xref: /dflybsd-src/sys/kern/vfs_bio.c (revision 61344b4b38cb6bc73f4a35e8dfac38b43fd8eeae)
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: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15  * $DragonFly: src/sys/kern/vfs_bio.c,v 1.34 2005/03/23 20:37:03 dillon Exp $
16  */
17 
18 /*
19  * this file contains a new buffer I/O scheme implementing a coherent
20  * VM object and buffer cache scheme.  Pains have been taken to make
21  * sure that the performance degradation associated with schemes such
22  * as this is not realized.
23  *
24  * Author:  John S. Dyson
25  * Significant help during the development and debugging phases
26  * had been provided by David Greenman, also of the FreeBSD core team.
27  *
28  * see man buf(9) for more info.
29  */
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/buf.h>
34 #include <sys/conf.h>
35 #include <sys/eventhandler.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
41 #include <sys/proc.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 #include <sys/proc.h>
48 #include <vm/vm.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
56 
57 #include <sys/buf2.h>
58 #include <sys/thread2.h>
59 #include <vm/vm_page2.h>
60 
61 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
62 
63 struct	bio_ops bioops;		/* I/O operation notification */
64 
65 struct buf *buf;		/* buffer header pool */
66 struct swqueue bswlist;
67 
68 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
69 		vm_offset_t to);
70 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
71 		vm_offset_t to);
72 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
73 			       int pageno, vm_page_t m);
74 static void vfs_clean_pages(struct buf * bp);
75 static void vfs_setdirty(struct buf *bp);
76 static void vfs_vmio_release(struct buf *bp);
77 static void vfs_backgroundwritedone(struct buf *bp);
78 static int flushbufqueues(void);
79 
80 static int bd_request;
81 
82 static void buf_daemon (void);
83 /*
84  * bogus page -- for I/O to/from partially complete buffers
85  * this is a temporary solution to the problem, but it is not
86  * really that bad.  it would be better to split the buffer
87  * for input in the case of buffers partially already in memory,
88  * but the code is intricate enough already.
89  */
90 vm_page_t bogus_page;
91 int vmiodirenable = TRUE;
92 int runningbufspace;
93 struct lwkt_token buftimetoken;  /* Interlock on setting prio and timo */
94 
95 static vm_offset_t bogus_offset;
96 
97 static int bufspace, maxbufspace,
98 	bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
99 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
100 static int needsbuffer;
101 static int lorunningspace, hirunningspace, runningbufreq;
102 static int numdirtybuffers, lodirtybuffers, hidirtybuffers;
103 static int numfreebuffers, lofreebuffers, hifreebuffers;
104 static int getnewbufcalls;
105 static int getnewbufrestarts;
106 
107 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD,
108 	&numdirtybuffers, 0, "");
109 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW,
110 	&lodirtybuffers, 0, "");
111 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW,
112 	&hidirtybuffers, 0, "");
113 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD,
114 	&numfreebuffers, 0, "");
115 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW,
116 	&lofreebuffers, 0, "");
117 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW,
118 	&hifreebuffers, 0, "");
119 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD,
120 	&runningbufspace, 0, "");
121 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW,
122 	&lorunningspace, 0, "");
123 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW,
124 	&hirunningspace, 0, "");
125 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD,
126 	&maxbufspace, 0, "");
127 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD,
128 	&hibufspace, 0, "");
129 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD,
130 	&lobufspace, 0, "");
131 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD,
132 	&bufspace, 0, "");
133 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW,
134 	&maxbufmallocspace, 0, "");
135 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD,
136 	&bufmallocspace, 0, "");
137 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW,
138 	&getnewbufcalls, 0, "");
139 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW,
140 	&getnewbufrestarts, 0, "");
141 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW,
142 	&vmiodirenable, 0, "");
143 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW,
144 	&bufdefragcnt, 0, "");
145 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW,
146 	&buffreekvacnt, 0, "");
147 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW,
148 	&bufreusecnt, 0, "");
149 
150 /*
151  * Disable background writes for now.  There appear to be races in the
152  * flags tests and locking operations as well as races in the completion
153  * code modifying the original bp (origbp) without holding a lock, assuming
154  * splbio protection when there might not be splbio protection.
155  */
156 static int dobkgrdwrite = 0;
157 SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0,
158 	"Do background writes (honoring the BV_BKGRDWRITE flag)?");
159 
160 static int bufhashmask;
161 static int bufhashshift;
162 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
163 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } };
164 char *buf_wmesg = BUF_WMESG;
165 
166 extern int vm_swap_size;
167 
168 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
169 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
170 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
171 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
172 
173 /*
174  * Buffer hash table code.  Note that the logical block scans linearly, which
175  * gives us some L1 cache locality.
176  */
177 
178 static __inline
179 struct bufhashhdr *
180 bufhash(struct vnode *vnp, daddr_t bn)
181 {
182 	u_int64_t hashkey64;
183 	int hashkey;
184 
185 	/*
186 	 * A variation on the Fibonacci hash that Knuth credits to
187 	 * R. W. Floyd, see Knuth's _Art of Computer Programming,
188 	 * Volume 3 / Sorting and Searching_
189 	 *
190          * We reduce the argument to 32 bits before doing the hash to
191 	 * avoid the need for a slow 64x64 multiply on 32 bit platforms.
192 	 *
193 	 * sizeof(struct vnode) is 168 on i386, so toss some of the lower
194 	 * bits of the vnode address to reduce the key range, which
195 	 * improves the distribution of keys across buckets.
196 	 *
197 	 * The file system cylinder group blocks are very heavily
198 	 * used.  They are located at invervals of fbg, which is
199 	 * on the order of 89 to 94 * 2^10, depending on other
200 	 * filesystem parameters, for a 16k block size.  Smaller block
201 	 * sizes will reduce fpg approximately proportionally.  This
202 	 * will cause the cylinder group index to be hashed using the
203 	 * lower bits of the hash multiplier, which will not distribute
204 	 * the keys as uniformly in a classic Fibonacci hash where a
205 	 * relatively small number of the upper bits of the result
206 	 * are used.  Using 2^16 as a close-enough approximation to
207 	 * fpg, split the hash multiplier in half, with the upper 16
208 	 * bits being the inverse of the golden ratio, and the lower
209 	 * 16 bits being a fraction between 1/3 and 3/7 (closer to
210 	 * 3/7 in this case), that gives good experimental results.
211 	 */
212 	hashkey64 = ((u_int64_t)(uintptr_t)vnp >> 3) + (u_int64_t)bn;
213 	hashkey = (((u_int32_t)(hashkey64 + (hashkey64 >> 32)) * 0x9E376DB1u) >>
214 	    bufhashshift) & bufhashmask;
215 	return(&bufhashtbl[hashkey]);
216 }
217 
218 /*
219  *	numdirtywakeup:
220  *
221  *	If someone is blocked due to there being too many dirty buffers,
222  *	and numdirtybuffers is now reasonable, wake them up.
223  */
224 
225 static __inline void
226 numdirtywakeup(int level)
227 {
228 	if (numdirtybuffers <= level) {
229 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
230 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
231 			wakeup(&needsbuffer);
232 		}
233 	}
234 }
235 
236 /*
237  *	bufspacewakeup:
238  *
239  *	Called when buffer space is potentially available for recovery.
240  *	getnewbuf() will block on this flag when it is unable to free
241  *	sufficient buffer space.  Buffer space becomes recoverable when
242  *	bp's get placed back in the queues.
243  */
244 
245 static __inline void
246 bufspacewakeup(void)
247 {
248 	/*
249 	 * If someone is waiting for BUF space, wake them up.  Even
250 	 * though we haven't freed the kva space yet, the waiting
251 	 * process will be able to now.
252 	 */
253 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
254 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
255 		wakeup(&needsbuffer);
256 	}
257 }
258 
259 /*
260  * runningbufwakeup() - in-progress I/O accounting.
261  *
262  */
263 static __inline void
264 runningbufwakeup(struct buf *bp)
265 {
266 	if (bp->b_runningbufspace) {
267 		runningbufspace -= bp->b_runningbufspace;
268 		bp->b_runningbufspace = 0;
269 		if (runningbufreq && runningbufspace <= lorunningspace) {
270 			runningbufreq = 0;
271 			wakeup(&runningbufreq);
272 		}
273 	}
274 }
275 
276 /*
277  *	bufcountwakeup:
278  *
279  *	Called when a buffer has been added to one of the free queues to
280  *	account for the buffer and to wakeup anyone waiting for free buffers.
281  *	This typically occurs when large amounts of metadata are being handled
282  *	by the buffer cache ( else buffer space runs out first, usually ).
283  */
284 
285 static __inline void
286 bufcountwakeup(void)
287 {
288 	++numfreebuffers;
289 	if (needsbuffer) {
290 		needsbuffer &= ~VFS_BIO_NEED_ANY;
291 		if (numfreebuffers >= hifreebuffers)
292 			needsbuffer &= ~VFS_BIO_NEED_FREE;
293 		wakeup(&needsbuffer);
294 	}
295 }
296 
297 /*
298  *	waitrunningbufspace()
299  *
300  *	runningbufspace is a measure of the amount of I/O currently
301  *	running.  This routine is used in async-write situations to
302  *	prevent creating huge backups of pending writes to a device.
303  *	Only asynchronous writes are governed by this function.
304  *
305  *	Reads will adjust runningbufspace, but will not block based on it.
306  *	The read load has a side effect of reducing the allowed write load.
307  *
308  *	This does NOT turn an async write into a sync write.  It waits
309  *	for earlier writes to complete and generally returns before the
310  *	caller's write has reached the device.
311  */
312 static __inline void
313 waitrunningbufspace(void)
314 {
315 	while (runningbufspace > hirunningspace) {
316 		int s;
317 
318 		s = splbio();	/* fix race against interrupt/biodone() */
319 		++runningbufreq;
320 		tsleep(&runningbufreq, 0, "wdrain", 0);
321 		splx(s);
322 	}
323 }
324 
325 /*
326  *	vfs_buf_test_cache:
327  *
328  *	Called when a buffer is extended.  This function clears the B_CACHE
329  *	bit if the newly extended portion of the buffer does not contain
330  *	valid data.
331  */
332 static __inline__
333 void
334 vfs_buf_test_cache(struct buf *bp,
335 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
336 		  vm_page_t m)
337 {
338 	if (bp->b_flags & B_CACHE) {
339 		int base = (foff + off) & PAGE_MASK;
340 		if (vm_page_is_valid(m, base, size) == 0)
341 			bp->b_flags &= ~B_CACHE;
342 	}
343 }
344 
345 static __inline__
346 void
347 bd_wakeup(int dirtybuflevel)
348 {
349 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
350 		bd_request = 1;
351 		wakeup(&bd_request);
352 	}
353 }
354 
355 /*
356  * bd_speedup - speedup the buffer cache flushing code
357  */
358 
359 static __inline__
360 void
361 bd_speedup(void)
362 {
363 	bd_wakeup(1);
364 }
365 
366 /*
367  * Initialize buffer headers and related structures.
368  */
369 
370 caddr_t
371 bufhashinit(caddr_t vaddr)
372 {
373 	/* first, make a null hash table */
374 	bufhashshift = 29;
375 	for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1)
376 		bufhashshift--;
377 	bufhashtbl = (void *)vaddr;
378 	vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask;
379 	--bufhashmask;
380 	return(vaddr);
381 }
382 
383 void
384 bufinit(void)
385 {
386 	struct buf *bp;
387 	int i;
388 
389 	TAILQ_INIT(&bswlist);
390 	LIST_INIT(&invalhash);
391 	lwkt_token_init(&buftimetoken);
392 
393 	for (i = 0; i <= bufhashmask; i++)
394 		LIST_INIT(&bufhashtbl[i]);
395 
396 	/* next, make a null set of free lists */
397 	for (i = 0; i < BUFFER_QUEUES; i++)
398 		TAILQ_INIT(&bufqueues[i]);
399 
400 	/* finally, initialize each buffer header and stick on empty q */
401 	for (i = 0; i < nbuf; i++) {
402 		bp = &buf[i];
403 		bzero(bp, sizeof *bp);
404 		bp->b_flags = B_INVAL;	/* we're just an empty header */
405 		bp->b_dev = NODEV;
406 		bp->b_qindex = QUEUE_EMPTY;
407 		bp->b_xflags = 0;
408 		xio_init(&bp->b_xio);
409 		LIST_INIT(&bp->b_dep);
410 		BUF_LOCKINIT(bp);
411 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
412 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
413 	}
414 
415 	/*
416 	 * maxbufspace is the absolute maximum amount of buffer space we are
417 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
418 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
419 	 * used by most other processes.  The differential is required to
420 	 * ensure that buf_daemon is able to run when other processes might
421 	 * be blocked waiting for buffer space.
422 	 *
423 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
424 	 * this may result in KVM fragmentation which is not handled optimally
425 	 * by the system.
426 	 */
427 	maxbufspace = nbuf * BKVASIZE;
428 	hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
429 	lobufspace = hibufspace - MAXBSIZE;
430 
431 	lorunningspace = 512 * 1024;
432 	hirunningspace = 1024 * 1024;
433 
434 /*
435  * Limit the amount of malloc memory since it is wired permanently into
436  * the kernel space.  Even though this is accounted for in the buffer
437  * allocation, we don't want the malloced region to grow uncontrolled.
438  * The malloc scheme improves memory utilization significantly on average
439  * (small) directories.
440  */
441 	maxbufmallocspace = hibufspace / 20;
442 
443 /*
444  * Reduce the chance of a deadlock occuring by limiting the number
445  * of delayed-write dirty buffers we allow to stack up.
446  */
447 	hidirtybuffers = nbuf / 4 + 20;
448 	numdirtybuffers = 0;
449 /*
450  * To support extreme low-memory systems, make sure hidirtybuffers cannot
451  * eat up all available buffer space.  This occurs when our minimum cannot
452  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
453  * BKVASIZE'd (8K) buffers.
454  */
455 	while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
456 		hidirtybuffers >>= 1;
457 	}
458 	lodirtybuffers = hidirtybuffers / 2;
459 
460 /*
461  * Try to keep the number of free buffers in the specified range,
462  * and give special processes (e.g. like buf_daemon) access to an
463  * emergency reserve.
464  */
465 	lofreebuffers = nbuf / 18 + 5;
466 	hifreebuffers = 2 * lofreebuffers;
467 	numfreebuffers = nbuf;
468 
469 /*
470  * Maximum number of async ops initiated per buf_daemon loop.  This is
471  * somewhat of a hack at the moment, we really need to limit ourselves
472  * based on the number of bytes of I/O in-transit that were initiated
473  * from buf_daemon.
474  */
475 
476 	bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
477 	bogus_page = vm_page_alloc(kernel_object,
478 			((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
479 			VM_ALLOC_NORMAL);
480 	vmstats.v_wire_count++;
481 
482 }
483 
484 /*
485  * bfreekva() - free the kva allocation for a buffer.
486  *
487  *	Must be called at splbio() or higher as this is the only locking for
488  *	buffer_map.
489  *
490  *	Since this call frees up buffer space, we call bufspacewakeup().
491  */
492 static void
493 bfreekva(struct buf * bp)
494 {
495 	int count;
496 
497 	if (bp->b_kvasize) {
498 		++buffreekvacnt;
499 		count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
500 		vm_map_lock(buffer_map);
501 		bufspace -= bp->b_kvasize;
502 		vm_map_delete(buffer_map,
503 		    (vm_offset_t) bp->b_kvabase,
504 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
505 		    &count
506 		);
507 		vm_map_unlock(buffer_map);
508 		vm_map_entry_release(count);
509 		bp->b_kvasize = 0;
510 		bufspacewakeup();
511 	}
512 }
513 
514 /*
515  *	bremfree:
516  *
517  *	Remove the buffer from the appropriate free list.
518  */
519 void
520 bremfree(struct buf * bp)
521 {
522 	int s = splbio();
523 	int old_qindex = bp->b_qindex;
524 
525 	if (bp->b_qindex != QUEUE_NONE) {
526 		KASSERT(BUF_REFCNTNB(bp) == 1,
527 				("bremfree: bp %p not locked",bp));
528 		TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
529 		bp->b_qindex = QUEUE_NONE;
530 	} else {
531 		if (BUF_REFCNTNB(bp) <= 1)
532 			panic("bremfree: removing a buffer not on a queue");
533 	}
534 
535 	/*
536 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
537 	 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
538 	 * the buffer was free and we must decrement numfreebuffers.
539 	 */
540 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
541 		switch(old_qindex) {
542 		case QUEUE_DIRTY:
543 		case QUEUE_CLEAN:
544 		case QUEUE_EMPTY:
545 		case QUEUE_EMPTYKVA:
546 			--numfreebuffers;
547 			break;
548 		default:
549 			break;
550 		}
551 	}
552 	splx(s);
553 }
554 
555 
556 /*
557  * Get a buffer with the specified data.  Look in the cache first.  We
558  * must clear B_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
559  * is set, the buffer is valid and we do not have to do anything ( see
560  * getblk() ).
561  */
562 int
563 bread(struct vnode * vp, daddr_t blkno, int size, struct buf ** bpp)
564 {
565 	struct buf *bp;
566 
567 	bp = getblk(vp, blkno, size, 0, 0);
568 	*bpp = bp;
569 
570 	/* if not found in cache, do some I/O */
571 	if ((bp->b_flags & B_CACHE) == 0) {
572 		KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
573 		bp->b_flags |= B_READ;
574 		bp->b_flags &= ~(B_ERROR | B_INVAL);
575 		vfs_busy_pages(bp, 0);
576 		VOP_STRATEGY(vp, bp);
577 		return (biowait(bp));
578 	}
579 	return (0);
580 }
581 
582 /*
583  * Operates like bread, but also starts asynchronous I/O on
584  * read-ahead blocks.  We must clear B_ERROR and B_INVAL prior
585  * to initiating I/O . If B_CACHE is set, the buffer is valid
586  * and we do not have to do anything.
587  */
588 int
589 breadn(struct vnode * vp, daddr_t blkno, int size, daddr_t * rablkno,
590 	int *rabsize, int cnt, struct buf ** bpp)
591 {
592 	struct buf *bp, *rabp;
593 	int i;
594 	int rv = 0, readwait = 0;
595 
596 	*bpp = bp = getblk(vp, blkno, size, 0, 0);
597 
598 	/* if not found in cache, do some I/O */
599 	if ((bp->b_flags & B_CACHE) == 0) {
600 		bp->b_flags |= B_READ;
601 		bp->b_flags &= ~(B_ERROR | B_INVAL);
602 		vfs_busy_pages(bp, 0);
603 		VOP_STRATEGY(vp, bp);
604 		++readwait;
605 	}
606 
607 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
608 		if (inmem(vp, *rablkno))
609 			continue;
610 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0);
611 
612 		if ((rabp->b_flags & B_CACHE) == 0) {
613 			rabp->b_flags |= B_READ | B_ASYNC;
614 			rabp->b_flags &= ~(B_ERROR | B_INVAL);
615 			vfs_busy_pages(rabp, 0);
616 			BUF_KERNPROC(rabp);
617 			VOP_STRATEGY(vp, rabp);
618 		} else {
619 			brelse(rabp);
620 		}
621 	}
622 
623 	if (readwait) {
624 		rv = biowait(bp);
625 	}
626 	return (rv);
627 }
628 
629 /*
630  * Write, release buffer on completion.  (Done by iodone
631  * if async).  Do not bother writing anything if the buffer
632  * is invalid.
633  *
634  * Note that we set B_CACHE here, indicating that buffer is
635  * fully valid and thus cacheable.  This is true even of NFS
636  * now so we set it generally.  This could be set either here
637  * or in biodone() since the I/O is synchronous.  We put it
638  * here.
639  */
640 int
641 bwrite(struct buf * bp)
642 {
643 	int oldflags, s;
644 	struct buf *newbp;
645 
646 	if (bp->b_flags & B_INVAL) {
647 		brelse(bp);
648 		return (0);
649 	}
650 
651 	oldflags = bp->b_flags;
652 
653 	if (BUF_REFCNTNB(bp) == 0)
654 		panic("bwrite: buffer is not busy???");
655 	s = splbio();
656 	/*
657 	 * If a background write is already in progress, delay
658 	 * writing this block if it is asynchronous. Otherwise
659 	 * wait for the background write to complete.
660 	 */
661 	if (bp->b_xflags & BX_BKGRDINPROG) {
662 		if (bp->b_flags & B_ASYNC) {
663 			splx(s);
664 			bdwrite(bp);
665 			return (0);
666 		}
667 		bp->b_xflags |= BX_BKGRDWAIT;
668 		tsleep(&bp->b_xflags, 0, "biord", 0);
669 		if (bp->b_xflags & BX_BKGRDINPROG)
670 			panic("bwrite: still writing");
671 	}
672 
673 	/* Mark the buffer clean */
674 	bundirty(bp);
675 
676 	/*
677 	 * If this buffer is marked for background writing and we
678 	 * do not have to wait for it, make a copy and write the
679 	 * copy so as to leave this buffer ready for further use.
680 	 *
681 	 * This optimization eats a lot of memory.  If we have a page
682 	 * or buffer shortfull we can't do it.
683 	 */
684 	if (dobkgrdwrite &&
685 	    (bp->b_xflags & BX_BKGRDWRITE) &&
686 	    (bp->b_flags & B_ASYNC) &&
687 	    !vm_page_count_severe() &&
688 	    !buf_dirty_count_severe()) {
689 		if (bp->b_flags & B_CALL)
690 			panic("bwrite: need chained iodone");
691 
692 		/* get a new block */
693 		newbp = geteblk(bp->b_bufsize);
694 
695 		/* set it to be identical to the old block */
696 		memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
697 		bgetvp(bp->b_vp, newbp);
698 		newbp->b_lblkno = bp->b_lblkno;
699 		newbp->b_blkno = bp->b_blkno;
700 		newbp->b_offset = bp->b_offset;
701 		newbp->b_iodone = vfs_backgroundwritedone;
702 		newbp->b_flags |= B_ASYNC | B_CALL;
703 		newbp->b_flags &= ~B_INVAL;
704 
705 		/* move over the dependencies */
706 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
707 			(*bioops.io_movedeps)(bp, newbp);
708 
709 		/*
710 		 * Initiate write on the copy, release the original to
711 		 * the B_LOCKED queue so that it cannot go away until
712 		 * the background write completes. If not locked it could go
713 		 * away and then be reconstituted while it was being written.
714 		 * If the reconstituted buffer were written, we could end up
715 		 * with two background copies being written at the same time.
716 		 */
717 		bp->b_xflags |= BX_BKGRDINPROG;
718 		bp->b_flags |= B_LOCKED;
719 		bqrelse(bp);
720 		bp = newbp;
721 	}
722 
723 	bp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
724 	bp->b_flags |= B_WRITEINPROG | B_CACHE;
725 
726 	bp->b_vp->v_numoutput++;
727 	vfs_busy_pages(bp, 1);
728 
729 	/*
730 	 * Normal bwrites pipeline writes
731 	 */
732 	bp->b_runningbufspace = bp->b_bufsize;
733 	runningbufspace += bp->b_runningbufspace;
734 
735 	splx(s);
736 	if (oldflags & B_ASYNC)
737 		BUF_KERNPROC(bp);
738 	VOP_STRATEGY(bp->b_vp, bp);
739 
740 	if ((oldflags & B_ASYNC) == 0) {
741 		int rtval = biowait(bp);
742 		brelse(bp);
743 		return (rtval);
744 	} else if ((oldflags & B_NOWDRAIN) == 0) {
745 		/*
746 		 * don't allow the async write to saturate the I/O
747 		 * system.  Deadlocks can occur only if a device strategy
748 		 * routine (like in VN) turns around and issues another
749 		 * high-level write, in which case B_NOWDRAIN is expected
750 		 * to be set.   Otherwise we will not deadlock here because
751 		 * we are blocking waiting for I/O that is already in-progress
752 		 * to complete.
753 		 */
754 		waitrunningbufspace();
755 	}
756 
757 	return (0);
758 }
759 
760 /*
761  * Complete a background write started from bwrite.
762  */
763 static void
764 vfs_backgroundwritedone(struct buf *bp)
765 {
766 	struct buf *origbp;
767 
768 	/*
769 	 * Find the original buffer that we are writing.
770 	 */
771 	if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
772 		panic("backgroundwritedone: lost buffer");
773 	/*
774 	 * Process dependencies then return any unfinished ones.
775 	 */
776 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
777 		(*bioops.io_complete)(bp);
778 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
779 		(*bioops.io_movedeps)(bp, origbp);
780 	/*
781 	 * Clear the BX_BKGRDINPROG flag in the original buffer
782 	 * and awaken it if it is waiting for the write to complete.
783 	 * If BX_BKGRDINPROG is not set in the original buffer it must
784 	 * have been released and re-instantiated - which is not legal.
785 	 */
786 	KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2"));
787 	origbp->b_xflags &= ~BX_BKGRDINPROG;
788 	if (origbp->b_xflags & BX_BKGRDWAIT) {
789 		origbp->b_xflags &= ~BX_BKGRDWAIT;
790 		wakeup(&origbp->b_xflags);
791 	}
792 	/*
793 	 * Clear the B_LOCKED flag and remove it from the locked
794 	 * queue if it currently resides there.
795 	 */
796 	origbp->b_flags &= ~B_LOCKED;
797 	if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
798 		bremfree(origbp);
799 		bqrelse(origbp);
800 	}
801 	/*
802 	 * This buffer is marked B_NOCACHE, so when it is released
803 	 * by biodone, it will be tossed. We mark it with B_READ
804 	 * to avoid biodone doing a second vwakeup.
805 	 */
806 	bp->b_flags |= B_NOCACHE | B_READ;
807 	bp->b_flags &= ~(B_CACHE | B_CALL | B_DONE);
808 	bp->b_iodone = 0;
809 	biodone(bp);
810 }
811 
812 /*
813  * Delayed write. (Buffer is marked dirty).  Do not bother writing
814  * anything if the buffer is marked invalid.
815  *
816  * Note that since the buffer must be completely valid, we can safely
817  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
818  * biodone() in order to prevent getblk from writing the buffer
819  * out synchronously.
820  */
821 void
822 bdwrite(struct buf *bp)
823 {
824 	if (BUF_REFCNTNB(bp) == 0)
825 		panic("bdwrite: buffer is not busy");
826 
827 	if (bp->b_flags & B_INVAL) {
828 		brelse(bp);
829 		return;
830 	}
831 	bdirty(bp);
832 
833 	/*
834 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
835 	 * true even of NFS now.
836 	 */
837 	bp->b_flags |= B_CACHE;
838 
839 	/*
840 	 * This bmap keeps the system from needing to do the bmap later,
841 	 * perhaps when the system is attempting to do a sync.  Since it
842 	 * is likely that the indirect block -- or whatever other datastructure
843 	 * that the filesystem needs is still in memory now, it is a good
844 	 * thing to do this.  Note also, that if the pageout daemon is
845 	 * requesting a sync -- there might not be enough memory to do
846 	 * the bmap then...  So, this is important to do.
847 	 */
848 	if (bp->b_lblkno == bp->b_blkno) {
849 		VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
850 	}
851 
852 	/*
853 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
854 	 */
855 	vfs_setdirty(bp);
856 
857 	/*
858 	 * We need to do this here to satisfy the vnode_pager and the
859 	 * pageout daemon, so that it thinks that the pages have been
860 	 * "cleaned".  Note that since the pages are in a delayed write
861 	 * buffer -- the VFS layer "will" see that the pages get written
862 	 * out on the next sync, or perhaps the cluster will be completed.
863 	 */
864 	vfs_clean_pages(bp);
865 	bqrelse(bp);
866 
867 	/*
868 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
869 	 * buffers (midpoint between our recovery point and our stall
870 	 * point).
871 	 */
872 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
873 
874 	/*
875 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
876 	 * due to the softdep code.
877 	 */
878 }
879 
880 /*
881  *	bdirty:
882  *
883  *	Turn buffer into delayed write request.  We must clear B_READ and
884  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
885  *	itself to properly update it in the dirty/clean lists.  We mark it
886  *	B_DONE to ensure that any asynchronization of the buffer properly
887  *	clears B_DONE ( else a panic will occur later ).
888  *
889  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
890  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
891  *	should only be called if the buffer is known-good.
892  *
893  *	Since the buffer is not on a queue, we do not update the numfreebuffers
894  *	count.
895  *
896  *	Must be called at splbio().
897  *	The buffer must be on QUEUE_NONE.
898  */
899 void
900 bdirty(struct buf *bp)
901 {
902 	KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
903 	bp->b_flags &= ~(B_READ|B_RELBUF);
904 
905 	if ((bp->b_flags & B_DELWRI) == 0) {
906 		bp->b_flags |= B_DONE | B_DELWRI;
907 		reassignbuf(bp, bp->b_vp);
908 		++numdirtybuffers;
909 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
910 	}
911 }
912 
913 /*
914  *	bundirty:
915  *
916  *	Clear B_DELWRI for buffer.
917  *
918  *	Since the buffer is not on a queue, we do not update the numfreebuffers
919  *	count.
920  *
921  *	Must be called at splbio().
922  *
923  *	The buffer is typically on QUEUE_NONE but there is one case in
924  *	brelse() that calls this function after placing the buffer on
925  *	a different queue.
926  */
927 
928 void
929 bundirty(struct buf *bp)
930 {
931 	if (bp->b_flags & B_DELWRI) {
932 		bp->b_flags &= ~B_DELWRI;
933 		reassignbuf(bp, bp->b_vp);
934 		--numdirtybuffers;
935 		numdirtywakeup(lodirtybuffers);
936 	}
937 	/*
938 	 * Since it is now being written, we can clear its deferred write flag.
939 	 */
940 	bp->b_flags &= ~B_DEFERRED;
941 }
942 
943 /*
944  *	bawrite:
945  *
946  *	Asynchronous write.  Start output on a buffer, but do not wait for
947  *	it to complete.  The buffer is released when the output completes.
948  *
949  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
950  *	B_INVAL buffers.  Not us.
951  */
952 void
953 bawrite(struct buf * bp)
954 {
955 	bp->b_flags |= B_ASYNC;
956 	(void) VOP_BWRITE(bp->b_vp, bp);
957 }
958 
959 /*
960  *	bowrite:
961  *
962  *	Ordered write.  Start output on a buffer, and flag it so that the
963  *	device will write it in the order it was queued.  The buffer is
964  *	released when the output completes.  bwrite() ( or the VOP routine
965  *	anyway ) is responsible for handling B_INVAL buffers.
966  */
967 int
968 bowrite(struct buf * bp)
969 {
970 	bp->b_flags |= B_ORDERED | B_ASYNC;
971 	return (VOP_BWRITE(bp->b_vp, bp));
972 }
973 
974 /*
975  *	bwillwrite:
976  *
977  *	Called prior to the locking of any vnodes when we are expecting to
978  *	write.  We do not want to starve the buffer cache with too many
979  *	dirty buffers so we block here.  By blocking prior to the locking
980  *	of any vnodes we attempt to avoid the situation where a locked vnode
981  *	prevents the various system daemons from flushing related buffers.
982  */
983 
984 void
985 bwillwrite(void)
986 {
987 	if (numdirtybuffers >= hidirtybuffers) {
988 		int s;
989 
990 		s = splbio();
991 		while (numdirtybuffers >= hidirtybuffers) {
992 			bd_wakeup(1);
993 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
994 			tsleep(&needsbuffer, 0, "flswai", 0);
995 		}
996 		splx(s);
997 	}
998 }
999 
1000 /*
1001  * Return true if we have too many dirty buffers.
1002  */
1003 int
1004 buf_dirty_count_severe(void)
1005 {
1006 	return(numdirtybuffers >= hidirtybuffers);
1007 }
1008 
1009 /*
1010  *	brelse:
1011  *
1012  *	Release a busy buffer and, if requested, free its resources.  The
1013  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1014  *	to be accessed later as a cache entity or reused for other purposes.
1015  */
1016 void
1017 brelse(struct buf * bp)
1018 {
1019 	int s;
1020 
1021 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1022 
1023 	s = splbio();
1024 
1025 	if (bp->b_flags & B_LOCKED)
1026 		bp->b_flags &= ~B_ERROR;
1027 
1028 	if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) {
1029 		/*
1030 		 * Failed write, redirty.  Must clear B_ERROR to prevent
1031 		 * pages from being scrapped.  If B_INVAL is set then
1032 		 * this case is not run and the next case is run to
1033 		 * destroy the buffer.  B_INVAL can occur if the buffer
1034 		 * is outside the range supported by the underlying device.
1035 		 */
1036 		bp->b_flags &= ~B_ERROR;
1037 		bdirty(bp);
1038 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) ||
1039 	    (bp->b_bufsize <= 0)) {
1040 		/*
1041 		 * Either a failed I/O or we were asked to free or not
1042 		 * cache the buffer.
1043 		 */
1044 		bp->b_flags |= B_INVAL;
1045 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1046 			(*bioops.io_deallocate)(bp);
1047 		if (bp->b_flags & B_DELWRI) {
1048 			--numdirtybuffers;
1049 			numdirtywakeup(lodirtybuffers);
1050 		}
1051 		bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF);
1052 		if ((bp->b_flags & B_VMIO) == 0) {
1053 			if (bp->b_bufsize)
1054 				allocbuf(bp, 0);
1055 			if (bp->b_vp)
1056 				brelvp(bp);
1057 		}
1058 	}
1059 
1060 	/*
1061 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1062 	 * is called with B_DELWRI set, the underlying pages may wind up
1063 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1064 	 * because pages associated with a B_DELWRI bp are marked clean.
1065 	 *
1066 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1067 	 * if B_DELWRI is set.
1068 	 *
1069 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1070 	 * on pages to return pages to the VM page queues.
1071 	 */
1072 	if (bp->b_flags & B_DELWRI)
1073 		bp->b_flags &= ~B_RELBUF;
1074 	else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG))
1075 		bp->b_flags |= B_RELBUF;
1076 
1077 	/*
1078 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1079 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1080 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1081 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1082 	 *
1083 	 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be
1084 	 * invalidated.  B_ERROR cannot be set for a failed write unless the
1085 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1086 	 *
1087 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1088 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1089 	 * the commit state and we cannot afford to lose the buffer. If the
1090 	 * buffer has a background write in progress, we need to keep it
1091 	 * around to prevent it from being reconstituted and starting a second
1092 	 * background write.
1093 	 */
1094 	if ((bp->b_flags & B_VMIO)
1095 	    && !(bp->b_vp->v_tag == VT_NFS &&
1096 		 !vn_isdisk(bp->b_vp, NULL) &&
1097 		 (bp->b_flags & B_DELWRI))
1098 	    ) {
1099 
1100 		int i, j, resid;
1101 		vm_page_t m;
1102 		off_t foff;
1103 		vm_pindex_t poff;
1104 		vm_object_t obj;
1105 		struct vnode *vp;
1106 
1107 		vp = bp->b_vp;
1108 
1109 		/*
1110 		 * Get the base offset and length of the buffer.  Note that
1111 		 * in the VMIO case if the buffer block size is not
1112 		 * page-aligned then b_data pointer may not be page-aligned.
1113 		 * But our b_xio.xio_pages array *IS* page aligned.
1114 		 *
1115 		 * block sizes less then DEV_BSIZE (usually 512) are not
1116 		 * supported due to the page granularity bits (m->valid,
1117 		 * m->dirty, etc...).
1118 		 *
1119 		 * See man buf(9) for more information
1120 		 */
1121 
1122 		resid = bp->b_bufsize;
1123 		foff = bp->b_offset;
1124 
1125 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
1126 			m = bp->b_xio.xio_pages[i];
1127 			vm_page_flag_clear(m, PG_ZERO);
1128 			/*
1129 			 * If we hit a bogus page, fixup *all* of them
1130 			 * now.  Note that we left these pages wired
1131 			 * when we removed them so they had better exist,
1132 			 * and they cannot be ripped out from under us so
1133 			 * no splvm() protection is necessary.
1134 			 */
1135 			if (m == bogus_page) {
1136 				VOP_GETVOBJECT(vp, &obj);
1137 				poff = OFF_TO_IDX(bp->b_offset);
1138 
1139 				for (j = i; j < bp->b_xio.xio_npages; j++) {
1140 					vm_page_t mtmp;
1141 
1142 					mtmp = bp->b_xio.xio_pages[j];
1143 					if (mtmp == bogus_page) {
1144 						mtmp = vm_page_lookup(obj, poff + j);
1145 						if (!mtmp) {
1146 							panic("brelse: page missing");
1147 						}
1148 						bp->b_xio.xio_pages[j] = mtmp;
1149 					}
1150 				}
1151 
1152 				if ((bp->b_flags & B_INVAL) == 0) {
1153 					pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1154 						bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1155 				}
1156 				m = bp->b_xio.xio_pages[i];
1157 			}
1158 
1159 			/*
1160 			 * Invalidate the backing store if B_NOCACHE is set
1161 			 * (e.g. used with vinvalbuf()).  If this is NFS
1162 			 * we impose a requirement that the block size be
1163 			 * a multiple of PAGE_SIZE and create a temporary
1164 			 * hack to basically invalidate the whole page.  The
1165 			 * problem is that NFS uses really odd buffer sizes
1166 			 * especially when tracking piecemeal writes and
1167 			 * it also vinvalbuf()'s a lot, which would result
1168 			 * in only partial page validation and invalidation
1169 			 * here.  If the file page is mmap()'d, however,
1170 			 * all the valid bits get set so after we invalidate
1171 			 * here we would end up with weird m->valid values
1172 			 * like 0xfc.  nfs_getpages() can't handle this so
1173 			 * we clear all the valid bits for the NFS case
1174 			 * instead of just some of them.
1175 			 *
1176 			 * The real bug is the VM system having to set m->valid
1177 			 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1178 			 * itself is an artifact of the whole 512-byte
1179 			 * granular mess that exists to support odd block
1180 			 * sizes and UFS meta-data block sizes (e.g. 6144).
1181 			 * A complete rewrite is required.
1182 			 */
1183 			if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1184 				int poffset = foff & PAGE_MASK;
1185 				int presid;
1186 
1187 				presid = PAGE_SIZE - poffset;
1188 				if (bp->b_vp->v_tag == VT_NFS &&
1189 				    bp->b_vp->v_type == VREG) {
1190 					; /* entire page */
1191 				} else if (presid > resid) {
1192 					presid = resid;
1193 				}
1194 				KASSERT(presid >= 0, ("brelse: extra page"));
1195 				vm_page_set_invalid(m, poffset, presid);
1196 			}
1197 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1198 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1199 		}
1200 
1201 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1202 			vfs_vmio_release(bp);
1203 
1204 	} else if (bp->b_flags & B_VMIO) {
1205 
1206 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1207 			vfs_vmio_release(bp);
1208 
1209 	}
1210 
1211 	if (bp->b_qindex != QUEUE_NONE)
1212 		panic("brelse: free buffer onto another queue???");
1213 	if (BUF_REFCNTNB(bp) > 1) {
1214 		/* Temporary panic to verify exclusive locking */
1215 		/* This panic goes away when we allow shared refs */
1216 		panic("brelse: multiple refs");
1217 		/* do not release to free list */
1218 		BUF_UNLOCK(bp);
1219 		splx(s);
1220 		return;
1221 	}
1222 
1223 	/* enqueue */
1224 
1225 	/* buffers with no memory */
1226 	if (bp->b_bufsize == 0) {
1227 		bp->b_flags |= B_INVAL;
1228 		bp->b_xflags &= ~BX_BKGRDWRITE;
1229 		if (bp->b_xflags & BX_BKGRDINPROG)
1230 			panic("losing buffer 1");
1231 		if (bp->b_kvasize) {
1232 			bp->b_qindex = QUEUE_EMPTYKVA;
1233 		} else {
1234 			bp->b_qindex = QUEUE_EMPTY;
1235 		}
1236 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1237 		LIST_REMOVE(bp, b_hash);
1238 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1239 		bp->b_dev = NODEV;
1240 	/* buffers with junk contents */
1241 	} else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1242 		bp->b_flags |= B_INVAL;
1243 		bp->b_xflags &= ~BX_BKGRDWRITE;
1244 		if (bp->b_xflags & BX_BKGRDINPROG)
1245 			panic("losing buffer 2");
1246 		bp->b_qindex = QUEUE_CLEAN;
1247 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1248 		LIST_REMOVE(bp, b_hash);
1249 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1250 		bp->b_dev = NODEV;
1251 
1252 	/* buffers that are locked */
1253 	} else if (bp->b_flags & B_LOCKED) {
1254 		bp->b_qindex = QUEUE_LOCKED;
1255 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1256 
1257 	/* remaining buffers */
1258 	} else {
1259 		switch(bp->b_flags & (B_DELWRI|B_AGE)) {
1260 		case B_DELWRI | B_AGE:
1261 		    bp->b_qindex = QUEUE_DIRTY;
1262 		    TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1263 		    break;
1264 		case B_DELWRI:
1265 		    bp->b_qindex = QUEUE_DIRTY;
1266 		    TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1267 		    break;
1268 		case B_AGE:
1269 		    bp->b_qindex = QUEUE_CLEAN;
1270 		    TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1271 		    break;
1272 		default:
1273 		    bp->b_qindex = QUEUE_CLEAN;
1274 		    TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1275 		    break;
1276 		}
1277 	}
1278 
1279 	/*
1280 	 * If B_INVAL, clear B_DELWRI.  We've already placed the buffer
1281 	 * on the correct queue.
1282 	 */
1283 	if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1284 		bundirty(bp);
1285 
1286 	/*
1287 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1288 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1289 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1290 	 * if B_INVAL is set ).
1291 	 */
1292 
1293 	if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1294 		bufcountwakeup();
1295 
1296 	/*
1297 	 * Something we can maybe free or reuse
1298 	 */
1299 	if (bp->b_bufsize || bp->b_kvasize)
1300 		bufspacewakeup();
1301 
1302 	/* unlock */
1303 	BUF_UNLOCK(bp);
1304 	bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1305 			B_DIRECT | B_NOWDRAIN);
1306 	splx(s);
1307 }
1308 
1309 /*
1310  * Release a buffer back to the appropriate queue but do not try to free
1311  * it.  The buffer is expected to be used again soon.
1312  *
1313  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1314  * biodone() to requeue an async I/O on completion.  It is also used when
1315  * known good buffers need to be requeued but we think we may need the data
1316  * again soon.
1317  *
1318  * XXX we should be able to leave the B_RELBUF hint set on completion.
1319  */
1320 void
1321 bqrelse(struct buf * bp)
1322 {
1323 	int s;
1324 
1325 	s = splbio();
1326 
1327 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1328 
1329 	if (bp->b_qindex != QUEUE_NONE)
1330 		panic("bqrelse: free buffer onto another queue???");
1331 	if (BUF_REFCNTNB(bp) > 1) {
1332 		/* do not release to free list */
1333 		panic("bqrelse: multiple refs");
1334 		BUF_UNLOCK(bp);
1335 		splx(s);
1336 		return;
1337 	}
1338 	if (bp->b_flags & B_LOCKED) {
1339 		bp->b_flags &= ~B_ERROR;
1340 		bp->b_qindex = QUEUE_LOCKED;
1341 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1342 		/* buffers with stale but valid contents */
1343 	} else if (bp->b_flags & B_DELWRI) {
1344 		bp->b_qindex = QUEUE_DIRTY;
1345 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1346 	} else if (vm_page_count_severe()) {
1347 		/*
1348 		 * We are too low on memory, we have to try to free the
1349 		 * buffer (most importantly: the wired pages making up its
1350 		 * backing store) *now*.
1351 		 */
1352 		splx(s);
1353 		brelse(bp);
1354 		return;
1355 	} else {
1356 		bp->b_qindex = QUEUE_CLEAN;
1357 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1358 	}
1359 
1360 	if ((bp->b_flags & B_LOCKED) == 0 &&
1361 	    ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1362 		bufcountwakeup();
1363 	}
1364 
1365 	/*
1366 	 * Something we can maybe free or reuse.
1367 	 */
1368 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1369 		bufspacewakeup();
1370 
1371 	/* unlock */
1372 	BUF_UNLOCK(bp);
1373 	bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1374 	splx(s);
1375 }
1376 
1377 static void
1378 vfs_vmio_release(struct buf *bp)
1379 {
1380 	int i, s;
1381 	vm_page_t m;
1382 
1383 	s = splvm();
1384 	for (i = 0; i < bp->b_xio.xio_npages; i++) {
1385 		m = bp->b_xio.xio_pages[i];
1386 		bp->b_xio.xio_pages[i] = NULL;
1387 		/*
1388 		 * In order to keep page LRU ordering consistent, put
1389 		 * everything on the inactive queue.
1390 		 */
1391 		vm_page_unwire(m, 0);
1392 		/*
1393 		 * We don't mess with busy pages, it is
1394 		 * the responsibility of the process that
1395 		 * busied the pages to deal with them.
1396 		 */
1397 		if ((m->flags & PG_BUSY) || (m->busy != 0))
1398 			continue;
1399 
1400 		if (m->wire_count == 0) {
1401 			vm_page_flag_clear(m, PG_ZERO);
1402 			/*
1403 			 * Might as well free the page if we can and it has
1404 			 * no valid data.  We also free the page if the
1405 			 * buffer was used for direct I/O.
1406 			 */
1407 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) {
1408 				vm_page_busy(m);
1409 				vm_page_protect(m, VM_PROT_NONE);
1410 				vm_page_free(m);
1411 			} else if (bp->b_flags & B_DIRECT) {
1412 				vm_page_try_to_free(m);
1413 			} else if (vm_page_count_severe()) {
1414 				vm_page_try_to_cache(m);
1415 			}
1416 		}
1417 	}
1418 	splx(s);
1419 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1420 	if (bp->b_bufsize) {
1421 		bufspacewakeup();
1422 		bp->b_bufsize = 0;
1423 	}
1424 	bp->b_xio.xio_npages = 0;
1425 	bp->b_flags &= ~B_VMIO;
1426 	if (bp->b_vp)
1427 		brelvp(bp);
1428 }
1429 
1430 /*
1431  * Check to see if a block is currently memory resident.
1432  */
1433 struct buf *
1434 gbincore(struct vnode * vp, daddr_t blkno)
1435 {
1436 	struct buf *bp;
1437 	struct bufhashhdr *bh;
1438 
1439 	bh = bufhash(vp, blkno);
1440 
1441 	/* Search hash chain */
1442 	LIST_FOREACH(bp, bh, b_hash) {
1443 		/* hit */
1444 		if (bp->b_vp == vp && bp->b_lblkno == blkno &&
1445 		    (bp->b_flags & B_INVAL) == 0) {
1446 			break;
1447 		}
1448 	}
1449 	return (bp);
1450 }
1451 
1452 /*
1453  *	vfs_bio_awrite:
1454  *
1455  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1456  *	This is much better then the old way of writing only one buffer at
1457  *	a time.  Note that we may not be presented with the buffers in the
1458  *	correct order, so we search for the cluster in both directions.
1459  */
1460 int
1461 vfs_bio_awrite(struct buf * bp)
1462 {
1463 	int i;
1464 	int j;
1465 	daddr_t lblkno = bp->b_lblkno;
1466 	struct vnode *vp = bp->b_vp;
1467 	int s;
1468 	int ncl;
1469 	struct buf *bpa;
1470 	int nwritten;
1471 	int size;
1472 	int maxcl;
1473 
1474 	s = splbio();
1475 	/*
1476 	 * right now we support clustered writing only to regular files.  If
1477 	 * we find a clusterable block we could be in the middle of a cluster
1478 	 * rather then at the beginning.
1479 	 */
1480 	if ((vp->v_type == VREG) &&
1481 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1482 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1483 
1484 		size = vp->v_mount->mnt_stat.f_iosize;
1485 		maxcl = MAXPHYS / size;
1486 
1487 		for (i = 1; i < maxcl; i++) {
1488 			if ((bpa = gbincore(vp, lblkno + i)) &&
1489 			    BUF_REFCNT(bpa) == 0 &&
1490 			    ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1491 			    (B_DELWRI | B_CLUSTEROK)) &&
1492 			    (bpa->b_bufsize == size)) {
1493 				if ((bpa->b_blkno == bpa->b_lblkno) ||
1494 				    (bpa->b_blkno !=
1495 				     bp->b_blkno + ((i * size) >> DEV_BSHIFT)))
1496 					break;
1497 			} else {
1498 				break;
1499 			}
1500 		}
1501 		for (j = 1; i + j <= maxcl && j <= lblkno; j++) {
1502 			if ((bpa = gbincore(vp, lblkno - j)) &&
1503 			    BUF_REFCNT(bpa) == 0 &&
1504 			    ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1505 			    (B_DELWRI | B_CLUSTEROK)) &&
1506 			    (bpa->b_bufsize == size)) {
1507 				if ((bpa->b_blkno == bpa->b_lblkno) ||
1508 				    (bpa->b_blkno !=
1509 				     bp->b_blkno - ((j * size) >> DEV_BSHIFT)))
1510 					break;
1511 			} else {
1512 				break;
1513 			}
1514 		}
1515 		--j;
1516 		ncl = i + j;
1517 		/*
1518 		 * this is a possible cluster write
1519 		 */
1520 		if (ncl != 1) {
1521 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1522 			splx(s);
1523 			return nwritten;
1524 		}
1525 	}
1526 
1527 	BUF_LOCK(bp, LK_EXCLUSIVE);
1528 	bremfree(bp);
1529 	bp->b_flags |= B_ASYNC;
1530 
1531 	splx(s);
1532 	/*
1533 	 * default (old) behavior, writing out only one block
1534 	 *
1535 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1536 	 */
1537 	nwritten = bp->b_bufsize;
1538 	(void) VOP_BWRITE(bp->b_vp, bp);
1539 
1540 	return nwritten;
1541 }
1542 
1543 /*
1544  *	getnewbuf:
1545  *
1546  *	Find and initialize a new buffer header, freeing up existing buffers
1547  *	in the bufqueues as necessary.  The new buffer is returned locked.
1548  *
1549  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1550  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1551  *
1552  *	We block if:
1553  *		We have insufficient buffer headers
1554  *		We have insufficient buffer space
1555  *		buffer_map is too fragmented ( space reservation fails )
1556  *		If we have to flush dirty buffers ( but we try to avoid this )
1557  *
1558  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1559  *	Instead we ask the buf daemon to do it for us.  We attempt to
1560  *	avoid piecemeal wakeups of the pageout daemon.
1561  */
1562 
1563 static struct buf *
1564 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1565 {
1566 	struct buf *bp;
1567 	struct buf *nbp;
1568 	int defrag = 0;
1569 	int nqindex;
1570 	static int flushingbufs;
1571 
1572 	/*
1573 	 * We can't afford to block since we might be holding a vnode lock,
1574 	 * which may prevent system daemons from running.  We deal with
1575 	 * low-memory situations by proactively returning memory and running
1576 	 * async I/O rather then sync I/O.
1577 	 */
1578 
1579 	++getnewbufcalls;
1580 	--getnewbufrestarts;
1581 restart:
1582 	++getnewbufrestarts;
1583 
1584 	/*
1585 	 * Setup for scan.  If we do not have enough free buffers,
1586 	 * we setup a degenerate case that immediately fails.  Note
1587 	 * that if we are specially marked process, we are allowed to
1588 	 * dip into our reserves.
1589 	 *
1590 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1591 	 *
1592 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1593 	 * However, there are a number of cases (defragging, reusing, ...)
1594 	 * where we cannot backup.
1595 	 */
1596 	nqindex = QUEUE_EMPTYKVA;
1597 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1598 
1599 	if (nbp == NULL) {
1600 		/*
1601 		 * If no EMPTYKVA buffers and we are either
1602 		 * defragging or reusing, locate a CLEAN buffer
1603 		 * to free or reuse.  If bufspace useage is low
1604 		 * skip this step so we can allocate a new buffer.
1605 		 */
1606 		if (defrag || bufspace >= lobufspace) {
1607 			nqindex = QUEUE_CLEAN;
1608 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1609 		}
1610 
1611 		/*
1612 		 * If we could not find or were not allowed to reuse a
1613 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1614 		 * buffer.  We can only use an EMPTY buffer if allocating
1615 		 * its KVA would not otherwise run us out of buffer space.
1616 		 */
1617 		if (nbp == NULL && defrag == 0 &&
1618 		    bufspace + maxsize < hibufspace) {
1619 			nqindex = QUEUE_EMPTY;
1620 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1621 		}
1622 	}
1623 
1624 	/*
1625 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1626 	 * depending.
1627 	 */
1628 
1629 	while ((bp = nbp) != NULL) {
1630 		int qindex = nqindex;
1631 
1632 		/*
1633 		 * Calculate next bp ( we can only use it if we do not block
1634 		 * or do other fancy things ).
1635 		 */
1636 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1637 			switch(qindex) {
1638 			case QUEUE_EMPTY:
1639 				nqindex = QUEUE_EMPTYKVA;
1640 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1641 					break;
1642 				/* fall through */
1643 			case QUEUE_EMPTYKVA:
1644 				nqindex = QUEUE_CLEAN;
1645 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1646 					break;
1647 				/* fall through */
1648 			case QUEUE_CLEAN:
1649 				/*
1650 				 * nbp is NULL.
1651 				 */
1652 				break;
1653 			}
1654 		}
1655 
1656 		/*
1657 		 * Sanity Checks
1658 		 */
1659 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1660 
1661 		/*
1662 		 * Note: we no longer distinguish between VMIO and non-VMIO
1663 		 * buffers.
1664 		 */
1665 
1666 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1667 
1668 		/*
1669 		 * If we are defragging then we need a buffer with
1670 		 * b_kvasize != 0.  XXX this situation should no longer
1671 		 * occur, if defrag is non-zero the buffer's b_kvasize
1672 		 * should also be non-zero at this point.  XXX
1673 		 */
1674 		if (defrag && bp->b_kvasize == 0) {
1675 			printf("Warning: defrag empty buffer %p\n", bp);
1676 			continue;
1677 		}
1678 
1679 		/*
1680 		 * Start freeing the bp.  This is somewhat involved.  nbp
1681 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1682 		 */
1683 
1684 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1685 			panic("getnewbuf: locked buf");
1686 		bremfree(bp);
1687 
1688 		if (qindex == QUEUE_CLEAN) {
1689 			if (bp->b_flags & B_VMIO) {
1690 				bp->b_flags &= ~B_ASYNC;
1691 				vfs_vmio_release(bp);
1692 			}
1693 			if (bp->b_vp)
1694 				brelvp(bp);
1695 		}
1696 
1697 		/*
1698 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1699 		 * the scan from this point on.
1700 		 *
1701 		 * Get the rest of the buffer freed up.  b_kva* is still
1702 		 * valid after this operation.
1703 		 */
1704 
1705 		if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1706 			(*bioops.io_deallocate)(bp);
1707 		if (bp->b_xflags & BX_BKGRDINPROG)
1708 			panic("losing buffer 3");
1709 		LIST_REMOVE(bp, b_hash);
1710 		LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1711 
1712 		/*
1713 		 * spl protection not required when scrapping a buffer's
1714 		 * contents because it is already wired.
1715 		 */
1716 		if (bp->b_bufsize)
1717 			allocbuf(bp, 0);
1718 
1719 		bp->b_flags = 0;
1720 		bp->b_xflags = 0;
1721 		bp->b_dev = NODEV;
1722 		bp->b_vp = NULL;
1723 		bp->b_blkno = bp->b_lblkno = 0;
1724 		bp->b_offset = NOOFFSET;
1725 		bp->b_iodone = 0;
1726 		bp->b_error = 0;
1727 		bp->b_resid = 0;
1728 		bp->b_bcount = 0;
1729 		bp->b_xio.xio_npages = 0;
1730 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1731 
1732 		LIST_INIT(&bp->b_dep);
1733 
1734 		/*
1735 		 * If we are defragging then free the buffer.
1736 		 */
1737 		if (defrag) {
1738 			bp->b_flags |= B_INVAL;
1739 			bfreekva(bp);
1740 			brelse(bp);
1741 			defrag = 0;
1742 			goto restart;
1743 		}
1744 
1745 		/*
1746 		 * If we are overcomitted then recover the buffer and its
1747 		 * KVM space.  This occurs in rare situations when multiple
1748 		 * processes are blocked in getnewbuf() or allocbuf().
1749 		 */
1750 		if (bufspace >= hibufspace)
1751 			flushingbufs = 1;
1752 		if (flushingbufs && bp->b_kvasize != 0) {
1753 			bp->b_flags |= B_INVAL;
1754 			bfreekva(bp);
1755 			brelse(bp);
1756 			goto restart;
1757 		}
1758 		if (bufspace < lobufspace)
1759 			flushingbufs = 0;
1760 		break;
1761 	}
1762 
1763 	/*
1764 	 * If we exhausted our list, sleep as appropriate.  We may have to
1765 	 * wakeup various daemons and write out some dirty buffers.
1766 	 *
1767 	 * Generally we are sleeping due to insufficient buffer space.
1768 	 */
1769 
1770 	if (bp == NULL) {
1771 		int flags;
1772 		char *waitmsg;
1773 
1774 		if (defrag) {
1775 			flags = VFS_BIO_NEED_BUFSPACE;
1776 			waitmsg = "nbufkv";
1777 		} else if (bufspace >= hibufspace) {
1778 			waitmsg = "nbufbs";
1779 			flags = VFS_BIO_NEED_BUFSPACE;
1780 		} else {
1781 			waitmsg = "newbuf";
1782 			flags = VFS_BIO_NEED_ANY;
1783 		}
1784 
1785 		bd_speedup();	/* heeeelp */
1786 
1787 		needsbuffer |= flags;
1788 		while (needsbuffer & flags) {
1789 			if (tsleep(&needsbuffer, slpflag, waitmsg, slptimeo))
1790 				return (NULL);
1791 		}
1792 	} else {
1793 		/*
1794 		 * We finally have a valid bp.  We aren't quite out of the
1795 		 * woods, we still have to reserve kva space.  In order
1796 		 * to keep fragmentation sane we only allocate kva in
1797 		 * BKVASIZE chunks.
1798 		 */
1799 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1800 
1801 		if (maxsize != bp->b_kvasize) {
1802 			vm_offset_t addr = 0;
1803 			int count;
1804 
1805 			bfreekva(bp);
1806 
1807 			count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1808 			vm_map_lock(buffer_map);
1809 
1810 			if (vm_map_findspace(buffer_map,
1811 				    vm_map_min(buffer_map), maxsize,
1812 				    maxsize, &addr)) {
1813 				/*
1814 				 * Uh oh.  Buffer map is to fragmented.  We
1815 				 * must defragment the map.
1816 				 */
1817 				vm_map_unlock(buffer_map);
1818 				vm_map_entry_release(count);
1819 				++bufdefragcnt;
1820 				defrag = 1;
1821 				bp->b_flags |= B_INVAL;
1822 				brelse(bp);
1823 				goto restart;
1824 			}
1825 			if (addr) {
1826 				vm_map_insert(buffer_map, &count,
1827 					NULL, 0,
1828 					addr, addr + maxsize,
1829 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1830 
1831 				bp->b_kvabase = (caddr_t) addr;
1832 				bp->b_kvasize = maxsize;
1833 				bufspace += bp->b_kvasize;
1834 				++bufreusecnt;
1835 			}
1836 			vm_map_unlock(buffer_map);
1837 			vm_map_entry_release(count);
1838 		}
1839 		bp->b_data = bp->b_kvabase;
1840 	}
1841 	return(bp);
1842 }
1843 
1844 /*
1845  *	buf_daemon:
1846  *
1847  *	buffer flushing daemon.  Buffers are normally flushed by the
1848  *	update daemon but if it cannot keep up this process starts to
1849  *	take the load in an attempt to prevent getnewbuf() from blocking.
1850  */
1851 
1852 static struct thread *bufdaemonthread;
1853 
1854 static struct kproc_desc buf_kp = {
1855 	"bufdaemon",
1856 	buf_daemon,
1857 	&bufdaemonthread
1858 };
1859 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1860 
1861 static void
1862 buf_daemon()
1863 {
1864 	int s;
1865 
1866 	/*
1867 	 * This process needs to be suspended prior to shutdown sync.
1868 	 */
1869 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
1870 	    bufdaemonthread, SHUTDOWN_PRI_LAST);
1871 
1872 	/*
1873 	 * This process is allowed to take the buffer cache to the limit
1874 	 */
1875 	s = splbio();
1876 
1877 	for (;;) {
1878 		kproc_suspend_loop();
1879 
1880 		/*
1881 		 * Do the flush.  Limit the amount of in-transit I/O we
1882 		 * allow to build up, otherwise we would completely saturate
1883 		 * the I/O system.  Wakeup any waiting processes before we
1884 		 * normally would so they can run in parallel with our drain.
1885 		 */
1886 		while (numdirtybuffers > lodirtybuffers) {
1887 			if (flushbufqueues() == 0)
1888 				break;
1889 			waitrunningbufspace();
1890 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1891 		}
1892 
1893 		/*
1894 		 * Only clear bd_request if we have reached our low water
1895 		 * mark.  The buf_daemon normally waits 5 seconds and
1896 		 * then incrementally flushes any dirty buffers that have
1897 		 * built up, within reason.
1898 		 *
1899 		 * If we were unable to hit our low water mark and couldn't
1900 		 * find any flushable buffers, we sleep half a second.
1901 		 * Otherwise we loop immediately.
1902 		 */
1903 		if (numdirtybuffers <= lodirtybuffers) {
1904 			/*
1905 			 * We reached our low water mark, reset the
1906 			 * request and sleep until we are needed again.
1907 			 * The sleep is just so the suspend code works.
1908 			 */
1909 			bd_request = 0;
1910 			tsleep(&bd_request, 0, "psleep", hz);
1911 		} else {
1912 			/*
1913 			 * We couldn't find any flushable dirty buffers but
1914 			 * still have too many dirty buffers, we
1915 			 * have to sleep and try again.  (rare)
1916 			 */
1917 			tsleep(&bd_request, 0, "qsleep", hz / 2);
1918 		}
1919 	}
1920 }
1921 
1922 /*
1923  *	flushbufqueues:
1924  *
1925  *	Try to flush a buffer in the dirty queue.  We must be careful to
1926  *	free up B_INVAL buffers instead of write them, which NFS is
1927  *	particularly sensitive to.
1928  */
1929 
1930 static int
1931 flushbufqueues(void)
1932 {
1933 	struct buf *bp;
1934 	int r = 0;
1935 
1936 	bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1937 
1938 	while (bp) {
1939 		KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
1940 		if ((bp->b_flags & B_DELWRI) != 0 &&
1941 		    (bp->b_xflags & BX_BKGRDINPROG) == 0) {
1942 			if (bp->b_flags & B_INVAL) {
1943 				if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1944 					panic("flushbufqueues: locked buf");
1945 				bremfree(bp);
1946 				brelse(bp);
1947 				++r;
1948 				break;
1949 			}
1950 			if (LIST_FIRST(&bp->b_dep) != NULL &&
1951 			    bioops.io_countdeps &&
1952 			    (bp->b_flags & B_DEFERRED) == 0 &&
1953 			    (*bioops.io_countdeps)(bp, 0)) {
1954 				TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY],
1955 				    bp, b_freelist);
1956 				TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY],
1957 				    bp, b_freelist);
1958 				bp->b_flags |= B_DEFERRED;
1959 				bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1960 				continue;
1961 			}
1962 			vfs_bio_awrite(bp);
1963 			++r;
1964 			break;
1965 		}
1966 		bp = TAILQ_NEXT(bp, b_freelist);
1967 	}
1968 	return (r);
1969 }
1970 
1971 /*
1972  * Check to see if a block is currently memory resident.
1973  */
1974 struct buf *
1975 incore(struct vnode * vp, daddr_t blkno)
1976 {
1977 	struct buf *bp;
1978 
1979 	crit_enter();
1980 	bp = gbincore(vp, blkno);
1981 	crit_exit();
1982 	return (bp);
1983 }
1984 
1985 /*
1986  * Returns true if no I/O is needed to access the associated VM object.
1987  * This is like incore except it also hunts around in the VM system for
1988  * the data.
1989  *
1990  * Note that we ignore vm_page_free() races from interrupts against our
1991  * lookup, since if the caller is not protected our return value will not
1992  * be any more valid then otherwise once we splx().
1993  */
1994 int
1995 inmem(struct vnode * vp, daddr_t blkno)
1996 {
1997 	vm_object_t obj;
1998 	vm_offset_t toff, tinc, size;
1999 	vm_page_t m;
2000 	vm_ooffset_t off;
2001 
2002 	if (incore(vp, blkno))
2003 		return 1;
2004 	if (vp->v_mount == NULL)
2005 		return 0;
2006 	if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0)
2007  		return 0;
2008 
2009 	size = PAGE_SIZE;
2010 	if (size > vp->v_mount->mnt_stat.f_iosize)
2011 		size = vp->v_mount->mnt_stat.f_iosize;
2012 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2013 
2014 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2015 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2016 		if (!m)
2017 			return 0;
2018 		tinc = size;
2019 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2020 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2021 		if (vm_page_is_valid(m,
2022 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2023 			return 0;
2024 	}
2025 	return 1;
2026 }
2027 
2028 /*
2029  *	vfs_setdirty:
2030  *
2031  *	Sets the dirty range for a buffer based on the status of the dirty
2032  *	bits in the pages comprising the buffer.
2033  *
2034  *	The range is limited to the size of the buffer.
2035  *
2036  *	This routine is primarily used by NFS, but is generalized for the
2037  *	B_VMIO case.
2038  */
2039 static void
2040 vfs_setdirty(struct buf *bp)
2041 {
2042 	int i;
2043 	vm_object_t object;
2044 
2045 	/*
2046 	 * Degenerate case - empty buffer
2047 	 */
2048 
2049 	if (bp->b_bufsize == 0)
2050 		return;
2051 
2052 	/*
2053 	 * We qualify the scan for modified pages on whether the
2054 	 * object has been flushed yet.  The OBJ_WRITEABLE flag
2055 	 * is not cleared simply by protecting pages off.
2056 	 */
2057 
2058 	if ((bp->b_flags & B_VMIO) == 0)
2059 		return;
2060 
2061 	object = bp->b_xio.xio_pages[0]->object;
2062 
2063 	if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2064 		printf("Warning: object %p writeable but not mightbedirty\n", object);
2065 	if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2066 		printf("Warning: object %p mightbedirty but not writeable\n", object);
2067 
2068 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2069 		vm_offset_t boffset;
2070 		vm_offset_t eoffset;
2071 
2072 		/*
2073 		 * test the pages to see if they have been modified directly
2074 		 * by users through the VM system.
2075 		 */
2076 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
2077 			vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2078 			vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2079 		}
2080 
2081 		/*
2082 		 * Calculate the encompassing dirty range, boffset and eoffset,
2083 		 * (eoffset - boffset) bytes.
2084 		 */
2085 
2086 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
2087 			if (bp->b_xio.xio_pages[i]->dirty)
2088 				break;
2089 		}
2090 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2091 
2092 		for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2093 			if (bp->b_xio.xio_pages[i]->dirty) {
2094 				break;
2095 			}
2096 		}
2097 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2098 
2099 		/*
2100 		 * Fit it to the buffer.
2101 		 */
2102 
2103 		if (eoffset > bp->b_bcount)
2104 			eoffset = bp->b_bcount;
2105 
2106 		/*
2107 		 * If we have a good dirty range, merge with the existing
2108 		 * dirty range.
2109 		 */
2110 
2111 		if (boffset < eoffset) {
2112 			if (bp->b_dirtyoff > boffset)
2113 				bp->b_dirtyoff = boffset;
2114 			if (bp->b_dirtyend < eoffset)
2115 				bp->b_dirtyend = eoffset;
2116 		}
2117 	}
2118 }
2119 
2120 /*
2121  *	getblk:
2122  *
2123  *	Get a block given a specified block and offset into a file/device.
2124  *	The buffers B_DONE bit will be cleared on return, making it almost
2125  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2126  *	return.  The caller should clear B_INVAL prior to initiating a
2127  *	READ.
2128  *
2129  *	IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2130  *	IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2131  *	OR SET B_INVAL BEFORE RETIRING IT.  If you retire a getblk'd buffer
2132  *	without doing any of those things the system will likely believe
2133  *	the buffer to be valid (especially if it is not B_VMIO), and the
2134  *	next getblk() will return the buffer with B_CACHE set.
2135  *
2136  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2137  *	an existing buffer.
2138  *
2139  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2140  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2141  *	and then cleared based on the backing VM.  If the previous buffer is
2142  *	non-0-sized but invalid, B_CACHE will be cleared.
2143  *
2144  *	If getblk() must create a new buffer, the new buffer is returned with
2145  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2146  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2147  *	backing VM.
2148  *
2149  *	getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos
2150  *	B_CACHE bit is clear.
2151  *
2152  *	What this means, basically, is that the caller should use B_CACHE to
2153  *	determine whether the buffer is fully valid or not and should clear
2154  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2155  *	the buffer by loading its data area with something, the caller needs
2156  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2157  *	the caller should set B_CACHE ( as an optimization ), else the caller
2158  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2159  *	a write attempt or if it was a successfull read.  If the caller
2160  *	intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2161  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2162  */
2163 struct buf *
2164 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
2165 {
2166 	struct buf *bp;
2167 	int s;
2168 	struct bufhashhdr *bh;
2169 
2170 	if (size > MAXBSIZE)
2171 		panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2172 
2173 	s = splbio();
2174 loop:
2175 	/*
2176 	 * Block if we are low on buffers.   Certain processes are allowed
2177 	 * to completely exhaust the buffer cache.
2178          *
2179          * If this check ever becomes a bottleneck it may be better to
2180          * move it into the else, when gbincore() fails.  At the moment
2181          * it isn't a problem.
2182 	 *
2183 	 * XXX remove, we cannot afford to block anywhere if holding a vnode
2184 	 * lock in low-memory situation, so take it to the max.
2185          */
2186 	if (numfreebuffers == 0) {
2187 		if (!curproc)
2188 			return NULL;
2189 		needsbuffer |= VFS_BIO_NEED_ANY;
2190 		tsleep(&needsbuffer, slpflag, "newbuf", slptimeo);
2191 	}
2192 
2193 	if ((bp = gbincore(vp, blkno))) {
2194 		/*
2195 		 * Buffer is in-core.  If the buffer is not busy, it must
2196 		 * be on a queue.
2197 		 */
2198 
2199 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2200 			if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL,
2201 			    "getblk", slpflag, slptimeo) == ENOLCK)
2202 				goto loop;
2203 			splx(s);
2204 			return (struct buf *) NULL;
2205 		}
2206 
2207 		/*
2208 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2209 		 * invalid.  Ohterwise, for a non-VMIO buffer, B_CACHE is set
2210 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2211 		 * backing VM cache.
2212 		 */
2213 		if (bp->b_flags & B_INVAL)
2214 			bp->b_flags &= ~B_CACHE;
2215 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2216 			bp->b_flags |= B_CACHE;
2217 		bremfree(bp);
2218 
2219 		/*
2220 		 * check for size inconsistancies for non-VMIO case.
2221 		 */
2222 
2223 		if (bp->b_bcount != size) {
2224 			if ((bp->b_flags & B_VMIO) == 0 ||
2225 			    (size > bp->b_kvasize)) {
2226 				if (bp->b_flags & B_DELWRI) {
2227 					bp->b_flags |= B_NOCACHE;
2228 					VOP_BWRITE(bp->b_vp, bp);
2229 				} else {
2230 					if ((bp->b_flags & B_VMIO) &&
2231 					   (LIST_FIRST(&bp->b_dep) == NULL)) {
2232 						bp->b_flags |= B_RELBUF;
2233 						brelse(bp);
2234 					} else {
2235 						bp->b_flags |= B_NOCACHE;
2236 						VOP_BWRITE(bp->b_vp, bp);
2237 					}
2238 				}
2239 				goto loop;
2240 			}
2241 		}
2242 
2243 		/*
2244 		 * If the size is inconsistant in the VMIO case, we can resize
2245 		 * the buffer.  This might lead to B_CACHE getting set or
2246 		 * cleared.  If the size has not changed, B_CACHE remains
2247 		 * unchanged from its previous state.
2248 		 */
2249 
2250 		if (bp->b_bcount != size)
2251 			allocbuf(bp, size);
2252 
2253 		KASSERT(bp->b_offset != NOOFFSET,
2254 		    ("getblk: no buffer offset"));
2255 
2256 		/*
2257 		 * A buffer with B_DELWRI set and B_CACHE clear must
2258 		 * be committed before we can return the buffer in
2259 		 * order to prevent the caller from issuing a read
2260 		 * ( due to B_CACHE not being set ) and overwriting
2261 		 * it.
2262 		 *
2263 		 * Most callers, including NFS and FFS, need this to
2264 		 * operate properly either because they assume they
2265 		 * can issue a read if B_CACHE is not set, or because
2266 		 * ( for example ) an uncached B_DELWRI might loop due
2267 		 * to softupdates re-dirtying the buffer.  In the latter
2268 		 * case, B_CACHE is set after the first write completes,
2269 		 * preventing further loops.
2270 		 *
2271 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2272 		 * above while extending the buffer, we cannot allow the
2273 		 * buffer to remain with B_CACHE set after the write
2274 		 * completes or it will represent a corrupt state.  To
2275 		 * deal with this we set B_NOCACHE to scrap the buffer
2276 		 * after the write.
2277 		 *
2278 		 * We might be able to do something fancy, like setting
2279 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2280 		 * so the below call doesn't set B_CACHE, but that gets real
2281 		 * confusing.  This is much easier.
2282 		 */
2283 
2284 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2285 			bp->b_flags |= B_NOCACHE;
2286 			VOP_BWRITE(bp->b_vp, bp);
2287 			goto loop;
2288 		}
2289 
2290 		splx(s);
2291 		bp->b_flags &= ~B_DONE;
2292 	} else {
2293 		/*
2294 		 * Buffer is not in-core, create new buffer.  The buffer
2295 		 * returned by getnewbuf() is locked.  Note that the returned
2296 		 * buffer is also considered valid (not marked B_INVAL).
2297 		 *
2298 		 * Calculating the offset for the I/O requires figuring out
2299 		 * the block size.  We use DEV_BSIZE for VBLK or VCHR and
2300 		 * the mount's f_iosize otherwise.  If the vnode does not
2301 		 * have an associated mount we assume that the passed size is
2302 		 * the block size.
2303 		 *
2304 		 * Note that vn_isdisk() cannot be used here since it may
2305 		 * return a failure for numerous reasons.   Note that the
2306 		 * buffer size may be larger then the block size (the caller
2307 		 * will use block numbers with the proper multiple).  Beware
2308 		 * of using any v_* fields which are part of unions.  In
2309 		 * particular, in DragonFly the mount point overloading
2310 		 * mechanism is such that the underlying directory (with a
2311 		 * non-NULL v_mountedhere) is not a special case.
2312 		 */
2313 		int bsize, maxsize, vmio;
2314 		off_t offset;
2315 
2316 		if (vp->v_type == VBLK || vp->v_type == VCHR)
2317 			bsize = DEV_BSIZE;
2318 		else if (vp->v_mount)
2319 			bsize = vp->v_mount->mnt_stat.f_iosize;
2320 		else
2321 			bsize = size;
2322 
2323 		offset = (off_t)blkno * bsize;
2324 		vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF);
2325 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2326 		maxsize = imax(maxsize, bsize);
2327 
2328 		if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2329 			if (slpflag || slptimeo) {
2330 				splx(s);
2331 				return NULL;
2332 			}
2333 			goto loop;
2334 		}
2335 
2336 		/*
2337 		 * This code is used to make sure that a buffer is not
2338 		 * created while the getnewbuf routine is blocked.
2339 		 * This can be a problem whether the vnode is locked or not.
2340 		 * If the buffer is created out from under us, we have to
2341 		 * throw away the one we just created.  There is now window
2342 		 * race because we are safely running at splbio() from the
2343 		 * point of the duplicate buffer creation through to here,
2344 		 * and we've locked the buffer.
2345 		 */
2346 		if (gbincore(vp, blkno)) {
2347 			bp->b_flags |= B_INVAL;
2348 			brelse(bp);
2349 			goto loop;
2350 		}
2351 
2352 		/*
2353 		 * Insert the buffer into the hash, so that it can
2354 		 * be found by incore.
2355 		 */
2356 		bp->b_blkno = bp->b_lblkno = blkno;
2357 		bp->b_offset = offset;
2358 
2359 		bgetvp(vp, bp);
2360 		LIST_REMOVE(bp, b_hash);
2361 		bh = bufhash(vp, blkno);
2362 		LIST_INSERT_HEAD(bh, bp, b_hash);
2363 
2364 		/*
2365 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2366 		 * buffer size starts out as 0, B_CACHE will be set by
2367 		 * allocbuf() for the VMIO case prior to it testing the
2368 		 * backing store for validity.
2369 		 */
2370 
2371 		if (vmio) {
2372 			bp->b_flags |= B_VMIO;
2373 #if defined(VFS_BIO_DEBUG)
2374 			if (vn_canvmio(vp) != TRUE)
2375 				printf("getblk: vmioing file type %d???\n", vp->v_type);
2376 #endif
2377 		} else {
2378 			bp->b_flags &= ~B_VMIO;
2379 		}
2380 
2381 		allocbuf(bp, size);
2382 
2383 		splx(s);
2384 		bp->b_flags &= ~B_DONE;
2385 	}
2386 	return (bp);
2387 }
2388 
2389 /*
2390  * Get an empty, disassociated buffer of given size.  The buffer is initially
2391  * set to B_INVAL.
2392  *
2393  * spl protection is not required for the allocbuf() call because races are
2394  * impossible here.
2395  */
2396 struct buf *
2397 geteblk(int size)
2398 {
2399 	struct buf *bp;
2400 	int s;
2401 	int maxsize;
2402 
2403 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2404 
2405 	s = splbio();
2406 	while ((bp = getnewbuf(0, 0, size, maxsize)) == 0);
2407 	splx(s);
2408 	allocbuf(bp, size);
2409 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2410 	return (bp);
2411 }
2412 
2413 
2414 /*
2415  * This code constitutes the buffer memory from either anonymous system
2416  * memory (in the case of non-VMIO operations) or from an associated
2417  * VM object (in the case of VMIO operations).  This code is able to
2418  * resize a buffer up or down.
2419  *
2420  * Note that this code is tricky, and has many complications to resolve
2421  * deadlock or inconsistant data situations.  Tread lightly!!!
2422  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2423  * the caller.  Calling this code willy nilly can result in the loss of data.
2424  *
2425  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2426  * B_CACHE for the non-VMIO case.
2427  *
2428  * This routine does not need to be called at splbio() but you must own the
2429  * buffer.
2430  */
2431 int
2432 allocbuf(struct buf *bp, int size)
2433 {
2434 	int newbsize, mbsize;
2435 	int i;
2436 
2437 	if (BUF_REFCNT(bp) == 0)
2438 		panic("allocbuf: buffer not busy");
2439 
2440 	if (bp->b_kvasize < size)
2441 		panic("allocbuf: buffer too small");
2442 
2443 	if ((bp->b_flags & B_VMIO) == 0) {
2444 		caddr_t origbuf;
2445 		int origbufsize;
2446 		/*
2447 		 * Just get anonymous memory from the kernel.  Don't
2448 		 * mess with B_CACHE.
2449 		 */
2450 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2451 #if !defined(NO_B_MALLOC)
2452 		if (bp->b_flags & B_MALLOC)
2453 			newbsize = mbsize;
2454 		else
2455 #endif
2456 			newbsize = round_page(size);
2457 
2458 		if (newbsize < bp->b_bufsize) {
2459 #if !defined(NO_B_MALLOC)
2460 			/*
2461 			 * malloced buffers are not shrunk
2462 			 */
2463 			if (bp->b_flags & B_MALLOC) {
2464 				if (newbsize) {
2465 					bp->b_bcount = size;
2466 				} else {
2467 					free(bp->b_data, M_BIOBUF);
2468 					if (bp->b_bufsize) {
2469 						bufmallocspace -= bp->b_bufsize;
2470 						bufspacewakeup();
2471 						bp->b_bufsize = 0;
2472 					}
2473 					bp->b_data = bp->b_kvabase;
2474 					bp->b_bcount = 0;
2475 					bp->b_flags &= ~B_MALLOC;
2476 				}
2477 				return 1;
2478 			}
2479 #endif
2480 			vm_hold_free_pages(
2481 			    bp,
2482 			    (vm_offset_t) bp->b_data + newbsize,
2483 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2484 		} else if (newbsize > bp->b_bufsize) {
2485 #if !defined(NO_B_MALLOC)
2486 			/*
2487 			 * We only use malloced memory on the first allocation.
2488 			 * and revert to page-allocated memory when the buffer
2489 			 * grows.
2490 			 */
2491 			if ( (bufmallocspace < maxbufmallocspace) &&
2492 				(bp->b_bufsize == 0) &&
2493 				(mbsize <= PAGE_SIZE/2)) {
2494 
2495 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2496 				bp->b_bufsize = mbsize;
2497 				bp->b_bcount = size;
2498 				bp->b_flags |= B_MALLOC;
2499 				bufmallocspace += mbsize;
2500 				return 1;
2501 			}
2502 #endif
2503 			origbuf = NULL;
2504 			origbufsize = 0;
2505 #if !defined(NO_B_MALLOC)
2506 			/*
2507 			 * If the buffer is growing on its other-than-first allocation,
2508 			 * then we revert to the page-allocation scheme.
2509 			 */
2510 			if (bp->b_flags & B_MALLOC) {
2511 				origbuf = bp->b_data;
2512 				origbufsize = bp->b_bufsize;
2513 				bp->b_data = bp->b_kvabase;
2514 				if (bp->b_bufsize) {
2515 					bufmallocspace -= bp->b_bufsize;
2516 					bufspacewakeup();
2517 					bp->b_bufsize = 0;
2518 				}
2519 				bp->b_flags &= ~B_MALLOC;
2520 				newbsize = round_page(newbsize);
2521 			}
2522 #endif
2523 			vm_hold_load_pages(
2524 			    bp,
2525 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2526 			    (vm_offset_t) bp->b_data + newbsize);
2527 #if !defined(NO_B_MALLOC)
2528 			if (origbuf) {
2529 				bcopy(origbuf, bp->b_data, origbufsize);
2530 				free(origbuf, M_BIOBUF);
2531 			}
2532 #endif
2533 		}
2534 	} else {
2535 		vm_page_t m;
2536 		int desiredpages;
2537 
2538 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2539 		desiredpages = (size == 0) ? 0 :
2540 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2541 
2542 #if !defined(NO_B_MALLOC)
2543 		if (bp->b_flags & B_MALLOC)
2544 			panic("allocbuf: VMIO buffer can't be malloced");
2545 #endif
2546 		/*
2547 		 * Set B_CACHE initially if buffer is 0 length or will become
2548 		 * 0-length.
2549 		 */
2550 		if (size == 0 || bp->b_bufsize == 0)
2551 			bp->b_flags |= B_CACHE;
2552 
2553 		if (newbsize < bp->b_bufsize) {
2554 			/*
2555 			 * DEV_BSIZE aligned new buffer size is less then the
2556 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2557 			 * if we have to remove any pages.
2558 			 */
2559 			if (desiredpages < bp->b_xio.xio_npages) {
2560 				for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2561 					/*
2562 					 * the page is not freed here -- it
2563 					 * is the responsibility of
2564 					 * vnode_pager_setsize
2565 					 */
2566 					m = bp->b_xio.xio_pages[i];
2567 					KASSERT(m != bogus_page,
2568 					    ("allocbuf: bogus page found"));
2569 					while (vm_page_sleep_busy(m, TRUE, "biodep"))
2570 						;
2571 
2572 					bp->b_xio.xio_pages[i] = NULL;
2573 					vm_page_unwire(m, 0);
2574 				}
2575 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2576 				    (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2577 				bp->b_xio.xio_npages = desiredpages;
2578 			}
2579 		} else if (size > bp->b_bcount) {
2580 			/*
2581 			 * We are growing the buffer, possibly in a
2582 			 * byte-granular fashion.
2583 			 */
2584 			struct vnode *vp;
2585 			vm_object_t obj;
2586 			vm_offset_t toff;
2587 			vm_offset_t tinc;
2588 
2589 			/*
2590 			 * Step 1, bring in the VM pages from the object,
2591 			 * allocating them if necessary.  We must clear
2592 			 * B_CACHE if these pages are not valid for the
2593 			 * range covered by the buffer.
2594 			 *
2595 			 * spl protection is required to protect against
2596 			 * interrupts unbusying and freeing pages between
2597 			 * our vm_page_lookup() and our busycheck/wiring
2598 			 * call.
2599 			 */
2600 			vp = bp->b_vp;
2601 			VOP_GETVOBJECT(vp, &obj);
2602 
2603 			crit_enter();
2604 			while (bp->b_xio.xio_npages < desiredpages) {
2605 				vm_page_t m;
2606 				vm_pindex_t pi;
2607 
2608 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_xio.xio_npages;
2609 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2610 					/*
2611 					 * note: must allocate system pages
2612 					 * since blocking here could intefere
2613 					 * with paging I/O, no matter which
2614 					 * process we are.
2615 					 */
2616 					m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
2617 					if (m == NULL) {
2618 						vm_wait();
2619 						vm_pageout_deficit += desiredpages -
2620 							bp->b_xio.xio_npages;
2621 					} else {
2622 						vm_page_wire(m);
2623 						vm_page_wakeup(m);
2624 						bp->b_flags &= ~B_CACHE;
2625 						bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2626 						++bp->b_xio.xio_npages;
2627 					}
2628 					continue;
2629 				}
2630 
2631 				/*
2632 				 * We found a page.  If we have to sleep on it,
2633 				 * retry because it might have gotten freed out
2634 				 * from under us.
2635 				 *
2636 				 * We can only test PG_BUSY here.  Blocking on
2637 				 * m->busy might lead to a deadlock:
2638 				 *
2639 				 *  vm_fault->getpages->cluster_read->allocbuf
2640 				 *
2641 				 */
2642 
2643 				if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2644 					continue;
2645 
2646 				/*
2647 				 * We have a good page.  Should we wakeup the
2648 				 * page daemon?
2649 				 */
2650 				if ((curthread != pagethread) &&
2651 				    ((m->queue - m->pc) == PQ_CACHE) &&
2652 				    ((vmstats.v_free_count + vmstats.v_cache_count) <
2653 					(vmstats.v_free_min + vmstats.v_cache_min))) {
2654 					pagedaemon_wakeup();
2655 				}
2656 				vm_page_flag_clear(m, PG_ZERO);
2657 				vm_page_wire(m);
2658 				bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2659 				++bp->b_xio.xio_npages;
2660 			}
2661 			crit_exit();
2662 
2663 			/*
2664 			 * Step 2.  We've loaded the pages into the buffer,
2665 			 * we have to figure out if we can still have B_CACHE
2666 			 * set.  Note that B_CACHE is set according to the
2667 			 * byte-granular range ( bcount and size ), new the
2668 			 * aligned range ( newbsize ).
2669 			 *
2670 			 * The VM test is against m->valid, which is DEV_BSIZE
2671 			 * aligned.  Needless to say, the validity of the data
2672 			 * needs to also be DEV_BSIZE aligned.  Note that this
2673 			 * fails with NFS if the server or some other client
2674 			 * extends the file's EOF.  If our buffer is resized,
2675 			 * B_CACHE may remain set! XXX
2676 			 */
2677 
2678 			toff = bp->b_bcount;
2679 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2680 
2681 			while ((bp->b_flags & B_CACHE) && toff < size) {
2682 				vm_pindex_t pi;
2683 
2684 				if (tinc > (size - toff))
2685 					tinc = size - toff;
2686 
2687 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2688 				    PAGE_SHIFT;
2689 
2690 				vfs_buf_test_cache(
2691 				    bp,
2692 				    bp->b_offset,
2693 				    toff,
2694 				    tinc,
2695 				    bp->b_xio.xio_pages[pi]
2696 				);
2697 				toff += tinc;
2698 				tinc = PAGE_SIZE;
2699 			}
2700 
2701 			/*
2702 			 * Step 3, fixup the KVM pmap.  Remember that
2703 			 * bp->b_data is relative to bp->b_offset, but
2704 			 * bp->b_offset may be offset into the first page.
2705 			 */
2706 
2707 			bp->b_data = (caddr_t)
2708 			    trunc_page((vm_offset_t)bp->b_data);
2709 			pmap_qenter(
2710 			    (vm_offset_t)bp->b_data,
2711 			    bp->b_xio.xio_pages,
2712 			    bp->b_xio.xio_npages
2713 			);
2714 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2715 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
2716 		}
2717 	}
2718 	if (newbsize < bp->b_bufsize)
2719 		bufspacewakeup();
2720 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
2721 	bp->b_bcount = size;		/* requested buffer size	*/
2722 	return 1;
2723 }
2724 
2725 /*
2726  *	biowait:
2727  *
2728  *	Wait for buffer I/O completion, returning error status.  The buffer
2729  *	is left locked and B_DONE on return.  B_EINTR is converted into a EINTR
2730  *	error and cleared.
2731  */
2732 int
2733 biowait(struct buf * bp)
2734 {
2735 	int s;
2736 
2737 	s = splbio();
2738 	while ((bp->b_flags & B_DONE) == 0) {
2739 #if defined(NO_SCHEDULE_MODS)
2740 		tsleep(bp, 0, "biowait", 0);
2741 #else
2742 		if (bp->b_flags & B_READ)
2743 			tsleep(bp, 0, "biord", 0);
2744 		else
2745 			tsleep(bp, 0, "biowr", 0);
2746 #endif
2747 	}
2748 	splx(s);
2749 	if (bp->b_flags & B_EINTR) {
2750 		bp->b_flags &= ~B_EINTR;
2751 		return (EINTR);
2752 	}
2753 	if (bp->b_flags & B_ERROR) {
2754 		return (bp->b_error ? bp->b_error : EIO);
2755 	} else {
2756 		return (0);
2757 	}
2758 }
2759 
2760 /*
2761  *	biodone:
2762  *
2763  *	Finish I/O on a buffer, optionally calling a completion function.
2764  *	This is usually called from an interrupt so process blocking is
2765  *	not allowed.
2766  *
2767  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2768  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
2769  *	assuming B_INVAL is clear.
2770  *
2771  *	For the VMIO case, we set B_CACHE if the op was a read and no
2772  *	read error occured, or if the op was a write.  B_CACHE is never
2773  *	set if the buffer is invalid or otherwise uncacheable.
2774  *
2775  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
2776  *	initiator to leave B_INVAL set to brelse the buffer out of existance
2777  *	in the biodone routine.
2778  *
2779  *	b_dev is required to be reinitialized prior to the top level strategy
2780  *	call in a device stack.  To avoid improper reuse, biodone() sets
2781  *	b_dev to NODEV.
2782  */
2783 void
2784 biodone(struct buf *bp)
2785 {
2786 	int s, error;
2787 
2788 	s = splbio();
2789 
2790 	KASSERT(BUF_REFCNTNB(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
2791 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
2792 
2793 	bp->b_flags |= B_DONE;
2794 	bp->b_dev = NODEV;
2795 	runningbufwakeup(bp);
2796 
2797 	if (bp->b_flags & B_FREEBUF) {
2798 		brelse(bp);
2799 		splx(s);
2800 		return;
2801 	}
2802 
2803 	if ((bp->b_flags & B_READ) == 0) {
2804 		vwakeup(bp);
2805 	}
2806 
2807 	/* call optional completion function if requested */
2808 	if (bp->b_flags & B_CALL) {
2809 		bp->b_flags &= ~B_CALL;
2810 		(*bp->b_iodone) (bp);
2811 		splx(s);
2812 		return;
2813 	}
2814 	if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
2815 		(*bioops.io_complete)(bp);
2816 
2817 	if (bp->b_flags & B_VMIO) {
2818 		int i;
2819 		vm_ooffset_t foff;
2820 		vm_page_t m;
2821 		vm_object_t obj;
2822 		int iosize;
2823 		struct vnode *vp = bp->b_vp;
2824 
2825 		error = VOP_GETVOBJECT(vp, &obj);
2826 
2827 #if defined(VFS_BIO_DEBUG)
2828 		if (vp->v_holdcnt == 0) {
2829 			panic("biodone: zero vnode hold count");
2830 		}
2831 
2832 		if (error) {
2833 			panic("biodone: missing VM object");
2834 		}
2835 
2836 		if ((vp->v_flag & VOBJBUF) == 0) {
2837 			panic("biodone: vnode is not setup for merged cache");
2838 		}
2839 #endif
2840 
2841 		foff = bp->b_offset;
2842 		KASSERT(bp->b_offset != NOOFFSET,
2843 		    ("biodone: no buffer offset"));
2844 
2845 		if (error) {
2846 			panic("biodone: no object");
2847 		}
2848 #if defined(VFS_BIO_DEBUG)
2849 		if (obj->paging_in_progress < bp->b_xio.xio_npages) {
2850 			printf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
2851 			    obj->paging_in_progress, bp->b_xio.xio_npages);
2852 		}
2853 #endif
2854 
2855 		/*
2856 		 * Set B_CACHE if the op was a normal read and no error
2857 		 * occured.  B_CACHE is set for writes in the b*write()
2858 		 * routines.
2859 		 */
2860 		iosize = bp->b_bcount - bp->b_resid;
2861 		if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) {
2862 			bp->b_flags |= B_CACHE;
2863 		}
2864 
2865 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
2866 			int bogusflag = 0;
2867 			int resid;
2868 
2869 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2870 			if (resid > iosize)
2871 				resid = iosize;
2872 
2873 			/*
2874 			 * cleanup bogus pages, restoring the originals.  Since
2875 			 * the originals should still be wired, we don't have
2876 			 * to worry about interrupt/freeing races destroying
2877 			 * the VM object association.
2878 			 */
2879 			m = bp->b_xio.xio_pages[i];
2880 			if (m == bogus_page) {
2881 				bogusflag = 1;
2882 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2883 				if (m == NULL)
2884 					panic("biodone: page disappeared");
2885 				bp->b_xio.xio_pages[i] = m;
2886 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2887 					bp->b_xio.xio_pages, bp->b_xio.xio_npages);
2888 			}
2889 #if defined(VFS_BIO_DEBUG)
2890 			if (OFF_TO_IDX(foff) != m->pindex) {
2891 				printf(
2892 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
2893 				    (unsigned long)foff, m->pindex);
2894 			}
2895 #endif
2896 
2897 			/*
2898 			 * In the write case, the valid and clean bits are
2899 			 * already changed correctly ( see bdwrite() ), so we
2900 			 * only need to do this here in the read case.
2901 			 */
2902 			if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) {
2903 				vfs_page_set_valid(bp, foff, i, m);
2904 			}
2905 			vm_page_flag_clear(m, PG_ZERO);
2906 
2907 			/*
2908 			 * when debugging new filesystems or buffer I/O methods, this
2909 			 * is the most common error that pops up.  if you see this, you
2910 			 * have not set the page busy flag correctly!!!
2911 			 */
2912 			if (m->busy == 0) {
2913 				printf("biodone: page busy < 0, "
2914 				    "pindex: %d, foff: 0x(%x,%x), "
2915 				    "resid: %d, index: %d\n",
2916 				    (int) m->pindex, (int)(foff >> 32),
2917 						(int) foff & 0xffffffff, resid, i);
2918 				if (!vn_isdisk(vp, NULL))
2919 					printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n",
2920 					    bp->b_vp->v_mount->mnt_stat.f_iosize,
2921 					    (int) bp->b_lblkno,
2922 					    bp->b_flags, bp->b_xio.xio_npages);
2923 				else
2924 					printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n",
2925 					    (int) bp->b_lblkno,
2926 					    bp->b_flags, bp->b_xio.xio_npages);
2927 				printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
2928 				    m->valid, m->dirty, m->wire_count);
2929 				panic("biodone: page busy < 0");
2930 			}
2931 			vm_page_io_finish(m);
2932 			vm_object_pip_subtract(obj, 1);
2933 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2934 			iosize -= resid;
2935 		}
2936 		if (obj)
2937 			vm_object_pip_wakeupn(obj, 0);
2938 	}
2939 
2940 	/*
2941 	 * For asynchronous completions, release the buffer now. The brelse
2942 	 * will do a wakeup there if necessary - so no need to do a wakeup
2943 	 * here in the async case. The sync case always needs to do a wakeup.
2944 	 */
2945 
2946 	if (bp->b_flags & B_ASYNC) {
2947 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
2948 			brelse(bp);
2949 		else
2950 			bqrelse(bp);
2951 	} else {
2952 		wakeup(bp);
2953 	}
2954 	splx(s);
2955 }
2956 
2957 /*
2958  * This routine is called in lieu of iodone in the case of
2959  * incomplete I/O.  This keeps the busy status for pages
2960  * consistant.
2961  */
2962 void
2963 vfs_unbusy_pages(struct buf *bp)
2964 {
2965 	int i;
2966 
2967 	runningbufwakeup(bp);
2968 	if (bp->b_flags & B_VMIO) {
2969 		struct vnode *vp = bp->b_vp;
2970 		vm_object_t obj;
2971 
2972 		VOP_GETVOBJECT(vp, &obj);
2973 
2974 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
2975 			vm_page_t m = bp->b_xio.xio_pages[i];
2976 
2977 			/*
2978 			 * When restoring bogus changes the original pages
2979 			 * should still be wired, so we are in no danger of
2980 			 * losing the object association and do not need
2981 			 * spl protection particularly.
2982 			 */
2983 			if (m == bogus_page) {
2984 				m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
2985 				if (!m) {
2986 					panic("vfs_unbusy_pages: page missing");
2987 				}
2988 				bp->b_xio.xio_pages[i] = m;
2989 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2990 					bp->b_xio.xio_pages, bp->b_xio.xio_npages);
2991 			}
2992 			vm_object_pip_subtract(obj, 1);
2993 			vm_page_flag_clear(m, PG_ZERO);
2994 			vm_page_io_finish(m);
2995 		}
2996 		vm_object_pip_wakeupn(obj, 0);
2997 	}
2998 }
2999 
3000 /*
3001  * vfs_page_set_valid:
3002  *
3003  *	Set the valid bits in a page based on the supplied offset.   The
3004  *	range is restricted to the buffer's size.
3005  *
3006  *	This routine is typically called after a read completes.
3007  */
3008 static void
3009 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3010 {
3011 	vm_ooffset_t soff, eoff;
3012 
3013 	/*
3014 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3015 	 * page boundry or cross the end of the buffer.  The end of the
3016 	 * buffer, in this case, is our file EOF, not the allocation size
3017 	 * of the buffer.
3018 	 */
3019 	soff = off;
3020 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3021 	if (eoff > bp->b_offset + bp->b_bcount)
3022 		eoff = bp->b_offset + bp->b_bcount;
3023 
3024 	/*
3025 	 * Set valid range.  This is typically the entire buffer and thus the
3026 	 * entire page.
3027 	 */
3028 	if (eoff > soff) {
3029 		vm_page_set_validclean(
3030 		    m,
3031 		   (vm_offset_t) (soff & PAGE_MASK),
3032 		   (vm_offset_t) (eoff - soff)
3033 		);
3034 	}
3035 }
3036 
3037 /*
3038  * This routine is called before a device strategy routine.
3039  * It is used to tell the VM system that paging I/O is in
3040  * progress, and treat the pages associated with the buffer
3041  * almost as being PG_BUSY.  Also the object paging_in_progress
3042  * flag is handled to make sure that the object doesn't become
3043  * inconsistant.
3044  *
3045  * Since I/O has not been initiated yet, certain buffer flags
3046  * such as B_ERROR or B_INVAL may be in an inconsistant state
3047  * and should be ignored.
3048  */
3049 void
3050 vfs_busy_pages(struct buf *bp, int clear_modify)
3051 {
3052 	int i, bogus;
3053 
3054 	if (bp->b_flags & B_VMIO) {
3055 		struct vnode *vp = bp->b_vp;
3056 		vm_object_t obj;
3057 		vm_ooffset_t foff;
3058 
3059 		VOP_GETVOBJECT(vp, &obj);
3060 		foff = bp->b_offset;
3061 		KASSERT(bp->b_offset != NOOFFSET,
3062 		    ("vfs_busy_pages: no buffer offset"));
3063 		vfs_setdirty(bp);
3064 
3065 retry:
3066 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3067 			vm_page_t m = bp->b_xio.xio_pages[i];
3068 			if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3069 				goto retry;
3070 		}
3071 
3072 		bogus = 0;
3073 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3074 			vm_page_t m = bp->b_xio.xio_pages[i];
3075 
3076 			vm_page_flag_clear(m, PG_ZERO);
3077 			if ((bp->b_flags & B_CLUSTER) == 0) {
3078 				vm_object_pip_add(obj, 1);
3079 				vm_page_io_start(m);
3080 			}
3081 
3082 			/*
3083 			 * When readying a buffer for a read ( i.e
3084 			 * clear_modify == 0 ), it is important to do
3085 			 * bogus_page replacement for valid pages in
3086 			 * partially instantiated buffers.  Partially
3087 			 * instantiated buffers can, in turn, occur when
3088 			 * reconstituting a buffer from its VM backing store
3089 			 * base.  We only have to do this if B_CACHE is
3090 			 * clear ( which causes the I/O to occur in the
3091 			 * first place ).  The replacement prevents the read
3092 			 * I/O from overwriting potentially dirty VM-backed
3093 			 * pages.  XXX bogus page replacement is, uh, bogus.
3094 			 * It may not work properly with small-block devices.
3095 			 * We need to find a better way.
3096 			 */
3097 
3098 			vm_page_protect(m, VM_PROT_NONE);
3099 			if (clear_modify)
3100 				vfs_page_set_valid(bp, foff, i, m);
3101 			else if (m->valid == VM_PAGE_BITS_ALL &&
3102 				(bp->b_flags & B_CACHE) == 0) {
3103 				bp->b_xio.xio_pages[i] = bogus_page;
3104 				bogus++;
3105 			}
3106 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3107 		}
3108 		if (bogus)
3109 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3110 				bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3111 	}
3112 
3113 	/*
3114 	 * This is the easiest place to put the process accounting for the I/O
3115 	 * for now.
3116 	 */
3117 	{
3118 		struct proc *p;
3119 
3120 		if ((p = curthread->td_proc) != NULL) {
3121 			if (bp->b_flags & B_READ)
3122 				p->p_stats->p_ru.ru_inblock++;
3123 			else
3124 				p->p_stats->p_ru.ru_oublock++;
3125 		}
3126 	}
3127 }
3128 
3129 /*
3130  * Tell the VM system that the pages associated with this buffer
3131  * are clean.  This is used for delayed writes where the data is
3132  * going to go to disk eventually without additional VM intevention.
3133  *
3134  * Note that while we only really need to clean through to b_bcount, we
3135  * just go ahead and clean through to b_bufsize.
3136  */
3137 static void
3138 vfs_clean_pages(struct buf *bp)
3139 {
3140 	int i;
3141 
3142 	if (bp->b_flags & B_VMIO) {
3143 		vm_ooffset_t foff;
3144 
3145 		foff = bp->b_offset;
3146 		KASSERT(bp->b_offset != NOOFFSET,
3147 		    ("vfs_clean_pages: no buffer offset"));
3148 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3149 			vm_page_t m = bp->b_xio.xio_pages[i];
3150 			vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3151 			vm_ooffset_t eoff = noff;
3152 
3153 			if (eoff > bp->b_offset + bp->b_bufsize)
3154 				eoff = bp->b_offset + bp->b_bufsize;
3155 			vfs_page_set_valid(bp, foff, i, m);
3156 			/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3157 			foff = noff;
3158 		}
3159 	}
3160 }
3161 
3162 /*
3163  *	vfs_bio_set_validclean:
3164  *
3165  *	Set the range within the buffer to valid and clean.  The range is
3166  *	relative to the beginning of the buffer, b_offset.  Note that b_offset
3167  *	itself may be offset from the beginning of the first page.
3168  */
3169 
3170 void
3171 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3172 {
3173 	if (bp->b_flags & B_VMIO) {
3174 		int i;
3175 		int n;
3176 
3177 		/*
3178 		 * Fixup base to be relative to beginning of first page.
3179 		 * Set initial n to be the maximum number of bytes in the
3180 		 * first page that can be validated.
3181 		 */
3182 
3183 		base += (bp->b_offset & PAGE_MASK);
3184 		n = PAGE_SIZE - (base & PAGE_MASK);
3185 
3186 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3187 			vm_page_t m = bp->b_xio.xio_pages[i];
3188 
3189 			if (n > size)
3190 				n = size;
3191 
3192 			vm_page_set_validclean(m, base & PAGE_MASK, n);
3193 			base += n;
3194 			size -= n;
3195 			n = PAGE_SIZE;
3196 		}
3197 	}
3198 }
3199 
3200 /*
3201  *	vfs_bio_clrbuf:
3202  *
3203  *	clear a buffer.  This routine essentially fakes an I/O, so we need
3204  *	to clear B_ERROR and B_INVAL.
3205  *
3206  *	Note that while we only theoretically need to clear through b_bcount,
3207  *	we go ahead and clear through b_bufsize.
3208  */
3209 
3210 void
3211 vfs_bio_clrbuf(struct buf *bp)
3212 {
3213 	int i, mask = 0;
3214 	caddr_t sa, ea;
3215 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3216 		bp->b_flags &= ~(B_INVAL|B_ERROR);
3217 		if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3218 		    (bp->b_offset & PAGE_MASK) == 0) {
3219 			mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3220 			if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3221 				bp->b_resid = 0;
3222 				return;
3223 			}
3224 			if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3225 			    ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3226 				bzero(bp->b_data, bp->b_bufsize);
3227 				bp->b_xio.xio_pages[0]->valid |= mask;
3228 				bp->b_resid = 0;
3229 				return;
3230 			}
3231 		}
3232 		ea = sa = bp->b_data;
3233 		for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3234 			int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3235 			ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3236 			ea = (caddr_t)(vm_offset_t)ulmin(
3237 			    (u_long)(vm_offset_t)ea,
3238 			    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3239 			mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3240 			if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3241 				continue;
3242 			if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3243 				if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3244 					bzero(sa, ea - sa);
3245 				}
3246 			} else {
3247 				for (; sa < ea; sa += DEV_BSIZE, j++) {
3248 					if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3249 						(bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3250 						bzero(sa, DEV_BSIZE);
3251 				}
3252 			}
3253 			bp->b_xio.xio_pages[i]->valid |= mask;
3254 			vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3255 		}
3256 		bp->b_resid = 0;
3257 	} else {
3258 		clrbuf(bp);
3259 	}
3260 }
3261 
3262 /*
3263  * vm_hold_load_pages and vm_hold_unload pages get pages into
3264  * a buffers address space.  The pages are anonymous and are
3265  * not associated with a file object.
3266  */
3267 void
3268 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3269 {
3270 	vm_offset_t pg;
3271 	vm_page_t p;
3272 	int index;
3273 
3274 	to = round_page(to);
3275 	from = round_page(from);
3276 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3277 
3278 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3279 
3280 tryagain:
3281 
3282 		/*
3283 		 * note: must allocate system pages since blocking here
3284 		 * could intefere with paging I/O, no matter which
3285 		 * process we are.
3286 		 */
3287 		p = vm_page_alloc(kernel_object,
3288 			((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3289 			VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3290 		if (!p) {
3291 			vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3292 			vm_wait();
3293 			goto tryagain;
3294 		}
3295 		vm_page_wire(p);
3296 		p->valid = VM_PAGE_BITS_ALL;
3297 		vm_page_flag_clear(p, PG_ZERO);
3298 		pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3299 		bp->b_xio.xio_pages[index] = p;
3300 		vm_page_wakeup(p);
3301 	}
3302 	bp->b_xio.xio_npages = index;
3303 }
3304 
3305 void
3306 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3307 {
3308 	vm_offset_t pg;
3309 	vm_page_t p;
3310 	int index, newnpages;
3311 
3312 	from = round_page(from);
3313 	to = round_page(to);
3314 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3315 
3316 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3317 		p = bp->b_xio.xio_pages[index];
3318 		if (p && (index < bp->b_xio.xio_npages)) {
3319 			if (p->busy) {
3320 				printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n",
3321 					bp->b_blkno, bp->b_lblkno);
3322 			}
3323 			bp->b_xio.xio_pages[index] = NULL;
3324 			pmap_kremove(pg);
3325 			vm_page_busy(p);
3326 			vm_page_unwire(p, 0);
3327 			vm_page_free(p);
3328 		}
3329 	}
3330 	bp->b_xio.xio_npages = newnpages;
3331 }
3332 
3333 /*
3334  * Map an IO request into kernel virtual address space.
3335  *
3336  * All requests are (re)mapped into kernel VA space.
3337  * Notice that we use b_bufsize for the size of the buffer
3338  * to be mapped.  b_bcount might be modified by the driver.
3339  */
3340 int
3341 vmapbuf(struct buf *bp)
3342 {
3343 	caddr_t addr, v, kva;
3344 	vm_paddr_t pa;
3345 	int pidx;
3346 	int i;
3347 	struct vm_page *m;
3348 
3349 	if ((bp->b_flags & B_PHYS) == 0)
3350 		panic("vmapbuf");
3351 	if (bp->b_bufsize < 0)
3352 		return (-1);
3353 	for (v = bp->b_saveaddr,
3354 		     addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data),
3355 		     pidx = 0;
3356 	     addr < bp->b_data + bp->b_bufsize;
3357 	     addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) {
3358 		/*
3359 		 * Do the vm_fault if needed; do the copy-on-write thing
3360 		 * when reading stuff off device into memory.
3361 		 */
3362 retry:
3363 		i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data,
3364 			(bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ);
3365 		if (i < 0) {
3366 			for (i = 0; i < pidx; ++i) {
3367 			    vm_page_unhold(bp->b_xio.xio_pages[i]);
3368 			    bp->b_xio.xio_pages[i] = NULL;
3369 			}
3370 			return(-1);
3371 		}
3372 
3373 		/*
3374 		 * WARNING!  If sparc support is MFCd in the future this will
3375 		 * have to be changed from pmap_kextract() to pmap_extract()
3376 		 * ala -current.
3377 		 */
3378 #ifdef __sparc64__
3379 #error "If MFCing sparc support use pmap_extract"
3380 #endif
3381 		pa = pmap_kextract((vm_offset_t)addr);
3382 		if (pa == 0) {
3383 			printf("vmapbuf: warning, race against user address during I/O");
3384 			goto retry;
3385 		}
3386 		m = PHYS_TO_VM_PAGE(pa);
3387 		vm_page_hold(m);
3388 		bp->b_xio.xio_pages[pidx] = m;
3389 	}
3390 	if (pidx > btoc(MAXPHYS))
3391 		panic("vmapbuf: mapped more than MAXPHYS");
3392 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_xio.xio_pages, pidx);
3393 
3394 	kva = bp->b_saveaddr;
3395 	bp->b_xio.xio_npages = pidx;
3396 	bp->b_saveaddr = bp->b_data;
3397 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3398 	return(0);
3399 }
3400 
3401 /*
3402  * Free the io map PTEs associated with this IO operation.
3403  * We also invalidate the TLB entries and restore the original b_addr.
3404  */
3405 void
3406 vunmapbuf(struct buf *bp)
3407 {
3408 	int pidx;
3409 	int npages;
3410 	vm_page_t *m;
3411 
3412 	if ((bp->b_flags & B_PHYS) == 0)
3413 		panic("vunmapbuf");
3414 
3415 	npages = bp->b_xio.xio_npages;
3416 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3417 		     npages);
3418 	m = bp->b_xio.xio_pages;
3419 	for (pidx = 0; pidx < npages; pidx++)
3420 		vm_page_unhold(*m++);
3421 
3422 	bp->b_data = bp->b_saveaddr;
3423 }
3424 
3425 #include "opt_ddb.h"
3426 #ifdef DDB
3427 #include <ddb/ddb.h>
3428 
3429 DB_SHOW_COMMAND(buffer, db_show_buffer)
3430 {
3431 	/* get args */
3432 	struct buf *bp = (struct buf *)addr;
3433 
3434 	if (!have_addr) {
3435 		db_printf("usage: show buffer <addr>\n");
3436 		return;
3437 	}
3438 
3439 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3440 	db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, "
3441 		  "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, "
3442 		  "b_blkno = %d, b_pblkno = %d\n",
3443 		  bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3444 		  major(bp->b_dev), minor(bp->b_dev),
3445 		  bp->b_data, bp->b_blkno, bp->b_pblkno);
3446 	if (bp->b_xio.xio_npages) {
3447 		int i;
3448 		db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3449 			bp->b_xio.xio_npages);
3450 		for (i = 0; i < bp->b_xio.xio_npages; i++) {
3451 			vm_page_t m;
3452 			m = bp->b_xio.xio_pages[i];
3453 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3454 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3455 			if ((i + 1) < bp->b_xio.xio_npages)
3456 				db_printf(",");
3457 		}
3458 		db_printf("\n");
3459 	}
3460 }
3461 #endif /* DDB */
3462