xref: /freebsd/sys/kern/vfs_bio.c (revision 8e6b01171e30297084bb0b4457c4183c2746aacc)
1 /*
2  * Copyright (c) 1994 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. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. Absolutely no warranty of function or purpose is made by the author
15  *    John S. Dyson.
16  * 4. This work was done expressly for inclusion into FreeBSD.  Other use
17  *    is allowed if this notation is included.
18  * 5. Modifications may be freely made to this file if the above conditions
19  *    are met.
20  *
21  * $Id: vfs_bio.c,v 1.66 1995/10/08 00:06:08 swallace Exp $
22  */
23 
24 /*
25  * this file contains a new buffer I/O scheme implementing a coherent
26  * VM object and buffer cache scheme.  Pains have been taken to make
27  * sure that the performance degradation associated with schemes such
28  * as this is not realized.
29  *
30  * Author:  John S. Dyson
31  * Significant help during the development and debugging phases
32  * had been provided by David Greenman, also of the FreeBSD core team.
33  */
34 
35 #define VMIO
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/sysproto.h>
39 #include <sys/kernel.h>
40 #include <sys/proc.h>
41 #include <sys/vnode.h>
42 #include <vm/vm.h>
43 #include <vm/vm_kern.h>
44 #include <vm/vm_pageout.h>
45 #include <vm/vm_page.h>
46 #include <vm/vm_object.h>
47 #include <sys/buf.h>
48 #include <sys/mount.h>
49 #include <sys/malloc.h>
50 #include <sys/resourcevar.h>
51 #include <sys/proc.h>
52 
53 #include <miscfs/specfs/specdev.h>
54 
55 /*
56  * System initialization
57  */
58 
59 static void vfs_update __P((void));
60 struct	proc *updateproc;
61 
62 static struct kproc_desc up_kp = {
63 	"update",
64 	vfs_update,
65 	&updateproc
66 };
67 SYSINIT_KT(update, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
68 
69 
70 struct buf *buf;		/* buffer header pool */
71 struct swqueue bswlist;
72 
73 void vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to);
74 void vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to);
75 void vfs_clean_pages(struct buf * bp);
76 static void vfs_setdirty(struct buf *bp);
77 static __inline struct buf * gbincore(struct vnode * vp, daddr_t blkno);
78 
79 int needsbuffer;
80 
81 /*
82  * Internal update daemon, process 3
83  *	The variable vfs_update_wakeup allows for internal syncs.
84  */
85 int vfs_update_wakeup;
86 
87 
88 /*
89  * buffers base kva
90  */
91 caddr_t buffers_kva;
92 
93 /*
94  * bogus page -- for I/O to/from partially complete buffers
95  * this is a temporary solution to the problem, but it is not
96  * really that bad.  it would be better to split the buffer
97  * for input in the case of buffers partially already in memory,
98  * but the code is intricate enough already.
99  */
100 vm_page_t bogus_page;
101 vm_offset_t bogus_offset;
102 
103 int bufspace, maxbufspace;
104 
105 /*
106  * advisory minimum for size of LRU queue or VMIO queue
107  */
108 int minbuf;
109 
110 struct bufhashhdr bufhashtbl[BUFHSZ], invalhash;
111 struct bqueues bufqueues[BUFFER_QUEUES];
112 
113 /*
114  * Initialize buffer headers and related structures.
115  */
116 void
117 bufinit()
118 {
119 	struct buf *bp;
120 	int i;
121 
122 	TAILQ_INIT(&bswlist);
123 	LIST_INIT(&invalhash);
124 
125 	/* first, make a null hash table */
126 	for (i = 0; i < BUFHSZ; i++)
127 		LIST_INIT(&bufhashtbl[i]);
128 
129 	/* next, make a null set of free lists */
130 	for (i = 0; i < BUFFER_QUEUES; i++)
131 		TAILQ_INIT(&bufqueues[i]);
132 
133 	buffers_kva = (caddr_t) kmem_alloc_pageable(buffer_map, MAXBSIZE * nbuf);
134 	/* finally, initialize each buffer header and stick on empty q */
135 	for (i = 0; i < nbuf; i++) {
136 		bp = &buf[i];
137 		bzero(bp, sizeof *bp);
138 		bp->b_flags = B_INVAL;	/* we're just an empty header */
139 		bp->b_dev = NODEV;
140 		bp->b_rcred = NOCRED;
141 		bp->b_wcred = NOCRED;
142 		bp->b_qindex = QUEUE_EMPTY;
143 		bp->b_vnbufs.le_next = NOLIST;
144 		bp->b_data = buffers_kva + i * MAXBSIZE;
145 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
146 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
147 	}
148 /*
149  * maxbufspace is currently calculated to support all filesystem blocks
150  * to be 8K.  If you happen to use a 16K filesystem, the size of the buffer
151  * cache is still the same as it would be for 8K filesystems.  This
152  * keeps the size of the buffer cache "in check" for big block filesystems.
153  */
154 	minbuf = nbuf / 3;
155 	maxbufspace = 2 * (nbuf + 8) * PAGE_SIZE;
156 
157 	bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
158 	bogus_page = vm_page_alloc(kernel_object,
159 			bogus_offset - VM_MIN_KERNEL_ADDRESS, VM_ALLOC_NORMAL);
160 
161 }
162 
163 /*
164  * remove the buffer from the appropriate free list
165  */
166 void
167 bremfree(struct buf * bp)
168 {
169 	int s = splbio();
170 
171 	if (bp->b_qindex != QUEUE_NONE) {
172 		TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
173 		bp->b_qindex = QUEUE_NONE;
174 	} else {
175 		panic("bremfree: removing a buffer when not on a queue");
176 	}
177 	splx(s);
178 }
179 
180 /*
181  * Get a buffer with the specified data.  Look in the cache first.
182  */
183 int
184 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
185     struct buf ** bpp)
186 {
187 	struct buf *bp;
188 
189 	bp = getblk(vp, blkno, size, 0, 0);
190 	*bpp = bp;
191 
192 	/* if not found in cache, do some I/O */
193 	if ((bp->b_flags & B_CACHE) == 0) {
194 		if (curproc != NULL)
195 			curproc->p_stats->p_ru.ru_inblock++;
196 		bp->b_flags |= B_READ;
197 		bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL);
198 		if (bp->b_rcred == NOCRED) {
199 			if (cred != NOCRED)
200 				crhold(cred);
201 			bp->b_rcred = cred;
202 		}
203 		vfs_busy_pages(bp, 0);
204 		VOP_STRATEGY(bp);
205 		return (biowait(bp));
206 	}
207 	return (0);
208 }
209 
210 /*
211  * Operates like bread, but also starts asynchronous I/O on
212  * read-ahead blocks.
213  */
214 int
215 breadn(struct vnode * vp, daddr_t blkno, int size,
216     daddr_t * rablkno, int *rabsize,
217     int cnt, struct ucred * cred, struct buf ** bpp)
218 {
219 	struct buf *bp, *rabp;
220 	int i;
221 	int rv = 0, readwait = 0;
222 
223 	*bpp = bp = getblk(vp, blkno, size, 0, 0);
224 
225 	/* if not found in cache, do some I/O */
226 	if ((bp->b_flags & B_CACHE) == 0) {
227 		if (curproc != NULL)
228 			curproc->p_stats->p_ru.ru_inblock++;
229 		bp->b_flags |= B_READ;
230 		bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL);
231 		if (bp->b_rcred == NOCRED) {
232 			if (cred != NOCRED)
233 				crhold(cred);
234 			bp->b_rcred = cred;
235 		}
236 		vfs_busy_pages(bp, 0);
237 		VOP_STRATEGY(bp);
238 		++readwait;
239 	}
240 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
241 		if (inmem(vp, *rablkno))
242 			continue;
243 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0);
244 
245 		if ((rabp->b_flags & B_CACHE) == 0) {
246 			if (curproc != NULL)
247 				curproc->p_stats->p_ru.ru_inblock++;
248 			rabp->b_flags |= B_READ | B_ASYNC;
249 			rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL);
250 			if (rabp->b_rcred == NOCRED) {
251 				if (cred != NOCRED)
252 					crhold(cred);
253 				rabp->b_rcred = cred;
254 			}
255 			vfs_busy_pages(rabp, 0);
256 			VOP_STRATEGY(rabp);
257 		} else {
258 			brelse(rabp);
259 		}
260 	}
261 
262 	if (readwait) {
263 		rv = biowait(bp);
264 	}
265 	return (rv);
266 }
267 
268 /*
269  * Write, release buffer on completion.  (Done by iodone
270  * if async.)
271  */
272 int
273 bwrite(struct buf * bp)
274 {
275 	int oldflags = bp->b_flags;
276 
277 	if (bp->b_flags & B_INVAL) {
278 		brelse(bp);
279 		return (0);
280 	}
281 	if (!(bp->b_flags & B_BUSY))
282 		panic("bwrite: buffer is not busy???");
283 
284 	bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI);
285 	bp->b_flags |= B_WRITEINPROG;
286 
287 	if ((oldflags & (B_ASYNC|B_DELWRI)) == (B_ASYNC|B_DELWRI)) {
288 		reassignbuf(bp, bp->b_vp);
289 	}
290 
291 	bp->b_vp->v_numoutput++;
292 	vfs_busy_pages(bp, 1);
293 	if (curproc != NULL)
294 		curproc->p_stats->p_ru.ru_oublock++;
295 	VOP_STRATEGY(bp);
296 
297 	if ((oldflags & B_ASYNC) == 0) {
298 		int rtval = biowait(bp);
299 
300 		if (oldflags & B_DELWRI) {
301 			reassignbuf(bp, bp->b_vp);
302 		}
303 		brelse(bp);
304 		return (rtval);
305 	}
306 	return (0);
307 }
308 
309 int
310 vn_bwrite(ap)
311 	struct vop_bwrite_args *ap;
312 {
313 	return (bwrite(ap->a_bp));
314 }
315 
316 /*
317  * Delayed write. (Buffer is marked dirty).
318  */
319 void
320 bdwrite(struct buf * bp)
321 {
322 
323 	if ((bp->b_flags & B_BUSY) == 0) {
324 		panic("bdwrite: buffer is not busy");
325 	}
326 	if (bp->b_flags & B_INVAL) {
327 		brelse(bp);
328 		return;
329 	}
330 	if (bp->b_flags & B_TAPE) {
331 		bawrite(bp);
332 		return;
333 	}
334 	bp->b_flags &= ~(B_READ|B_RELBUF);
335 	if ((bp->b_flags & B_DELWRI) == 0) {
336 		bp->b_flags |= B_DONE | B_DELWRI;
337 		reassignbuf(bp, bp->b_vp);
338 	}
339 
340 	/*
341 	 * This bmap keeps the system from needing to do the bmap later,
342 	 * perhaps when the system is attempting to do a sync.  Since it
343 	 * is likely that the indirect block -- or whatever other datastructure
344 	 * that the filesystem needs is still in memory now, it is a good
345 	 * thing to do this.  Note also, that if the pageout daemon is
346 	 * requesting a sync -- there might not be enough memory to do
347 	 * the bmap then...  So, this is important to do.
348 	 */
349 	if( bp->b_lblkno == bp->b_blkno) {
350 		VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
351 	}
352 
353 	/*
354 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
355 	 */
356 	vfs_setdirty(bp);
357 
358 	/*
359 	 * We need to do this here to satisfy the vnode_pager and the
360 	 * pageout daemon, so that it thinks that the pages have been
361 	 * "cleaned".  Note that since the pages are in a delayed write
362 	 * buffer -- the VFS layer "will" see that the pages get written
363 	 * out on the next sync, or perhaps the cluster will be completed.
364 	 */
365 	vfs_clean_pages(bp);
366 	brelse(bp);
367 	return;
368 }
369 
370 /*
371  * Asynchronous write.
372  * Start output on a buffer, but do not wait for it to complete.
373  * The buffer is released when the output completes.
374  */
375 void
376 bawrite(struct buf * bp)
377 {
378 	bp->b_flags |= B_ASYNC;
379 	(void) VOP_BWRITE(bp);
380 }
381 
382 /*
383  * Release a buffer.
384  */
385 void
386 brelse(struct buf * bp)
387 {
388 	int s;
389 
390 	if (bp->b_flags & B_CLUSTER) {
391 		relpbuf(bp);
392 		return;
393 	}
394 	/* anyone need a "free" block? */
395 	s = splbio();
396 
397 	if (needsbuffer) {
398 		needsbuffer = 0;
399 		wakeup(&needsbuffer);
400 	}
401 
402 	/* anyone need this block? */
403 	if (bp->b_flags & B_WANTED) {
404 		bp->b_flags &= ~(B_WANTED | B_AGE);
405 		wakeup(bp);
406 	} else if (bp->b_flags & B_VMIO) {
407 		bp->b_flags &= ~B_WANTED;
408 		wakeup(bp);
409 	}
410 	if (bp->b_flags & B_LOCKED)
411 		bp->b_flags &= ~B_ERROR;
412 
413 	if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
414 	    (bp->b_bufsize <= 0)) {
415 		bp->b_flags |= B_INVAL;
416 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
417 		if (((bp->b_flags & B_VMIO) == 0) && bp->b_vp)
418 			brelvp(bp);
419 	}
420 
421 	/*
422 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
423 	 * constituted, so the B_INVAL flag is used to *invalidate* the buffer,
424 	 * but the VM object is kept around.  The B_NOCACHE flag is used to
425 	 * invalidate the pages in the VM object.
426 	 */
427 	if (bp->b_flags & B_VMIO) {
428 		vm_offset_t foff;
429 		vm_object_t obj;
430 		int i, resid;
431 		vm_page_t m;
432 		struct vnode *vp;
433 		int iototal = bp->b_bufsize;
434 
435 		vp = bp->b_vp;
436 		if (!vp)
437 			panic("brelse: missing vp");
438 		if (!vp->v_mount)
439 			panic("brelse: missing mount info");
440 
441 		if (bp->b_npages) {
442 			obj = (vm_object_t) vp->v_object;
443 			foff = trunc_page(vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno);
444 			for (i = 0; i < bp->b_npages; i++) {
445 				m = bp->b_pages[i];
446 				if (m == bogus_page) {
447 					m = vm_page_lookup(obj, foff);
448 					if (!m) {
449 						panic("brelse: page missing\n");
450 					}
451 					bp->b_pages[i] = m;
452 					pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages);
453 				}
454 				resid = (m->offset + PAGE_SIZE) - foff;
455 				if (resid > iototal)
456 					resid = iototal;
457 				if (resid > 0) {
458 					/*
459 					 * Don't invalidate the page if the local machine has already
460 					 * modified it.  This is the lesser of two evils, and should
461 					 * be fixed.
462 					 */
463 					if (bp->b_flags & (B_NOCACHE | B_ERROR)) {
464 						vm_page_test_dirty(m);
465 						if (m->dirty == 0) {
466 							vm_page_set_invalid(m, foff, resid);
467 							if (m->valid == 0)
468 								vm_page_protect(m, VM_PROT_NONE);
469 						}
470 					}
471 				}
472 				foff += resid;
473 				iototal -= resid;
474 			}
475 		}
476 
477 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
478 			for(i = 0; i < bp->b_npages; i++) {
479 				m = bp->b_pages[i];
480 				--m->bmapped;
481 				if (m->bmapped == 0) {
482 					if (m->flags & PG_WANTED) {
483 						wakeup(m);
484 						m->flags &= ~PG_WANTED;
485 					}
486 					if ((m->busy == 0) && ((m->flags & PG_BUSY) == 0)) {
487 						vm_page_test_dirty(m);
488 						/*
489 						 * if page isn't valid, no sense in keeping it around
490 						 */
491 						if (m->valid == 0) {
492 							vm_page_protect(m, VM_PROT_NONE);
493 							vm_page_free(m);
494 						/*
495 						 * if page isn't dirty and hasn't been referenced by
496 						 * a process, then cache it
497 						 */
498 						} else if ((m->dirty & m->valid) == 0 &&
499 						    (m->flags & PG_REFERENCED) == 0 &&
500 						    !pmap_is_referenced(VM_PAGE_TO_PHYS(m))) {
501 							vm_page_cache(m);
502 						/*
503 						 * otherwise activate it
504 						 */
505 						} else if ((m->flags & PG_ACTIVE) == 0) {
506 							vm_page_activate(m);
507 							m->act_count = 0;
508 						}
509 					}
510 				}
511 			}
512 			bufspace -= bp->b_bufsize;
513 			pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
514 			bp->b_npages = 0;
515 			bp->b_bufsize = 0;
516 			bp->b_flags &= ~B_VMIO;
517 			if (bp->b_vp)
518 				brelvp(bp);
519 		}
520 	}
521 	if (bp->b_qindex != QUEUE_NONE)
522 		panic("brelse: free buffer onto another queue???");
523 
524 	/* enqueue */
525 	/* buffers with no memory */
526 	if (bp->b_bufsize == 0) {
527 		bp->b_qindex = QUEUE_EMPTY;
528 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
529 		LIST_REMOVE(bp, b_hash);
530 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
531 		bp->b_dev = NODEV;
532 		/* buffers with junk contents */
533 	} else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
534 		bp->b_qindex = QUEUE_AGE;
535 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist);
536 		LIST_REMOVE(bp, b_hash);
537 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
538 		bp->b_dev = NODEV;
539 		/* buffers that are locked */
540 	} else if (bp->b_flags & B_LOCKED) {
541 		bp->b_qindex = QUEUE_LOCKED;
542 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
543 		/* buffers with stale but valid contents */
544 	} else if (bp->b_flags & B_AGE) {
545 		bp->b_qindex = QUEUE_AGE;
546 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist);
547 		/* buffers with valid and quite potentially reuseable contents */
548 	} else {
549 		bp->b_qindex = QUEUE_LRU;
550 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist);
551 	}
552 
553 	/* unlock */
554 	bp->b_flags &= ~(B_WANTED | B_BUSY | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
555 	splx(s);
556 }
557 
558 /*
559  * Check to see if a block is currently memory resident.
560  */
561 static __inline struct buf *
562 gbincore(struct vnode * vp, daddr_t blkno)
563 {
564 	struct buf *bp;
565 	struct bufhashhdr *bh;
566 
567 	bh = BUFHASH(vp, blkno);
568 	bp = bh->lh_first;
569 
570 	/* Search hash chain */
571 	while (bp != NULL) {
572 		/* hit */
573 		if (bp->b_vp == vp && bp->b_lblkno == blkno) {
574 			break;
575 		}
576 		bp = bp->b_hash.le_next;
577 	}
578 	return (bp);
579 }
580 
581 /*
582  * this routine implements clustered async writes for
583  * clearing out B_DELWRI buffers...  This is much better
584  * than the old way of writing only one buffer at a time.
585  */
586 void
587 vfs_bio_awrite(struct buf * bp)
588 {
589 	int i;
590 	daddr_t lblkno = bp->b_lblkno;
591 	struct vnode *vp = bp->b_vp;
592 	int s;
593 	int ncl;
594 	struct buf *bpa;
595 
596 	s = splbio();
597 	if (vp->v_mount && (vp->v_flag & VVMIO) &&
598 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
599 		int size = vp->v_mount->mnt_stat.f_iosize;
600 		int maxcl = MAXPHYS / size;
601 
602 		for (i = 1; i < maxcl; i++) {
603 			if ((bpa = gbincore(vp, lblkno + i)) &&
604 			    ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
605 			    (B_DELWRI | B_CLUSTEROK)) &&
606 			    (bpa->b_bufsize == size)) {
607 				if ((bpa->b_blkno == bpa->b_lblkno) ||
608 				    (bpa->b_blkno != bp->b_blkno + (i * size) / DEV_BSIZE))
609 					break;
610 			} else {
611 				break;
612 			}
613 		}
614 		ncl = i;
615 		/*
616 		 * this is a possible cluster write
617 		 */
618 		if (ncl != 1) {
619 			bremfree(bp);
620 			cluster_wbuild(vp, bp, size, lblkno, ncl, -1);
621 			splx(s);
622 			return;
623 		}
624 	}
625 	/*
626 	 * default (old) behavior, writing out only one block
627 	 */
628 	bremfree(bp);
629 	bp->b_flags |= B_BUSY | B_ASYNC;
630 	(void) VOP_BWRITE(bp);
631 	splx(s);
632 }
633 
634 
635 /*
636  * Find a buffer header which is available for use.
637  */
638 static struct buf *
639 getnewbuf(int slpflag, int slptimeo, int doingvmio)
640 {
641 	struct buf *bp;
642 	int s;
643 	int firstbp = 1;
644 
645 	s = splbio();
646 start:
647 	if (bufspace >= maxbufspace)
648 		goto trytofreespace;
649 
650 	/* can we constitute a new buffer? */
651 	if ((bp = bufqueues[QUEUE_EMPTY].tqh_first)) {
652 		if (bp->b_qindex != QUEUE_EMPTY)
653 			panic("getnewbuf: inconsistent EMPTY queue");
654 		bremfree(bp);
655 		goto fillbuf;
656 	}
657 trytofreespace:
658 	/*
659 	 * We keep the file I/O from hogging metadata I/O
660 	 * This is desirable because file data is cached in the
661 	 * VM/Buffer cache even if a buffer is freed.
662 	 */
663 	if ((bp = bufqueues[QUEUE_AGE].tqh_first)) {
664 		if (bp->b_qindex != QUEUE_AGE)
665 			panic("getnewbuf: inconsistent AGE queue");
666 	} else if ((bp = bufqueues[QUEUE_LRU].tqh_first)) {
667 		if (bp->b_qindex != QUEUE_LRU)
668 			panic("getnewbuf: inconsistent LRU queue");
669 	}
670 	if (!bp) {
671 		/* wait for a free buffer of any kind */
672 		needsbuffer = 1;
673 		tsleep(&needsbuffer, PRIBIO | slpflag, "newbuf", slptimeo);
674 		splx(s);
675 		return (0);
676 	}
677 
678 	/* if we are a delayed write, convert to an async write */
679 	if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) {
680 		vfs_bio_awrite(bp);
681 		if (!slpflag && !slptimeo) {
682 			splx(s);
683 			return (0);
684 		}
685 		goto start;
686 	}
687 
688 	if (bp->b_flags & B_WANTED) {
689 		bp->b_flags &= ~B_WANTED;
690 		wakeup(bp);
691 	}
692 	bremfree(bp);
693 
694 	if (bp->b_flags & B_VMIO) {
695 		bp->b_flags |= B_RELBUF | B_BUSY | B_DONE;
696 		brelse(bp);
697 		bremfree(bp);
698 	}
699 
700 	if (bp->b_vp)
701 		brelvp(bp);
702 
703 	/* we are not free, nor do we contain interesting data */
704 	if (bp->b_rcred != NOCRED)
705 		crfree(bp->b_rcred);
706 	if (bp->b_wcred != NOCRED)
707 		crfree(bp->b_wcred);
708 fillbuf:
709 	bp->b_flags |= B_BUSY;
710 	LIST_REMOVE(bp, b_hash);
711 	LIST_INSERT_HEAD(&invalhash, bp, b_hash);
712 	splx(s);
713 	if (bp->b_bufsize) {
714 		allocbuf(bp, 0);
715 	}
716 	bp->b_flags = B_BUSY;
717 	bp->b_dev = NODEV;
718 	bp->b_vp = NULL;
719 	bp->b_blkno = bp->b_lblkno = 0;
720 	bp->b_iodone = 0;
721 	bp->b_error = 0;
722 	bp->b_resid = 0;
723 	bp->b_bcount = 0;
724 	bp->b_npages = 0;
725 	bp->b_wcred = bp->b_rcred = NOCRED;
726 	bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE;
727 	bp->b_dirtyoff = bp->b_dirtyend = 0;
728 	bp->b_validoff = bp->b_validend = 0;
729 	if (bufspace >= maxbufspace) {
730 		s = splbio();
731 		bp->b_flags |= B_INVAL;
732 		brelse(bp);
733 		goto trytofreespace;
734 	}
735 	return (bp);
736 }
737 
738 /*
739  * Check to see if a block is currently memory resident.
740  */
741 struct buf *
742 incore(struct vnode * vp, daddr_t blkno)
743 {
744 	struct buf *bp;
745 	struct bufhashhdr *bh;
746 
747 	int s = splbio();
748 
749 	bh = BUFHASH(vp, blkno);
750 	bp = bh->lh_first;
751 
752 	/* Search hash chain */
753 	while (bp != NULL) {
754 		/* hit */
755 		if (bp->b_vp == vp && bp->b_lblkno == blkno &&
756 		    (bp->b_flags & B_INVAL) == 0) {
757 			break;
758 		}
759 		bp = bp->b_hash.le_next;
760 	}
761 	splx(s);
762 	return (bp);
763 }
764 
765 /*
766  * Returns true if no I/O is needed to access the
767  * associated VM object.  This is like incore except
768  * it also hunts around in the VM system for the data.
769  */
770 
771 int
772 inmem(struct vnode * vp, daddr_t blkno)
773 {
774 	vm_object_t obj;
775 	vm_offset_t off, toff, tinc;
776 	vm_page_t m;
777 
778 	if (incore(vp, blkno))
779 		return 1;
780 	if (vp->v_mount == NULL)
781 		return 0;
782 	if ((vp->v_object == NULL) || (vp->v_flag & VVMIO) == 0)
783 		return 0;
784 
785 	obj = vp->v_object;
786 	tinc = PAGE_SIZE;
787 	if (tinc > vp->v_mount->mnt_stat.f_iosize)
788 		tinc = vp->v_mount->mnt_stat.f_iosize;
789 	off = blkno * vp->v_mount->mnt_stat.f_iosize;
790 
791 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
792 		int mask;
793 
794 		m = vm_page_lookup(obj, trunc_page(toff + off));
795 		if (!m)
796 			return 0;
797 		if (vm_page_is_valid(m, toff + off, tinc) == 0)
798 			return 0;
799 	}
800 	return 1;
801 }
802 
803 /*
804  * now we set the dirty range for the buffer --
805  * for NFS -- if the file is mapped and pages have
806  * been written to, let it know.  We want the
807  * entire range of the buffer to be marked dirty if
808  * any of the pages have been written to for consistancy
809  * with the b_validoff, b_validend set in the nfs write
810  * code, and used by the nfs read code.
811  */
812 static void
813 vfs_setdirty(struct buf *bp) {
814 	int i;
815 	vm_object_t object;
816 	vm_offset_t boffset, offset;
817 	/*
818 	 * We qualify the scan for modified pages on whether the
819 	 * object has been flushed yet.  The OBJ_WRITEABLE flag
820 	 * is not cleared simply by protecting pages off.
821 	 */
822 	if ((bp->b_flags & B_VMIO) &&
823 		((object = bp->b_pages[0]->object)->flags & OBJ_WRITEABLE)) {
824 		/*
825 		 * test the pages to see if they have been modified directly
826 		 * by users through the VM system.
827 		 */
828 		for (i = 0; i < bp->b_npages; i++)
829 			vm_page_test_dirty(bp->b_pages[i]);
830 
831 		/*
832 		 * scan forwards for the first page modified
833 		 */
834 		for (i = 0; i < bp->b_npages; i++) {
835 			if (bp->b_pages[i]->dirty) {
836 				break;
837 			}
838 		}
839 		boffset = i * PAGE_SIZE;
840 		if (boffset < bp->b_dirtyoff) {
841 			bp->b_dirtyoff = boffset;
842 		}
843 
844 		/*
845 		 * scan backwards for the last page modified
846 		 */
847 		for (i = bp->b_npages - 1; i >= 0; --i) {
848 			if (bp->b_pages[i]->dirty) {
849 				break;
850 			}
851 		}
852 		boffset = (i + 1) * PAGE_SIZE;
853 		offset = boffset + bp->b_pages[0]->offset;
854 		if (offset >= object->size) {
855 			boffset = object->size - bp->b_pages[0]->offset;
856 		}
857 		if (bp->b_dirtyend < boffset) {
858 			bp->b_dirtyend = boffset;
859 		}
860 	}
861 }
862 
863 /*
864  * Get a block given a specified block and offset into a file/device.
865  */
866 struct buf *
867 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
868 {
869 	struct buf *bp;
870 	int s;
871 	struct bufhashhdr *bh;
872 	vm_offset_t off;
873 	int nleft;
874 
875 	s = splbio();
876 loop:
877 	if (bp = gbincore(vp, blkno)) {
878 		if (bp->b_flags & (B_BUSY|B_INVAL)) {
879 			bp->b_flags |= B_WANTED;
880 			if (!tsleep(bp, PRIBIO | slpflag, "getblk", slptimeo))
881 				goto loop;
882 
883 			splx(s);
884 			return (struct buf *) NULL;
885 		}
886 		bp->b_flags |= B_BUSY | B_CACHE;
887 		bremfree(bp);
888 
889 		/*
890 		 * check for size inconsistancies (note that they shouldn't happen
891 		 * but do when filesystems don't handle the size changes correctly.)
892 		 * We are conservative on metadata and don't just extend the buffer
893 		 * but write and re-constitute it.
894 		 */
895 		if (bp->b_bcount != size) {
896 			if (bp->b_flags & B_VMIO) {
897 				allocbuf(bp, size);
898 			} else {
899 				bp->b_flags |= B_NOCACHE;
900 				VOP_BWRITE(bp);
901 				goto loop;
902 			}
903 		}
904 		/*
905 		 * make sure that all pages in the buffer are valid, if they
906 		 * aren't, clear the cache flag.
907 		 * ASSUMPTION:
908 		 *  if the buffer is greater than 1 page in size, it is assumed
909 		 *  that the buffer address starts on a page boundary...
910 		 */
911 		if (bp->b_flags & B_VMIO) {
912 			int szleft, i;
913 			szleft = size;
914 			for (i=0;i<bp->b_npages;i++) {
915 				if (szleft > PAGE_SIZE) {
916 					if ((bp->b_pages[i]->valid & VM_PAGE_BITS_ALL) !=
917 						VM_PAGE_BITS_ALL) {
918 						bp->b_flags &= ~(B_CACHE|B_DONE);
919 						break;
920 					}
921 					szleft -= PAGE_SIZE;
922 				} else {
923 					if (!vm_page_is_valid(bp->b_pages[i],
924 						(((vm_offset_t) bp->b_data) & PAGE_MASK),
925 						szleft)) {
926 						bp->b_flags &= ~(B_CACHE|B_DONE);
927 						break;
928 					}
929 					szleft = 0;
930 				}
931 			}
932 		}
933 		splx(s);
934 		return (bp);
935 	} else {
936 		vm_object_t obj;
937 		int doingvmio;
938 
939 		if ((obj = vp->v_object) && (vp->v_flag & VVMIO)) {
940 			doingvmio = 1;
941 		} else {
942 			doingvmio = 0;
943 		}
944 		if ((bp = getnewbuf(slpflag, slptimeo, doingvmio)) == 0) {
945 			if (slpflag || slptimeo)
946 				return NULL;
947 			goto loop;
948 		}
949 
950 		/*
951 		 * This code is used to make sure that a buffer is not
952 		 * created while the getnewbuf routine is blocked.
953 		 * Normally the vnode is locked so this isn't a problem.
954 		 * VBLK type I/O requests, however, don't lock the vnode.
955 		 */
956 		if (!VOP_ISLOCKED(vp) && gbincore(vp, blkno)) {
957 			bp->b_flags |= B_INVAL;
958 			brelse(bp);
959 			goto loop;
960 		}
961 
962 		/*
963 		 * Insert the buffer into the hash, so that it can
964 		 * be found by incore.
965 		 */
966 		bp->b_blkno = bp->b_lblkno = blkno;
967 		bgetvp(vp, bp);
968 		LIST_REMOVE(bp, b_hash);
969 		bh = BUFHASH(vp, blkno);
970 		LIST_INSERT_HEAD(bh, bp, b_hash);
971 
972 		if (doingvmio) {
973 			bp->b_flags |= (B_VMIO | B_CACHE);
974 #if defined(VFS_BIO_DEBUG)
975 			if (vp->v_type != VREG)
976 				printf("getblk: vmioing file type %d???\n", vp->v_type);
977 #endif
978 		} else {
979 			bp->b_flags &= ~B_VMIO;
980 		}
981 		splx(s);
982 
983 		allocbuf(bp, size);
984 		return (bp);
985 	}
986 }
987 
988 /*
989  * Get an empty, disassociated buffer of given size.
990  */
991 struct buf *
992 geteblk(int size)
993 {
994 	struct buf *bp;
995 
996 	while ((bp = getnewbuf(0, 0, 0)) == 0);
997 	allocbuf(bp, size);
998 	bp->b_flags |= B_INVAL;
999 	return (bp);
1000 }
1001 
1002 /*
1003  * This code constitutes the buffer memory from either anonymous system
1004  * memory (in the case of non-VMIO operations) or from an associated
1005  * VM object (in the case of VMIO operations).
1006  *
1007  * Note that this code is tricky, and has many complications to resolve
1008  * deadlock or inconsistant data situations.  Tread lightly!!!
1009  *
1010  * Modify the length of a buffer's underlying buffer storage without
1011  * destroying information (unless, of course the buffer is shrinking).
1012  */
1013 int
1014 allocbuf(struct buf * bp, int size)
1015 {
1016 
1017 	int s;
1018 	int newbsize, mbsize;
1019 	int i;
1020 
1021 	if (!(bp->b_flags & B_BUSY))
1022 		panic("allocbuf: buffer not busy");
1023 
1024 	if ((bp->b_flags & B_VMIO) == 0) {
1025 		/*
1026 		 * Just get anonymous memory from the kernel
1027 		 */
1028 		mbsize = ((size + DEV_BSIZE - 1) / DEV_BSIZE) * DEV_BSIZE;
1029 		newbsize = round_page(size);
1030 
1031 		if (newbsize < bp->b_bufsize) {
1032 			vm_hold_free_pages(
1033 			    bp,
1034 			    (vm_offset_t) bp->b_data + newbsize,
1035 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
1036 		} else if (newbsize > bp->b_bufsize) {
1037 			vm_hold_load_pages(
1038 			    bp,
1039 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
1040 			    (vm_offset_t) bp->b_data + newbsize);
1041 		}
1042 	} else {
1043 		vm_page_t m;
1044 		int desiredpages;
1045 
1046 		newbsize = ((size + DEV_BSIZE - 1) / DEV_BSIZE) * DEV_BSIZE;
1047 		desiredpages = round_page(newbsize) / PAGE_SIZE;
1048 
1049 		if (newbsize < bp->b_bufsize) {
1050 			if (desiredpages < bp->b_npages) {
1051 				pmap_qremove((vm_offset_t) trunc_page(bp->b_data) +
1052 				    desiredpages * PAGE_SIZE, (bp->b_npages - desiredpages));
1053 				for (i = desiredpages; i < bp->b_npages; i++) {
1054 					m = bp->b_pages[i];
1055 					s = splhigh();
1056 					while ((m->flags & PG_BUSY) || (m->busy != 0)) {
1057 						m->flags |= PG_WANTED;
1058 						tsleep(m, PVM, "biodep", 0);
1059 					}
1060 					splx(s);
1061 
1062 					if (m->bmapped == 0) {
1063 						printf("allocbuf: bmapped is zero for page %d\n", i);
1064 						panic("allocbuf: error");
1065 					}
1066 					--m->bmapped;
1067 					if (m->bmapped == 0) {
1068 						vm_page_protect(m, VM_PROT_NONE);
1069 						vm_page_free(m);
1070 					}
1071 					bp->b_pages[i] = NULL;
1072 				}
1073 				bp->b_npages = desiredpages;
1074 			}
1075 		} else if (newbsize > bp->b_bufsize) {
1076 			vm_object_t obj;
1077 			vm_offset_t tinc, off, toff, objoff;
1078 			int pageindex, curbpnpages;
1079 			struct vnode *vp;
1080 			int bsize;
1081 
1082 			vp = bp->b_vp;
1083 			bsize = vp->v_mount->mnt_stat.f_iosize;
1084 
1085 			if (bp->b_npages < desiredpages) {
1086 				obj = vp->v_object;
1087 				tinc = PAGE_SIZE;
1088 				if (tinc > bsize)
1089 					tinc = bsize;
1090 				off = bp->b_lblkno * bsize;
1091 		doretry:
1092 				curbpnpages = bp->b_npages;
1093 				bp->b_flags |= B_CACHE;
1094 				for (toff = 0; toff < newbsize; toff += tinc) {
1095 					int mask;
1096 					int bytesinpage;
1097 
1098 					pageindex = toff / PAGE_SIZE;
1099 					objoff = trunc_page(toff + off);
1100 					if (pageindex < curbpnpages) {
1101 						int pb;
1102 
1103 						m = bp->b_pages[pageindex];
1104 						if (m->offset != objoff)
1105 							panic("allocbuf: page changed offset??!!!?");
1106 						bytesinpage = tinc;
1107 						if (tinc > (newbsize - toff))
1108 							bytesinpage = newbsize - toff;
1109 						if (!vm_page_is_valid(m, toff + off, bytesinpage)) {
1110 							bp->b_flags &= ~B_CACHE;
1111 						}
1112 						if ((m->flags & PG_ACTIVE) == 0) {
1113 							vm_page_activate(m);
1114 							m->act_count = 0;
1115 						}
1116 						continue;
1117 					}
1118 					m = vm_page_lookup(obj, objoff);
1119 					if (!m) {
1120 						m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL);
1121 						if (!m) {
1122 							int j;
1123 
1124 							for (j = bp->b_npages; j < pageindex; j++) {
1125 								PAGE_WAKEUP(bp->b_pages[j]);
1126 							}
1127 							VM_WAIT;
1128 							goto doretry;
1129 						}
1130 						vm_page_activate(m);
1131 						m->act_count = 0;
1132 						m->valid = 0;
1133 						bp->b_flags &= ~B_CACHE;
1134 					} else if (m->flags & PG_BUSY) {
1135 						int j;
1136 
1137 						for (j = bp->b_npages; j < pageindex; j++) {
1138 							PAGE_WAKEUP(bp->b_pages[j]);
1139 						}
1140 
1141 						s = splbio();
1142 						m->flags |= PG_WANTED;
1143 						tsleep(m, PRIBIO, "pgtblk", 0);
1144 						splx(s);
1145 
1146 						goto doretry;
1147 					} else {
1148 						int pb;
1149 						if ((curproc != pageproc) &&
1150 							(m->flags & PG_CACHE) &&
1151 						    (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min) {
1152 							pagedaemon_wakeup();
1153 						}
1154 						bytesinpage = tinc;
1155 						if (tinc > (newbsize - toff))
1156 							bytesinpage = newbsize - toff;
1157 						if (!vm_page_is_valid(m, toff + off, bytesinpage)) {
1158 							bp->b_flags &= ~B_CACHE;
1159 						}
1160 						if ((m->flags & PG_ACTIVE) == 0) {
1161 							vm_page_activate(m);
1162 							m->act_count = 0;
1163 						}
1164 						m->flags |= PG_BUSY;
1165 					}
1166 					bp->b_pages[pageindex] = m;
1167 					curbpnpages = pageindex + 1;
1168 				}
1169 				for (i = bp->b_npages; i < curbpnpages; i++) {
1170 					m = bp->b_pages[i];
1171 					m->bmapped++;
1172 					PAGE_WAKEUP(m);
1173 				}
1174 				bp->b_npages = curbpnpages;
1175 				bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE;
1176 				pmap_qenter((vm_offset_t) bp->b_data, bp->b_pages, bp->b_npages);
1177 				bp->b_data += off % PAGE_SIZE;
1178 			}
1179 		}
1180 	}
1181 	bufspace += (newbsize - bp->b_bufsize);
1182 	bp->b_bufsize = newbsize;
1183 	bp->b_bcount = size;
1184 	return 1;
1185 }
1186 
1187 /*
1188  * Wait for buffer I/O completion, returning error status.
1189  */
1190 int
1191 biowait(register struct buf * bp)
1192 {
1193 	int s;
1194 
1195 	s = splbio();
1196 	while ((bp->b_flags & B_DONE) == 0)
1197 		tsleep(bp, PRIBIO, "biowait", 0);
1198 	splx(s);
1199 	if (bp->b_flags & B_EINTR) {
1200 		bp->b_flags &= ~B_EINTR;
1201 		return (EINTR);
1202 	}
1203 	if (bp->b_flags & B_ERROR) {
1204 		return (bp->b_error ? bp->b_error : EIO);
1205 	} else {
1206 		return (0);
1207 	}
1208 }
1209 
1210 /*
1211  * Finish I/O on a buffer, calling an optional function.
1212  * This is usually called from interrupt level, so process blocking
1213  * is not *a good idea*.
1214  */
1215 void
1216 biodone(register struct buf * bp)
1217 {
1218 	int s;
1219 
1220 	s = splbio();
1221 	if (!(bp->b_flags & B_BUSY))
1222 		panic("biodone: buffer not busy");
1223 
1224 	if (bp->b_flags & B_DONE) {
1225 		splx(s);
1226 		printf("biodone: buffer already done\n");
1227 		return;
1228 	}
1229 	bp->b_flags |= B_DONE;
1230 
1231 	if ((bp->b_flags & B_READ) == 0) {
1232 		struct vnode *vp = bp->b_vp;
1233 		vwakeup(bp);
1234 	}
1235 #ifdef BOUNCE_BUFFERS
1236 	if (bp->b_flags & B_BOUNCE)
1237 		vm_bounce_free(bp);
1238 #endif
1239 
1240 	/* call optional completion function if requested */
1241 	if (bp->b_flags & B_CALL) {
1242 		bp->b_flags &= ~B_CALL;
1243 		(*bp->b_iodone) (bp);
1244 		splx(s);
1245 		return;
1246 	}
1247 	if (bp->b_flags & B_VMIO) {
1248 		int i, resid;
1249 		vm_offset_t foff;
1250 		vm_page_t m;
1251 		vm_object_t obj;
1252 		int iosize;
1253 		struct vnode *vp = bp->b_vp;
1254 
1255 		foff = vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
1256 		obj = vp->v_object;
1257 		if (!obj) {
1258 			panic("biodone: no object");
1259 		}
1260 #if defined(VFS_BIO_DEBUG)
1261 		if (obj->paging_in_progress < bp->b_npages) {
1262 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
1263 			    obj->paging_in_progress, bp->b_npages);
1264 		}
1265 #endif
1266 		iosize = bp->b_bufsize;
1267 		for (i = 0; i < bp->b_npages; i++) {
1268 			int bogusflag = 0;
1269 			m = bp->b_pages[i];
1270 			if (m == bogus_page) {
1271 				bogusflag = 1;
1272 				m = vm_page_lookup(obj, foff);
1273 				if (!m) {
1274 #if defined(VFS_BIO_DEBUG)
1275 					printf("biodone: page disappeared\n");
1276 #endif
1277 					--obj->paging_in_progress;
1278 					continue;
1279 				}
1280 				bp->b_pages[i] = m;
1281 				pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages);
1282 			}
1283 #if defined(VFS_BIO_DEBUG)
1284 			if (trunc_page(foff) != m->offset) {
1285 				printf("biodone: foff(%d)/m->offset(%d) mismatch\n", foff, m->offset);
1286 			}
1287 #endif
1288 			resid = (m->offset + PAGE_SIZE) - foff;
1289 			if (resid > iosize)
1290 				resid = iosize;
1291 			/*
1292 			 * In the write case, the valid and clean bits are
1293 			 * already changed correctly, so we only need to do this
1294 			 * here in the read case.
1295 			 */
1296 			if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) {
1297 				vm_page_set_validclean(m, foff & (PAGE_SIZE-1), resid);
1298 			}
1299 
1300 			/*
1301 			 * when debugging new filesystems or buffer I/O methods, this
1302 			 * is the most common error that pops up.  if you see this, you
1303 			 * have not set the page busy flag correctly!!!
1304 			 */
1305 			if (m->busy == 0) {
1306 				printf("biodone: page busy < 0, "
1307 				    "off: %ld, foff: %ld, "
1308 				    "resid: %d, index: %d\n",
1309 				    m->offset, foff, resid, i);
1310 				printf(" iosize: %ld, lblkno: %ld, flags: 0x%x, npages: %d\n",
1311 				    bp->b_vp->v_mount->mnt_stat.f_iosize,
1312 				    bp->b_lblkno, bp->b_flags, bp->b_npages);
1313 				printf(" valid: 0x%x, dirty: 0x%x, mapped: %d\n",
1314 				    m->valid, m->dirty, m->bmapped);
1315 				panic("biodone: page busy < 0\n");
1316 			}
1317 			--m->busy;
1318 			if ((m->busy == 0) && (m->flags & PG_WANTED)) {
1319 				m->flags &= ~PG_WANTED;
1320 				wakeup(m);
1321 			}
1322 			--obj->paging_in_progress;
1323 			foff += resid;
1324 			iosize -= resid;
1325 		}
1326 		if (obj && obj->paging_in_progress == 0 &&
1327 		    (obj->flags & OBJ_PIPWNT)) {
1328 			obj->flags &= ~OBJ_PIPWNT;
1329 			wakeup(obj);
1330 		}
1331 	}
1332 	/*
1333 	 * For asynchronous completions, release the buffer now. The brelse
1334 	 * checks for B_WANTED and will do the wakeup there if necessary - so
1335 	 * no need to do a wakeup here in the async case.
1336 	 */
1337 
1338 	if (bp->b_flags & B_ASYNC) {
1339 		brelse(bp);
1340 	} else {
1341 		bp->b_flags &= ~B_WANTED;
1342 		wakeup(bp);
1343 	}
1344 	splx(s);
1345 }
1346 
1347 int
1348 count_lock_queue()
1349 {
1350 	int count;
1351 	struct buf *bp;
1352 
1353 	count = 0;
1354 	for (bp = bufqueues[QUEUE_LOCKED].tqh_first;
1355 	    bp != NULL;
1356 	    bp = bp->b_freelist.tqe_next)
1357 		count++;
1358 	return (count);
1359 }
1360 
1361 int vfs_update_interval = 30;
1362 
1363 void
1364 vfs_update()
1365 {
1366 	(void) spl0();
1367 	while (1) {
1368 		tsleep(&vfs_update_wakeup, PRIBIO, "update",
1369 		    hz * vfs_update_interval);
1370 		vfs_update_wakeup = 0;
1371 		sync(curproc, NULL, NULL);
1372 	}
1373 }
1374 
1375 /*
1376  * This routine is called in lieu of iodone in the case of
1377  * incomplete I/O.  This keeps the busy status for pages
1378  * consistant.
1379  */
1380 void
1381 vfs_unbusy_pages(struct buf * bp)
1382 {
1383 	int i;
1384 
1385 	if (bp->b_flags & B_VMIO) {
1386 		struct vnode *vp = bp->b_vp;
1387 		vm_object_t obj = vp->v_object;
1388 		vm_offset_t foff;
1389 
1390 		foff = trunc_page(vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno);
1391 
1392 		for (i = 0; i < bp->b_npages; i++) {
1393 			vm_page_t m = bp->b_pages[i];
1394 
1395 			if (m == bogus_page) {
1396 				m = vm_page_lookup(obj, foff + i * PAGE_SIZE);
1397 				if (!m) {
1398 					panic("vfs_unbusy_pages: page missing\n");
1399 				}
1400 				bp->b_pages[i] = m;
1401 				pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages);
1402 			}
1403 			--obj->paging_in_progress;
1404 			--m->busy;
1405 			if ((m->busy == 0) && (m->flags & PG_WANTED)) {
1406 				m->flags &= ~PG_WANTED;
1407 				wakeup(m);
1408 			}
1409 		}
1410 		if (obj->paging_in_progress == 0 &&
1411 		    (obj->flags & OBJ_PIPWNT)) {
1412 			obj->flags &= ~OBJ_PIPWNT;
1413 			wakeup(obj);
1414 		}
1415 	}
1416 }
1417 
1418 /*
1419  * This routine is called before a device strategy routine.
1420  * It is used to tell the VM system that paging I/O is in
1421  * progress, and treat the pages associated with the buffer
1422  * almost as being PG_BUSY.  Also the object paging_in_progress
1423  * flag is handled to make sure that the object doesn't become
1424  * inconsistant.
1425  */
1426 void
1427 vfs_busy_pages(struct buf * bp, int clear_modify)
1428 {
1429 	int i;
1430 
1431 	if (bp->b_flags & B_VMIO) {
1432 		vm_object_t obj = bp->b_vp->v_object;
1433 		vm_offset_t foff = bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
1434 		int iocount = bp->b_bufsize;
1435 
1436 		vfs_setdirty(bp);
1437 		for (i = 0; i < bp->b_npages; i++) {
1438 			vm_page_t m = bp->b_pages[i];
1439 			int resid = (m->offset + PAGE_SIZE) - foff;
1440 
1441 			if (resid > iocount)
1442 				resid = iocount;
1443 			if ((bp->b_flags & B_CLUSTER) == 0) {
1444 				obj->paging_in_progress++;
1445 				m->busy++;
1446 			}
1447 			if (clear_modify) {
1448 				vm_page_protect(m, VM_PROT_READ);
1449 				vm_page_set_validclean(m,
1450 					foff & (PAGE_SIZE-1), resid);
1451 			} else if (bp->b_bcount >= PAGE_SIZE) {
1452 				if (m->valid && (bp->b_flags & B_CACHE) == 0) {
1453 					bp->b_pages[i] = bogus_page;
1454 					pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages);
1455 				}
1456 			}
1457 			foff += resid;
1458 			iocount -= resid;
1459 		}
1460 	}
1461 }
1462 
1463 /*
1464  * Tell the VM system that the pages associated with this buffer
1465  * are clean.  This is used for delayed writes where the data is
1466  * going to go to disk eventually without additional VM intevention.
1467  */
1468 void
1469 vfs_clean_pages(struct buf * bp)
1470 {
1471 	int i;
1472 
1473 	if (bp->b_flags & B_VMIO) {
1474 		vm_offset_t foff =
1475 			bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
1476 		int iocount = bp->b_bufsize;
1477 
1478 		for (i = 0; i < bp->b_npages; i++) {
1479 			vm_page_t m = bp->b_pages[i];
1480 			int resid = (m->offset + PAGE_SIZE) - foff;
1481 
1482 			if (resid > iocount)
1483 				resid = iocount;
1484 			if (resid > 0) {
1485 				vm_page_set_validclean(m,
1486 					foff & (PAGE_SIZE-1), resid);
1487 			}
1488 			foff += resid;
1489 			iocount -= resid;
1490 		}
1491 	}
1492 }
1493 
1494 void
1495 vfs_bio_clrbuf(struct buf *bp) {
1496 	int i;
1497 	if( bp->b_flags & B_VMIO) {
1498 		if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE)) {
1499 			int j;
1500 			if( bp->b_pages[0]->valid != VM_PAGE_BITS_ALL) {
1501 				for(j=0; j < bp->b_bufsize / DEV_BSIZE;j++) {
1502 					bzero(bp->b_data + j * DEV_BSIZE, DEV_BSIZE);
1503 				}
1504 			}
1505 			bp->b_resid = 0;
1506 			return;
1507 		}
1508 		for(i=0;i<bp->b_npages;i++) {
1509 			if( bp->b_pages[i]->valid == VM_PAGE_BITS_ALL)
1510 				continue;
1511 			if( bp->b_pages[i]->valid == 0) {
1512 				bzero(bp->b_data + i * PAGE_SIZE, PAGE_SIZE);
1513 			} else {
1514 				int j;
1515 				for(j=0;j<PAGE_SIZE/DEV_BSIZE;j++) {
1516 					if( (bp->b_pages[i]->valid & (1<<j)) == 0)
1517 						bzero(bp->b_data + i * PAGE_SIZE + j * DEV_BSIZE, DEV_BSIZE);
1518 				}
1519 			}
1520 			bp->b_pages[i]->valid = VM_PAGE_BITS_ALL;
1521 		}
1522 		bp->b_resid = 0;
1523 	} else {
1524 		clrbuf(bp);
1525 	}
1526 }
1527 
1528 /*
1529  * vm_hold_load_pages and vm_hold_unload pages get pages into
1530  * a buffers address space.  The pages are anonymous and are
1531  * not associated with a file object.
1532  */
1533 void
1534 vm_hold_load_pages(struct buf * bp, vm_offset_t froma, vm_offset_t toa)
1535 {
1536 	vm_offset_t pg;
1537 	vm_page_t p;
1538 	vm_offset_t from = round_page(froma);
1539 	vm_offset_t to = round_page(toa);
1540 
1541 	for (pg = from; pg < to; pg += PAGE_SIZE) {
1542 
1543 tryagain:
1544 
1545 		p = vm_page_alloc(kernel_object, pg - VM_MIN_KERNEL_ADDRESS,
1546 		    VM_ALLOC_NORMAL);
1547 		if (!p) {
1548 			VM_WAIT;
1549 			goto tryagain;
1550 		}
1551 		vm_page_wire(p);
1552 		pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
1553 		bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE] = p;
1554 		PAGE_WAKEUP(p);
1555 		bp->b_npages++;
1556 	}
1557 }
1558 
1559 void
1560 vm_hold_free_pages(struct buf * bp, vm_offset_t froma, vm_offset_t toa)
1561 {
1562 	vm_offset_t pg;
1563 	vm_page_t p;
1564 	vm_offset_t from = round_page(froma);
1565 	vm_offset_t to = round_page(toa);
1566 
1567 	for (pg = from; pg < to; pg += PAGE_SIZE) {
1568 		p = bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE];
1569 		bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE] = 0;
1570 		pmap_kremove(pg);
1571 		vm_page_free(p);
1572 		--bp->b_npages;
1573 	}
1574 }
1575