
303 		bp->b_dev = NODEV;
304 		bp->b_rcred = NOCRED;
305 		bp->b_wcred = NOCRED;
306 		bp->b_qindex = QUEUE_EMPTY;
307 		bp->b_xflags = 0;
308 		LIST_INIT(&bp->b_dep);
309 		BUF_LOCKINIT(bp);
310 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
311 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
312 	}
313 
314 	/*
315 	 * maxbufspace is the absolute maximum amount of buffer space we are
316 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
317 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
318 	 * used by most other processes.  The differential is required to
319 	 * ensure that buf_daemon is able to run when other processes might
320 	 * be blocked waiting for buffer space.
321 	 *
322 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
323 	 * this may result in KVM fragmentation which is not handled optimally
324 	 * by the system.
325 	 */
326 	maxbufspace = nbuf * BKVASIZE;
327 	hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
328 	lobufspace = hibufspace - MAXBSIZE;
329 
330 /*
331  * Limit the amount of malloc memory since it is wired permanently into
332  * the kernel space.  Even though this is accounted for in the buffer
333  * allocation, we don't want the malloced region to grow uncontrolled.
334  * The malloc scheme improves memory utilization significantly on average
335  * (small) directories.
336  */
337 	maxbufmallocspace = hibufspace / 20;
338 
339 /*
340  * Reduce the chance of a deadlock occuring by limiting the number
341  * of delayed-write dirty buffers we allow to stack up.
342  */
343 	hidirtybuffers = nbuf / 4 + 20;
344 	numdirtybuffers = 0;
345 /*
346  * To support extreme low-memory systems, make sure hidirtybuffers cannot
347  * eat up all available buffer space.  This occurs when our minimum cannot
348  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
349  * BKVASIZE'd (8K) buffers.
350  */
351 	while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
352 		hidirtybuffers >>= 1;
353 	}
354 
355 /*
356  * Try to keep the number of free buffers in the specified range,
357  * and give special processes (e.g. like buf_daemon) access to an
358  * emergency reserve.
359  */
360 	lofreebuffers = nbuf / 18 + 5;
361 	hifreebuffers = 2 * lofreebuffers;
362 	numfreebuffers = nbuf;
363 
364 /*
365  * Maximum number of async ops initiated per buf_daemon loop.  This is
366  * somewhat of a hack at the moment, we really need to limit ourselves
367  * based on the number of bytes of I/O in-transit that were initiated
368  * from buf_daemon.
369  */
370 	if ((maxbdrun = nswbuf / 4) < 4)
371 		maxbdrun = 4;
372 
373 	bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
374 	bogus_page = vm_page_alloc(kernel_object,
375 			((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
376 			VM_ALLOC_NORMAL);
377 	cnt.v_wire_count++;
378 
379 }
380 
381 /*
382  * bfreekva() - free the kva allocation for a buffer.
383  *
384  *	Must be called at splbio() or higher as this is the only locking for
385  *	buffer_map.
386  *
387  *	Since this call frees up buffer space, we call bufspacewakeup().
388  */
389 static void
390 bfreekva(struct buf * bp)
391 {
392 	if (bp->b_kvasize) {
393 		++buffreekvacnt;
394 		bufspace -= bp->b_kvasize;
395 		vm_map_delete(buffer_map,
396 		    (vm_offset_t) bp->b_kvabase,
397 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize
398 		);
399 		bp->b_kvasize = 0;
400 		bufspacewakeup();
401 	}
402 }
403 
404 /*
405  *	bremfree:
406  *
407  *	Remove the buffer from the appropriate free list.
408  */
409 void
410 bremfree(struct buf * bp)
411 {
412 	int s = splbio();
413 	int old_qindex = bp->b_qindex;
414 
415 	if (bp->b_qindex != QUEUE_NONE) {
416 		KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
417 		TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
418 		bp->b_qindex = QUEUE_NONE;
419 		runningbufspace += bp->b_bufsize;
420 	} else {
421 		if (BUF_REFCNT(bp) <= 1)
422 			panic("bremfree: removing a buffer not on a queue");
423 	}
424 
425 	/*
426 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
427 	 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
428 	 * the buffer was free and we must decrement numfreebuffers.
429 	 */
430 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
431 		switch(old_qindex) {
432 		case QUEUE_DIRTY:
433 		case QUEUE_CLEAN:
434 		case QUEUE_EMPTY:
435 		case QUEUE_EMPTYKVA:
436 			--numfreebuffers;
437 			break;
438 		default:
439 			break;
440 		}
441 	}
442 	splx(s);
443 }
444 
445 
446 /*
447  * Get a buffer with the specified data.  Look in the cache first.  We
448  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
449  * is set, the buffer is valid and we do not have to do anything ( see
450  * getblk() ).
451  */
452 int
453 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
454     struct buf ** bpp)
455 {
456 	struct buf *bp;
457 
458 	bp = getblk(vp, blkno, size, 0, 0);
459 	*bpp = bp;
460 
461 	/* if not found in cache, do some I/O */
462 	if ((bp->b_flags & B_CACHE) == 0) {
463 		if (curproc != NULL)
464 			curproc->p_stats->p_ru.ru_inblock++;
465 		KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
466 		bp->b_iocmd = BIO_READ;
467 		bp->b_flags &= ~B_INVAL;
468 		bp->b_ioflags &= ~BIO_ERROR;
469 		if (bp->b_rcred == NOCRED) {
470 			if (cred != NOCRED)
471 				crhold(cred);
472 			bp->b_rcred = cred;
473 		}
474 		vfs_busy_pages(bp, 0);
475 		VOP_STRATEGY(vp, bp);
476 		return (bufwait(bp));
477 	}
478 	return (0);
479 }
480 
481 /*
482  * Operates like bread, but also starts asynchronous I/O on
483  * read-ahead blocks.  We must clear BIO_ERROR and B_INVAL prior
484  * to initiating I/O . If B_CACHE is set, the buffer is valid
485  * and we do not have to do anything.
486  */
487 int
488 breadn(struct vnode * vp, daddr_t blkno, int size,
489     daddr_t * rablkno, int *rabsize,
490     int cnt, struct ucred * cred, struct buf ** bpp)
491 {
492 	struct buf *bp, *rabp;
493 	int i;
494 	int rv = 0, readwait = 0;
495 
496 	*bpp = bp = getblk(vp, blkno, size, 0, 0);
497 
498 	/* if not found in cache, do some I/O */
499 	if ((bp->b_flags & B_CACHE) == 0) {
500 		if (curproc != NULL)
501 			curproc->p_stats->p_ru.ru_inblock++;
502 		bp->b_iocmd = BIO_READ;
503 		bp->b_flags &= ~B_INVAL;
504 		bp->b_ioflags &= ~BIO_ERROR;
505 		if (bp->b_rcred == NOCRED) {
506 			if (cred != NOCRED)
507 				crhold(cred);
508 			bp->b_rcred = cred;
509 		}
510 		vfs_busy_pages(bp, 0);
511 		VOP_STRATEGY(vp, bp);
512 		++readwait;
513 	}
514 
515 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
516 		if (inmem(vp, *rablkno))
517 			continue;
518 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0);
519 
520 		if ((rabp->b_flags & B_CACHE) == 0) {
521 			if (curproc != NULL)
522 				curproc->p_stats->p_ru.ru_inblock++;
523 			rabp->b_flags |= B_ASYNC;
524 			rabp->b_flags &= ~B_INVAL;
525 			rabp->b_ioflags &= ~BIO_ERROR;
526 			rabp->b_iocmd = BIO_READ;
527 			if (rabp->b_rcred == NOCRED) {
528 				if (cred != NOCRED)
529 					crhold(cred);
530 				rabp->b_rcred = cred;
531 			}
532 			vfs_busy_pages(rabp, 0);
533 			BUF_KERNPROC(rabp);
534 			VOP_STRATEGY(vp, rabp);
535 		} else {
536 			brelse(rabp);
537 		}
538 	}
539 
540 	if (readwait) {
541 		rv = bufwait(bp);
542 	}
543 	return (rv);
544 }
545 
546 /*
547  * Write, release buffer on completion.  (Done by iodone
548  * if async).  Do not bother writing anything if the buffer
549  * is invalid.
550  *
551  * Note that we set B_CACHE here, indicating that buffer is
552  * fully valid and thus cacheable.  This is true even of NFS
553  * now so we set it generally.  This could be set either here
554  * or in biodone() since the I/O is synchronous.  We put it
555  * here.
556  */
557 int
558 bwrite(struct buf * bp)
559 {
560 	int oldflags, s;
561 	struct buf *newbp;
562 
563 	if (bp->b_flags & B_INVAL) {
564 		brelse(bp);
565 		return (0);
566 	}
567 
568 	oldflags = bp->b_flags;
569 
570 	if (BUF_REFCNT(bp) == 0)
571 		panic("bwrite: buffer is not busy???");
572 	s = splbio();
573 	/*
574 	 * If a background write is already in progress, delay
575 	 * writing this block if it is asynchronous. Otherwise
576 	 * wait for the background write to complete.
577 	 */
578 	if (bp->b_xflags & BX_BKGRDINPROG) {
579 		if (bp->b_flags & B_ASYNC) {
580 			splx(s);
581 			bdwrite(bp);
582 			return (0);
583 		}
584 		bp->b_xflags |= BX_BKGRDWAIT;
585 		tsleep(&bp->b_xflags, PRIBIO, "biord", 0);
586 		if (bp->b_xflags & BX_BKGRDINPROG)
587 			panic("bwrite: still writing");
588 	}
589 
590 	/* Mark the buffer clean */
591 	bundirty(bp);
592 
593 	/*
594 	 * If this buffer is marked for background writing and we
595 	 * do not have to wait for it, make a copy and write the
596 	 * copy so as to leave this buffer ready for further use.
597 	 */
598 	if ((bp->b_xflags & BX_BKGRDWRITE) && (bp->b_flags & B_ASYNC)) {
599 		if (bp->b_iodone != NULL) {
600 			printf("bp->b_iodone = %p\n", bp->b_iodone);
601 			panic("bwrite: need chained iodone");
602 		}
603 
604 		/* get a new block */
605 		newbp = geteblk(bp->b_bufsize);
606 
607 		/* set it to be identical to the old block */
608 		memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
609 		bgetvp(bp->b_vp, newbp);
610 		newbp->b_lblkno = bp->b_lblkno;
611 		newbp->b_blkno = bp->b_blkno;
612 		newbp->b_offset = bp->b_offset;
613 		newbp->b_iodone = vfs_backgroundwritedone;
614 		newbp->b_flags |= B_ASYNC;
615 		newbp->b_flags &= ~B_INVAL;
616 
617 		/* move over the dependencies */
618 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
619 			(*bioops.io_movedeps)(bp, newbp);
620 
621 		/*
622 		 * Initiate write on the copy, release the original to
623 		 * the B_LOCKED queue so that it cannot go away until
624 		 * the background write completes. If not locked it could go
625 		 * away and then be reconstituted while it was being written.
626 		 * If the reconstituted buffer were written, we could end up
627 		 * with two background copies being written at the same time.
628 		 */
629 		bp->b_xflags |= BX_BKGRDINPROG;
630 		bp->b_flags |= B_LOCKED;
631 		bqrelse(bp);
632 		bp = newbp;
633 	}
634 
635 	bp->b_flags &= ~B_DONE;
636 	bp->b_ioflags &= ~BIO_ERROR;
637 	bp->b_flags |= B_WRITEINPROG | B_CACHE;
638 	bp->b_iocmd = BIO_WRITE;
639 
640 	bp->b_vp->v_numoutput++;
641 	vfs_busy_pages(bp, 1);
642 	if (curproc != NULL)
643 		curproc->p_stats->p_ru.ru_oublock++;
644 	splx(s);
645 	if (oldflags & B_ASYNC)
646 		BUF_KERNPROC(bp);
647 	BUF_STRATEGY(bp);
648 
649 	if ((oldflags & B_ASYNC) == 0) {
650 		int rtval = bufwait(bp);
651 		brelse(bp);
652 		return (rtval);
653 	}
654 
655 	return (0);
656 }
657 
658 /*
659  * Complete a background write started from bwrite.
660  */
661 static void
662 vfs_backgroundwritedone(bp)
663 	struct buf *bp;
664 {
665 	struct buf *origbp;
666 
667 	/*
668 	 * Find the original buffer that we are writing.
669 	 */
670 	if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
671 		panic("backgroundwritedone: lost buffer");
672 	/*
673 	 * Process dependencies then return any unfinished ones.
674 	 */
675 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
676 		(*bioops.io_complete)(bp);
677 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
678 		(*bioops.io_movedeps)(bp, origbp);
679 	/*
680 	 * Clear the BX_BKGRDINPROG flag in the original buffer
681 	 * and awaken it if it is waiting for the write to complete.
682 	 */
683 	origbp->b_xflags &= ~BX_BKGRDINPROG;
684 	if (origbp->b_xflags & BX_BKGRDWAIT) {
685 		origbp->b_xflags &= ~BX_BKGRDWAIT;
686 		wakeup(&origbp->b_xflags);
687 	}
688 	/*
689 	 * Clear the B_LOCKED flag and remove it from the locked
690 	 * queue if it currently resides there.
691 	 */
692 	origbp->b_flags &= ~B_LOCKED;
693 	if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
694 		bremfree(origbp);
695 		bqrelse(origbp);
696 	}
697 	/*
698 	 * This buffer is marked B_NOCACHE, so when it is released
699 	 * by biodone, it will be tossed. We mark it with BIO_READ
700 	 * to avoid biodone doing a second vwakeup.
701 	 */
702 	bp->b_flags |= B_NOCACHE;
703 	bp->b_iocmd = BIO_READ;
704 	bp->b_flags &= ~(B_CACHE | B_DONE);
705 	bp->b_iodone = 0;
706 	bufdone(bp);
707 }
708 
709 /*
710  * Delayed write. (Buffer is marked dirty).  Do not bother writing
711  * anything if the buffer is marked invalid.
712  *
713  * Note that since the buffer must be completely valid, we can safely
714  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
715  * biodone() in order to prevent getblk from writing the buffer
716  * out synchronously.
717  */
718 void
719 bdwrite(struct buf * bp)
720 {
721 	if (BUF_REFCNT(bp) == 0)
722 		panic("bdwrite: buffer is not busy");
723 
724 	if (bp->b_flags & B_INVAL) {
725 		brelse(bp);
726 		return;
727 	}
728 	bdirty(bp);
729 
730 	/*
731 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
732 	 * true even of NFS now.
733 	 */
734 	bp->b_flags |= B_CACHE;
735 
736 	/*
737 	 * This bmap keeps the system from needing to do the bmap later,
738 	 * perhaps when the system is attempting to do a sync.  Since it
739 	 * is likely that the indirect block -- or whatever other datastructure
740 	 * that the filesystem needs is still in memory now, it is a good
741 	 * thing to do this.  Note also, that if the pageout daemon is
742 	 * requesting a sync -- there might not be enough memory to do
743 	 * the bmap then...  So, this is important to do.
744 	 */
745 	if (bp->b_lblkno == bp->b_blkno) {
746 		VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
747 	}
748 
749 	/*
750 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
751 	 */
752 	vfs_setdirty(bp);
753 
754 	/*
755 	 * We need to do this here to satisfy the vnode_pager and the
756 	 * pageout daemon, so that it thinks that the pages have been
757 	 * "cleaned".  Note that since the pages are in a delayed write
758 	 * buffer -- the VFS layer "will" see that the pages get written
759 	 * out on the next sync, or perhaps the cluster will be completed.
760 	 */
761 	vfs_clean_pages(bp);
762 	bqrelse(bp);
763 
764 	/*
765 	 * Wakeup the buffer flushing daemon if we have saturated the
766 	 * buffer cache.
767 	 */
768 
769 	bd_wakeup(hidirtybuffers);
770 
771 	/*
772 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
773 	 * due to the softdep code.
774 	 */
775 }
776 
777 /*
778  *	bdirty:
779  *
780  *	Turn buffer into delayed write request.  We must clear BIO_READ and
781  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
782  *	itself to properly update it in the dirty/clean lists.  We mark it
783  *	B_DONE to ensure that any asynchronization of the buffer properly
784  *	clears B_DONE ( else a panic will occur later ).
785  *
786  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
787  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
788  *	should only be called if the buffer is known-good.
789  *
790  *	Since the buffer is not on a queue, we do not update the numfreebuffers
791  *	count.
792  *
793  *	Must be called at splbio().
794  *	The buffer must be on QUEUE_NONE.
795  */
796 void
797 bdirty(bp)
798 	struct buf *bp;
799 {
800 	KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
801 	bp->b_flags &= ~(B_RELBUF);
802 	bp->b_iocmd = BIO_WRITE;
803 
804 	if ((bp->b_flags & B_DELWRI) == 0) {
805 		bp->b_flags |= B_DONE | B_DELWRI;
806 		reassignbuf(bp, bp->b_vp);
807 		++numdirtybuffers;
808 		bd_wakeup(hidirtybuffers);
809 	}
810 }
811 
812 /*
813  *	bundirty:
814  *
815  *	Clear B_DELWRI for buffer.
816  *
817  *	Since the buffer is not on a queue, we do not update the numfreebuffers
818  *	count.
819  *
820  *	Must be called at splbio().
821  *	The buffer must be on QUEUE_NONE.
822  */
823 
824 void
825 bundirty(bp)
826 	struct buf *bp;
827 {
828 	KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
829 
830 	if (bp->b_flags & B_DELWRI) {
831 		bp->b_flags &= ~B_DELWRI;
832 		reassignbuf(bp, bp->b_vp);
833 		--numdirtybuffers;
834 		numdirtywakeup();
835 	}
836 	/*
837 	 * Since it is now being written, we can clear its deferred write flag.
838 	 */
839 	bp->b_flags &= ~B_DEFERRED;
840 }
841 
842 /*
843  *	bawrite:
844  *
845  *	Asynchronous write.  Start output on a buffer, but do not wait for
846  *	it to complete.  The buffer is released when the output completes.
847  *
848  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
849  *	B_INVAL buffers.  Not us.
850  */
851 void
852 bawrite(struct buf * bp)
853 {
854 	bp->b_flags |= B_ASYNC;
855 	(void) BUF_WRITE(bp);
856 }
857 
858 /*
859  *	bowrite:
860  *
861  *	Ordered write.  Start output on a buffer, and flag it so that the
862  *	device will write it in the order it was queued.  The buffer is
863  *	released when the output completes.  bwrite() ( or the VOP routine
864  *	anyway ) is responsible for handling B_INVAL buffers.
865  */
866 int
867 bowrite(struct buf * bp)
868 {
869 	bp->b_ioflags |= BIO_ORDERED;
870 	bp->b_flags |= B_ASYNC;
871 	return (BUF_WRITE(bp));
872 }
873 
874 /*
875  *	bwillwrite:
876  *
877  *	Called prior to the locking of any vnodes when we are expecting to
878  *	write.  We do not want to starve the buffer cache with too many
879  *	dirty buffers so we block here.  By blocking prior to the locking
880  *	of any vnodes we attempt to avoid the situation where a locked vnode
881  *	prevents the various system daemons from flushing related buffers.
882  */
883 
884 void
885 bwillwrite(void)
886 {
887 	int slop = hidirtybuffers / 10;
888 
889 	if (numdirtybuffers > hidirtybuffers + slop) {
890 		int s;
891 
892 		s = splbio();
893 		while (numdirtybuffers > hidirtybuffers) {
894 			bd_wakeup(hidirtybuffers);
895 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
896 			tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0);
897 		}
898 		splx(s);
899 	}
900 }
901 
902 /*
903  *	brelse:
904  *
905  *	Release a busy buffer and, if requested, free its resources.  The
906  *	buffer will be stashed in the appropriate bufqueue[] allowing it
907  *	to be accessed later as a cache entity or reused for other purposes.
908  */
909 void
910 brelse(struct buf * bp)
911 {
912 	int s;
913 
914 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
915 
916 	s = splbio();
917 
918 	if (bp->b_flags & B_LOCKED)
919 		bp->b_ioflags &= ~BIO_ERROR;
920 
921 	if (bp->b_iocmd == BIO_WRITE &&
922 	    (bp->b_ioflags & BIO_ERROR) &&
923 	    !(bp->b_flags & B_INVAL)) {
924 		/*
925 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
926 		 * pages from being scrapped.  If B_INVAL is set then
927 		 * this case is not run and the next case is run to
928 		 * destroy the buffer.  B_INVAL can occur if the buffer
929 		 * is outside the range supported by the underlying device.
930 		 */
931 		bp->b_ioflags &= ~BIO_ERROR;
932 		bdirty(bp);
933 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
934 	    (bp->b_ioflags & BIO_ERROR) ||
935 	    bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) {
936 		/*
937 		 * Either a failed I/O or we were asked to free or not
938 		 * cache the buffer.
939 		 */
940 		bp->b_flags |= B_INVAL;
941 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
942 			(*bioops.io_deallocate)(bp);
943 		if (bp->b_flags & B_DELWRI) {
944 			--numdirtybuffers;
945 			numdirtywakeup();
946 		}
947 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
948 		if ((bp->b_flags & B_VMIO) == 0) {
949 			if (bp->b_bufsize)
950 				allocbuf(bp, 0);
951 			if (bp->b_vp)
952 				brelvp(bp);
953 		}
954 	}
955 
956 	/*
957 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
958 	 * is called with B_DELWRI set, the underlying pages may wind up
959 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
960 	 * because pages associated with a B_DELWRI bp are marked clean.
961 	 *
962 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
963 	 * if B_DELWRI is set.
964 	 */
965 
966 	if (bp->b_flags & B_DELWRI)
967 		bp->b_flags &= ~B_RELBUF;
968 
969 	/*
970 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
971 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
972 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
973 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
974 	 *
975 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
976 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
977 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
978 	 *
979 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
980 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
981 	 * the commit state and we cannot afford to lose the buffer. If the
982 	 * buffer has a background write in progress, we need to keep it
983 	 * around to prevent it from being reconstituted and starting a second
984 	 * background write.
985 	 */
986 	if ((bp->b_flags & B_VMIO)
987 	    && !(bp->b_vp->v_tag == VT_NFS &&
988 		 !vn_isdisk(bp->b_vp, NULL) &&
989 		 (bp->b_flags & B_DELWRI) &&
990 		 (bp->b_xflags & BX_BKGRDINPROG))
991 	    ) {
992 
993 		int i, j, resid;
994 		vm_page_t m;
995 		off_t foff;
996 		vm_pindex_t poff;
997 		vm_object_t obj;
998 		struct vnode *vp;
999 
1000 		vp = bp->b_vp;
1001 
1002 		/*
1003 		 * Get the base offset and length of the buffer.  Note that
1004 		 * for block sizes that are less then PAGE_SIZE, the b_data
1005 		 * base of the buffer does not represent exactly b_offset and
1006 		 * neither b_offset nor b_size are necessarily page aligned.
1007 		 * Instead, the starting position of b_offset is:
1008 		 *
1009 		 * 	b_data + (b_offset & PAGE_MASK)
1010 		 *
1011 		 * block sizes less then DEV_BSIZE (usually 512) are not
1012 		 * supported due to the page granularity bits (m->valid,
1013 		 * m->dirty, etc...).
1014 		 *
1015 		 * See man buf(9) for more information
1016 		 */
1017 
1018 		resid = bp->b_bufsize;
1019 		foff = bp->b_offset;
1020 
1021 		for (i = 0; i < bp->b_npages; i++) {
1022 			m = bp->b_pages[i];
1023 			vm_page_flag_clear(m, PG_ZERO);
1024 			if (m == bogus_page) {
1025 
1026 				obj = (vm_object_t) vp->v_object;
1027 				poff = OFF_TO_IDX(bp->b_offset);
1028 
1029 				for (j = i; j < bp->b_npages; j++) {
1030 					m = bp->b_pages[j];
1031 					if (m == bogus_page) {
1032 						m = vm_page_lookup(obj, poff + j);
1033 						if (!m) {
1034 							panic("brelse: page missing\n");
1035 						}
1036 						bp->b_pages[j] = m;
1037 					}
1038 				}
1039 
1040 				if ((bp->b_flags & B_INVAL) == 0) {
1041 					pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1042 				}
1043 			}
1044 			if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) {
1045 				int poffset = foff & PAGE_MASK;
1046 				int presid = resid > (PAGE_SIZE - poffset) ?
1047 					(PAGE_SIZE - poffset) : resid;
1048 
1049 				KASSERT(presid >= 0, ("brelse: extra page"));
1050 				vm_page_set_invalid(m, poffset, presid);
1051 			}
1052 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1053 			foff = (foff + PAGE_SIZE) & ~PAGE_MASK;
1054 		}
1055 
1056 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1057 			vfs_vmio_release(bp);
1058 
1059 	} else if (bp->b_flags & B_VMIO) {
1060 
1061 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1062 			vfs_vmio_release(bp);
1063 
1064 	}
1065 
1066 	if (bp->b_qindex != QUEUE_NONE)
1067 		panic("brelse: free buffer onto another queue???");
1068 	if (BUF_REFCNT(bp) > 1) {
1069 		/* Temporary panic to verify exclusive locking */
1070 		/* This panic goes away when we allow shared refs */
1071 		panic("brelse: multiple refs");
1072 		/* do not release to free list */
1073 		BUF_UNLOCK(bp);
1074 		splx(s);
1075 		return;
1076 	}
1077 
1078 	/* enqueue */
1079 
1080 	/* buffers with no memory */
1081 	if (bp->b_bufsize == 0) {
1082 		bp->b_flags |= B_INVAL;
1083 		bp->b_xflags &= ~BX_BKGRDWRITE;
1084 		if (bp->b_xflags & BX_BKGRDINPROG)
1085 			panic("losing buffer 1");
1086 		if (bp->b_kvasize) {
1087 			bp->b_qindex = QUEUE_EMPTYKVA;
1088 		} else {
1089 			bp->b_qindex = QUEUE_EMPTY;
1090 		}
1091 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1092 		LIST_REMOVE(bp, b_hash);
1093 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1094 		bp->b_dev = NODEV;
1095 	/* buffers with junk contents */
1096 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) {
1097 		bp->b_flags |= B_INVAL;
1098 		bp->b_xflags &= ~BX_BKGRDWRITE;
1099 		if (bp->b_xflags & BX_BKGRDINPROG)
1100 			panic("losing buffer 2");
1101 		bp->b_qindex = QUEUE_CLEAN;
1102 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1103 		LIST_REMOVE(bp, b_hash);
1104 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1105 		bp->b_dev = NODEV;
1106 
1107 	/* buffers that are locked */
1108 	} else if (bp->b_flags & B_LOCKED) {
1109 		bp->b_qindex = QUEUE_LOCKED;
1110 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1111 
1112 	/* remaining buffers */
1113 	} else {
1114 		switch(bp->b_flags & (B_DELWRI|B_AGE)) {
1115 		case B_DELWRI | B_AGE:
1116 		    bp->b_qindex = QUEUE_DIRTY;
1117 		    TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1118 		    break;
1119 		case B_DELWRI:
1120 		    bp->b_qindex = QUEUE_DIRTY;
1121 		    TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1122 		    break;
1123 		case B_AGE:
1124 		    bp->b_qindex = QUEUE_CLEAN;
1125 		    TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1126 		    break;
1127 		default:
1128 		    bp->b_qindex = QUEUE_CLEAN;
1129 		    TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1130 		    break;
1131 		}
1132 	}
1133 
1134 	/*
1135 	 * If B_INVAL, clear B_DELWRI.  We've already placed the buffer
1136 	 * on the correct queue.
1137 	 */
1138 	if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) {
1139 		bp->b_flags &= ~B_DELWRI;
1140 		--numdirtybuffers;
1141 		numdirtywakeup();
1142 	}
1143 
1144 	runningbufspace -= bp->b_bufsize;
1145 
1146 	/*
1147 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1148 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1149 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1150 	 * if B_INVAL is set ).
1151 	 */
1152 
1153 	if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1154 		bufcountwakeup();
1155 
1156 	/*
1157 	 * Something we can maybe free.
1158 	 */
1159 
1160 	if (bp->b_bufsize || bp->b_kvasize)
1161 		bufspacewakeup();
1162 
1163 	/* unlock */
1164 	BUF_UNLOCK(bp);
1165 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1166 	bp->b_ioflags &= ~BIO_ORDERED;
1167 	splx(s);
1168 }
1169 
1170 /*
1171  * Release a buffer back to the appropriate queue but do not try to free
1172  * it.
1173  *
1174  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1175  * biodone() to requeue an async I/O on completion.  It is also used when
1176  * known good buffers need to be requeued but we think we may need the data
1177  * again soon.
1178  */
1179 void
1180 bqrelse(struct buf * bp)
1181 {
1182 	int s;
1183 
1184 	s = splbio();
1185 
1186 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1187 
1188 	if (bp->b_qindex != QUEUE_NONE)
1189 		panic("bqrelse: free buffer onto another queue???");
1190 	if (BUF_REFCNT(bp) > 1) {
1191 		/* do not release to free list */
1192 		panic("bqrelse: multiple refs");
1193 		BUF_UNLOCK(bp);
1194 		splx(s);
1195 		return;
1196 	}
1197 	if (bp->b_flags & B_LOCKED) {
1198 		bp->b_ioflags &= ~BIO_ERROR;
1199 		bp->b_qindex = QUEUE_LOCKED;
1200 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1201 		/* buffers with stale but valid contents */
1202 	} else if (bp->b_flags & B_DELWRI) {
1203 		bp->b_qindex = QUEUE_DIRTY;
1204 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1205 	} else {
1206 		bp->b_qindex = QUEUE_CLEAN;
1207 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1208 	}
1209 
1210 	runningbufspace -= bp->b_bufsize;
1211 
1212 	if ((bp->b_flags & B_LOCKED) == 0 &&
1213 	    ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1214 		bufcountwakeup();
1215 	}
1216 
1217 	/*
1218 	 * Something we can maybe wakeup
1219 	 */
1220 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1221 		bufspacewakeup();
1222 
1223 	/* unlock */
1224 	BUF_UNLOCK(bp);
1225 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1226 	bp->b_ioflags &= ~BIO_ORDERED;
1227 	splx(s);
1228 }
1229 
1230 static void
1231 vfs_vmio_release(bp)
1232 	struct buf *bp;
1233 {
1234 	int i, s;
1235 	vm_page_t m;
1236 
1237 	s = splvm();
1238 	for (i = 0; i < bp->b_npages; i++) {
1239 		m = bp->b_pages[i];
1240 		bp->b_pages[i] = NULL;
1241 		/*
1242 		 * In order to keep page LRU ordering consistent, put
1243 		 * everything on the inactive queue.
1244 		 */
1245 		vm_page_unwire(m, 0);
1246 		/*
1247 		 * We don't mess with busy pages, it is
1248 		 * the responsibility of the process that
1249 		 * busied the pages to deal with them.
1250 		 */
1251 		if ((m->flags & PG_BUSY) || (m->busy != 0))
1252 			continue;
1253 
1254 		if (m->wire_count == 0) {
1255 			vm_page_flag_clear(m, PG_ZERO);
1256 			/*
1257 			 * Might as well free the page if we can and it has
1258 			 * no valid data.
1259 			 */
1260 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) {
1261 				vm_page_busy(m);
1262 				vm_page_protect(m, VM_PROT_NONE);
1263 				vm_page_free(m);
1264 			}
1265 		}
1266 	}
1267 	runningbufspace -= bp->b_bufsize;
1268 	splx(s);
1269 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1270 	if (bp->b_bufsize)
1271 		bufspacewakeup();
1272 	bp->b_npages = 0;
1273 	bp->b_bufsize = 0;
1274 	bp->b_flags &= ~B_VMIO;
1275 	if (bp->b_vp)
1276 		brelvp(bp);
1277 }
1278 
1279 /*
1280  * Check to see if a block is currently memory resident.
1281  */
1282 struct buf *
1283 gbincore(struct vnode * vp, daddr_t blkno)
1284 {
1285 	struct buf *bp;
1286 	struct bufhashhdr *bh;
1287 
1288 	bh = bufhash(vp, blkno);
1289 
1290 	/* Search hash chain */
1291 	LIST_FOREACH(bp, bh, b_hash) {
1292 		/* hit */
1293 		if (bp->b_vp == vp && bp->b_lblkno == blkno &&
1294 		    (bp->b_flags & B_INVAL) == 0) {
1295 			break;
1296 		}
1297 	}
1298 	return (bp);
1299 }
1300 
1301 /*
1302  *	vfs_bio_awrite:
1303  *
1304  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1305  *	This is much better then the old way of writing only one buffer at
1306  *	a time.  Note that we may not be presented with the buffers in the
1307  *	correct order, so we search for the cluster in both directions.
1308  */
1309 int
1310 vfs_bio_awrite(struct buf * bp)
1311 {
1312 	int i;
1313 	int j;
1314 	daddr_t lblkno = bp->b_lblkno;
1315 	struct vnode *vp = bp->b_vp;
1316 	int s;
1317 	int ncl;
1318 	struct buf *bpa;
1319 	int nwritten;
1320 	int size;
1321 	int maxcl;
1322 
1323 	s = splbio();
1324 	/*
1325 	 * right now we support clustered writing only to regular files.  If
1326 	 * we find a clusterable block we could be in the middle of a cluster
1327 	 * rather then at the beginning.
1328 	 */
1329 	if ((vp->v_type == VREG) &&
1330 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1331 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1332 
1333 		size = vp->v_mount->mnt_stat.f_iosize;
1334 		maxcl = MAXPHYS / size;
1335 
1336 		for (i = 1; i < maxcl; i++) {
1337 			if ((bpa = gbincore(vp, lblkno + i)) &&
1338 			    BUF_REFCNT(bpa) == 0 &&
1339 			    ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1340 			    (B_DELWRI | B_CLUSTEROK)) &&
1341 			    (bpa->b_bufsize == size)) {
1342 				if ((bpa->b_blkno == bpa->b_lblkno) ||
1343 				    (bpa->b_blkno !=
1344 				     bp->b_blkno + ((i * size) >> DEV_BSHIFT)))
1345 					break;
1346 			} else {
1347 				break;
1348 			}
1349 		}
1350 		for (j = 1; i + j <= maxcl && j <= lblkno; j++) {
1351 			if ((bpa = gbincore(vp, lblkno - j)) &&
1352 			    BUF_REFCNT(bpa) == 0 &&
1353 			    ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1354 			    (B_DELWRI | B_CLUSTEROK)) &&
1355 			    (bpa->b_bufsize == size)) {
1356 				if ((bpa->b_blkno == bpa->b_lblkno) ||
1357 				    (bpa->b_blkno !=
1358 				     bp->b_blkno - ((j * size) >> DEV_BSHIFT)))
1359 					break;
1360 			} else {
1361 				break;
1362 			}
1363 		}
1364 		--j;
1365 		ncl = i + j;
1366 		/*
1367 		 * this is a possible cluster write
1368 		 */
1369 		if (ncl != 1) {
1370 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1371 			splx(s);
1372 			return nwritten;
1373 		}
1374 	}
1375 
1376 	BUF_LOCK(bp, LK_EXCLUSIVE);
1377 	bremfree(bp);
1378 	bp->b_flags |= B_ASYNC;
1379 
1380 	splx(s);
1381 	/*
1382 	 * default (old) behavior, writing out only one block
1383 	 *
1384 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1385 	 */
1386 	nwritten = bp->b_bufsize;
1387 	(void) BUF_WRITE(bp);
1388 
1389 	return nwritten;
1390 }
1391 
1392 /*
1393  *	getnewbuf:
1394  *
1395  *	Find and initialize a new buffer header, freeing up existing buffers
1396  *	in the bufqueues as necessary.  The new buffer is returned locked.
1397  *
1398  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1399  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1400  *
1401  *	We block if:
1402  *		We have insufficient buffer headers
1403  *		We have insufficient buffer space
1404  *		buffer_map is too fragmented ( space reservation fails )
1405  *		If we have to flush dirty buffers ( but we try to avoid this )
1406  *
1407  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1408  *	Instead we ask the buf daemon to do it for us.  We attempt to
1409  *	avoid piecemeal wakeups of the pageout daemon.
1410  */
1411 
1412 static struct buf *
1413 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1414 {
1415 	struct buf *bp;
1416 	struct buf *nbp;
1417 	int defrag = 0;
1418 	int nqindex;
1419 	int isspecial;
1420 	static int flushingbufs;
1421 
1422 	if (curproc && (curproc->p_flag & P_BUFEXHAUST) == 0)
1423 		isspecial = 0;
1424 	else
1425 		isspecial = 1;
1426 
1427 	++getnewbufcalls;
1428 	--getnewbufrestarts;
1429 restart:
1430 	++getnewbufrestarts;
1431 
1432 	/*
1433 	 * Setup for scan.  If we do not have enough free buffers,
1434 	 * we setup a degenerate case that immediately fails.  Note
1435 	 * that if we are specially marked process, we are allowed to
1436 	 * dip into our reserves.
1437 	 *
1438 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1439 	 *
1440 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1441 	 * However, there are a number of cases (defragging, reusing, ...)
1442 	 * where we cannot backup.
1443 	 */
1444 
1445 	if (isspecial == 0 && numfreebuffers < lofreebuffers) {
1446 		/*
1447 		 * This will cause an immediate failure
1448 		 */
1449 		nqindex = QUEUE_CLEAN;
1450 		nbp = NULL;
1451 	} else {
1452 		/*
1453 		 * Locate a buffer which already has KVA assigned.  First
1454 		 * try EMPTYKVA buffers.
1455 		 */
1456 		nqindex = QUEUE_EMPTYKVA;
1457 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1458 
1459 		if (nbp == NULL) {
1460 			/*
1461 			 * If no EMPTYKVA buffers and we are either
1462 			 * defragging or reusing, locate a CLEAN buffer
1463 			 * to free or reuse.  If bufspace useage is low
1464 			 * skip this step so we can allocate a new buffer.
1465 			 */
1466 			if (defrag || bufspace >= lobufspace) {
1467 				nqindex = QUEUE_CLEAN;
1468 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1469 			}
1470 
1471 			/*
1472 			 * Nada.  If we are allowed to allocate an EMPTY
1473 			 * buffer, go get one.
1474 			 */
1475 			if (nbp == NULL && defrag == 0 &&
1476 			    (isspecial || bufspace < hibufspace)) {
1477 				nqindex = QUEUE_EMPTY;
1478 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1479 			}
1480 		}
1481 	}
1482 
1483 	/*
1484 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1485 	 * depending.
1486 	 */
1487 
1488 	while ((bp = nbp) != NULL) {
1489 		int qindex = nqindex;
1490 
1491 		/*
1492 		 * Calculate next bp ( we can only use it if we do not block
1493 		 * or do other fancy things ).
1494 		 */
1495 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1496 			switch(qindex) {
1497 			case QUEUE_EMPTY:
1498 				nqindex = QUEUE_EMPTYKVA;
1499 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1500 					break;
1501 				/* fall through */
1502 			case QUEUE_EMPTYKVA:
1503 				nqindex = QUEUE_CLEAN;
1504 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1505 					break;
1506 				/* fall through */
1507 			case QUEUE_CLEAN:
1508 				/*
1509 				 * nbp is NULL.
1510 				 */
1511 				break;
1512 			}
1513 		}
1514 
1515 		/*
1516 		 * Sanity Checks
1517 		 */
1518 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1519 
1520 		/*
1521 		 * Note: we no longer distinguish between VMIO and non-VMIO
1522 		 * buffers.
1523 		 */
1524 
1525 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1526 
1527 		/*
1528 		 * If we are defragging then we need a buffer with
1529 		 * b_kvasize != 0.  XXX this situation should no longer
1530 		 * occur, if defrag is non-zero the buffer's b_kvasize
1531 		 * should also be non-zero at this point.  XXX
1532 		 */
1533 		if (defrag && bp->b_kvasize == 0) {
1534 			printf("Warning: defrag empty buffer %p\n", bp);
1535 			continue;
1536 		}
1537 
1538 		/*
1539 		 * Start freeing the bp.  This is somewhat involved.  nbp
1540 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1541 		 */
1542 
1543 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1544 			panic("getnewbuf: locked buf");
1545 		bremfree(bp);
1546 
1547 		if (qindex == QUEUE_CLEAN) {
1548 			if (bp->b_flags & B_VMIO) {
1549 				bp->b_flags &= ~B_ASYNC;
1550 				vfs_vmio_release(bp);
1551 			}
1552 			if (bp->b_vp)
1553 				brelvp(bp);
1554 		}
1555 
1556 		/*
1557 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1558 		 * the scan from this point on.
1559 		 *
1560 		 * Get the rest of the buffer freed up.  b_kva* is still
1561 		 * valid after this operation.
1562 		 */
1563 
1564 		if (bp->b_rcred != NOCRED) {
1565 			crfree(bp->b_rcred);
1566 			bp->b_rcred = NOCRED;
1567 		}
1568 		if (bp->b_wcred != NOCRED) {
1569 			crfree(bp->b_wcred);
1570 			bp->b_wcred = NOCRED;
1571 		}
1572 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1573 			(*bioops.io_deallocate)(bp);
1574 		if (bp->b_xflags & BX_BKGRDINPROG)
1575 			panic("losing buffer 3");
1576 		LIST_REMOVE(bp, b_hash);
1577 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1578 
1579 		if (bp->b_bufsize)
1580 			allocbuf(bp, 0);
1581 
1582 		bp->b_flags = 0;
1583 		bp->b_ioflags = 0;
1584 		bp->b_xflags = 0;
1585 		bp->b_dev = NODEV;
1586 		bp->b_vp = NULL;
1587 		bp->b_blkno = bp->b_lblkno = 0;
1588 		bp->b_offset = NOOFFSET;
1589 		bp->b_iodone = 0;
1590 		bp->b_error = 0;
1591 		bp->b_resid = 0;
1592 		bp->b_bcount = 0;
1593 		bp->b_npages = 0;
1594 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1595 
1596 		LIST_INIT(&bp->b_dep);
1597 
1598 		/*
1599 		 * If we are defragging then free the buffer.
1600 		 */
1601 		if (defrag) {
1602 			bp->b_flags |= B_INVAL;
1603 			bfreekva(bp);
1604 			brelse(bp);
1605 			defrag = 0;
1606 			goto restart;
1607 		}
1608 
1609 		/*
1610 		 * If we are a normal process then deal with bufspace
1611 		 * hysteresis.  A normal process tries to keep bufspace
1612 		 * between lobufspace and hibufspace.  Note: if we encounter
1613 		 * a buffer with b_kvasize == 0 then it means we started
1614 		 * our scan on the EMPTY list and should allocate a new
1615 		 * buffer.
1616 		 */
1617 		if (isspecial == 0) {
1618 			if (bufspace > hibufspace)
1619 				flushingbufs = 1;
1620 			if (flushingbufs && bp->b_kvasize != 0) {
1621 				bp->b_flags |= B_INVAL;
1622 				bfreekva(bp);
1623 				brelse(bp);
1624 				goto restart;
1625 			}
1626 			if (bufspace < lobufspace)
1627 				flushingbufs = 0;
1628 		}
1629 		break;
1630 	}
1631 
1632 	/*
1633 	 * If we exhausted our list, sleep as appropriate.  We may have to
1634 	 * wakeup various daemons and write out some dirty buffers.
1635 	 *
1636 	 * Generally we are sleeping due to insufficient buffer space.
1637 	 */
1638 
1639 	if (bp == NULL) {
1640 		int flags;
1641 		char *waitmsg;
1642 
1643 		if (defrag) {
1644 			flags = VFS_BIO_NEED_BUFSPACE;
1645 			waitmsg = "nbufkv";
1646 		} else if (bufspace >= hibufspace) {
1647 			waitmsg = "nbufbs";
1648 			flags = VFS_BIO_NEED_BUFSPACE;
1649 		} else {
1650 			waitmsg = "newbuf";
1651 			flags = VFS_BIO_NEED_ANY;
1652 		}
1653 
1654 		bd_speedup();	/* heeeelp */
1655 
1656 		needsbuffer |= flags;
1657 		while (needsbuffer & flags) {
1658 			if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag,
1659 			    waitmsg, slptimeo))
1660 				return (NULL);
1661 		}
1662 	} else {
1663 		/*
1664 		 * We finally have a valid bp.  We aren't quite out of the
1665 		 * woods, we still have to reserve kva space.  In order
1666 		 * to keep fragmentation sane we only allocate kva in
1667 		 * BKVASIZE chunks.
1668 		 */
1669 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1670 
1671 		if (maxsize != bp->b_kvasize) {
1672 			vm_offset_t addr = 0;
1673 
1674 			bfreekva(bp);
1675 
1676 			if (vm_map_findspace(buffer_map,
1677 				vm_map_min(buffer_map), maxsize, &addr)) {
1678 				/*
1679 				 * Uh oh.  Buffer map is to fragmented.  We
1680 				 * must defragment the map.
1681 				 */
1682 				++bufdefragcnt;
1683 				defrag = 1;
1684 				bp->b_flags |= B_INVAL;
1685 				brelse(bp);
1686 				goto restart;
1687 			}
1688 			if (addr) {
1689 				vm_map_insert(buffer_map, NULL, 0,
1690 					addr, addr + maxsize,
1691 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1692 
1693 				bp->b_kvabase = (caddr_t) addr;
1694 				bp->b_kvasize = maxsize;
1695 				bufspace += bp->b_kvasize;
1696 				++bufreusecnt;
1697 			}
1698 		}
1699 		bp->b_data = bp->b_kvabase;
1700 	}
1701 	return(bp);
1702 }
1703 
1704 /*
1705  *	waitfreebuffers:
1706  *
1707  *	Wait for sufficient free buffers.  Only called from normal processes.
1708  */
1709 
1710 static void
1711 waitfreebuffers(int slpflag, int slptimeo)
1712 {
1713 	while (numfreebuffers < hifreebuffers) {
1714 		if (numfreebuffers >= hifreebuffers)
1715 			break;
1716 		needsbuffer |= VFS_BIO_NEED_FREE;
1717 		if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo))
1718 			break;
1719 	}
1720 }
1721 
1722 /*
1723  *	buf_daemon:
1724  *
1725  *	buffer flushing daemon.  Buffers are normally flushed by the
1726  *	update daemon but if it cannot keep up this process starts to
1727  *	take the load in an attempt to prevent getnewbuf() from blocking.
1728  */
1729 
1730 static struct proc *bufdaemonproc;
1731 static int bd_interval;
1732 static int bd_flushto;
1733 static int bd_flushinc;
1734 
1735 static struct kproc_desc buf_kp = {
1736 	"bufdaemon",
1737 	buf_daemon,
1738 	&bufdaemonproc
1739 };
1740 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1741 
1742 static void
1743 buf_daemon()
1744 {
1745 	int s;
1746 
1747 	/*
1748 	 * This process needs to be suspended prior to shutdown sync.
1749 	 */
1750 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, bufdaemonproc,
1751 	    SHUTDOWN_PRI_LAST);
1752 
1753 	/*
1754 	 * This process is allowed to take the buffer cache to the limit
1755 	 */
1756 	curproc->p_flag |= P_BUFEXHAUST;
1757 	s = splbio();
1758 
1759 	bd_interval = 5 * hz;	/* dynamically adjusted */
1760 	bd_flushto = hidirtybuffers;	/* dynamically adjusted */
1761 	bd_flushinc = 1;
1762 
1763 	for (;;) {
1764 		kproc_suspend_loop(bufdaemonproc);
1765 
1766 		bd_request = 0;
1767 
1768 		/*
1769 		 * Do the flush.  Limit the number of buffers we flush in one
1770 		 * go.  The failure condition occurs when processes are writing
1771 		 * buffers faster then we can dispose of them.  In this case
1772 		 * we may be flushing so often that the previous set of flushes
1773 		 * have not had time to complete, causing us to run out of
1774 		 * physical buffers and block.
1775 		 */
1776 		{
1777 			int runcount = maxbdrun;
1778 
1779 			while (numdirtybuffers > bd_flushto && runcount) {
1780 				--runcount;
1781 				if (flushbufqueues() == 0)
1782 					break;
1783 			}
1784 		}
1785 
1786 		if (bd_request ||
1787 		    tsleep(&bd_request, PVM, "psleep", bd_interval) == 0) {
1788 			/*
1789 			 * Another request is pending or we were woken up
1790 			 * without timing out.  Flush more.
1791 			 */
1792 			--bd_flushto;
1793 			if (bd_flushto >= numdirtybuffers - 5) {
1794 				bd_flushto = numdirtybuffers - 10;
1795 				bd_flushinc = 1;
1796 			}
1797 			if (bd_flushto < 2)
1798 				bd_flushto = 2;
1799 		} else {
1800 			/*
1801 			 * We slept and timed out, we can slow down.
1802 			 */
1803 			bd_flushto += bd_flushinc;
1804 			if (bd_flushto > hidirtybuffers)
1805 				bd_flushto = hidirtybuffers;
1806 			++bd_flushinc;
1807 			if (bd_flushinc > hidirtybuffers / 20 + 1)
1808 				bd_flushinc = hidirtybuffers / 20 + 1;
1809 		}
1810 
1811 		/*
1812 		 * Set the interval on a linear scale based on hidirtybuffers
1813 		 * with a maximum frequency of 1/10 second.
1814 		 */
1815 		bd_interval = bd_flushto * 5 * hz / hidirtybuffers;
1816 		if (bd_interval < hz / 10)
1817 			bd_interval = hz / 10;
1818 	}
1819 }
1820 
1821 /*
1822  *	flushbufqueues:
1823  *
1824  *	Try to flush a buffer in the dirty queue.  We must be careful to
1825  *	free up B_INVAL buffers instead of write them, which NFS is
1826  *	particularly sensitive to.
1827  */
1828 
1829 static int
1830 flushbufqueues(void)
1831 {
1832 	struct buf *bp;
1833 	int r = 0;
1834 
1835 	bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1836 
1837 	while (bp) {
1838 		KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
1839 		if ((bp->b_flags & B_DELWRI) != 0 &&
1840 		    (bp->b_xflags & BX_BKGRDINPROG) == 0) {
1841 			if (bp->b_flags & B_INVAL) {
1842 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1843 					panic("flushbufqueues: locked buf");
1844 				bremfree(bp);
1845 				brelse(bp);
1846 				++r;
1847 				break;
1848 			}
1849 			if (LIST_FIRST(&bp->b_dep) != NULL &&
1850 			    bioops.io_countdeps &&
1851 			    (bp->b_flags & B_DEFERRED) == 0 &&
1852 			    (*bioops.io_countdeps)(bp, 0)) {
1853 				TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY],
1854 				    bp, b_freelist);
1855 				TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY],
1856 				    bp, b_freelist);
1857 				bp->b_flags |= B_DEFERRED;
1858 				bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1859 				continue;
1860 			}
1861 			vfs_bio_awrite(bp);
1862 			++r;
1863 			break;
1864 		}
1865 		bp = TAILQ_NEXT(bp, b_freelist);
1866 	}
1867 	return (r);
1868 }
1869 
1870 /*
1871  * Check to see if a block is currently memory resident.
1872  */
1873 struct buf *
1874 incore(struct vnode * vp, daddr_t blkno)
1875 {
1876 	struct buf *bp;
1877 
1878 	int s = splbio();
1879 	bp = gbincore(vp, blkno);
1880 	splx(s);
1881 	return (bp);
1882 }
1883 
1884 /*
1885  * Returns true if no I/O is needed to access the
1886  * associated VM object.  This is like incore except
1887  * it also hunts around in the VM system for the data.
1888  */
1889 
1890 int
1891 inmem(struct vnode * vp, daddr_t blkno)
1892 {
1893 	vm_object_t obj;
1894 	vm_offset_t toff, tinc, size;
1895 	vm_page_t m;
1896 	vm_ooffset_t off;
1897 
1898 	if (incore(vp, blkno))
1899 		return 1;
1900 	if (vp->v_mount == NULL)
1901 		return 0;
1902 	if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0)
1903 		return 0;
1904 
1905 	obj = vp->v_object;
1906 	size = PAGE_SIZE;
1907 	if (size > vp->v_mount->mnt_stat.f_iosize)
1908 		size = vp->v_mount->mnt_stat.f_iosize;
1909 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
1910 
1911 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
1912 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
1913 		if (!m)
1914 			return 0;
1915 		tinc = size;
1916 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
1917 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
1918 		if (vm_page_is_valid(m,
1919 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
1920 			return 0;
1921 	}
1922 	return 1;
1923 }
1924 
1925 /*
1926  *	vfs_setdirty:
1927  *
1928  *	Sets the dirty range for a buffer based on the status of the dirty
1929  *	bits in the pages comprising the buffer.
1930  *
1931  *	The range is limited to the size of the buffer.
1932  *
1933  *	This routine is primarily used by NFS, but is generalized for the
1934  *	B_VMIO case.
1935  */
1936 static void
1937 vfs_setdirty(struct buf *bp)
1938 {
1939 	int i;
1940 	vm_object_t object;
1941 
1942 	/*
1943 	 * Degenerate case - empty buffer
1944 	 */
1945 
1946 	if (bp->b_bufsize == 0)
1947 		return;
1948 
1949 	/*
1950 	 * We qualify the scan for modified pages on whether the
1951 	 * object has been flushed yet.  The OBJ_WRITEABLE flag
1952 	 * is not cleared simply by protecting pages off.
1953 	 */
1954 
1955 	if ((bp->b_flags & B_VMIO) == 0)
1956 		return;
1957 
1958 	object = bp->b_pages[0]->object;
1959 
1960 	if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
1961 		printf("Warning: object %p writeable but not mightbedirty\n", object);
1962 	if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
1963 		printf("Warning: object %p mightbedirty but not writeable\n", object);
1964 
1965 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
1966 		vm_offset_t boffset;
1967 		vm_offset_t eoffset;
1968 
1969 		/*
1970 		 * test the pages to see if they have been modified directly
1971 		 * by users through the VM system.
1972 		 */
1973 		for (i = 0; i < bp->b_npages; i++) {
1974 			vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
1975 			vm_page_test_dirty(bp->b_pages[i]);
1976 		}
1977 
1978 		/*
1979 		 * Calculate the encompassing dirty range, boffset and eoffset,
1980 		 * (eoffset - boffset) bytes.
1981 		 */
1982 
1983 		for (i = 0; i < bp->b_npages; i++) {
1984 			if (bp->b_pages[i]->dirty)
1985 				break;
1986 		}
1987 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
1988 
1989 		for (i = bp->b_npages - 1; i >= 0; --i) {
1990 			if (bp->b_pages[i]->dirty) {
1991 				break;
1992 			}
1993 		}
1994 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
1995 
1996 		/*
1997 		 * Fit it to the buffer.
1998 		 */
1999 
2000 		if (eoffset > bp->b_bcount)
2001 			eoffset = bp->b_bcount;
2002 
2003 		/*
2004 		 * If we have a good dirty range, merge with the existing
2005 		 * dirty range.
2006 		 */
2007 
2008 		if (boffset < eoffset) {
2009 			if (bp->b_dirtyoff > boffset)
2010 				bp->b_dirtyoff = boffset;
2011 			if (bp->b_dirtyend < eoffset)
2012 				bp->b_dirtyend = eoffset;
2013 		}
2014 	}
2015 }
2016 
2017 /*
2018  *	getblk:
2019  *
2020  *	Get a block given a specified block and offset into a file/device.
2021  *	The buffers B_DONE bit will be cleared on return, making it almost
2022  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2023  *	return.  The caller should clear B_INVAL prior to initiating a
2024  *	READ.
2025  *
2026  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2027  *	an existing buffer.
2028  *
2029  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2030  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2031  *	and then cleared based on the backing VM.  If the previous buffer is
2032  *	non-0-sized but invalid, B_CACHE will be cleared.
2033  *
2034  *	If getblk() must create a new buffer, the new buffer is returned with
2035  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2036  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2037  *	backing VM.
2038  *
2039  *	getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos
2040  *	B_CACHE bit is clear.
2041  *
2042  *	What this means, basically, is that the caller should use B_CACHE to
2043  *	determine whether the buffer is fully valid or not and should clear
2044  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2045  *	the buffer by loading its data area with something, the caller needs
2046  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2047  *	the caller should set B_CACHE ( as an optimization ), else the caller
2048  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2049  *	a write attempt or if it was a successfull read.  If the caller
2050  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2051  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2052  */
2053 struct buf *
2054 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
2055 {
2056 	struct buf *bp;
2057 	int s;
2058 	struct bufhashhdr *bh;
2059 
2060 	if (size > MAXBSIZE)
2061 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2062 
2063 	s = splbio();
2064 loop:
2065 	/*
2066 	 * Block if we are low on buffers.   Certain processes are allowed
2067 	 * to completely exhaust the buffer cache.
2068          *
2069          * If this check ever becomes a bottleneck it may be better to
2070          * move it into the else, when gbincore() fails.  At the moment
2071          * it isn't a problem.
2072          */
2073 	if (!curproc || (curproc->p_flag & P_BUFEXHAUST)) {
2074 		if (numfreebuffers == 0) {
2075 			if (!curproc)
2076 				return NULL;
2077 			needsbuffer |= VFS_BIO_NEED_ANY;
2078 			tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf",
2079 			    slptimeo);
2080 		}
2081 	} else if (numfreebuffers < lofreebuffers) {
2082 		waitfreebuffers(slpflag, slptimeo);
2083 	}
2084 
2085 	if ((bp = gbincore(vp, blkno))) {
2086 		/*
2087 		 * Buffer is in-core.  If the buffer is not busy, it must
2088 		 * be on a queue.
2089 		 */
2090 
2091 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2092 			if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL,
2093 			    "getblk", slpflag, slptimeo) == ENOLCK)
2094 				goto loop;
2095 			splx(s);
2096 			return (struct buf *) NULL;
2097 		}
2098 
2099 		/*
2100 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2101 		 * invalid.  Ohterwise, for a non-VMIO buffer, B_CACHE is set
2102 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2103 		 * backing VM cache.
2104 		 */
2105 		if (bp->b_flags & B_INVAL)
2106 			bp->b_flags &= ~B_CACHE;
2107 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2108 			bp->b_flags |= B_CACHE;
2109 		bremfree(bp);
2110 
2111 		/*
2112 		 * check for size inconsistancies for non-VMIO case.
2113 		 */
2114 
2115 		if (bp->b_bcount != size) {
2116 			if ((bp->b_flags & B_VMIO) == 0 ||
2117 			    (size > bp->b_kvasize)) {
2118 				if (bp->b_flags & B_DELWRI) {
2119 					bp->b_flags |= B_NOCACHE;
2120 					BUF_WRITE(bp);
2121 				} else {
2122 					if ((bp->b_flags & B_VMIO) &&
2123 					   (LIST_FIRST(&bp->b_dep) == NULL)) {
2124 						bp->b_flags |= B_RELBUF;
2125 						brelse(bp);
2126 					} else {
2127 						bp->b_flags |= B_NOCACHE;
2128 						BUF_WRITE(bp);
2129 					}
2130 				}
2131 				goto loop;
2132 			}
2133 		}
2134 
2135 		/*
2136 		 * If the size is inconsistant in the VMIO case, we can resize
2137 		 * the buffer.  This might lead to B_CACHE getting set or
2138 		 * cleared.  If the size has not changed, B_CACHE remains
2139 		 * unchanged from its previous state.
2140 		 */
2141 
2142 		if (bp->b_bcount != size)
2143 			allocbuf(bp, size);
2144 
2145 		KASSERT(bp->b_offset != NOOFFSET,
2146 		    ("getblk: no buffer offset"));
2147 
2148 		/*
2149 		 * A buffer with B_DELWRI set and B_CACHE clear must
2150 		 * be committed before we can return the buffer in
2151 		 * order to prevent the caller from issuing a read
2152 		 * ( due to B_CACHE not being set ) and overwriting
2153 		 * it.
2154 		 *
2155 		 * Most callers, including NFS and FFS, need this to
2156 		 * operate properly either because they assume they
2157 		 * can issue a read if B_CACHE is not set, or because
2158 		 * ( for example ) an uncached B_DELWRI might loop due
2159 		 * to softupdates re-dirtying the buffer.  In the latter
2160 		 * case, B_CACHE is set after the first write completes,
2161 		 * preventing further loops.
2162 		 */
2163 
2164 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2165 			BUF_WRITE(bp);
2166 			goto loop;
2167 		}
2168 
2169 		splx(s);
2170 		bp->b_flags &= ~B_DONE;
2171 	} else {
2172 		/*
2173 		 * Buffer is not in-core, create new buffer.  The buffer
2174 		 * returned by getnewbuf() is locked.  Note that the returned
2175 		 * buffer is also considered valid (not marked B_INVAL).
2176 		 */
2177 		int bsize, maxsize, vmio;
2178 		off_t offset;
2179 
2180 		if (vn_isdisk(vp, NULL))
2181 			bsize = DEV_BSIZE;
2182 		else if (vp->v_mountedhere)
2183 			bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2184 		else if (vp->v_mount)
2185 			bsize = vp->v_mount->mnt_stat.f_iosize;
2186 		else
2187 			bsize = size;
2188 
2189 		offset = (off_t)blkno * bsize;
2190 		vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF);
2191 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2192 		maxsize = imax(maxsize, bsize);
2193 
2194 		if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2195 			if (slpflag || slptimeo) {
2196 				splx(s);
2197 				return NULL;
2198 			}
2199 			goto loop;
2200 		}
2201 
2202 		/*
2203 		 * This code is used to make sure that a buffer is not
2204 		 * created while the getnewbuf routine is blocked.
2205 		 * This can be a problem whether the vnode is locked or not.
2206 		 * If the buffer is created out from under us, we have to
2207 		 * throw away the one we just created.  There is now window
2208 		 * race because we are safely running at splbio() from the
2209 		 * point of the duplicate buffer creation through to here,
2210 		 * and we've locked the buffer.
2211 		 */
2212 		if (gbincore(vp, blkno)) {
2213 			bp->b_flags |= B_INVAL;
2214 			brelse(bp);
2215 			goto loop;
2216 		}
2217 
2218 		/*
2219 		 * Insert the buffer into the hash, so that it can
2220 		 * be found by incore.
2221 		 */
2222 		bp->b_blkno = bp->b_lblkno = blkno;
2223 		bp->b_offset = offset;
2224 
2225 		bgetvp(vp, bp);
2226 		LIST_REMOVE(bp, b_hash);
2227 		bh = bufhash(vp, blkno);
2228 		LIST_INSERT_HEAD(bh, bp, b_hash);
2229 
2230 		/*
2231 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2232 		 * buffer size starts out as 0, B_CACHE will be set by
2233 		 * allocbuf() for the VMIO case prior to it testing the
2234 		 * backing store for validity.
2235 		 */
2236 
2237 		if (vmio) {
2238 			bp->b_flags |= B_VMIO;
2239 #if defined(VFS_BIO_DEBUG)
2240 			if (vp->v_type != VREG && vp->v_type != VBLK)
2241 				printf("getblk: vmioing file type %d???\n", vp->v_type);
2242 #endif
2243 		} else {
2244 			bp->b_flags &= ~B_VMIO;
2245 		}
2246 
2247 		allocbuf(bp, size);
2248 
2249 		splx(s);
2250 		bp->b_flags &= ~B_DONE;
2251 	}
2252 	return (bp);
2253 }
2254 
2255 /*
2256  * Get an empty, disassociated buffer of given size.  The buffer is initially
2257  * set to B_INVAL.
2258  */
2259 struct buf *
2260 geteblk(int size)
2261 {
2262 	struct buf *bp;
2263 	int s;
2264 	int maxsize;
2265 
2266 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2267 
2268 	s = splbio();
2269 	while ((bp = getnewbuf(0, 0, size, maxsize)) == 0);
2270 	splx(s);
2271 	allocbuf(bp, size);
2272 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2273 	return (bp);
2274 }
2275 
2276 
2277 /*
2278  * This code constitutes the buffer memory from either anonymous system
2279  * memory (in the case of non-VMIO operations) or from an associated
2280  * VM object (in the case of VMIO operations).  This code is able to
2281  * resize a buffer up or down.
2282  *
2283  * Note that this code is tricky, and has many complications to resolve
2284  * deadlock or inconsistant data situations.  Tread lightly!!!
2285  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2286  * the caller.  Calling this code willy nilly can result in the loss of data.
2287  *
2288  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2289  * B_CACHE for the non-VMIO case.
2290  */
2291 
2292 int
2293 allocbuf(struct buf *bp, int size)
2294 {
2295 	int newbsize, mbsize;
2296 	int i;
2297 
2298 	if (BUF_REFCNT(bp) == 0)
2299 		panic("allocbuf: buffer not busy");
2300 
2301 	if (bp->b_kvasize < size)
2302 		panic("allocbuf: buffer too small");
2303 
2304 	if ((bp->b_flags & B_VMIO) == 0) {
2305 		caddr_t origbuf;
2306 		int origbufsize;
2307 		/*
2308 		 * Just get anonymous memory from the kernel.  Don't
2309 		 * mess with B_CACHE.
2310 		 */
2311 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2312 #if !defined(NO_B_MALLOC)
2313 		if (bp->b_flags & B_MALLOC)
2314 			newbsize = mbsize;
2315 		else
2316 #endif
2317 			newbsize = round_page(size);
2318 
2319 		if (newbsize < bp->b_bufsize) {
2320 #if !defined(NO_B_MALLOC)
2321 			/*
2322 			 * malloced buffers are not shrunk
2323 			 */
2324 			if (bp->b_flags & B_MALLOC) {
2325 				if (newbsize) {
2326 					bp->b_bcount = size;
2327 				} else {
2328 					free(bp->b_data, M_BIOBUF);
2329 					bufmallocspace -= bp->b_bufsize;
2330 					runningbufspace -= bp->b_bufsize;
2331 					if (bp->b_bufsize)
2332 						bufspacewakeup();
2333 					bp->b_data = bp->b_kvabase;
2334 					bp->b_bufsize = 0;
2335 					bp->b_bcount = 0;
2336 					bp->b_flags &= ~B_MALLOC;
2337 				}
2338 				return 1;
2339 			}
2340 #endif
2341 			vm_hold_free_pages(
2342 			    bp,
2343 			    (vm_offset_t) bp->b_data + newbsize,
2344 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2345 		} else if (newbsize > bp->b_bufsize) {
2346 #if !defined(NO_B_MALLOC)
2347 			/*
2348 			 * We only use malloced memory on the first allocation.
2349 			 * and revert to page-allocated memory when the buffer
2350 			 * grows.
2351 			 */
2352 			if ( (bufmallocspace < maxbufmallocspace) &&
2353 				(bp->b_bufsize == 0) &&
2354 				(mbsize <= PAGE_SIZE/2)) {
2355 
2356 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2357 				bp->b_bufsize = mbsize;
2358 				bp->b_bcount = size;
2359 				bp->b_flags |= B_MALLOC;
2360 				bufmallocspace += mbsize;
2361 				runningbufspace += bp->b_bufsize;
2362 				return 1;
2363 			}
2364 #endif
2365 			origbuf = NULL;
2366 			origbufsize = 0;
2367 #if !defined(NO_B_MALLOC)
2368 			/*
2369 			 * If the buffer is growing on its other-than-first allocation,
2370 			 * then we revert to the page-allocation scheme.
2371 			 */
2372 			if (bp->b_flags & B_MALLOC) {
2373 				origbuf = bp->b_data;
2374 				origbufsize = bp->b_bufsize;
2375 				bp->b_data = bp->b_kvabase;
2376 				bufmallocspace -= bp->b_bufsize;
2377 				runningbufspace -= bp->b_bufsize;
2378 				if (bp->b_bufsize)
2379 					bufspacewakeup();
2380 				bp->b_bufsize = 0;
2381 				bp->b_flags &= ~B_MALLOC;
2382 				newbsize = round_page(newbsize);
2383 			}
2384 #endif
2385 			vm_hold_load_pages(
2386 			    bp,
2387 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2388 			    (vm_offset_t) bp->b_data + newbsize);
2389 #if !defined(NO_B_MALLOC)
2390 			if (origbuf) {
2391 				bcopy(origbuf, bp->b_data, origbufsize);
2392 				free(origbuf, M_BIOBUF);
2393 			}
2394 #endif
2395 		}
2396 	} else {
2397 		vm_page_t m;
2398 		int desiredpages;
2399 
2400 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2401 		desiredpages = (size == 0) ? 0 :
2402 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2403 
2404 #if !defined(NO_B_MALLOC)
2405 		if (bp->b_flags & B_MALLOC)
2406 			panic("allocbuf: VMIO buffer can't be malloced");
2407 #endif
2408 		/*
2409 		 * Set B_CACHE initially if buffer is 0 length or will become
2410 		 * 0-length.
2411 		 */
2412 		if (size == 0 || bp->b_bufsize == 0)
2413 			bp->b_flags |= B_CACHE;
2414 
2415 		if (newbsize < bp->b_bufsize) {
2416 			/*
2417 			 * DEV_BSIZE aligned new buffer size is less then the
2418 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2419 			 * if we have to remove any pages.
2420 			 */
2421 			if (desiredpages < bp->b_npages) {
2422 				for (i = desiredpages; i < bp->b_npages; i++) {
2423 					/*
2424 					 * the page is not freed here -- it
2425 					 * is the responsibility of
2426 					 * vnode_pager_setsize
2427 					 */
2428 					m = bp->b_pages[i];
2429 					KASSERT(m != bogus_page,
2430 					    ("allocbuf: bogus page found"));
2431 					while (vm_page_sleep_busy(m, TRUE, "biodep"))
2432 						;
2433 
2434 					bp->b_pages[i] = NULL;
2435 					vm_page_unwire(m, 0);
2436 				}
2437 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2438 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2439 				bp->b_npages = desiredpages;
2440 			}
2441 		} else if (size > bp->b_bcount) {
2442 			/*
2443 			 * We are growing the buffer, possibly in a
2444 			 * byte-granular fashion.
2445 			 */
2446 			struct vnode *vp;
2447 			vm_object_t obj;
2448 			vm_offset_t toff;
2449 			vm_offset_t tinc;
2450 
2451 			/*
2452 			 * Step 1, bring in the VM pages from the object,
2453 			 * allocating them if necessary.  We must clear
2454 			 * B_CACHE if these pages are not valid for the
2455 			 * range covered by the buffer.
2456 			 */
2457 
2458 			vp = bp->b_vp;
2459 			obj = vp->v_object;
2460 
2461 			while (bp->b_npages < desiredpages) {
2462 				vm_page_t m;
2463 				vm_pindex_t pi;
2464 
2465 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2466 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2467 					m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL);
2468 					if (m == NULL) {
2469 						VM_WAIT;
2470 						vm_pageout_deficit += desiredpages - bp->b_npages;
2471 					} else {
2472 						vm_page_wire(m);
2473 						vm_page_wakeup(m);
2474 						bp->b_flags &= ~B_CACHE;
2475 						bp->b_pages[bp->b_npages] = m;
2476 						++bp->b_npages;
2477 					}
2478 					continue;
2479 				}
2480 
2481 				/*
2482 				 * We found a page.  If we have to sleep on it,
2483 				 * retry because it might have gotten freed out
2484 				 * from under us.
2485 				 *
2486 				 * We can only test PG_BUSY here.  Blocking on
2487 				 * m->busy might lead to a deadlock:
2488 				 *
2489 				 *  vm_fault->getpages->cluster_read->allocbuf
2490 				 *
2491 				 */
2492 
2493 				if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2494 					continue;
2495 
2496 				/*
2497 				 * We have a good page.  Should we wakeup the
2498 				 * page daemon?
2499 				 */
2500 				if ((curproc != pageproc) &&
2501 				    ((m->queue - m->pc) == PQ_CACHE) &&
2502 				    ((cnt.v_free_count + cnt.v_cache_count) <
2503 					(cnt.v_free_min + cnt.v_cache_min))) {
2504 					pagedaemon_wakeup();
2505 				}
2506 				vm_page_flag_clear(m, PG_ZERO);
2507 				vm_page_wire(m);
2508 				bp->b_pages[bp->b_npages] = m;
2509 				++bp->b_npages;
2510 			}
2511 
2512 			/*
2513 			 * Step 2.  We've loaded the pages into the buffer,
2514 			 * we have to figure out if we can still have B_CACHE
2515 			 * set.  Note that B_CACHE is set according to the
2516 			 * byte-granular range ( bcount and size ), new the
2517 			 * aligned range ( newbsize ).
2518 			 *
2519 			 * The VM test is against m->valid, which is DEV_BSIZE
2520 			 * aligned.  Needless to say, the validity of the data
2521 			 * needs to also be DEV_BSIZE aligned.  Note that this
2522 			 * fails with NFS if the server or some other client
2523 			 * extends the file's EOF.  If our buffer is resized,
2524 			 * B_CACHE may remain set! XXX
2525 			 */
2526 
2527 			toff = bp->b_bcount;
2528 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2529 
2530 			while ((bp->b_flags & B_CACHE) && toff < size) {
2531 				vm_pindex_t pi;
2532 
2533 				if (tinc > (size - toff))
2534 					tinc = size - toff;
2535 
2536 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2537 				    PAGE_SHIFT;
2538 
2539 				vfs_buf_test_cache(
2540 				    bp,
2541 				    bp->b_offset,
2542 				    toff,
2543 				    tinc,
2544 				    bp->b_pages[pi]
2545 				);
2546 				toff += tinc;
2547 				tinc = PAGE_SIZE;
2548 			}
2549 
2550 			/*
2551 			 * Step 3, fixup the KVM pmap.  Remember that
2552 			 * bp->b_data is relative to bp->b_offset, but
2553 			 * bp->b_offset may be offset into the first page.
2554 			 */
2555 
2556 			bp->b_data = (caddr_t)
2557 			    trunc_page((vm_offset_t)bp->b_data);
2558 			pmap_qenter(
2559 			    (vm_offset_t)bp->b_data,
2560 			    bp->b_pages,
2561 			    bp->b_npages
2562 			);
2563 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2564 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
2565 		}
2566 	}
2567 	runningbufspace += (newbsize - bp->b_bufsize);
2568 	if (newbsize < bp->b_bufsize)
2569 		bufspacewakeup();
2570 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
2571 	bp->b_bcount = size;		/* requested buffer size	*/
2572 	return 1;
2573 }
2574 
2575 /*
2576  *	bufwait:
2577  *
2578  *	Wait for buffer I/O completion, returning error status.  The buffer
2579  *	is left locked and B_DONE on return.  B_EINTR is converted into a EINTR
2580  *	error and cleared.
2581  */
2582 int
2583 bufwait(register struct buf * bp)
2584 {
2585 	int s;
2586 
2587 	s = splbio();
2588 	while ((bp->b_flags & B_DONE) == 0) {
2589 		if (bp->b_iocmd == BIO_READ)
2590 			tsleep(bp, PRIBIO, "biord", 0);
2591 		else
2592 			tsleep(bp, PRIBIO, "biowr", 0);
2593 	}
2594 	splx(s);
2595 	if (bp->b_flags & B_EINTR) {
2596 		bp->b_flags &= ~B_EINTR;
2597 		return (EINTR);
2598 	}
2599 	if (bp->b_ioflags & BIO_ERROR) {
2600 		return (bp->b_error ? bp->b_error : EIO);
2601 	} else {
2602 		return (0);
2603 	}
2604 }
2605 
2606  /*
2607   * Call back function from struct bio back up to struct buf.
2608   * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY().
2609   */
2610 void
2611 bufdonebio(struct bio *bp)
2612 {
2613 	bufdone(bp->bio_caller2);
2614 }
2615 
2616 /*
2617  *	bufdone:
2618  *
2619  *	Finish I/O on a buffer, optionally calling a completion function.
2620  *	This is usually called from an interrupt so process blocking is
2621  *	not allowed.
2622  *
2623  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2624  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
2625  *	assuming B_INVAL is clear.
2626  *
2627  *	For the VMIO case, we set B_CACHE if the op was a read and no
2628  *	read error occured, or if the op was a write.  B_CACHE is never
2629  *	set if the buffer is invalid or otherwise uncacheable.
2630  *
2631  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
2632  *	initiator to leave B_INVAL set to brelse the buffer out of existance
2633  *	in the biodone routine.
2634  */
2635 void
2636 bufdone(struct buf *bp)
2637 {
2638 	int s;
2639 	void    (*biodone) __P((struct buf *));
2640 
2641 	s = splbio();
2642 
2643 	KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
2644 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
2645 
2646 	bp->b_flags |= B_DONE;
2647 
2648 	if (bp->b_iocmd == BIO_DELETE) {
2649 		brelse(bp);
2650 		splx(s);
2651 		return;
2652 	}
2653 
2654 	if (bp->b_iocmd == BIO_WRITE) {
2655 		vwakeup(bp);
2656 	}
2657 
2658 	/* call optional completion function if requested */
2659 	if (bp->b_iodone != NULL) {
2660 		biodone = bp->b_iodone;
2661 		bp->b_iodone = NULL;
2662 		(*biodone) (bp);
2663 		splx(s);
2664 		return;
2665 	}
2666 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
2667 		(*bioops.io_complete)(bp);
2668 
2669 	if (bp->b_flags & B_VMIO) {
2670 		int i, resid;
2671 		vm_ooffset_t foff;
2672 		vm_page_t m;
2673 		vm_object_t obj;
2674 		int iosize;
2675 		struct vnode *vp = bp->b_vp;
2676 
2677 		obj = vp->v_object;
2678 
2679 #if defined(VFS_BIO_DEBUG)
2680 		if (vp->v_usecount == 0) {
2681 			panic("biodone: zero vnode ref count");
2682 		}
2683 
2684 		if (vp->v_object == NULL) {
2685 			panic("biodone: missing VM object");
2686 		}
2687 
2688 		if ((vp->v_flag & VOBJBUF) == 0) {
2689 			panic("biodone: vnode is not setup for merged cache");
2690 		}
2691 #endif
2692 
2693 		foff = bp->b_offset;
2694 		KASSERT(bp->b_offset != NOOFFSET,
2695 		    ("biodone: no buffer offset"));
2696 
2697 		if (!obj) {
2698 			panic("biodone: no object");
2699 		}
2700 #if defined(VFS_BIO_DEBUG)
2701 		if (obj->paging_in_progress < bp->b_npages) {
2702 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
2703 			    obj->paging_in_progress, bp->b_npages);
2704 		}
2705 #endif
2706 
2707 		/*
2708 		 * Set B_CACHE if the op was a normal read and no error
2709 		 * occured.  B_CACHE is set for writes in the b*write()
2710 		 * routines.
2711 		 */
2712 		iosize = bp->b_bcount - bp->b_resid;
2713 		if (bp->b_iocmd == BIO_READ &&
2714 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
2715 		    !(bp->b_ioflags & BIO_ERROR)) {
2716 			bp->b_flags |= B_CACHE;
2717 		}
2718 
2719 		for (i = 0; i < bp->b_npages; i++) {
2720 			int bogusflag = 0;
2721 			m = bp->b_pages[i];
2722 			if (m == bogus_page) {
2723 				bogusflag = 1;
2724 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2725 				if (!m) {
2726 #if defined(VFS_BIO_DEBUG)
2727 					printf("biodone: page disappeared\n");
2728 #endif
2729 					vm_object_pip_subtract(obj, 1);
2730 					bp->b_flags &= ~B_CACHE;
2731 					continue;
2732 				}
2733 				bp->b_pages[i] = m;
2734 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2735 			}
2736 #if defined(VFS_BIO_DEBUG)
2737 			if (OFF_TO_IDX(foff) != m->pindex) {
2738 				printf(
2739 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
2740 				    (unsigned long)foff, m->pindex);
2741 			}
2742 #endif
2743 			resid = IDX_TO_OFF(m->pindex + 1) - foff;
2744 			if (resid > iosize)
2745 				resid = iosize;
2746 
2747 			/*
2748 			 * In the write case, the valid and clean bits are
2749 			 * already changed correctly ( see bdwrite() ), so we
2750 			 * only need to do this here in the read case.
2751 			 */
2752 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
2753 				vfs_page_set_valid(bp, foff, i, m);
2754 			}
2755 			vm_page_flag_clear(m, PG_ZERO);
2756 
2757 			/*
2758 			 * when debugging new filesystems or buffer I/O methods, this
2759 			 * is the most common error that pops up.  if you see this, you
2760 			 * have not set the page busy flag correctly!!!
2761 			 */
2762 			if (m->busy == 0) {
2763 				printf("biodone: page busy < 0, "
2764 				    "pindex: %d, foff: 0x(%x,%x), "
2765 				    "resid: %d, index: %d\n",
2766 				    (int) m->pindex, (int)(foff >> 32),
2767 						(int) foff & 0xffffffff, resid, i);
2768 				if (!vn_isdisk(vp, NULL))
2769 					printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n",
2770 					    bp->b_vp->v_mount->mnt_stat.f_iosize,
2771 					    (int) bp->b_lblkno,
2772 					    bp->b_flags, bp->b_npages);
2773 				else
2774 					printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n",
2775 					    (int) bp->b_lblkno,
2776 					    bp->b_flags, bp->b_npages);
2777 				printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
2778 				    m->valid, m->dirty, m->wire_count);
2779 				panic("biodone: page busy < 0\n");
2780 			}
2781 			vm_page_io_finish(m);
2782 			vm_object_pip_subtract(obj, 1);
2783 			foff += resid;
2784 			iosize -= resid;
2785 		}
2786 		if (obj)
2787 			vm_object_pip_wakeupn(obj, 0);
2788 	}
2789 	/*
2790 	 * For asynchronous completions, release the buffer now. The brelse
2791 	 * will do a wakeup there if necessary - so no need to do a wakeup
2792 	 * here in the async case. The sync case always needs to do a wakeup.
2793 	 */
2794 
2795 	if (bp->b_flags & B_ASYNC) {
2796 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
2797 			brelse(bp);
2798 		else
2799 			bqrelse(bp);
2800 	} else {
2801 		wakeup(bp);
2802 	}
2803 	splx(s);
2804 }
2805 
2806 /*
2807  * This routine is called in lieu of iodone in the case of
2808  * incomplete I/O.  This keeps the busy status for pages
2809  * consistant.
2810  */
2811 void
2812 vfs_unbusy_pages(struct buf * bp)
2813 {
2814 	int i;
2815 
2816 	if (bp->b_flags & B_VMIO) {
2817 		struct vnode *vp = bp->b_vp;
2818 		vm_object_t obj = vp->v_object;
2819 
2820 		for (i = 0; i < bp->b_npages; i++) {
2821 			vm_page_t m = bp->b_pages[i];
2822 
2823 			if (m == bogus_page) {
2824 				m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
2825 				if (!m) {
2826 					panic("vfs_unbusy_pages: page missing\n");
2827 				}
2828 				bp->b_pages[i] = m;
2829 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2830 			}
2831 			vm_object_pip_subtract(obj, 1);
2832 			vm_page_flag_clear(m, PG_ZERO);
2833 			vm_page_io_finish(m);
2834 		}
2835 		vm_object_pip_wakeupn(obj, 0);
2836 	}
2837 }
2838 
2839 /*
2840  * vfs_page_set_valid:
2841  *
2842  *	Set the valid bits in a page based on the supplied offset.   The
2843  *	range is restricted to the buffer's size.
2844  *
2845  *	This routine is typically called after a read completes.
2846  */
2847 static void
2848 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
2849 {
2850 	vm_ooffset_t soff, eoff;
2851 
2852 	/*
2853 	 * Start and end offsets in buffer.  eoff - soff may not cross a
2854 	 * page boundry or cross the end of the buffer.  The end of the
2855 	 * buffer, in this case, is our file EOF, not the allocation size
2856 	 * of the buffer.
2857 	 */
2858 	soff = off;
2859 	eoff = (off + PAGE_SIZE) & ~PAGE_MASK;
2860 	if (eoff > bp->b_offset + bp->b_bcount)
2861 		eoff = bp->b_offset + bp->b_bcount;
2862 
2863 	/*
2864 	 * Set valid range.  This is typically the entire buffer and thus the
2865 	 * entire page.
2866 	 */
2867 	if (eoff > soff) {
2868 		vm_page_set_validclean(
2869 		    m,
2870 		   (vm_offset_t) (soff & PAGE_MASK),
2871 		   (vm_offset_t) (eoff - soff)
2872 		);
2873 	}
2874 }
2875 
2876 /*
2877  * This routine is called before a device strategy routine.
2878  * It is used to tell the VM system that paging I/O is in
2879  * progress, and treat the pages associated with the buffer
2880  * almost as being PG_BUSY.  Also the object paging_in_progress
2881  * flag is handled to make sure that the object doesn't become
2882  * inconsistant.
2883  *
2884  * Since I/O has not been initiated yet, certain buffer flags
2885  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
2886  * and should be ignored.
2887  */
2888 void
2889 vfs_busy_pages(struct buf * bp, int clear_modify)
2890 {
2891 	int i, bogus;
2892 
2893 	if (bp->b_flags & B_VMIO) {
2894 		struct vnode *vp = bp->b_vp;
2895 		vm_object_t obj = vp->v_object;
2896 		vm_ooffset_t foff;
2897 
2898 		foff = bp->b_offset;
2899 		KASSERT(bp->b_offset != NOOFFSET,
2900 		    ("vfs_busy_pages: no buffer offset"));
2901 		vfs_setdirty(bp);
2902 
2903 retry:
2904 		for (i = 0; i < bp->b_npages; i++) {
2905 			vm_page_t m = bp->b_pages[i];
2906 			if (vm_page_sleep_busy(m, FALSE, "vbpage"))
2907 				goto retry;
2908 		}
2909 
2910 		bogus = 0;
2911 		for (i = 0; i < bp->b_npages; i++) {
2912 			vm_page_t m = bp->b_pages[i];
2913 
2914 			vm_page_flag_clear(m, PG_ZERO);
2915 			if ((bp->b_flags & B_CLUSTER) == 0) {
2916 				vm_object_pip_add(obj, 1);
2917 				vm_page_io_start(m);
2918 			}
2919 
2920 			/*
2921 			 * When readying a buffer for a read ( i.e
2922 			 * clear_modify == 0 ), it is important to do
2923 			 * bogus_page replacement for valid pages in
2924 			 * partially instantiated buffers.  Partially
2925 			 * instantiated buffers can, in turn, occur when
2926 			 * reconstituting a buffer from its VM backing store
2927 			 * base.  We only have to do this if B_CACHE is
2928 			 * clear ( which causes the I/O to occur in the
2929 			 * first place ).  The replacement prevents the read
2930 			 * I/O from overwriting potentially dirty VM-backed
2931 			 * pages.  XXX bogus page replacement is, uh, bogus.
2932 			 * It may not work properly with small-block devices.
2933 			 * We need to find a better way.
2934 			 */
2935 
2936 			vm_page_protect(m, VM_PROT_NONE);
2937 			if (clear_modify)
2938 				vfs_page_set_valid(bp, foff, i, m);
2939 			else if (m->valid == VM_PAGE_BITS_ALL &&
2940 				(bp->b_flags & B_CACHE) == 0) {
2941 				bp->b_pages[i] = bogus_page;
2942 				bogus++;
2943 			}
2944 			foff = (foff + PAGE_SIZE) & ~PAGE_MASK;
2945 		}
2946 		if (bogus)
2947 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2948 	}
2949 }
2950 
2951 /*
2952  * Tell the VM system that the pages associated with this buffer
2953  * are clean.  This is used for delayed writes where the data is
2954  * going to go to disk eventually without additional VM intevention.
2955  *
2956  * Note that while we only really need to clean through to b_bcount, we
2957  * just go ahead and clean through to b_bufsize.
2958  */
2959 static void
2960 vfs_clean_pages(struct buf * bp)
2961 {
2962 	int i;
2963 
2964 	if (bp->b_flags & B_VMIO) {
2965 		vm_ooffset_t foff;
2966 
2967 		foff = bp->b_offset;
2968 		KASSERT(bp->b_offset != NOOFFSET,
2969 		    ("vfs_clean_pages: no buffer offset"));
2970 		for (i = 0; i < bp->b_npages; i++) {
2971 			vm_page_t m = bp->b_pages[i];
2972 			vm_ooffset_t noff = (foff + PAGE_SIZE) & ~PAGE_MASK;
2973 			vm_ooffset_t eoff = noff;
2974 
2975 			if (eoff > bp->b_offset + bp->b_bufsize)
2976 				eoff = bp->b_offset + bp->b_bufsize;
2977 			vfs_page_set_valid(bp, foff, i, m);
2978 			/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2979 			foff = noff;
2980 		}
2981 	}
2982 }
2983 
2984 /*
2985  *	vfs_bio_set_validclean:
2986  *
2987  *	Set the range within the buffer to valid and clean.  The range is
2988  *	relative to the beginning of the buffer, b_offset.  Note that b_offset
2989  *	itself may be offset from the beginning of the first page.
2990  */
2991 
2992 void
2993 vfs_bio_set_validclean(struct buf *bp, int base, int size)
2994 {
2995 	if (bp->b_flags & B_VMIO) {
2996 		int i;
2997 		int n;
2998 
2999 		/*
3000 		 * Fixup base to be relative to beginning of first page.
3001 		 * Set initial n to be the maximum number of bytes in the
3002 		 * first page that can be validated.
3003 		 */
3004 
3005 		base += (bp->b_offset & PAGE_MASK);
3006 		n = PAGE_SIZE - (base & PAGE_MASK);
3007 
3008 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3009 			vm_page_t m = bp->b_pages[i];
3010 
3011 			if (n > size)
3012 				n = size;
3013 
3014 			vm_page_set_validclean(m, base & PAGE_MASK, n);
3015 			base += n;
3016 			size -= n;
3017 			n = PAGE_SIZE;
3018 		}
3019 	}
3020 }
3021 
3022 /*
3023  *	vfs_bio_clrbuf:
3024  *
3025  *	clear a buffer.  This routine essentially fakes an I/O, so we need
3026  *	to clear BIO_ERROR and B_INVAL.
3027  *
3028  *	Note that while we only theoretically need to clear through b_bcount,
3029  *	we go ahead and clear through b_bufsize.
3030  */
3031 
3032 void
3033 vfs_bio_clrbuf(struct buf *bp) {
3034 	int i, mask = 0;
3035 	caddr_t sa, ea;
3036 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3037 		bp->b_flags &= ~B_INVAL;
3038 		bp->b_ioflags &= ~BIO_ERROR;
3039 		if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3040 		    (bp->b_offset & PAGE_MASK) == 0) {
3041 			mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3042 			if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3043 			    ((bp->b_pages[0]->valid & mask) != mask)) {
3044 				bzero(bp->b_data, bp->b_bufsize);
3045 			}
3046 			bp->b_pages[0]->valid |= mask;
3047 			bp->b_resid = 0;
3048 			return;
3049 		}
3050 		ea = sa = bp->b_data;
3051 		for(i=0;i<bp->b_npages;i++,sa=ea) {
3052 			int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3053 			ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3054 			ea = (caddr_t)(vm_offset_t)ulmin(
3055 			    (u_long)(vm_offset_t)ea,
3056 			    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3057 			mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3058 			if ((bp->b_pages[i]->valid & mask) == mask)
3059 				continue;
3060 			if ((bp->b_pages[i]->valid & mask) == 0) {
3061 				if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3062 					bzero(sa, ea - sa);
3063 				}
3064 			} else {
3065 				for (; sa < ea; sa += DEV_BSIZE, j++) {
3066 					if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3067 						(bp->b_pages[i]->valid & (1<<j)) == 0)
3068 						bzero(sa, DEV_BSIZE);
3069 				}
3070 			}
3071 			bp->b_pages[i]->valid |= mask;
3072 			vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
3073 		}
3074 		bp->b_resid = 0;
3075 	} else {
3076 		clrbuf(bp);
3077 	}
3078 }
3079 
3080 /*
3081  * vm_hold_load_pages and vm_hold_unload pages get pages into
3082  * a buffers address space.  The pages are anonymous and are
3083  * not associated with a file object.
3084  */
3085 void
3086 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3087 {
3088 	vm_offset_t pg;
3089 	vm_page_t p;
3090 	int index;
3091 
3092 	to = round_page(to);
3093 	from = round_page(from);
3094 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3095 
3096 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3097 
3098 tryagain:
3099 
3100 		p = vm_page_alloc(kernel_object,
3101 			((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3102 		    VM_ALLOC_NORMAL);
3103 		if (!p) {
3104 			vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3105 			VM_WAIT;
3106 			goto tryagain;
3107 		}
3108 		vm_page_wire(p);
3109 		p->valid = VM_PAGE_BITS_ALL;
3110 		vm_page_flag_clear(p, PG_ZERO);
3111 		pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3112 		bp->b_pages[index] = p;
3113 		vm_page_wakeup(p);
3114 	}
3115 	bp->b_npages = index;
3116 }
3117 
3118 void
3119 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3120 {
3121 	vm_offset_t pg;
3122 	vm_page_t p;
3123 	int index, newnpages;
3124 
3125 	from = round_page(from);
3126 	to = round_page(to);
3127 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3128 
3129 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3130 		p = bp->b_pages[index];
3131 		if (p && (index < bp->b_npages)) {
3132 			if (p->busy) {
3133 				printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n",
3134 					bp->b_blkno, bp->b_lblkno);
3135 			}
3136 			bp->b_pages[index] = NULL;
3137 			pmap_kremove(pg);
3138 			vm_page_busy(p);
3139 			vm_page_unwire(p, 0);
3140 			vm_page_free(p);
3141 		}
3142 	}
3143 	bp->b_npages = newnpages;
3144 }
3145 
3146 
3147 #include "opt_ddb.h"
3148 #ifdef DDB
3149 #include <ddb/ddb.h>
3150 
3151 DB_SHOW_COMMAND(buffer, db_show_buffer)
3152 {
3153 	/* get args */
3154 	struct buf *bp = (struct buf *)addr;
3155 
3156 	if (!have_addr) {
3157 		db_printf("usage: show buffer <addr>\n");
3158 		return;
3159 	}
3160 
3161 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3162 	db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, "
3163 		  "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, "
3164 		  "b_blkno = %d, b_pblkno = %d\n",
3165 		  bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3166 		  major(bp->b_dev), minor(bp->b_dev),
3167 		  bp->b_data, bp->b_blkno, bp->b_pblkno);
3168 	if (bp->b_npages) {
3169 		int i;
3170 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3171 		for (i = 0; i < bp->b_npages; i++) {
3172 			vm_page_t m;
3173 			m = bp->b_pages[i];
3174 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3175 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3176 			if ((i + 1) < bp->b_npages)
3177 				db_printf(",");
3178 		}
3179 		db_printf("\n");
3180 	}
3181 }
3182 #endif /* DDB */
3183