xref: /freebsd/sys/kern/vfs_bio.c (revision 9336e0699bda8a301cd2bfa37106b6ec5e32012e)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * this file contains a new buffer I/O scheme implementing a coherent
30  * VM object and buffer cache scheme.  Pains have been taken to make
31  * sure that the performance degradation associated with schemes such
32  * as this is not realized.
33  *
34  * Author:  John S. Dyson
35  * Significant help during the development and debugging phases
36  * had been provided by David Greenman, also of the FreeBSD core team.
37  *
38  * see man buf(9) for more info.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bio.h>
47 #include <sys/conf.h>
48 #include <sys/buf.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/malloc.h>
54 #include <sys/mount.h>
55 #include <sys/mutex.h>
56 #include <sys/kernel.h>
57 #include <sys/kthread.h>
58 #include <sys/proc.h>
59 #include <sys/resourcevar.h>
60 #include <sys/sysctl.h>
61 #include <sys/vmmeter.h>
62 #include <sys/vnode.h>
63 #include <geom/geom.h>
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_pageout.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_map.h>
72 #include "opt_directio.h"
73 #include "opt_swap.h"
74 
75 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
76 
77 struct	bio_ops bioops;		/* I/O operation notification */
78 
79 struct	buf_ops buf_ops_bio = {
80 	.bop_name	=	"buf_ops_bio",
81 	.bop_write	=	bufwrite,
82 	.bop_strategy	=	bufstrategy,
83 	.bop_sync	=	bufsync,
84 	.bop_bdflush	=	bufbdflush,
85 };
86 
87 /*
88  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
89  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
90  */
91 struct buf *buf;		/* buffer header pool */
92 
93 static struct proc *bufdaemonproc;
94 
95 static int inmem(struct vnode *vp, daddr_t blkno);
96 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from,
97 		vm_offset_t to);
98 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
99 		vm_offset_t to);
100 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
101 		vm_page_t m);
102 static void vfs_clean_pages(struct buf *bp);
103 static void vfs_setdirty(struct buf *bp);
104 static void vfs_setdirty_locked_object(struct buf *bp);
105 static void vfs_vmio_release(struct buf *bp);
106 static int vfs_bio_clcheck(struct vnode *vp, int size,
107 		daddr_t lblkno, daddr_t blkno);
108 static int flushbufqueues(int, int);
109 static void buf_daemon(void);
110 static void bremfreel(struct buf *bp);
111 
112 int vmiodirenable = TRUE;
113 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
114     "Use the VM system for directory writes");
115 int runningbufspace;
116 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
117     "Amount of presently outstanding async buffer io");
118 static int bufspace;
119 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
120     "KVA memory used for bufs");
121 static int maxbufspace;
122 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
123     "Maximum allowed value of bufspace (including buf_daemon)");
124 static int bufmallocspace;
125 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
126     "Amount of malloced memory for buffers");
127 static int maxbufmallocspace;
128 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
129     "Maximum amount of malloced memory for buffers");
130 static int lobufspace;
131 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
132     "Minimum amount of buffers we want to have");
133 int hibufspace;
134 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
135     "Maximum allowed value of bufspace (excluding buf_daemon)");
136 static int bufreusecnt;
137 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
138     "Number of times we have reused a buffer");
139 static int buffreekvacnt;
140 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
141     "Number of times we have freed the KVA space from some buffer");
142 static int bufdefragcnt;
143 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
144     "Number of times we have had to repeat buffer allocation to defragment");
145 static int lorunningspace;
146 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
147     "Minimum preferred space used for in-progress I/O");
148 static int hirunningspace;
149 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
150     "Maximum amount of space to use for in-progress I/O");
151 int dirtybufferflushes;
152 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
153     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
154 int bdwriteskip;
155 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
156     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
157 int altbufferflushes;
158 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
159     0, "Number of fsync flushes to limit dirty buffers");
160 static int recursiveflushes;
161 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
162     0, "Number of flushes skipped due to being recursive");
163 static int numdirtybuffers;
164 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
165     "Number of buffers that are dirty (has unwritten changes) at the moment");
166 static int lodirtybuffers;
167 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
168     "How many buffers we want to have free before bufdaemon can sleep");
169 static int hidirtybuffers;
170 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
171     "When the number of dirty buffers is considered severe");
172 int dirtybufthresh;
173 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
174     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
175 static int numfreebuffers;
176 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
177     "Number of free buffers");
178 static int lofreebuffers;
179 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
180    "XXX Unused");
181 static int hifreebuffers;
182 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
183    "XXX Complicatedly unused");
184 static int getnewbufcalls;
185 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
186    "Number of calls to getnewbuf");
187 static int getnewbufrestarts;
188 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
189     "Number of times getnewbuf has had to restart a buffer aquisition");
190 
191 /*
192  * Wakeup point for bufdaemon, as well as indicator of whether it is already
193  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
194  * is idling.
195  */
196 static int bd_request;
197 
198 /*
199  * This lock synchronizes access to bd_request.
200  */
201 static struct mtx bdlock;
202 
203 /*
204  * bogus page -- for I/O to/from partially complete buffers
205  * this is a temporary solution to the problem, but it is not
206  * really that bad.  it would be better to split the buffer
207  * for input in the case of buffers partially already in memory,
208  * but the code is intricate enough already.
209  */
210 vm_page_t bogus_page;
211 
212 /*
213  * Synchronization (sleep/wakeup) variable for active buffer space requests.
214  * Set when wait starts, cleared prior to wakeup().
215  * Used in runningbufwakeup() and waitrunningbufspace().
216  */
217 static int runningbufreq;
218 
219 /*
220  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
221  * waitrunningbufspace().
222  */
223 static struct mtx rbreqlock;
224 
225 /*
226  * Synchronization (sleep/wakeup) variable for buffer requests.
227  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
228  * by and/or.
229  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
230  * getnewbuf(), and getblk().
231  */
232 static int needsbuffer;
233 
234 /*
235  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
236  */
237 static struct mtx nblock;
238 
239 /*
240  * Lock that protects against bwait()/bdone()/B_DONE races.
241  */
242 
243 static struct mtx bdonelock;
244 
245 /*
246  * Lock that protects against bwait()/bdone()/B_DONE races.
247  */
248 static struct mtx bpinlock;
249 
250 /*
251  * Definitions for the buffer free lists.
252  */
253 #define BUFFER_QUEUES	6	/* number of free buffer queues */
254 
255 #define QUEUE_NONE	0	/* on no queue */
256 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
257 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
258 #define QUEUE_DIRTY_GIANT 3	/* B_DELWRI buffers that need giant */
259 #define QUEUE_EMPTYKVA	4	/* empty buffer headers w/KVA assignment */
260 #define QUEUE_EMPTY	5	/* empty buffer headers */
261 
262 /* Queues for free buffers with various properties */
263 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
264 
265 /* Lock for the bufqueues */
266 static struct mtx bqlock;
267 
268 /*
269  * Single global constant for BUF_WMESG, to avoid getting multiple references.
270  * buf_wmesg is referred from macros.
271  */
272 const char *buf_wmesg = BUF_WMESG;
273 
274 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
275 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
276 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
277 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
278 
279 #ifdef DIRECTIO
280 extern void ffs_rawread_setup(void);
281 #endif /* DIRECTIO */
282 /*
283  *	numdirtywakeup:
284  *
285  *	If someone is blocked due to there being too many dirty buffers,
286  *	and numdirtybuffers is now reasonable, wake them up.
287  */
288 
289 static __inline void
290 numdirtywakeup(int level)
291 {
292 
293 	if (numdirtybuffers <= level) {
294 		mtx_lock(&nblock);
295 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
296 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
297 			wakeup(&needsbuffer);
298 		}
299 		mtx_unlock(&nblock);
300 	}
301 }
302 
303 /*
304  *	bufspacewakeup:
305  *
306  *	Called when buffer space is potentially available for recovery.
307  *	getnewbuf() will block on this flag when it is unable to free
308  *	sufficient buffer space.  Buffer space becomes recoverable when
309  *	bp's get placed back in the queues.
310  */
311 
312 static __inline void
313 bufspacewakeup(void)
314 {
315 
316 	/*
317 	 * If someone is waiting for BUF space, wake them up.  Even
318 	 * though we haven't freed the kva space yet, the waiting
319 	 * process will be able to now.
320 	 */
321 	mtx_lock(&nblock);
322 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
323 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
324 		wakeup(&needsbuffer);
325 	}
326 	mtx_unlock(&nblock);
327 }
328 
329 /*
330  * runningbufwakeup() - in-progress I/O accounting.
331  *
332  */
333 void
334 runningbufwakeup(struct buf *bp)
335 {
336 
337 	if (bp->b_runningbufspace) {
338 		atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
339 		bp->b_runningbufspace = 0;
340 		mtx_lock(&rbreqlock);
341 		if (runningbufreq && runningbufspace <= lorunningspace) {
342 			runningbufreq = 0;
343 			wakeup(&runningbufreq);
344 		}
345 		mtx_unlock(&rbreqlock);
346 	}
347 }
348 
349 /*
350  *	bufcountwakeup:
351  *
352  *	Called when a buffer has been added to one of the free queues to
353  *	account for the buffer and to wakeup anyone waiting for free buffers.
354  *	This typically occurs when large amounts of metadata are being handled
355  *	by the buffer cache ( else buffer space runs out first, usually ).
356  */
357 
358 static __inline void
359 bufcountwakeup(void)
360 {
361 
362 	atomic_add_int(&numfreebuffers, 1);
363 	mtx_lock(&nblock);
364 	if (needsbuffer) {
365 		needsbuffer &= ~VFS_BIO_NEED_ANY;
366 		if (numfreebuffers >= hifreebuffers)
367 			needsbuffer &= ~VFS_BIO_NEED_FREE;
368 		wakeup(&needsbuffer);
369 	}
370 	mtx_unlock(&nblock);
371 }
372 
373 /*
374  *	waitrunningbufspace()
375  *
376  *	runningbufspace is a measure of the amount of I/O currently
377  *	running.  This routine is used in async-write situations to
378  *	prevent creating huge backups of pending writes to a device.
379  *	Only asynchronous writes are governed by this function.
380  *
381  *	Reads will adjust runningbufspace, but will not block based on it.
382  *	The read load has a side effect of reducing the allowed write load.
383  *
384  *	This does NOT turn an async write into a sync write.  It waits
385  *	for earlier writes to complete and generally returns before the
386  *	caller's write has reached the device.
387  */
388 void
389 waitrunningbufspace(void)
390 {
391 
392 	mtx_lock(&rbreqlock);
393 	while (runningbufspace > hirunningspace) {
394 		++runningbufreq;
395 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
396 	}
397 	mtx_unlock(&rbreqlock);
398 }
399 
400 
401 /*
402  *	vfs_buf_test_cache:
403  *
404  *	Called when a buffer is extended.  This function clears the B_CACHE
405  *	bit if the newly extended portion of the buffer does not contain
406  *	valid data.
407  */
408 static __inline
409 void
410 vfs_buf_test_cache(struct buf *bp,
411 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
412 		  vm_page_t m)
413 {
414 
415 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
416 	if (bp->b_flags & B_CACHE) {
417 		int base = (foff + off) & PAGE_MASK;
418 		if (vm_page_is_valid(m, base, size) == 0)
419 			bp->b_flags &= ~B_CACHE;
420 	}
421 }
422 
423 /* Wake up the buffer daemon if necessary */
424 static __inline
425 void
426 bd_wakeup(int dirtybuflevel)
427 {
428 
429 	mtx_lock(&bdlock);
430 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
431 		bd_request = 1;
432 		wakeup(&bd_request);
433 	}
434 	mtx_unlock(&bdlock);
435 }
436 
437 /*
438  * bd_speedup - speedup the buffer cache flushing code
439  */
440 
441 static __inline
442 void
443 bd_speedup(void)
444 {
445 
446 	bd_wakeup(1);
447 }
448 
449 /*
450  * Calculating buffer cache scaling values and reserve space for buffer
451  * headers.  This is called during low level kernel initialization and
452  * may be called more then once.  We CANNOT write to the memory area
453  * being reserved at this time.
454  */
455 caddr_t
456 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
457 {
458 	int maxbuf;
459 
460 	/*
461 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
462 	 * PAGE_SIZE is >= 1K)
463 	 */
464 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
465 
466 	/*
467 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
468 	 * For the first 64MB of ram nominally allocate sufficient buffers to
469 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
470 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
471 	 * the buffer cache we limit the eventual kva reservation to
472 	 * maxbcache bytes.
473 	 *
474 	 * factor represents the 1/4 x ram conversion.
475 	 */
476 	if (nbuf == 0) {
477 		int factor = 4 * BKVASIZE / 1024;
478 
479 		nbuf = 50;
480 		if (physmem_est > 4096)
481 			nbuf += min((physmem_est - 4096) / factor,
482 			    65536 / factor);
483 		if (physmem_est > 65536)
484 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
485 
486 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
487 			nbuf = maxbcache / BKVASIZE;
488 
489 		/* XXX Avoid integer overflows later on with maxbufspace. */
490 		maxbuf = (INT_MAX / 3) / BKVASIZE;
491 		if (nbuf > maxbuf)
492 			nbuf = maxbuf;
493 	}
494 
495 #if 0
496 	/*
497 	 * Do not allow the buffer_map to be more then 1/2 the size of the
498 	 * kernel_map.
499 	 */
500 	if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
501 	    (BKVASIZE * 2)) {
502 		nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
503 		    (BKVASIZE * 2);
504 		printf("Warning: nbufs capped at %d\n", nbuf);
505 	}
506 #endif
507 
508 	/*
509 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
510 	 * We have no less then 16 and no more then 256.
511 	 */
512 	nswbuf = max(min(nbuf/4, 256), 16);
513 #ifdef NSWBUF_MIN
514 	if (nswbuf < NSWBUF_MIN)
515 		nswbuf = NSWBUF_MIN;
516 #endif
517 #ifdef DIRECTIO
518 	ffs_rawread_setup();
519 #endif
520 
521 	/*
522 	 * Reserve space for the buffer cache buffers
523 	 */
524 	swbuf = (void *)v;
525 	v = (caddr_t)(swbuf + nswbuf);
526 	buf = (void *)v;
527 	v = (caddr_t)(buf + nbuf);
528 
529 	return(v);
530 }
531 
532 /* Initialize the buffer subsystem.  Called before use of any buffers. */
533 void
534 bufinit(void)
535 {
536 	struct buf *bp;
537 	int i;
538 
539 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
540 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
541 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
542 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
543 	mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
544 	mtx_init(&bpinlock, "bpin lock", NULL, MTX_DEF);
545 
546 	/* next, make a null set of free lists */
547 	for (i = 0; i < BUFFER_QUEUES; i++)
548 		TAILQ_INIT(&bufqueues[i]);
549 
550 	/* finally, initialize each buffer header and stick on empty q */
551 	for (i = 0; i < nbuf; i++) {
552 		bp = &buf[i];
553 		bzero(bp, sizeof *bp);
554 		bp->b_flags = B_INVAL;	/* we're just an empty header */
555 		bp->b_rcred = NOCRED;
556 		bp->b_wcred = NOCRED;
557 		bp->b_qindex = QUEUE_EMPTY;
558 		bp->b_vflags = 0;
559 		bp->b_xflags = 0;
560 		LIST_INIT(&bp->b_dep);
561 		BUF_LOCKINIT(bp);
562 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
563 	}
564 
565 	/*
566 	 * maxbufspace is the absolute maximum amount of buffer space we are
567 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
568 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
569 	 * used by most other processes.  The differential is required to
570 	 * ensure that buf_daemon is able to run when other processes might
571 	 * be blocked waiting for buffer space.
572 	 *
573 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
574 	 * this may result in KVM fragmentation which is not handled optimally
575 	 * by the system.
576 	 */
577 	maxbufspace = nbuf * BKVASIZE;
578 	hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
579 	lobufspace = hibufspace - MAXBSIZE;
580 
581 	lorunningspace = 512 * 1024;
582 	hirunningspace = 1024 * 1024;
583 
584 /*
585  * Limit the amount of malloc memory since it is wired permanently into
586  * the kernel space.  Even though this is accounted for in the buffer
587  * allocation, we don't want the malloced region to grow uncontrolled.
588  * The malloc scheme improves memory utilization significantly on average
589  * (small) directories.
590  */
591 	maxbufmallocspace = hibufspace / 20;
592 
593 /*
594  * Reduce the chance of a deadlock occuring by limiting the number
595  * of delayed-write dirty buffers we allow to stack up.
596  */
597 	hidirtybuffers = nbuf / 4 + 20;
598 	dirtybufthresh = hidirtybuffers * 9 / 10;
599 	numdirtybuffers = 0;
600 /*
601  * To support extreme low-memory systems, make sure hidirtybuffers cannot
602  * eat up all available buffer space.  This occurs when our minimum cannot
603  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
604  * BKVASIZE'd (8K) buffers.
605  */
606 	while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
607 		hidirtybuffers >>= 1;
608 	}
609 	lodirtybuffers = hidirtybuffers / 2;
610 
611 /*
612  * Try to keep the number of free buffers in the specified range,
613  * and give special processes (e.g. like buf_daemon) access to an
614  * emergency reserve.
615  */
616 	lofreebuffers = nbuf / 18 + 5;
617 	hifreebuffers = 2 * lofreebuffers;
618 	numfreebuffers = nbuf;
619 
620 /*
621  * Maximum number of async ops initiated per buf_daemon loop.  This is
622  * somewhat of a hack at the moment, we really need to limit ourselves
623  * based on the number of bytes of I/O in-transit that were initiated
624  * from buf_daemon.
625  */
626 
627 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
628 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
629 }
630 
631 /*
632  * bfreekva() - free the kva allocation for a buffer.
633  *
634  *	Since this call frees up buffer space, we call bufspacewakeup().
635  */
636 static void
637 bfreekva(struct buf *bp)
638 {
639 
640 	if (bp->b_kvasize) {
641 		atomic_add_int(&buffreekvacnt, 1);
642 		atomic_subtract_int(&bufspace, bp->b_kvasize);
643 		vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
644 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
645 		bp->b_kvasize = 0;
646 		bufspacewakeup();
647 	}
648 }
649 
650 /*
651  *	bremfree:
652  *
653  *	Mark the buffer for removal from the appropriate free list in brelse.
654  *
655  */
656 void
657 bremfree(struct buf *bp)
658 {
659 
660 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
661 	KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked."));
662 	KASSERT((bp->b_flags & B_REMFREE) == 0,
663 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
664 	KASSERT(bp->b_qindex != QUEUE_NONE,
665 	    ("bremfree: buffer %p not on a queue.", bp));
666 
667 	bp->b_flags |= B_REMFREE;
668 	/* Fixup numfreebuffers count.  */
669 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
670 		atomic_subtract_int(&numfreebuffers, 1);
671 }
672 
673 /*
674  *	bremfreef:
675  *
676  *	Force an immediate removal from a free list.  Used only in nfs when
677  *	it abuses the b_freelist pointer.
678  */
679 void
680 bremfreef(struct buf *bp)
681 {
682 	mtx_lock(&bqlock);
683 	bremfreel(bp);
684 	mtx_unlock(&bqlock);
685 }
686 
687 /*
688  *	bremfreel:
689  *
690  *	Removes a buffer from the free list, must be called with the
691  *	bqlock held.
692  */
693 static void
694 bremfreel(struct buf *bp)
695 {
696 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
697 	    bp, bp->b_vp, bp->b_flags);
698 	KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp));
699 	KASSERT(bp->b_qindex != QUEUE_NONE,
700 	    ("bremfreel: buffer %p not on a queue.", bp));
701 	mtx_assert(&bqlock, MA_OWNED);
702 
703 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
704 	bp->b_qindex = QUEUE_NONE;
705 	/*
706 	 * If this was a delayed bremfree() we only need to remove the buffer
707 	 * from the queue and return the stats are already done.
708 	 */
709 	if (bp->b_flags & B_REMFREE) {
710 		bp->b_flags &= ~B_REMFREE;
711 		return;
712 	}
713 	/*
714 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
715 	 * delayed-write, the buffer was free and we must decrement
716 	 * numfreebuffers.
717 	 */
718 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
719 		atomic_subtract_int(&numfreebuffers, 1);
720 }
721 
722 
723 /*
724  * Get a buffer with the specified data.  Look in the cache first.  We
725  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
726  * is set, the buffer is valid and we do not have to do anything ( see
727  * getblk() ).  This is really just a special case of breadn().
728  */
729 int
730 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
731     struct buf **bpp)
732 {
733 
734 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
735 }
736 
737 /*
738  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
739  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
740  * the buffer is valid and we do not have to do anything.
741  */
742 void
743 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
744     int cnt, struct ucred * cred)
745 {
746 	struct buf *rabp;
747 	int i;
748 
749 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
750 		if (inmem(vp, *rablkno))
751 			continue;
752 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
753 
754 		if ((rabp->b_flags & B_CACHE) == 0) {
755 			if (!TD_IS_IDLETHREAD(curthread))
756 				curthread->td_ru.ru_inblock++;
757 			rabp->b_flags |= B_ASYNC;
758 			rabp->b_flags &= ~B_INVAL;
759 			rabp->b_ioflags &= ~BIO_ERROR;
760 			rabp->b_iocmd = BIO_READ;
761 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
762 				rabp->b_rcred = crhold(cred);
763 			vfs_busy_pages(rabp, 0);
764 			BUF_KERNPROC(rabp);
765 			rabp->b_iooffset = dbtob(rabp->b_blkno);
766 			bstrategy(rabp);
767 		} else {
768 			brelse(rabp);
769 		}
770 	}
771 }
772 
773 /*
774  * Operates like bread, but also starts asynchronous I/O on
775  * read-ahead blocks.
776  */
777 int
778 breadn(struct vnode * vp, daddr_t blkno, int size,
779     daddr_t * rablkno, int *rabsize,
780     int cnt, struct ucred * cred, struct buf **bpp)
781 {
782 	struct buf *bp;
783 	int rv = 0, readwait = 0;
784 
785 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
786 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
787 
788 	/* if not found in cache, do some I/O */
789 	if ((bp->b_flags & B_CACHE) == 0) {
790 		if (!TD_IS_IDLETHREAD(curthread))
791 			curthread->td_ru.ru_inblock++;
792 		bp->b_iocmd = BIO_READ;
793 		bp->b_flags &= ~B_INVAL;
794 		bp->b_ioflags &= ~BIO_ERROR;
795 		if (bp->b_rcred == NOCRED && cred != NOCRED)
796 			bp->b_rcred = crhold(cred);
797 		vfs_busy_pages(bp, 0);
798 		bp->b_iooffset = dbtob(bp->b_blkno);
799 		bstrategy(bp);
800 		++readwait;
801 	}
802 
803 	breada(vp, rablkno, rabsize, cnt, cred);
804 
805 	if (readwait) {
806 		rv = bufwait(bp);
807 	}
808 	return (rv);
809 }
810 
811 /*
812  * Write, release buffer on completion.  (Done by iodone
813  * if async).  Do not bother writing anything if the buffer
814  * is invalid.
815  *
816  * Note that we set B_CACHE here, indicating that buffer is
817  * fully valid and thus cacheable.  This is true even of NFS
818  * now so we set it generally.  This could be set either here
819  * or in biodone() since the I/O is synchronous.  We put it
820  * here.
821  */
822 int
823 bufwrite(struct buf *bp)
824 {
825 	int oldflags;
826 	struct vnode *vp;
827 	int vp_md;
828 
829 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
830 	if (bp->b_flags & B_INVAL) {
831 		brelse(bp);
832 		return (0);
833 	}
834 
835 	oldflags = bp->b_flags;
836 
837 	if (BUF_REFCNT(bp) == 0)
838 		panic("bufwrite: buffer is not busy???");
839 
840 	if (bp->b_pin_count > 0)
841 		bunpin_wait(bp);
842 
843 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
844 	    ("FFS background buffer should not get here %p", bp));
845 
846 	vp = bp->b_vp;
847 	if (vp)
848 		vp_md = vp->v_vflag & VV_MD;
849 	else
850 		vp_md = 0;
851 
852 	/* Mark the buffer clean */
853 	bundirty(bp);
854 
855 	bp->b_flags &= ~B_DONE;
856 	bp->b_ioflags &= ~BIO_ERROR;
857 	bp->b_flags |= B_CACHE;
858 	bp->b_iocmd = BIO_WRITE;
859 
860 	bufobj_wref(bp->b_bufobj);
861 	vfs_busy_pages(bp, 1);
862 
863 	/*
864 	 * Normal bwrites pipeline writes
865 	 */
866 	bp->b_runningbufspace = bp->b_bufsize;
867 	atomic_add_int(&runningbufspace, bp->b_runningbufspace);
868 
869 	if (!TD_IS_IDLETHREAD(curthread))
870 		curthread->td_ru.ru_oublock++;
871 	if (oldflags & B_ASYNC)
872 		BUF_KERNPROC(bp);
873 	bp->b_iooffset = dbtob(bp->b_blkno);
874 	bstrategy(bp);
875 
876 	if ((oldflags & B_ASYNC) == 0) {
877 		int rtval = bufwait(bp);
878 		brelse(bp);
879 		return (rtval);
880 	} else {
881 		/*
882 		 * don't allow the async write to saturate the I/O
883 		 * system.  We will not deadlock here because
884 		 * we are blocking waiting for I/O that is already in-progress
885 		 * to complete. We do not block here if it is the update
886 		 * or syncer daemon trying to clean up as that can lead
887 		 * to deadlock.
888 		 */
889 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
890 			waitrunningbufspace();
891 	}
892 
893 	return (0);
894 }
895 
896 void
897 bufbdflush(struct bufobj *bo, struct buf *bp)
898 {
899 	struct buf *nbp;
900 
901 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
902 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
903 		altbufferflushes++;
904 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
905 		BO_LOCK(bo);
906 		/*
907 		 * Try to find a buffer to flush.
908 		 */
909 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
910 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
911 			    BUF_LOCK(nbp,
912 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
913 				continue;
914 			if (bp == nbp)
915 				panic("bdwrite: found ourselves");
916 			BO_UNLOCK(bo);
917 			/* Don't countdeps with the bo lock held. */
918 			if (buf_countdeps(nbp, 0)) {
919 				BO_LOCK(bo);
920 				BUF_UNLOCK(nbp);
921 				continue;
922 			}
923 			if (nbp->b_flags & B_CLUSTEROK) {
924 				vfs_bio_awrite(nbp);
925 			} else {
926 				bremfree(nbp);
927 				bawrite(nbp);
928 			}
929 			dirtybufferflushes++;
930 			break;
931 		}
932 		if (nbp == NULL)
933 			BO_UNLOCK(bo);
934 	}
935 }
936 
937 /*
938  * Delayed write. (Buffer is marked dirty).  Do not bother writing
939  * anything if the buffer is marked invalid.
940  *
941  * Note that since the buffer must be completely valid, we can safely
942  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
943  * biodone() in order to prevent getblk from writing the buffer
944  * out synchronously.
945  */
946 void
947 bdwrite(struct buf *bp)
948 {
949 	struct thread *td = curthread;
950 	struct vnode *vp;
951 	struct bufobj *bo;
952 
953 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
954 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
955 	KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy"));
956 
957 	if (bp->b_flags & B_INVAL) {
958 		brelse(bp);
959 		return;
960 	}
961 
962 	/*
963 	 * If we have too many dirty buffers, don't create any more.
964 	 * If we are wildly over our limit, then force a complete
965 	 * cleanup. Otherwise, just keep the situation from getting
966 	 * out of control. Note that we have to avoid a recursive
967 	 * disaster and not try to clean up after our own cleanup!
968 	 */
969 	vp = bp->b_vp;
970 	bo = bp->b_bufobj;
971 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
972 		td->td_pflags |= TDP_INBDFLUSH;
973 		BO_BDFLUSH(bo, bp);
974 		td->td_pflags &= ~TDP_INBDFLUSH;
975 	} else
976 		recursiveflushes++;
977 
978 	bdirty(bp);
979 	/*
980 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
981 	 * true even of NFS now.
982 	 */
983 	bp->b_flags |= B_CACHE;
984 
985 	/*
986 	 * This bmap keeps the system from needing to do the bmap later,
987 	 * perhaps when the system is attempting to do a sync.  Since it
988 	 * is likely that the indirect block -- or whatever other datastructure
989 	 * that the filesystem needs is still in memory now, it is a good
990 	 * thing to do this.  Note also, that if the pageout daemon is
991 	 * requesting a sync -- there might not be enough memory to do
992 	 * the bmap then...  So, this is important to do.
993 	 */
994 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
995 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
996 	}
997 
998 	/*
999 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
1000 	 */
1001 	vfs_setdirty(bp);
1002 
1003 	/*
1004 	 * We need to do this here to satisfy the vnode_pager and the
1005 	 * pageout daemon, so that it thinks that the pages have been
1006 	 * "cleaned".  Note that since the pages are in a delayed write
1007 	 * buffer -- the VFS layer "will" see that the pages get written
1008 	 * out on the next sync, or perhaps the cluster will be completed.
1009 	 */
1010 	vfs_clean_pages(bp);
1011 	bqrelse(bp);
1012 
1013 	/*
1014 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
1015 	 * buffers (midpoint between our recovery point and our stall
1016 	 * point).
1017 	 */
1018 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1019 
1020 	/*
1021 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1022 	 * due to the softdep code.
1023 	 */
1024 }
1025 
1026 /*
1027  *	bdirty:
1028  *
1029  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1030  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1031  *	itself to properly update it in the dirty/clean lists.  We mark it
1032  *	B_DONE to ensure that any asynchronization of the buffer properly
1033  *	clears B_DONE ( else a panic will occur later ).
1034  *
1035  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1036  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1037  *	should only be called if the buffer is known-good.
1038  *
1039  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1040  *	count.
1041  *
1042  *	The buffer must be on QUEUE_NONE.
1043  */
1044 void
1045 bdirty(struct buf *bp)
1046 {
1047 
1048 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1049 	    bp, bp->b_vp, bp->b_flags);
1050 	KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp));
1051 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1052 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1053 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1054 	bp->b_flags &= ~(B_RELBUF);
1055 	bp->b_iocmd = BIO_WRITE;
1056 
1057 	if ((bp->b_flags & B_DELWRI) == 0) {
1058 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1059 		reassignbuf(bp);
1060 		atomic_add_int(&numdirtybuffers, 1);
1061 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1062 	}
1063 }
1064 
1065 /*
1066  *	bundirty:
1067  *
1068  *	Clear B_DELWRI for buffer.
1069  *
1070  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1071  *	count.
1072  *
1073  *	The buffer must be on QUEUE_NONE.
1074  */
1075 
1076 void
1077 bundirty(struct buf *bp)
1078 {
1079 
1080 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1081 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1082 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1083 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1084 	KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp));
1085 
1086 	if (bp->b_flags & B_DELWRI) {
1087 		bp->b_flags &= ~B_DELWRI;
1088 		reassignbuf(bp);
1089 		atomic_subtract_int(&numdirtybuffers, 1);
1090 		numdirtywakeup(lodirtybuffers);
1091 	}
1092 	/*
1093 	 * Since it is now being written, we can clear its deferred write flag.
1094 	 */
1095 	bp->b_flags &= ~B_DEFERRED;
1096 }
1097 
1098 /*
1099  *	bawrite:
1100  *
1101  *	Asynchronous write.  Start output on a buffer, but do not wait for
1102  *	it to complete.  The buffer is released when the output completes.
1103  *
1104  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1105  *	B_INVAL buffers.  Not us.
1106  */
1107 void
1108 bawrite(struct buf *bp)
1109 {
1110 
1111 	bp->b_flags |= B_ASYNC;
1112 	(void) bwrite(bp);
1113 }
1114 
1115 /*
1116  *	bwillwrite:
1117  *
1118  *	Called prior to the locking of any vnodes when we are expecting to
1119  *	write.  We do not want to starve the buffer cache with too many
1120  *	dirty buffers so we block here.  By blocking prior to the locking
1121  *	of any vnodes we attempt to avoid the situation where a locked vnode
1122  *	prevents the various system daemons from flushing related buffers.
1123  */
1124 
1125 void
1126 bwillwrite(void)
1127 {
1128 
1129 	if (numdirtybuffers >= hidirtybuffers) {
1130 		mtx_lock(&nblock);
1131 		while (numdirtybuffers >= hidirtybuffers) {
1132 			bd_wakeup(1);
1133 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1134 			msleep(&needsbuffer, &nblock,
1135 			    (PRIBIO + 4), "flswai", 0);
1136 		}
1137 		mtx_unlock(&nblock);
1138 	}
1139 }
1140 
1141 /*
1142  * Return true if we have too many dirty buffers.
1143  */
1144 int
1145 buf_dirty_count_severe(void)
1146 {
1147 
1148 	return(numdirtybuffers >= hidirtybuffers);
1149 }
1150 
1151 /*
1152  *	brelse:
1153  *
1154  *	Release a busy buffer and, if requested, free its resources.  The
1155  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1156  *	to be accessed later as a cache entity or reused for other purposes.
1157  */
1158 void
1159 brelse(struct buf *bp)
1160 {
1161 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1162 	    bp, bp->b_vp, bp->b_flags);
1163 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1164 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1165 
1166 	if (bp->b_flags & B_MANAGED) {
1167 		bqrelse(bp);
1168 		return;
1169 	}
1170 
1171 	if (bp->b_iocmd == BIO_WRITE &&
1172 	    (bp->b_ioflags & BIO_ERROR) &&
1173 	    bp->b_error != ENXIO &&
1174 	    !(bp->b_flags & B_INVAL)) {
1175 		/*
1176 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1177 		 * pages from being scrapped.  If B_INVAL is set then
1178 		 * this case is not run and the next case is run to
1179 		 * destroy the buffer.  B_INVAL can occur if the buffer
1180 		 * is outside the range supported by the underlying device.
1181 		 * If the error is that the device went away (ENXIO), we
1182 		 * shouldn't redirty the buffer either, but discard the
1183 		 * data too.
1184 		 */
1185 		bp->b_ioflags &= ~BIO_ERROR;
1186 		bdirty(bp);
1187 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1188 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1189 		/*
1190 		 * Either a failed I/O or we were asked to free or not
1191 		 * cache the buffer.
1192 		 */
1193 		bp->b_flags |= B_INVAL;
1194 		if (!LIST_EMPTY(&bp->b_dep))
1195 			buf_deallocate(bp);
1196 		if (bp->b_flags & B_DELWRI) {
1197 			atomic_subtract_int(&numdirtybuffers, 1);
1198 			numdirtywakeup(lodirtybuffers);
1199 		}
1200 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1201 		if ((bp->b_flags & B_VMIO) == 0) {
1202 			if (bp->b_bufsize)
1203 				allocbuf(bp, 0);
1204 			if (bp->b_vp)
1205 				brelvp(bp);
1206 		}
1207 	}
1208 
1209 	/*
1210 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1211 	 * is called with B_DELWRI set, the underlying pages may wind up
1212 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1213 	 * because pages associated with a B_DELWRI bp are marked clean.
1214 	 *
1215 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1216 	 * if B_DELWRI is set.
1217 	 *
1218 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1219 	 * on pages to return pages to the VM page queues.
1220 	 */
1221 	if (bp->b_flags & B_DELWRI)
1222 		bp->b_flags &= ~B_RELBUF;
1223 	else if (vm_page_count_severe()) {
1224 		/*
1225 		 * XXX This lock may not be necessary since BKGRDINPROG
1226 		 * cannot be set while we hold the buf lock, it can only be
1227 		 * cleared if it is already pending.
1228 		 */
1229 		if (bp->b_vp) {
1230 			BO_LOCK(bp->b_bufobj);
1231 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1232 				bp->b_flags |= B_RELBUF;
1233 			BO_UNLOCK(bp->b_bufobj);
1234 		} else
1235 			bp->b_flags |= B_RELBUF;
1236 	}
1237 
1238 	/*
1239 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1240 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1241 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1242 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1243 	 *
1244 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1245 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1246 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1247 	 *
1248 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1249 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1250 	 * the commit state and we cannot afford to lose the buffer. If the
1251 	 * buffer has a background write in progress, we need to keep it
1252 	 * around to prevent it from being reconstituted and starting a second
1253 	 * background write.
1254 	 */
1255 	if ((bp->b_flags & B_VMIO)
1256 	    && !(bp->b_vp->v_mount != NULL &&
1257 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1258 		 !vn_isdisk(bp->b_vp, NULL) &&
1259 		 (bp->b_flags & B_DELWRI))
1260 	    ) {
1261 
1262 		int i, j, resid;
1263 		vm_page_t m;
1264 		off_t foff;
1265 		vm_pindex_t poff;
1266 		vm_object_t obj;
1267 
1268 		obj = bp->b_bufobj->bo_object;
1269 
1270 		/*
1271 		 * Get the base offset and length of the buffer.  Note that
1272 		 * in the VMIO case if the buffer block size is not
1273 		 * page-aligned then b_data pointer may not be page-aligned.
1274 		 * But our b_pages[] array *IS* page aligned.
1275 		 *
1276 		 * block sizes less then DEV_BSIZE (usually 512) are not
1277 		 * supported due to the page granularity bits (m->valid,
1278 		 * m->dirty, etc...).
1279 		 *
1280 		 * See man buf(9) for more information
1281 		 */
1282 		resid = bp->b_bufsize;
1283 		foff = bp->b_offset;
1284 		VM_OBJECT_LOCK(obj);
1285 		for (i = 0; i < bp->b_npages; i++) {
1286 			int had_bogus = 0;
1287 
1288 			m = bp->b_pages[i];
1289 
1290 			/*
1291 			 * If we hit a bogus page, fixup *all* the bogus pages
1292 			 * now.
1293 			 */
1294 			if (m == bogus_page) {
1295 				poff = OFF_TO_IDX(bp->b_offset);
1296 				had_bogus = 1;
1297 
1298 				for (j = i; j < bp->b_npages; j++) {
1299 					vm_page_t mtmp;
1300 					mtmp = bp->b_pages[j];
1301 					if (mtmp == bogus_page) {
1302 						mtmp = vm_page_lookup(obj, poff + j);
1303 						if (!mtmp) {
1304 							panic("brelse: page missing\n");
1305 						}
1306 						bp->b_pages[j] = mtmp;
1307 					}
1308 				}
1309 
1310 				if ((bp->b_flags & B_INVAL) == 0) {
1311 					pmap_qenter(
1312 					    trunc_page((vm_offset_t)bp->b_data),
1313 					    bp->b_pages, bp->b_npages);
1314 				}
1315 				m = bp->b_pages[i];
1316 			}
1317 			if ((bp->b_flags & B_NOCACHE) ||
1318 			    (bp->b_ioflags & BIO_ERROR)) {
1319 				int poffset = foff & PAGE_MASK;
1320 				int presid = resid > (PAGE_SIZE - poffset) ?
1321 					(PAGE_SIZE - poffset) : resid;
1322 
1323 				KASSERT(presid >= 0, ("brelse: extra page"));
1324 				vm_page_lock_queues();
1325 				vm_page_set_invalid(m, poffset, presid);
1326 				vm_page_unlock_queues();
1327 				if (had_bogus)
1328 					printf("avoided corruption bug in bogus_page/brelse code\n");
1329 			}
1330 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1331 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1332 		}
1333 		VM_OBJECT_UNLOCK(obj);
1334 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1335 			vfs_vmio_release(bp);
1336 
1337 	} else if (bp->b_flags & B_VMIO) {
1338 
1339 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1340 			vfs_vmio_release(bp);
1341 		}
1342 
1343 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1344 		if (bp->b_bufsize != 0)
1345 			allocbuf(bp, 0);
1346 		if (bp->b_vp != NULL)
1347 			brelvp(bp);
1348 	}
1349 
1350 	if (BUF_REFCNT(bp) > 1) {
1351 		/* do not release to free list */
1352 		BUF_UNLOCK(bp);
1353 		return;
1354 	}
1355 
1356 	/* enqueue */
1357 	mtx_lock(&bqlock);
1358 	/* Handle delayed bremfree() processing. */
1359 	if (bp->b_flags & B_REMFREE)
1360 		bremfreel(bp);
1361 	if (bp->b_qindex != QUEUE_NONE)
1362 		panic("brelse: free buffer onto another queue???");
1363 
1364 	/* buffers with no memory */
1365 	if (bp->b_bufsize == 0) {
1366 		bp->b_flags |= B_INVAL;
1367 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1368 		if (bp->b_vflags & BV_BKGRDINPROG)
1369 			panic("losing buffer 1");
1370 		if (bp->b_kvasize) {
1371 			bp->b_qindex = QUEUE_EMPTYKVA;
1372 		} else {
1373 			bp->b_qindex = QUEUE_EMPTY;
1374 		}
1375 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1376 	/* buffers with junk contents */
1377 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1378 	    (bp->b_ioflags & BIO_ERROR)) {
1379 		bp->b_flags |= B_INVAL;
1380 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1381 		if (bp->b_vflags & BV_BKGRDINPROG)
1382 			panic("losing buffer 2");
1383 		bp->b_qindex = QUEUE_CLEAN;
1384 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1385 	/* remaining buffers */
1386 	} else {
1387 		if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1388 		    (B_DELWRI|B_NEEDSGIANT))
1389 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1390 		if (bp->b_flags & B_DELWRI)
1391 			bp->b_qindex = QUEUE_DIRTY;
1392 		else
1393 			bp->b_qindex = QUEUE_CLEAN;
1394 		if (bp->b_flags & B_AGE)
1395 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1396 		else
1397 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1398 	}
1399 	mtx_unlock(&bqlock);
1400 
1401 	/*
1402 	 * If B_INVAL and B_DELWRI is set, clear B_DELWRI.  We have already
1403 	 * placed the buffer on the correct queue.  We must also disassociate
1404 	 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1405 	 * find it.
1406 	 */
1407 	if (bp->b_flags & B_INVAL) {
1408 		if (bp->b_flags & B_DELWRI)
1409 			bundirty(bp);
1410 		if (bp->b_vp)
1411 			brelvp(bp);
1412 	}
1413 
1414 	/*
1415 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1416 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1417 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1418 	 * if B_INVAL is set ).
1419 	 */
1420 
1421 	if (!(bp->b_flags & B_DELWRI))
1422 		bufcountwakeup();
1423 
1424 	/*
1425 	 * Something we can maybe free or reuse
1426 	 */
1427 	if (bp->b_bufsize || bp->b_kvasize)
1428 		bufspacewakeup();
1429 
1430 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1431 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1432 		panic("brelse: not dirty");
1433 	/* unlock */
1434 	BUF_UNLOCK(bp);
1435 }
1436 
1437 /*
1438  * Release a buffer back to the appropriate queue but do not try to free
1439  * it.  The buffer is expected to be used again soon.
1440  *
1441  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1442  * biodone() to requeue an async I/O on completion.  It is also used when
1443  * known good buffers need to be requeued but we think we may need the data
1444  * again soon.
1445  *
1446  * XXX we should be able to leave the B_RELBUF hint set on completion.
1447  */
1448 void
1449 bqrelse(struct buf *bp)
1450 {
1451 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1452 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1453 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1454 
1455 	if (BUF_REFCNT(bp) > 1) {
1456 		/* do not release to free list */
1457 		BUF_UNLOCK(bp);
1458 		return;
1459 	}
1460 
1461 	if (bp->b_flags & B_MANAGED) {
1462 		if (bp->b_flags & B_REMFREE) {
1463 			mtx_lock(&bqlock);
1464 			bremfreel(bp);
1465 			mtx_unlock(&bqlock);
1466 		}
1467 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1468 		BUF_UNLOCK(bp);
1469 		return;
1470 	}
1471 
1472 	mtx_lock(&bqlock);
1473 	/* Handle delayed bremfree() processing. */
1474 	if (bp->b_flags & B_REMFREE)
1475 		bremfreel(bp);
1476 	if (bp->b_qindex != QUEUE_NONE)
1477 		panic("bqrelse: free buffer onto another queue???");
1478 	/* buffers with stale but valid contents */
1479 	if (bp->b_flags & B_DELWRI) {
1480 		if (bp->b_flags & B_NEEDSGIANT)
1481 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1482 		else
1483 			bp->b_qindex = QUEUE_DIRTY;
1484 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1485 	} else {
1486 		/*
1487 		 * XXX This lock may not be necessary since BKGRDINPROG
1488 		 * cannot be set while we hold the buf lock, it can only be
1489 		 * cleared if it is already pending.
1490 		 */
1491 		BO_LOCK(bp->b_bufobj);
1492 		if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) {
1493 			BO_UNLOCK(bp->b_bufobj);
1494 			bp->b_qindex = QUEUE_CLEAN;
1495 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1496 			    b_freelist);
1497 		} else {
1498 			/*
1499 			 * We are too low on memory, we have to try to free
1500 			 * the buffer (most importantly: the wired pages
1501 			 * making up its backing store) *now*.
1502 			 */
1503 			BO_UNLOCK(bp->b_bufobj);
1504 			mtx_unlock(&bqlock);
1505 			brelse(bp);
1506 			return;
1507 		}
1508 	}
1509 	mtx_unlock(&bqlock);
1510 
1511 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1512 		bufcountwakeup();
1513 
1514 	/*
1515 	 * Something we can maybe free or reuse.
1516 	 */
1517 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1518 		bufspacewakeup();
1519 
1520 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1521 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1522 		panic("bqrelse: not dirty");
1523 	/* unlock */
1524 	BUF_UNLOCK(bp);
1525 }
1526 
1527 /* Give pages used by the bp back to the VM system (where possible) */
1528 static void
1529 vfs_vmio_release(struct buf *bp)
1530 {
1531 	int i;
1532 	vm_page_t m;
1533 
1534 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1535 	vm_page_lock_queues();
1536 	for (i = 0; i < bp->b_npages; i++) {
1537 		m = bp->b_pages[i];
1538 		bp->b_pages[i] = NULL;
1539 		/*
1540 		 * In order to keep page LRU ordering consistent, put
1541 		 * everything on the inactive queue.
1542 		 */
1543 		vm_page_unwire(m, 0);
1544 		/*
1545 		 * We don't mess with busy pages, it is
1546 		 * the responsibility of the process that
1547 		 * busied the pages to deal with them.
1548 		 */
1549 		if ((m->oflags & VPO_BUSY) || (m->busy != 0))
1550 			continue;
1551 
1552 		if (m->wire_count == 0) {
1553 			/*
1554 			 * Might as well free the page if we can and it has
1555 			 * no valid data.  We also free the page if the
1556 			 * buffer was used for direct I/O
1557 			 */
1558 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1559 			    m->hold_count == 0) {
1560 				vm_page_free(m);
1561 			} else if (bp->b_flags & B_DIRECT) {
1562 				vm_page_try_to_free(m);
1563 			} else if (vm_page_count_severe()) {
1564 				vm_page_try_to_cache(m);
1565 			}
1566 		}
1567 	}
1568 	vm_page_unlock_queues();
1569 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1570 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1571 
1572 	if (bp->b_bufsize) {
1573 		bufspacewakeup();
1574 		bp->b_bufsize = 0;
1575 	}
1576 	bp->b_npages = 0;
1577 	bp->b_flags &= ~B_VMIO;
1578 	if (bp->b_vp)
1579 		brelvp(bp);
1580 }
1581 
1582 /*
1583  * Check to see if a block at a particular lbn is available for a clustered
1584  * write.
1585  */
1586 static int
1587 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1588 {
1589 	struct buf *bpa;
1590 	int match;
1591 
1592 	match = 0;
1593 
1594 	/* If the buf isn't in core skip it */
1595 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1596 		return (0);
1597 
1598 	/* If the buf is busy we don't want to wait for it */
1599 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1600 		return (0);
1601 
1602 	/* Only cluster with valid clusterable delayed write buffers */
1603 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1604 	    (B_DELWRI | B_CLUSTEROK))
1605 		goto done;
1606 
1607 	if (bpa->b_bufsize != size)
1608 		goto done;
1609 
1610 	/*
1611 	 * Check to see if it is in the expected place on disk and that the
1612 	 * block has been mapped.
1613 	 */
1614 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1615 		match = 1;
1616 done:
1617 	BUF_UNLOCK(bpa);
1618 	return (match);
1619 }
1620 
1621 /*
1622  *	vfs_bio_awrite:
1623  *
1624  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1625  *	This is much better then the old way of writing only one buffer at
1626  *	a time.  Note that we may not be presented with the buffers in the
1627  *	correct order, so we search for the cluster in both directions.
1628  */
1629 int
1630 vfs_bio_awrite(struct buf *bp)
1631 {
1632 	int i;
1633 	int j;
1634 	daddr_t lblkno = bp->b_lblkno;
1635 	struct vnode *vp = bp->b_vp;
1636 	int ncl;
1637 	int nwritten;
1638 	int size;
1639 	int maxcl;
1640 
1641 	/*
1642 	 * right now we support clustered writing only to regular files.  If
1643 	 * we find a clusterable block we could be in the middle of a cluster
1644 	 * rather then at the beginning.
1645 	 */
1646 	if ((vp->v_type == VREG) &&
1647 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1648 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1649 
1650 		size = vp->v_mount->mnt_stat.f_iosize;
1651 		maxcl = MAXPHYS / size;
1652 
1653 		VI_LOCK(vp);
1654 		for (i = 1; i < maxcl; i++)
1655 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1656 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1657 				break;
1658 
1659 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1660 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1661 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1662 				break;
1663 
1664 		VI_UNLOCK(vp);
1665 		--j;
1666 		ncl = i + j;
1667 		/*
1668 		 * this is a possible cluster write
1669 		 */
1670 		if (ncl != 1) {
1671 			BUF_UNLOCK(bp);
1672 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1673 			return nwritten;
1674 		}
1675 	}
1676 	bremfree(bp);
1677 	bp->b_flags |= B_ASYNC;
1678 	/*
1679 	 * default (old) behavior, writing out only one block
1680 	 *
1681 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1682 	 */
1683 	nwritten = bp->b_bufsize;
1684 	(void) bwrite(bp);
1685 
1686 	return nwritten;
1687 }
1688 
1689 /*
1690  *	getnewbuf:
1691  *
1692  *	Find and initialize a new buffer header, freeing up existing buffers
1693  *	in the bufqueues as necessary.  The new buffer is returned locked.
1694  *
1695  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1696  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1697  *
1698  *	We block if:
1699  *		We have insufficient buffer headers
1700  *		We have insufficient buffer space
1701  *		buffer_map is too fragmented ( space reservation fails )
1702  *		If we have to flush dirty buffers ( but we try to avoid this )
1703  *
1704  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1705  *	Instead we ask the buf daemon to do it for us.  We attempt to
1706  *	avoid piecemeal wakeups of the pageout daemon.
1707  */
1708 
1709 static struct buf *
1710 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1711 {
1712 	struct buf *bp;
1713 	struct buf *nbp;
1714 	int defrag = 0;
1715 	int nqindex;
1716 	static int flushingbufs;
1717 
1718 	/*
1719 	 * We can't afford to block since we might be holding a vnode lock,
1720 	 * which may prevent system daemons from running.  We deal with
1721 	 * low-memory situations by proactively returning memory and running
1722 	 * async I/O rather then sync I/O.
1723 	 */
1724 
1725 	atomic_add_int(&getnewbufcalls, 1);
1726 	atomic_subtract_int(&getnewbufrestarts, 1);
1727 restart:
1728 	atomic_add_int(&getnewbufrestarts, 1);
1729 
1730 	/*
1731 	 * Setup for scan.  If we do not have enough free buffers,
1732 	 * we setup a degenerate case that immediately fails.  Note
1733 	 * that if we are specially marked process, we are allowed to
1734 	 * dip into our reserves.
1735 	 *
1736 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1737 	 *
1738 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1739 	 * However, there are a number of cases (defragging, reusing, ...)
1740 	 * where we cannot backup.
1741 	 */
1742 	mtx_lock(&bqlock);
1743 	nqindex = QUEUE_EMPTYKVA;
1744 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1745 
1746 	if (nbp == NULL) {
1747 		/*
1748 		 * If no EMPTYKVA buffers and we are either
1749 		 * defragging or reusing, locate a CLEAN buffer
1750 		 * to free or reuse.  If bufspace useage is low
1751 		 * skip this step so we can allocate a new buffer.
1752 		 */
1753 		if (defrag || bufspace >= lobufspace) {
1754 			nqindex = QUEUE_CLEAN;
1755 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1756 		}
1757 
1758 		/*
1759 		 * If we could not find or were not allowed to reuse a
1760 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1761 		 * buffer.  We can only use an EMPTY buffer if allocating
1762 		 * its KVA would not otherwise run us out of buffer space.
1763 		 */
1764 		if (nbp == NULL && defrag == 0 &&
1765 		    bufspace + maxsize < hibufspace) {
1766 			nqindex = QUEUE_EMPTY;
1767 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1768 		}
1769 	}
1770 
1771 	/*
1772 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1773 	 * depending.
1774 	 */
1775 
1776 	while ((bp = nbp) != NULL) {
1777 		int qindex = nqindex;
1778 
1779 		/*
1780 		 * Calculate next bp ( we can only use it if we do not block
1781 		 * or do other fancy things ).
1782 		 */
1783 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1784 			switch(qindex) {
1785 			case QUEUE_EMPTY:
1786 				nqindex = QUEUE_EMPTYKVA;
1787 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1788 					break;
1789 				/* FALLTHROUGH */
1790 			case QUEUE_EMPTYKVA:
1791 				nqindex = QUEUE_CLEAN;
1792 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1793 					break;
1794 				/* FALLTHROUGH */
1795 			case QUEUE_CLEAN:
1796 				/*
1797 				 * nbp is NULL.
1798 				 */
1799 				break;
1800 			}
1801 		}
1802 		/*
1803 		 * If we are defragging then we need a buffer with
1804 		 * b_kvasize != 0.  XXX this situation should no longer
1805 		 * occur, if defrag is non-zero the buffer's b_kvasize
1806 		 * should also be non-zero at this point.  XXX
1807 		 */
1808 		if (defrag && bp->b_kvasize == 0) {
1809 			printf("Warning: defrag empty buffer %p\n", bp);
1810 			continue;
1811 		}
1812 
1813 		/*
1814 		 * Start freeing the bp.  This is somewhat involved.  nbp
1815 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1816 		 */
1817 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1818 			continue;
1819 		if (bp->b_vp) {
1820 			BO_LOCK(bp->b_bufobj);
1821 			if (bp->b_vflags & BV_BKGRDINPROG) {
1822 				BO_UNLOCK(bp->b_bufobj);
1823 				BUF_UNLOCK(bp);
1824 				continue;
1825 			}
1826 			BO_UNLOCK(bp->b_bufobj);
1827 		}
1828 		CTR6(KTR_BUF,
1829 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1830 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1831 		    bp->b_kvasize, bp->b_bufsize, qindex);
1832 
1833 		/*
1834 		 * Sanity Checks
1835 		 */
1836 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1837 
1838 		/*
1839 		 * Note: we no longer distinguish between VMIO and non-VMIO
1840 		 * buffers.
1841 		 */
1842 
1843 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1844 
1845 		bremfreel(bp);
1846 		mtx_unlock(&bqlock);
1847 
1848 		if (qindex == QUEUE_CLEAN) {
1849 			if (bp->b_flags & B_VMIO) {
1850 				bp->b_flags &= ~B_ASYNC;
1851 				vfs_vmio_release(bp);
1852 			}
1853 			if (bp->b_vp)
1854 				brelvp(bp);
1855 		}
1856 
1857 		/*
1858 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1859 		 * the scan from this point on.
1860 		 *
1861 		 * Get the rest of the buffer freed up.  b_kva* is still
1862 		 * valid after this operation.
1863 		 */
1864 
1865 		if (bp->b_rcred != NOCRED) {
1866 			crfree(bp->b_rcred);
1867 			bp->b_rcred = NOCRED;
1868 		}
1869 		if (bp->b_wcred != NOCRED) {
1870 			crfree(bp->b_wcred);
1871 			bp->b_wcred = NOCRED;
1872 		}
1873 		if (!LIST_EMPTY(&bp->b_dep))
1874 			buf_deallocate(bp);
1875 		if (bp->b_vflags & BV_BKGRDINPROG)
1876 			panic("losing buffer 3");
1877 		KASSERT(bp->b_vp == NULL,
1878 		    ("bp: %p still has vnode %p.  qindex: %d",
1879 		    bp, bp->b_vp, qindex));
1880 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1881 		   ("bp: %p still on a buffer list. xflags %X",
1882 		    bp, bp->b_xflags));
1883 
1884 		if (bp->b_bufsize)
1885 			allocbuf(bp, 0);
1886 
1887 		bp->b_flags = 0;
1888 		bp->b_ioflags = 0;
1889 		bp->b_xflags = 0;
1890 		bp->b_vflags = 0;
1891 		bp->b_vp = NULL;
1892 		bp->b_blkno = bp->b_lblkno = 0;
1893 		bp->b_offset = NOOFFSET;
1894 		bp->b_iodone = 0;
1895 		bp->b_error = 0;
1896 		bp->b_resid = 0;
1897 		bp->b_bcount = 0;
1898 		bp->b_npages = 0;
1899 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1900 		bp->b_bufobj = NULL;
1901 		bp->b_pin_count = 0;
1902 		bp->b_fsprivate1 = NULL;
1903 		bp->b_fsprivate2 = NULL;
1904 		bp->b_fsprivate3 = NULL;
1905 
1906 		LIST_INIT(&bp->b_dep);
1907 
1908 		/*
1909 		 * If we are defragging then free the buffer.
1910 		 */
1911 		if (defrag) {
1912 			bp->b_flags |= B_INVAL;
1913 			bfreekva(bp);
1914 			brelse(bp);
1915 			defrag = 0;
1916 			goto restart;
1917 		}
1918 
1919 		/*
1920 		 * Notify any waiters for the buffer lock about
1921 		 * identity change by freeing the buffer.
1922 		 */
1923 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp) > 0) {
1924 			bp->b_flags |= B_INVAL;
1925 			bfreekva(bp);
1926 			brelse(bp);
1927 			goto restart;
1928 		}
1929 
1930 		/*
1931 		 * If we are overcomitted then recover the buffer and its
1932 		 * KVM space.  This occurs in rare situations when multiple
1933 		 * processes are blocked in getnewbuf() or allocbuf().
1934 		 */
1935 		if (bufspace >= hibufspace)
1936 			flushingbufs = 1;
1937 		if (flushingbufs && bp->b_kvasize != 0) {
1938 			bp->b_flags |= B_INVAL;
1939 			bfreekva(bp);
1940 			brelse(bp);
1941 			goto restart;
1942 		}
1943 		if (bufspace < lobufspace)
1944 			flushingbufs = 0;
1945 		break;
1946 	}
1947 
1948 	/*
1949 	 * If we exhausted our list, sleep as appropriate.  We may have to
1950 	 * wakeup various daemons and write out some dirty buffers.
1951 	 *
1952 	 * Generally we are sleeping due to insufficient buffer space.
1953 	 */
1954 
1955 	if (bp == NULL) {
1956 		int flags;
1957 		char *waitmsg;
1958 
1959 		if (defrag) {
1960 			flags = VFS_BIO_NEED_BUFSPACE;
1961 			waitmsg = "nbufkv";
1962 		} else if (bufspace >= hibufspace) {
1963 			waitmsg = "nbufbs";
1964 			flags = VFS_BIO_NEED_BUFSPACE;
1965 		} else {
1966 			waitmsg = "newbuf";
1967 			flags = VFS_BIO_NEED_ANY;
1968 		}
1969 		mtx_lock(&nblock);
1970 		needsbuffer |= flags;
1971 		mtx_unlock(&nblock);
1972 		mtx_unlock(&bqlock);
1973 
1974 		bd_speedup();	/* heeeelp */
1975 
1976 		mtx_lock(&nblock);
1977 		while (needsbuffer & flags) {
1978 			if (msleep(&needsbuffer, &nblock,
1979 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1980 				mtx_unlock(&nblock);
1981 				return (NULL);
1982 			}
1983 		}
1984 		mtx_unlock(&nblock);
1985 	} else {
1986 		/*
1987 		 * We finally have a valid bp.  We aren't quite out of the
1988 		 * woods, we still have to reserve kva space.  In order
1989 		 * to keep fragmentation sane we only allocate kva in
1990 		 * BKVASIZE chunks.
1991 		 */
1992 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1993 
1994 		if (maxsize != bp->b_kvasize) {
1995 			vm_offset_t addr = 0;
1996 
1997 			bfreekva(bp);
1998 
1999 			vm_map_lock(buffer_map);
2000 			if (vm_map_findspace(buffer_map,
2001 				vm_map_min(buffer_map), maxsize, &addr)) {
2002 				/*
2003 				 * Uh oh.  Buffer map is to fragmented.  We
2004 				 * must defragment the map.
2005 				 */
2006 				atomic_add_int(&bufdefragcnt, 1);
2007 				vm_map_unlock(buffer_map);
2008 				defrag = 1;
2009 				bp->b_flags |= B_INVAL;
2010 				brelse(bp);
2011 				goto restart;
2012 			}
2013 			if (addr) {
2014 				vm_map_insert(buffer_map, NULL, 0,
2015 					addr, addr + maxsize,
2016 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2017 
2018 				bp->b_kvabase = (caddr_t) addr;
2019 				bp->b_kvasize = maxsize;
2020 				atomic_add_int(&bufspace, bp->b_kvasize);
2021 				atomic_add_int(&bufreusecnt, 1);
2022 			}
2023 			vm_map_unlock(buffer_map);
2024 		}
2025 		bp->b_saveaddr = bp->b_kvabase;
2026 		bp->b_data = bp->b_saveaddr;
2027 	}
2028 	return(bp);
2029 }
2030 
2031 /*
2032  *	buf_daemon:
2033  *
2034  *	buffer flushing daemon.  Buffers are normally flushed by the
2035  *	update daemon but if it cannot keep up this process starts to
2036  *	take the load in an attempt to prevent getnewbuf() from blocking.
2037  */
2038 
2039 static struct kproc_desc buf_kp = {
2040 	"bufdaemon",
2041 	buf_daemon,
2042 	&bufdaemonproc
2043 };
2044 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
2045 
2046 static void
2047 buf_daemon()
2048 {
2049 
2050 	/*
2051 	 * This process needs to be suspended prior to shutdown sync.
2052 	 */
2053 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2054 	    SHUTDOWN_PRI_LAST);
2055 
2056 	/*
2057 	 * This process is allowed to take the buffer cache to the limit
2058 	 */
2059 	curthread->td_pflags |= TDP_NORUNNINGBUF;
2060 	mtx_lock(&bdlock);
2061 	for (;;) {
2062 		bd_request = 0;
2063 		mtx_unlock(&bdlock);
2064 
2065 		kproc_suspend_check(bufdaemonproc);
2066 
2067 		/*
2068 		 * Do the flush.  Limit the amount of in-transit I/O we
2069 		 * allow to build up, otherwise we would completely saturate
2070 		 * the I/O system.  Wakeup any waiting processes before we
2071 		 * normally would so they can run in parallel with our drain.
2072 		 */
2073 		while (numdirtybuffers > lodirtybuffers) {
2074 			int flushed;
2075 
2076 			flushed = flushbufqueues(QUEUE_DIRTY, 0);
2077 			/* The list empty check here is slightly racy */
2078 			if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2079 				mtx_lock(&Giant);
2080 				flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0);
2081 				mtx_unlock(&Giant);
2082 			}
2083 			if (flushed == 0) {
2084 				/*
2085 				 * Could not find any buffers without rollback
2086 				 * dependencies, so just write the first one
2087 				 * in the hopes of eventually making progress.
2088 				 */
2089 				flushbufqueues(QUEUE_DIRTY, 1);
2090 				if (!TAILQ_EMPTY(
2091 				    &bufqueues[QUEUE_DIRTY_GIANT])) {
2092 					mtx_lock(&Giant);
2093 					flushbufqueues(QUEUE_DIRTY_GIANT, 1);
2094 					mtx_unlock(&Giant);
2095 				}
2096 				break;
2097 			}
2098 			uio_yield();
2099 		}
2100 
2101 		/*
2102 		 * Only clear bd_request if we have reached our low water
2103 		 * mark.  The buf_daemon normally waits 1 second and
2104 		 * then incrementally flushes any dirty buffers that have
2105 		 * built up, within reason.
2106 		 *
2107 		 * If we were unable to hit our low water mark and couldn't
2108 		 * find any flushable buffers, we sleep half a second.
2109 		 * Otherwise we loop immediately.
2110 		 */
2111 		mtx_lock(&bdlock);
2112 		if (numdirtybuffers <= lodirtybuffers) {
2113 			/*
2114 			 * We reached our low water mark, reset the
2115 			 * request and sleep until we are needed again.
2116 			 * The sleep is just so the suspend code works.
2117 			 */
2118 			bd_request = 0;
2119 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2120 		} else {
2121 			/*
2122 			 * We couldn't find any flushable dirty buffers but
2123 			 * still have too many dirty buffers, we
2124 			 * have to sleep and try again.  (rare)
2125 			 */
2126 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2127 		}
2128 	}
2129 }
2130 
2131 /*
2132  *	flushbufqueues:
2133  *
2134  *	Try to flush a buffer in the dirty queue.  We must be careful to
2135  *	free up B_INVAL buffers instead of write them, which NFS is
2136  *	particularly sensitive to.
2137  */
2138 static int flushwithdeps = 0;
2139 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2140     0, "Number of buffers flushed with dependecies that require rollbacks");
2141 
2142 static int
2143 flushbufqueues(int queue, int flushdeps)
2144 {
2145 	struct thread *td = curthread;
2146 	struct buf sentinel;
2147 	struct vnode *vp;
2148 	struct mount *mp;
2149 	struct buf *bp;
2150 	int hasdeps;
2151 	int flushed;
2152 	int target;
2153 
2154 	target = numdirtybuffers - lodirtybuffers;
2155 	if (flushdeps && target > 2)
2156 		target /= 2;
2157 	flushed = 0;
2158 	bp = NULL;
2159 	mtx_lock(&bqlock);
2160 	TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist);
2161 	while (flushed != target) {
2162 		bp = TAILQ_FIRST(&bufqueues[queue]);
2163 		if (bp == &sentinel)
2164 			break;
2165 		TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist);
2166 		TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist);
2167 
2168 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2169 			continue;
2170 		if (bp->b_pin_count > 0) {
2171 			BUF_UNLOCK(bp);
2172 			continue;
2173 		}
2174 		BO_LOCK(bp->b_bufobj);
2175 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2176 		    (bp->b_flags & B_DELWRI) == 0) {
2177 			BO_UNLOCK(bp->b_bufobj);
2178 			BUF_UNLOCK(bp);
2179 			continue;
2180 		}
2181 		BO_UNLOCK(bp->b_bufobj);
2182 		if (bp->b_flags & B_INVAL) {
2183 			bremfreel(bp);
2184 			mtx_unlock(&bqlock);
2185 			brelse(bp);
2186 			flushed++;
2187 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2188 			mtx_lock(&bqlock);
2189 			continue;
2190 		}
2191 
2192 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2193 			if (flushdeps == 0) {
2194 				BUF_UNLOCK(bp);
2195 				continue;
2196 			}
2197 			hasdeps = 1;
2198 		} else
2199 			hasdeps = 0;
2200 		/*
2201 		 * We must hold the lock on a vnode before writing
2202 		 * one of its buffers. Otherwise we may confuse, or
2203 		 * in the case of a snapshot vnode, deadlock the
2204 		 * system.
2205 		 *
2206 		 * The lock order here is the reverse of the normal
2207 		 * of vnode followed by buf lock.  This is ok because
2208 		 * the NOWAIT will prevent deadlock.
2209 		 */
2210 		vp = bp->b_vp;
2211 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2212 			BUF_UNLOCK(bp);
2213 			continue;
2214 		}
2215 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2216 			mtx_unlock(&bqlock);
2217 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2218 			    bp, bp->b_vp, bp->b_flags);
2219 			vfs_bio_awrite(bp);
2220 			vn_finished_write(mp);
2221 			VOP_UNLOCK(vp, 0, td);
2222 			flushwithdeps += hasdeps;
2223 			flushed++;
2224 			waitrunningbufspace();
2225 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2226 			mtx_lock(&bqlock);
2227 			continue;
2228 		}
2229 		vn_finished_write(mp);
2230 		BUF_UNLOCK(bp);
2231 	}
2232 	TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist);
2233 	mtx_unlock(&bqlock);
2234 	return (flushed);
2235 }
2236 
2237 /*
2238  * Check to see if a block is currently memory resident.
2239  */
2240 struct buf *
2241 incore(struct bufobj *bo, daddr_t blkno)
2242 {
2243 	struct buf *bp;
2244 
2245 	BO_LOCK(bo);
2246 	bp = gbincore(bo, blkno);
2247 	BO_UNLOCK(bo);
2248 	return (bp);
2249 }
2250 
2251 /*
2252  * Returns true if no I/O is needed to access the
2253  * associated VM object.  This is like incore except
2254  * it also hunts around in the VM system for the data.
2255  */
2256 
2257 static int
2258 inmem(struct vnode * vp, daddr_t blkno)
2259 {
2260 	vm_object_t obj;
2261 	vm_offset_t toff, tinc, size;
2262 	vm_page_t m;
2263 	vm_ooffset_t off;
2264 
2265 	ASSERT_VOP_LOCKED(vp, "inmem");
2266 
2267 	if (incore(&vp->v_bufobj, blkno))
2268 		return 1;
2269 	if (vp->v_mount == NULL)
2270 		return 0;
2271 	obj = vp->v_object;
2272 	if (obj == NULL)
2273 		return (0);
2274 
2275 	size = PAGE_SIZE;
2276 	if (size > vp->v_mount->mnt_stat.f_iosize)
2277 		size = vp->v_mount->mnt_stat.f_iosize;
2278 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2279 
2280 	VM_OBJECT_LOCK(obj);
2281 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2282 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2283 		if (!m)
2284 			goto notinmem;
2285 		tinc = size;
2286 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2287 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2288 		if (vm_page_is_valid(m,
2289 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2290 			goto notinmem;
2291 	}
2292 	VM_OBJECT_UNLOCK(obj);
2293 	return 1;
2294 
2295 notinmem:
2296 	VM_OBJECT_UNLOCK(obj);
2297 	return (0);
2298 }
2299 
2300 /*
2301  *	vfs_setdirty:
2302  *
2303  *	Sets the dirty range for a buffer based on the status of the dirty
2304  *	bits in the pages comprising the buffer.
2305  *
2306  *	The range is limited to the size of the buffer.
2307  *
2308  *	This routine is primarily used by NFS, but is generalized for the
2309  *	B_VMIO case.
2310  */
2311 static void
2312 vfs_setdirty(struct buf *bp)
2313 {
2314 
2315 	/*
2316 	 * Degenerate case - empty buffer
2317 	 */
2318 
2319 	if (bp->b_bufsize == 0)
2320 		return;
2321 
2322 	/*
2323 	 * We qualify the scan for modified pages on whether the
2324 	 * object has been flushed yet.
2325 	 */
2326 
2327 	if ((bp->b_flags & B_VMIO) == 0)
2328 		return;
2329 
2330 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2331 	vfs_setdirty_locked_object(bp);
2332 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2333 }
2334 
2335 static void
2336 vfs_setdirty_locked_object(struct buf *bp)
2337 {
2338 	vm_object_t object;
2339 	int i;
2340 
2341 	object = bp->b_bufobj->bo_object;
2342 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2343 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2344 		vm_offset_t boffset;
2345 		vm_offset_t eoffset;
2346 
2347 		vm_page_lock_queues();
2348 		/*
2349 		 * test the pages to see if they have been modified directly
2350 		 * by users through the VM system.
2351 		 */
2352 		for (i = 0; i < bp->b_npages; i++)
2353 			vm_page_test_dirty(bp->b_pages[i]);
2354 
2355 		/*
2356 		 * Calculate the encompassing dirty range, boffset and eoffset,
2357 		 * (eoffset - boffset) bytes.
2358 		 */
2359 
2360 		for (i = 0; i < bp->b_npages; i++) {
2361 			if (bp->b_pages[i]->dirty)
2362 				break;
2363 		}
2364 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2365 
2366 		for (i = bp->b_npages - 1; i >= 0; --i) {
2367 			if (bp->b_pages[i]->dirty) {
2368 				break;
2369 			}
2370 		}
2371 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2372 
2373 		vm_page_unlock_queues();
2374 		/*
2375 		 * Fit it to the buffer.
2376 		 */
2377 
2378 		if (eoffset > bp->b_bcount)
2379 			eoffset = bp->b_bcount;
2380 
2381 		/*
2382 		 * If we have a good dirty range, merge with the existing
2383 		 * dirty range.
2384 		 */
2385 
2386 		if (boffset < eoffset) {
2387 			if (bp->b_dirtyoff > boffset)
2388 				bp->b_dirtyoff = boffset;
2389 			if (bp->b_dirtyend < eoffset)
2390 				bp->b_dirtyend = eoffset;
2391 		}
2392 	}
2393 }
2394 
2395 /*
2396  *	getblk:
2397  *
2398  *	Get a block given a specified block and offset into a file/device.
2399  *	The buffers B_DONE bit will be cleared on return, making it almost
2400  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2401  *	return.  The caller should clear B_INVAL prior to initiating a
2402  *	READ.
2403  *
2404  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2405  *	an existing buffer.
2406  *
2407  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2408  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2409  *	and then cleared based on the backing VM.  If the previous buffer is
2410  *	non-0-sized but invalid, B_CACHE will be cleared.
2411  *
2412  *	If getblk() must create a new buffer, the new buffer is returned with
2413  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2414  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2415  *	backing VM.
2416  *
2417  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2418  *	B_CACHE bit is clear.
2419  *
2420  *	What this means, basically, is that the caller should use B_CACHE to
2421  *	determine whether the buffer is fully valid or not and should clear
2422  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2423  *	the buffer by loading its data area with something, the caller needs
2424  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2425  *	the caller should set B_CACHE ( as an optimization ), else the caller
2426  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2427  *	a write attempt or if it was a successfull read.  If the caller
2428  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2429  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2430  */
2431 struct buf *
2432 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2433     int flags)
2434 {
2435 	struct buf *bp;
2436 	struct bufobj *bo;
2437 	int error;
2438 
2439 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2440 	ASSERT_VOP_LOCKED(vp, "getblk");
2441 	if (size > MAXBSIZE)
2442 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2443 
2444 	bo = &vp->v_bufobj;
2445 loop:
2446 	/*
2447 	 * Block if we are low on buffers.   Certain processes are allowed
2448 	 * to completely exhaust the buffer cache.
2449          *
2450          * If this check ever becomes a bottleneck it may be better to
2451          * move it into the else, when gbincore() fails.  At the moment
2452          * it isn't a problem.
2453 	 *
2454 	 * XXX remove if 0 sections (clean this up after its proven)
2455          */
2456 	if (numfreebuffers == 0) {
2457 		if (TD_IS_IDLETHREAD(curthread))
2458 			return NULL;
2459 		mtx_lock(&nblock);
2460 		needsbuffer |= VFS_BIO_NEED_ANY;
2461 		mtx_unlock(&nblock);
2462 	}
2463 
2464 	BO_LOCK(bo);
2465 	bp = gbincore(bo, blkno);
2466 	if (bp != NULL) {
2467 		int lockflags;
2468 		/*
2469 		 * Buffer is in-core.  If the buffer is not busy, it must
2470 		 * be on a queue.
2471 		 */
2472 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2473 
2474 		if (flags & GB_LOCK_NOWAIT)
2475 			lockflags |= LK_NOWAIT;
2476 
2477 		error = BUF_TIMELOCK(bp, lockflags,
2478 		    VI_MTX(vp), "getblk", slpflag, slptimeo);
2479 
2480 		/*
2481 		 * If we slept and got the lock we have to restart in case
2482 		 * the buffer changed identities.
2483 		 */
2484 		if (error == ENOLCK)
2485 			goto loop;
2486 		/* We timed out or were interrupted. */
2487 		else if (error)
2488 			return (NULL);
2489 
2490 		/*
2491 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2492 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2493 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2494 		 * backing VM cache.
2495 		 */
2496 		if (bp->b_flags & B_INVAL)
2497 			bp->b_flags &= ~B_CACHE;
2498 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2499 			bp->b_flags |= B_CACHE;
2500 		bremfree(bp);
2501 
2502 		/*
2503 		 * check for size inconsistancies for non-VMIO case.
2504 		 */
2505 
2506 		if (bp->b_bcount != size) {
2507 			if ((bp->b_flags & B_VMIO) == 0 ||
2508 			    (size > bp->b_kvasize)) {
2509 				if (bp->b_flags & B_DELWRI) {
2510 					/*
2511 					 * If buffer is pinned and caller does
2512 					 * not want sleep  waiting for it to be
2513 					 * unpinned, bail out
2514 					 * */
2515 					if (bp->b_pin_count > 0) {
2516 						if (flags & GB_LOCK_NOWAIT) {
2517 							bqrelse(bp);
2518 							return (NULL);
2519 						} else {
2520 							bunpin_wait(bp);
2521 						}
2522 					}
2523 					bp->b_flags |= B_NOCACHE;
2524 					bwrite(bp);
2525 				} else {
2526 					if (LIST_EMPTY(&bp->b_dep)) {
2527 						bp->b_flags |= B_RELBUF;
2528 						brelse(bp);
2529 					} else {
2530 						bp->b_flags |= B_NOCACHE;
2531 						bwrite(bp);
2532 					}
2533 				}
2534 				goto loop;
2535 			}
2536 		}
2537 
2538 		/*
2539 		 * If the size is inconsistant in the VMIO case, we can resize
2540 		 * the buffer.  This might lead to B_CACHE getting set or
2541 		 * cleared.  If the size has not changed, B_CACHE remains
2542 		 * unchanged from its previous state.
2543 		 */
2544 
2545 		if (bp->b_bcount != size)
2546 			allocbuf(bp, size);
2547 
2548 		KASSERT(bp->b_offset != NOOFFSET,
2549 		    ("getblk: no buffer offset"));
2550 
2551 		/*
2552 		 * A buffer with B_DELWRI set and B_CACHE clear must
2553 		 * be committed before we can return the buffer in
2554 		 * order to prevent the caller from issuing a read
2555 		 * ( due to B_CACHE not being set ) and overwriting
2556 		 * it.
2557 		 *
2558 		 * Most callers, including NFS and FFS, need this to
2559 		 * operate properly either because they assume they
2560 		 * can issue a read if B_CACHE is not set, or because
2561 		 * ( for example ) an uncached B_DELWRI might loop due
2562 		 * to softupdates re-dirtying the buffer.  In the latter
2563 		 * case, B_CACHE is set after the first write completes,
2564 		 * preventing further loops.
2565 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2566 		 * above while extending the buffer, we cannot allow the
2567 		 * buffer to remain with B_CACHE set after the write
2568 		 * completes or it will represent a corrupt state.  To
2569 		 * deal with this we set B_NOCACHE to scrap the buffer
2570 		 * after the write.
2571 		 *
2572 		 * We might be able to do something fancy, like setting
2573 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2574 		 * so the below call doesn't set B_CACHE, but that gets real
2575 		 * confusing.  This is much easier.
2576 		 */
2577 
2578 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2579 			bp->b_flags |= B_NOCACHE;
2580 			bwrite(bp);
2581 			goto loop;
2582 		}
2583 		bp->b_flags &= ~B_DONE;
2584 	} else {
2585 		int bsize, maxsize, vmio;
2586 		off_t offset;
2587 
2588 		/*
2589 		 * Buffer is not in-core, create new buffer.  The buffer
2590 		 * returned by getnewbuf() is locked.  Note that the returned
2591 		 * buffer is also considered valid (not marked B_INVAL).
2592 		 */
2593 		BO_UNLOCK(bo);
2594 		/*
2595 		 * If the user does not want us to create the buffer, bail out
2596 		 * here.
2597 		 */
2598 		if (flags & GB_NOCREAT)
2599 			return NULL;
2600 		bsize = bo->bo_bsize;
2601 		offset = blkno * bsize;
2602 		vmio = vp->v_object != NULL;
2603 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2604 		maxsize = imax(maxsize, bsize);
2605 
2606 		bp = getnewbuf(slpflag, slptimeo, size, maxsize);
2607 		if (bp == NULL) {
2608 			if (slpflag || slptimeo)
2609 				return NULL;
2610 			goto loop;
2611 		}
2612 
2613 		/*
2614 		 * This code is used to make sure that a buffer is not
2615 		 * created while the getnewbuf routine is blocked.
2616 		 * This can be a problem whether the vnode is locked or not.
2617 		 * If the buffer is created out from under us, we have to
2618 		 * throw away the one we just created.
2619 		 *
2620 		 * Note: this must occur before we associate the buffer
2621 		 * with the vp especially considering limitations in
2622 		 * the splay tree implementation when dealing with duplicate
2623 		 * lblkno's.
2624 		 */
2625 		BO_LOCK(bo);
2626 		if (gbincore(bo, blkno)) {
2627 			BO_UNLOCK(bo);
2628 			bp->b_flags |= B_INVAL;
2629 			brelse(bp);
2630 			goto loop;
2631 		}
2632 
2633 		/*
2634 		 * Insert the buffer into the hash, so that it can
2635 		 * be found by incore.
2636 		 */
2637 		bp->b_blkno = bp->b_lblkno = blkno;
2638 		bp->b_offset = offset;
2639 		bgetvp(vp, bp);
2640 		BO_UNLOCK(bo);
2641 
2642 		/*
2643 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2644 		 * buffer size starts out as 0, B_CACHE will be set by
2645 		 * allocbuf() for the VMIO case prior to it testing the
2646 		 * backing store for validity.
2647 		 */
2648 
2649 		if (vmio) {
2650 			bp->b_flags |= B_VMIO;
2651 #if defined(VFS_BIO_DEBUG)
2652 			if (vn_canvmio(vp) != TRUE)
2653 				printf("getblk: VMIO on vnode type %d\n",
2654 					vp->v_type);
2655 #endif
2656 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2657 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2658 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2659 		} else {
2660 			bp->b_flags &= ~B_VMIO;
2661 			KASSERT(bp->b_bufobj->bo_object == NULL,
2662 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2663 			    bp, bp->b_bufobj->bo_object));
2664 		}
2665 
2666 		allocbuf(bp, size);
2667 		bp->b_flags &= ~B_DONE;
2668 	}
2669 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2670 	KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2671 	KASSERT(bp->b_bufobj == bo,
2672 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2673 	return (bp);
2674 }
2675 
2676 /*
2677  * Get an empty, disassociated buffer of given size.  The buffer is initially
2678  * set to B_INVAL.
2679  */
2680 struct buf *
2681 geteblk(int size)
2682 {
2683 	struct buf *bp;
2684 	int maxsize;
2685 
2686 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2687 	while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2688 		continue;
2689 	allocbuf(bp, size);
2690 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2691 	KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2692 	return (bp);
2693 }
2694 
2695 
2696 /*
2697  * This code constitutes the buffer memory from either anonymous system
2698  * memory (in the case of non-VMIO operations) or from an associated
2699  * VM object (in the case of VMIO operations).  This code is able to
2700  * resize a buffer up or down.
2701  *
2702  * Note that this code is tricky, and has many complications to resolve
2703  * deadlock or inconsistant data situations.  Tread lightly!!!
2704  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2705  * the caller.  Calling this code willy nilly can result in the loss of data.
2706  *
2707  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2708  * B_CACHE for the non-VMIO case.
2709  */
2710 
2711 int
2712 allocbuf(struct buf *bp, int size)
2713 {
2714 	int newbsize, mbsize;
2715 	int i;
2716 
2717 	if (BUF_REFCNT(bp) == 0)
2718 		panic("allocbuf: buffer not busy");
2719 
2720 	if (bp->b_kvasize < size)
2721 		panic("allocbuf: buffer too small");
2722 
2723 	if ((bp->b_flags & B_VMIO) == 0) {
2724 		caddr_t origbuf;
2725 		int origbufsize;
2726 		/*
2727 		 * Just get anonymous memory from the kernel.  Don't
2728 		 * mess with B_CACHE.
2729 		 */
2730 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2731 		if (bp->b_flags & B_MALLOC)
2732 			newbsize = mbsize;
2733 		else
2734 			newbsize = round_page(size);
2735 
2736 		if (newbsize < bp->b_bufsize) {
2737 			/*
2738 			 * malloced buffers are not shrunk
2739 			 */
2740 			if (bp->b_flags & B_MALLOC) {
2741 				if (newbsize) {
2742 					bp->b_bcount = size;
2743 				} else {
2744 					free(bp->b_data, M_BIOBUF);
2745 					if (bp->b_bufsize) {
2746 						atomic_subtract_int(
2747 						    &bufmallocspace,
2748 						    bp->b_bufsize);
2749 						bufspacewakeup();
2750 						bp->b_bufsize = 0;
2751 					}
2752 					bp->b_saveaddr = bp->b_kvabase;
2753 					bp->b_data = bp->b_saveaddr;
2754 					bp->b_bcount = 0;
2755 					bp->b_flags &= ~B_MALLOC;
2756 				}
2757 				return 1;
2758 			}
2759 			vm_hold_free_pages(
2760 			    bp,
2761 			    (vm_offset_t) bp->b_data + newbsize,
2762 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2763 		} else if (newbsize > bp->b_bufsize) {
2764 			/*
2765 			 * We only use malloced memory on the first allocation.
2766 			 * and revert to page-allocated memory when the buffer
2767 			 * grows.
2768 			 */
2769 			/*
2770 			 * There is a potential smp race here that could lead
2771 			 * to bufmallocspace slightly passing the max.  It
2772 			 * is probably extremely rare and not worth worrying
2773 			 * over.
2774 			 */
2775 			if ( (bufmallocspace < maxbufmallocspace) &&
2776 				(bp->b_bufsize == 0) &&
2777 				(mbsize <= PAGE_SIZE/2)) {
2778 
2779 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2780 				bp->b_bufsize = mbsize;
2781 				bp->b_bcount = size;
2782 				bp->b_flags |= B_MALLOC;
2783 				atomic_add_int(&bufmallocspace, mbsize);
2784 				return 1;
2785 			}
2786 			origbuf = NULL;
2787 			origbufsize = 0;
2788 			/*
2789 			 * If the buffer is growing on its other-than-first allocation,
2790 			 * then we revert to the page-allocation scheme.
2791 			 */
2792 			if (bp->b_flags & B_MALLOC) {
2793 				origbuf = bp->b_data;
2794 				origbufsize = bp->b_bufsize;
2795 				bp->b_data = bp->b_kvabase;
2796 				if (bp->b_bufsize) {
2797 					atomic_subtract_int(&bufmallocspace,
2798 					    bp->b_bufsize);
2799 					bufspacewakeup();
2800 					bp->b_bufsize = 0;
2801 				}
2802 				bp->b_flags &= ~B_MALLOC;
2803 				newbsize = round_page(newbsize);
2804 			}
2805 			vm_hold_load_pages(
2806 			    bp,
2807 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2808 			    (vm_offset_t) bp->b_data + newbsize);
2809 			if (origbuf) {
2810 				bcopy(origbuf, bp->b_data, origbufsize);
2811 				free(origbuf, M_BIOBUF);
2812 			}
2813 		}
2814 	} else {
2815 		int desiredpages;
2816 
2817 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2818 		desiredpages = (size == 0) ? 0 :
2819 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2820 
2821 		if (bp->b_flags & B_MALLOC)
2822 			panic("allocbuf: VMIO buffer can't be malloced");
2823 		/*
2824 		 * Set B_CACHE initially if buffer is 0 length or will become
2825 		 * 0-length.
2826 		 */
2827 		if (size == 0 || bp->b_bufsize == 0)
2828 			bp->b_flags |= B_CACHE;
2829 
2830 		if (newbsize < bp->b_bufsize) {
2831 			/*
2832 			 * DEV_BSIZE aligned new buffer size is less then the
2833 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2834 			 * if we have to remove any pages.
2835 			 */
2836 			if (desiredpages < bp->b_npages) {
2837 				vm_page_t m;
2838 
2839 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2840 				vm_page_lock_queues();
2841 				for (i = desiredpages; i < bp->b_npages; i++) {
2842 					/*
2843 					 * the page is not freed here -- it
2844 					 * is the responsibility of
2845 					 * vnode_pager_setsize
2846 					 */
2847 					m = bp->b_pages[i];
2848 					KASSERT(m != bogus_page,
2849 					    ("allocbuf: bogus page found"));
2850 					while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2851 						vm_page_lock_queues();
2852 
2853 					bp->b_pages[i] = NULL;
2854 					vm_page_unwire(m, 0);
2855 				}
2856 				vm_page_unlock_queues();
2857 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2858 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2859 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2860 				bp->b_npages = desiredpages;
2861 			}
2862 		} else if (size > bp->b_bcount) {
2863 			/*
2864 			 * We are growing the buffer, possibly in a
2865 			 * byte-granular fashion.
2866 			 */
2867 			struct vnode *vp;
2868 			vm_object_t obj;
2869 			vm_offset_t toff;
2870 			vm_offset_t tinc;
2871 
2872 			/*
2873 			 * Step 1, bring in the VM pages from the object,
2874 			 * allocating them if necessary.  We must clear
2875 			 * B_CACHE if these pages are not valid for the
2876 			 * range covered by the buffer.
2877 			 */
2878 
2879 			vp = bp->b_vp;
2880 			obj = bp->b_bufobj->bo_object;
2881 
2882 			VM_OBJECT_LOCK(obj);
2883 			while (bp->b_npages < desiredpages) {
2884 				vm_page_t m;
2885 				vm_pindex_t pi;
2886 
2887 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2888 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2889 					/*
2890 					 * note: must allocate system pages
2891 					 * since blocking here could intefere
2892 					 * with paging I/O, no matter which
2893 					 * process we are.
2894 					 */
2895 					m = vm_page_alloc(obj, pi,
2896 					    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2897 					    VM_ALLOC_WIRED);
2898 					if (m == NULL) {
2899 						atomic_add_int(&vm_pageout_deficit,
2900 						    desiredpages - bp->b_npages);
2901 						VM_OBJECT_UNLOCK(obj);
2902 						VM_WAIT;
2903 						VM_OBJECT_LOCK(obj);
2904 					} else {
2905 						if (m->valid == 0)
2906 							bp->b_flags &= ~B_CACHE;
2907 						bp->b_pages[bp->b_npages] = m;
2908 						++bp->b_npages;
2909 					}
2910 					continue;
2911 				}
2912 
2913 				/*
2914 				 * We found a page.  If we have to sleep on it,
2915 				 * retry because it might have gotten freed out
2916 				 * from under us.
2917 				 *
2918 				 * We can only test VPO_BUSY here.  Blocking on
2919 				 * m->busy might lead to a deadlock:
2920 				 *
2921 				 *  vm_fault->getpages->cluster_read->allocbuf
2922 				 *
2923 				 */
2924 				if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2925 					continue;
2926 
2927 				/*
2928 				 * We have a good page.
2929 				 */
2930 				vm_page_lock_queues();
2931 				vm_page_wire(m);
2932 				vm_page_unlock_queues();
2933 				bp->b_pages[bp->b_npages] = m;
2934 				++bp->b_npages;
2935 			}
2936 
2937 			/*
2938 			 * Step 2.  We've loaded the pages into the buffer,
2939 			 * we have to figure out if we can still have B_CACHE
2940 			 * set.  Note that B_CACHE is set according to the
2941 			 * byte-granular range ( bcount and size ), new the
2942 			 * aligned range ( newbsize ).
2943 			 *
2944 			 * The VM test is against m->valid, which is DEV_BSIZE
2945 			 * aligned.  Needless to say, the validity of the data
2946 			 * needs to also be DEV_BSIZE aligned.  Note that this
2947 			 * fails with NFS if the server or some other client
2948 			 * extends the file's EOF.  If our buffer is resized,
2949 			 * B_CACHE may remain set! XXX
2950 			 */
2951 
2952 			toff = bp->b_bcount;
2953 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2954 
2955 			while ((bp->b_flags & B_CACHE) && toff < size) {
2956 				vm_pindex_t pi;
2957 
2958 				if (tinc > (size - toff))
2959 					tinc = size - toff;
2960 
2961 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2962 				    PAGE_SHIFT;
2963 
2964 				vfs_buf_test_cache(
2965 				    bp,
2966 				    bp->b_offset,
2967 				    toff,
2968 				    tinc,
2969 				    bp->b_pages[pi]
2970 				);
2971 				toff += tinc;
2972 				tinc = PAGE_SIZE;
2973 			}
2974 			VM_OBJECT_UNLOCK(obj);
2975 
2976 			/*
2977 			 * Step 3, fixup the KVM pmap.  Remember that
2978 			 * bp->b_data is relative to bp->b_offset, but
2979 			 * bp->b_offset may be offset into the first page.
2980 			 */
2981 
2982 			bp->b_data = (caddr_t)
2983 			    trunc_page((vm_offset_t)bp->b_data);
2984 			pmap_qenter(
2985 			    (vm_offset_t)bp->b_data,
2986 			    bp->b_pages,
2987 			    bp->b_npages
2988 			);
2989 
2990 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2991 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
2992 		}
2993 	}
2994 	if (newbsize < bp->b_bufsize)
2995 		bufspacewakeup();
2996 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
2997 	bp->b_bcount = size;		/* requested buffer size	*/
2998 	return 1;
2999 }
3000 
3001 void
3002 biodone(struct bio *bp)
3003 {
3004 	void (*done)(struct bio *);
3005 
3006 	mtx_lock(&bdonelock);
3007 	bp->bio_flags |= BIO_DONE;
3008 	done = bp->bio_done;
3009 	if (done == NULL)
3010 		wakeup(bp);
3011 	mtx_unlock(&bdonelock);
3012 	if (done != NULL)
3013 		done(bp);
3014 }
3015 
3016 /*
3017  * Wait for a BIO to finish.
3018  *
3019  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3020  * case is not yet clear.
3021  */
3022 int
3023 biowait(struct bio *bp, const char *wchan)
3024 {
3025 
3026 	mtx_lock(&bdonelock);
3027 	while ((bp->bio_flags & BIO_DONE) == 0)
3028 		msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
3029 	mtx_unlock(&bdonelock);
3030 	if (bp->bio_error != 0)
3031 		return (bp->bio_error);
3032 	if (!(bp->bio_flags & BIO_ERROR))
3033 		return (0);
3034 	return (EIO);
3035 }
3036 
3037 void
3038 biofinish(struct bio *bp, struct devstat *stat, int error)
3039 {
3040 
3041 	if (error) {
3042 		bp->bio_error = error;
3043 		bp->bio_flags |= BIO_ERROR;
3044 	}
3045 	if (stat != NULL)
3046 		devstat_end_transaction_bio(stat, bp);
3047 	biodone(bp);
3048 }
3049 
3050 /*
3051  *	bufwait:
3052  *
3053  *	Wait for buffer I/O completion, returning error status.  The buffer
3054  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3055  *	error and cleared.
3056  */
3057 int
3058 bufwait(struct buf *bp)
3059 {
3060 	if (bp->b_iocmd == BIO_READ)
3061 		bwait(bp, PRIBIO, "biord");
3062 	else
3063 		bwait(bp, PRIBIO, "biowr");
3064 	if (bp->b_flags & B_EINTR) {
3065 		bp->b_flags &= ~B_EINTR;
3066 		return (EINTR);
3067 	}
3068 	if (bp->b_ioflags & BIO_ERROR) {
3069 		return (bp->b_error ? bp->b_error : EIO);
3070 	} else {
3071 		return (0);
3072 	}
3073 }
3074 
3075  /*
3076   * Call back function from struct bio back up to struct buf.
3077   */
3078 static void
3079 bufdonebio(struct bio *bip)
3080 {
3081 	struct buf *bp;
3082 
3083 	bp = bip->bio_caller2;
3084 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3085 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3086 	bp->b_ioflags = bip->bio_flags;
3087 	bp->b_error = bip->bio_error;
3088 	if (bp->b_error)
3089 		bp->b_ioflags |= BIO_ERROR;
3090 	bufdone(bp);
3091 	g_destroy_bio(bip);
3092 }
3093 
3094 void
3095 dev_strategy(struct cdev *dev, struct buf *bp)
3096 {
3097 	struct cdevsw *csw;
3098 	struct bio *bip;
3099 
3100 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3101 		panic("b_iocmd botch");
3102 	for (;;) {
3103 		bip = g_new_bio();
3104 		if (bip != NULL)
3105 			break;
3106 		/* Try again later */
3107 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3108 	}
3109 	bip->bio_cmd = bp->b_iocmd;
3110 	bip->bio_offset = bp->b_iooffset;
3111 	bip->bio_length = bp->b_bcount;
3112 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3113 	bip->bio_data = bp->b_data;
3114 	bip->bio_done = bufdonebio;
3115 	bip->bio_caller2 = bp;
3116 	bip->bio_dev = dev;
3117 	KASSERT(dev->si_refcount > 0,
3118 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3119 	    devtoname(dev)));
3120 	csw = dev_refthread(dev);
3121 	if (csw == NULL) {
3122 		g_destroy_bio(bip);
3123 		bp->b_error = ENXIO;
3124 		bp->b_ioflags = BIO_ERROR;
3125 		bufdone(bp);
3126 		return;
3127 	}
3128 	(*csw->d_strategy)(bip);
3129 	dev_relthread(dev);
3130 }
3131 
3132 /*
3133  *	bufdone:
3134  *
3135  *	Finish I/O on a buffer, optionally calling a completion function.
3136  *	This is usually called from an interrupt so process blocking is
3137  *	not allowed.
3138  *
3139  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3140  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3141  *	assuming B_INVAL is clear.
3142  *
3143  *	For the VMIO case, we set B_CACHE if the op was a read and no
3144  *	read error occured, or if the op was a write.  B_CACHE is never
3145  *	set if the buffer is invalid or otherwise uncacheable.
3146  *
3147  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3148  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3149  *	in the biodone routine.
3150  */
3151 void
3152 bufdone(struct buf *bp)
3153 {
3154 	struct bufobj *dropobj;
3155 	void    (*biodone)(struct buf *);
3156 
3157 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3158 	dropobj = NULL;
3159 
3160 	KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp,
3161 	    BUF_REFCNT(bp)));
3162 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3163 
3164 	runningbufwakeup(bp);
3165 	if (bp->b_iocmd == BIO_WRITE)
3166 		dropobj = bp->b_bufobj;
3167 	/* call optional completion function if requested */
3168 	if (bp->b_iodone != NULL) {
3169 		biodone = bp->b_iodone;
3170 		bp->b_iodone = NULL;
3171 		(*biodone) (bp);
3172 		if (dropobj)
3173 			bufobj_wdrop(dropobj);
3174 		return;
3175 	}
3176 
3177 	bufdone_finish(bp);
3178 
3179 	if (dropobj)
3180 		bufobj_wdrop(dropobj);
3181 }
3182 
3183 void
3184 bufdone_finish(struct buf *bp)
3185 {
3186 	KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp,
3187 	    BUF_REFCNT(bp)));
3188 
3189 	if (!LIST_EMPTY(&bp->b_dep))
3190 		buf_complete(bp);
3191 
3192 	if (bp->b_flags & B_VMIO) {
3193 		int i;
3194 		vm_ooffset_t foff;
3195 		vm_page_t m;
3196 		vm_object_t obj;
3197 		int iosize;
3198 		struct vnode *vp = bp->b_vp;
3199 		boolean_t are_queues_locked;
3200 
3201 		obj = bp->b_bufobj->bo_object;
3202 
3203 #if defined(VFS_BIO_DEBUG)
3204 		mp_fixme("usecount and vflag accessed without locks.");
3205 		if (vp->v_usecount == 0) {
3206 			panic("biodone: zero vnode ref count");
3207 		}
3208 
3209 		KASSERT(vp->v_object != NULL,
3210 			("biodone: vnode %p has no vm_object", vp));
3211 #endif
3212 
3213 		foff = bp->b_offset;
3214 		KASSERT(bp->b_offset != NOOFFSET,
3215 		    ("biodone: no buffer offset"));
3216 
3217 		VM_OBJECT_LOCK(obj);
3218 #if defined(VFS_BIO_DEBUG)
3219 		if (obj->paging_in_progress < bp->b_npages) {
3220 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3221 			    obj->paging_in_progress, bp->b_npages);
3222 		}
3223 #endif
3224 
3225 		/*
3226 		 * Set B_CACHE if the op was a normal read and no error
3227 		 * occured.  B_CACHE is set for writes in the b*write()
3228 		 * routines.
3229 		 */
3230 		iosize = bp->b_bcount - bp->b_resid;
3231 		if (bp->b_iocmd == BIO_READ &&
3232 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3233 		    !(bp->b_ioflags & BIO_ERROR)) {
3234 			bp->b_flags |= B_CACHE;
3235 		}
3236 		if (bp->b_iocmd == BIO_READ) {
3237 			vm_page_lock_queues();
3238 			are_queues_locked = TRUE;
3239 		} else
3240 			are_queues_locked = FALSE;
3241 		for (i = 0; i < bp->b_npages; i++) {
3242 			int bogusflag = 0;
3243 			int resid;
3244 
3245 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3246 			if (resid > iosize)
3247 				resid = iosize;
3248 
3249 			/*
3250 			 * cleanup bogus pages, restoring the originals
3251 			 */
3252 			m = bp->b_pages[i];
3253 			if (m == bogus_page) {
3254 				bogusflag = 1;
3255 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3256 				if (m == NULL)
3257 					panic("biodone: page disappeared!");
3258 				bp->b_pages[i] = m;
3259 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3260 				    bp->b_pages, bp->b_npages);
3261 			}
3262 #if defined(VFS_BIO_DEBUG)
3263 			if (OFF_TO_IDX(foff) != m->pindex) {
3264 				printf(
3265 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3266 				    (intmax_t)foff, (uintmax_t)m->pindex);
3267 			}
3268 #endif
3269 
3270 			/*
3271 			 * In the write case, the valid and clean bits are
3272 			 * already changed correctly ( see bdwrite() ), so we
3273 			 * only need to do this here in the read case.
3274 			 */
3275 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3276 				vfs_page_set_valid(bp, foff, m);
3277 			}
3278 
3279 			/*
3280 			 * when debugging new filesystems or buffer I/O methods, this
3281 			 * is the most common error that pops up.  if you see this, you
3282 			 * have not set the page busy flag correctly!!!
3283 			 */
3284 			if (m->busy == 0) {
3285 				printf("biodone: page busy < 0, "
3286 				    "pindex: %d, foff: 0x(%x,%x), "
3287 				    "resid: %d, index: %d\n",
3288 				    (int) m->pindex, (int)(foff >> 32),
3289 						(int) foff & 0xffffffff, resid, i);
3290 				if (!vn_isdisk(vp, NULL))
3291 					printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3292 					    (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3293 					    (intmax_t) bp->b_lblkno,
3294 					    bp->b_flags, bp->b_npages);
3295 				else
3296 					printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3297 					    (intmax_t) bp->b_lblkno,
3298 					    bp->b_flags, bp->b_npages);
3299 				printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3300 				    (u_long)m->valid, (u_long)m->dirty,
3301 				    m->wire_count);
3302 				panic("biodone: page busy < 0\n");
3303 			}
3304 			vm_page_io_finish(m);
3305 			vm_object_pip_subtract(obj, 1);
3306 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3307 			iosize -= resid;
3308 		}
3309 		if (are_queues_locked)
3310 			vm_page_unlock_queues();
3311 		vm_object_pip_wakeupn(obj, 0);
3312 		VM_OBJECT_UNLOCK(obj);
3313 	}
3314 
3315 	/*
3316 	 * For asynchronous completions, release the buffer now. The brelse
3317 	 * will do a wakeup there if necessary - so no need to do a wakeup
3318 	 * here in the async case. The sync case always needs to do a wakeup.
3319 	 */
3320 
3321 	if (bp->b_flags & B_ASYNC) {
3322 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3323 			brelse(bp);
3324 		else
3325 			bqrelse(bp);
3326 	} else
3327 		bdone(bp);
3328 }
3329 
3330 /*
3331  * This routine is called in lieu of iodone in the case of
3332  * incomplete I/O.  This keeps the busy status for pages
3333  * consistant.
3334  */
3335 void
3336 vfs_unbusy_pages(struct buf *bp)
3337 {
3338 	int i;
3339 	vm_object_t obj;
3340 	vm_page_t m;
3341 
3342 	runningbufwakeup(bp);
3343 	if (!(bp->b_flags & B_VMIO))
3344 		return;
3345 
3346 	obj = bp->b_bufobj->bo_object;
3347 	VM_OBJECT_LOCK(obj);
3348 	for (i = 0; i < bp->b_npages; i++) {
3349 		m = bp->b_pages[i];
3350 		if (m == bogus_page) {
3351 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3352 			if (!m)
3353 				panic("vfs_unbusy_pages: page missing\n");
3354 			bp->b_pages[i] = m;
3355 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3356 			    bp->b_pages, bp->b_npages);
3357 		}
3358 		vm_object_pip_subtract(obj, 1);
3359 		vm_page_io_finish(m);
3360 	}
3361 	vm_object_pip_wakeupn(obj, 0);
3362 	VM_OBJECT_UNLOCK(obj);
3363 }
3364 
3365 /*
3366  * vfs_page_set_valid:
3367  *
3368  *	Set the valid bits in a page based on the supplied offset.   The
3369  *	range is restricted to the buffer's size.
3370  *
3371  *	This routine is typically called after a read completes.
3372  */
3373 static void
3374 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3375 {
3376 	vm_ooffset_t soff, eoff;
3377 
3378 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3379 	/*
3380 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3381 	 * page boundry or cross the end of the buffer.  The end of the
3382 	 * buffer, in this case, is our file EOF, not the allocation size
3383 	 * of the buffer.
3384 	 */
3385 	soff = off;
3386 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3387 	if (eoff > bp->b_offset + bp->b_bcount)
3388 		eoff = bp->b_offset + bp->b_bcount;
3389 
3390 	/*
3391 	 * Set valid range.  This is typically the entire buffer and thus the
3392 	 * entire page.
3393 	 */
3394 	if (eoff > soff) {
3395 		vm_page_set_validclean(
3396 		    m,
3397 		   (vm_offset_t) (soff & PAGE_MASK),
3398 		   (vm_offset_t) (eoff - soff)
3399 		);
3400 	}
3401 }
3402 
3403 /*
3404  * This routine is called before a device strategy routine.
3405  * It is used to tell the VM system that paging I/O is in
3406  * progress, and treat the pages associated with the buffer
3407  * almost as being VPO_BUSY.  Also the object paging_in_progress
3408  * flag is handled to make sure that the object doesn't become
3409  * inconsistant.
3410  *
3411  * Since I/O has not been initiated yet, certain buffer flags
3412  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3413  * and should be ignored.
3414  */
3415 void
3416 vfs_busy_pages(struct buf *bp, int clear_modify)
3417 {
3418 	int i, bogus;
3419 	vm_object_t obj;
3420 	vm_ooffset_t foff;
3421 	vm_page_t m;
3422 
3423 	if (!(bp->b_flags & B_VMIO))
3424 		return;
3425 
3426 	obj = bp->b_bufobj->bo_object;
3427 	foff = bp->b_offset;
3428 	KASSERT(bp->b_offset != NOOFFSET,
3429 	    ("vfs_busy_pages: no buffer offset"));
3430 	VM_OBJECT_LOCK(obj);
3431 	if (bp->b_bufsize != 0)
3432 		vfs_setdirty_locked_object(bp);
3433 retry:
3434 	for (i = 0; i < bp->b_npages; i++) {
3435 		m = bp->b_pages[i];
3436 
3437 		if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3438 			goto retry;
3439 	}
3440 	bogus = 0;
3441 	vm_page_lock_queues();
3442 	for (i = 0; i < bp->b_npages; i++) {
3443 		m = bp->b_pages[i];
3444 
3445 		if ((bp->b_flags & B_CLUSTER) == 0) {
3446 			vm_object_pip_add(obj, 1);
3447 			vm_page_io_start(m);
3448 		}
3449 		/*
3450 		 * When readying a buffer for a read ( i.e
3451 		 * clear_modify == 0 ), it is important to do
3452 		 * bogus_page replacement for valid pages in
3453 		 * partially instantiated buffers.  Partially
3454 		 * instantiated buffers can, in turn, occur when
3455 		 * reconstituting a buffer from its VM backing store
3456 		 * base.  We only have to do this if B_CACHE is
3457 		 * clear ( which causes the I/O to occur in the
3458 		 * first place ).  The replacement prevents the read
3459 		 * I/O from overwriting potentially dirty VM-backed
3460 		 * pages.  XXX bogus page replacement is, uh, bogus.
3461 		 * It may not work properly with small-block devices.
3462 		 * We need to find a better way.
3463 		 */
3464 		pmap_remove_all(m);
3465 		if (clear_modify)
3466 			vfs_page_set_valid(bp, foff, m);
3467 		else if (m->valid == VM_PAGE_BITS_ALL &&
3468 		    (bp->b_flags & B_CACHE) == 0) {
3469 			bp->b_pages[i] = bogus_page;
3470 			bogus++;
3471 		}
3472 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3473 	}
3474 	vm_page_unlock_queues();
3475 	VM_OBJECT_UNLOCK(obj);
3476 	if (bogus)
3477 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3478 		    bp->b_pages, bp->b_npages);
3479 }
3480 
3481 /*
3482  * Tell the VM system that the pages associated with this buffer
3483  * are clean.  This is used for delayed writes where the data is
3484  * going to go to disk eventually without additional VM intevention.
3485  *
3486  * Note that while we only really need to clean through to b_bcount, we
3487  * just go ahead and clean through to b_bufsize.
3488  */
3489 static void
3490 vfs_clean_pages(struct buf *bp)
3491 {
3492 	int i;
3493 	vm_ooffset_t foff, noff, eoff;
3494 	vm_page_t m;
3495 
3496 	if (!(bp->b_flags & B_VMIO))
3497 		return;
3498 
3499 	foff = bp->b_offset;
3500 	KASSERT(bp->b_offset != NOOFFSET,
3501 	    ("vfs_clean_pages: no buffer offset"));
3502 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3503 	vm_page_lock_queues();
3504 	for (i = 0; i < bp->b_npages; i++) {
3505 		m = bp->b_pages[i];
3506 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3507 		eoff = noff;
3508 
3509 		if (eoff > bp->b_offset + bp->b_bufsize)
3510 			eoff = bp->b_offset + bp->b_bufsize;
3511 		vfs_page_set_valid(bp, foff, m);
3512 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3513 		foff = noff;
3514 	}
3515 	vm_page_unlock_queues();
3516 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3517 }
3518 
3519 /*
3520  *	vfs_bio_set_validclean:
3521  *
3522  *	Set the range within the buffer to valid and clean.  The range is
3523  *	relative to the beginning of the buffer, b_offset.  Note that b_offset
3524  *	itself may be offset from the beginning of the first page.
3525  *
3526  */
3527 
3528 void
3529 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3530 {
3531 	int i, n;
3532 	vm_page_t m;
3533 
3534 	if (!(bp->b_flags & B_VMIO))
3535 		return;
3536 	/*
3537 	 * Fixup base to be relative to beginning of first page.
3538 	 * Set initial n to be the maximum number of bytes in the
3539 	 * first page that can be validated.
3540 	 */
3541 
3542 	base += (bp->b_offset & PAGE_MASK);
3543 	n = PAGE_SIZE - (base & PAGE_MASK);
3544 
3545 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3546 	vm_page_lock_queues();
3547 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3548 		m = bp->b_pages[i];
3549 		if (n > size)
3550 			n = size;
3551 		vm_page_set_validclean(m, base & PAGE_MASK, n);
3552 		base += n;
3553 		size -= n;
3554 		n = PAGE_SIZE;
3555 	}
3556 	vm_page_unlock_queues();
3557 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3558 }
3559 
3560 /*
3561  *	vfs_bio_clrbuf:
3562  *
3563  *	clear a buffer.  This routine essentially fakes an I/O, so we need
3564  *	to clear BIO_ERROR and B_INVAL.
3565  *
3566  *	Note that while we only theoretically need to clear through b_bcount,
3567  *	we go ahead and clear through b_bufsize.
3568  */
3569 
3570 void
3571 vfs_bio_clrbuf(struct buf *bp)
3572 {
3573 	int i, j, mask = 0;
3574 	caddr_t sa, ea;
3575 
3576 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3577 		clrbuf(bp);
3578 		return;
3579 	}
3580 
3581 	bp->b_flags &= ~B_INVAL;
3582 	bp->b_ioflags &= ~BIO_ERROR;
3583 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3584 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3585 	    (bp->b_offset & PAGE_MASK) == 0) {
3586 		if (bp->b_pages[0] == bogus_page)
3587 			goto unlock;
3588 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3589 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3590 		if ((bp->b_pages[0]->valid & mask) == mask)
3591 			goto unlock;
3592 		if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3593 		    ((bp->b_pages[0]->valid & mask) == 0)) {
3594 			bzero(bp->b_data, bp->b_bufsize);
3595 			bp->b_pages[0]->valid |= mask;
3596 			goto unlock;
3597 		}
3598 	}
3599 	ea = sa = bp->b_data;
3600 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3601 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3602 		ea = (caddr_t)(vm_offset_t)ulmin(
3603 		    (u_long)(vm_offset_t)ea,
3604 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3605 		if (bp->b_pages[i] == bogus_page)
3606 			continue;
3607 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3608 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3609 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3610 		if ((bp->b_pages[i]->valid & mask) == mask)
3611 			continue;
3612 		if ((bp->b_pages[i]->valid & mask) == 0) {
3613 			if ((bp->b_pages[i]->flags & PG_ZERO) == 0)
3614 				bzero(sa, ea - sa);
3615 		} else {
3616 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3617 				if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3618 				    (bp->b_pages[i]->valid & (1 << j)) == 0)
3619 					bzero(sa, DEV_BSIZE);
3620 			}
3621 		}
3622 		bp->b_pages[i]->valid |= mask;
3623 	}
3624 unlock:
3625 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3626 	bp->b_resid = 0;
3627 }
3628 
3629 /*
3630  * vm_hold_load_pages and vm_hold_free_pages get pages into
3631  * a buffers address space.  The pages are anonymous and are
3632  * not associated with a file object.
3633  */
3634 static void
3635 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3636 {
3637 	vm_offset_t pg;
3638 	vm_page_t p;
3639 	int index;
3640 
3641 	to = round_page(to);
3642 	from = round_page(from);
3643 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3644 
3645 	VM_OBJECT_LOCK(kernel_object);
3646 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3647 tryagain:
3648 		/*
3649 		 * note: must allocate system pages since blocking here
3650 		 * could intefere with paging I/O, no matter which
3651 		 * process we are.
3652 		 */
3653 		p = vm_page_alloc(kernel_object,
3654 			((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3655 		    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3656 		if (!p) {
3657 			atomic_add_int(&vm_pageout_deficit,
3658 			    (to - pg) >> PAGE_SHIFT);
3659 			VM_OBJECT_UNLOCK(kernel_object);
3660 			VM_WAIT;
3661 			VM_OBJECT_LOCK(kernel_object);
3662 			goto tryagain;
3663 		}
3664 		p->valid = VM_PAGE_BITS_ALL;
3665 		pmap_qenter(pg, &p, 1);
3666 		bp->b_pages[index] = p;
3667 	}
3668 	VM_OBJECT_UNLOCK(kernel_object);
3669 	bp->b_npages = index;
3670 }
3671 
3672 /* Return pages associated with this buf to the vm system */
3673 static void
3674 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3675 {
3676 	vm_offset_t pg;
3677 	vm_page_t p;
3678 	int index, newnpages;
3679 
3680 	from = round_page(from);
3681 	to = round_page(to);
3682 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3683 
3684 	VM_OBJECT_LOCK(kernel_object);
3685 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3686 		p = bp->b_pages[index];
3687 		if (p && (index < bp->b_npages)) {
3688 			if (p->busy) {
3689 				printf(
3690 			    "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3691 				    (intmax_t)bp->b_blkno,
3692 				    (intmax_t)bp->b_lblkno);
3693 			}
3694 			bp->b_pages[index] = NULL;
3695 			pmap_qremove(pg, 1);
3696 			vm_page_lock_queues();
3697 			vm_page_unwire(p, 0);
3698 			vm_page_free(p);
3699 			vm_page_unlock_queues();
3700 		}
3701 	}
3702 	VM_OBJECT_UNLOCK(kernel_object);
3703 	bp->b_npages = newnpages;
3704 }
3705 
3706 /*
3707  * Map an IO request into kernel virtual address space.
3708  *
3709  * All requests are (re)mapped into kernel VA space.
3710  * Notice that we use b_bufsize for the size of the buffer
3711  * to be mapped.  b_bcount might be modified by the driver.
3712  *
3713  * Note that even if the caller determines that the address space should
3714  * be valid, a race or a smaller-file mapped into a larger space may
3715  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3716  * check the return value.
3717  */
3718 int
3719 vmapbuf(struct buf *bp)
3720 {
3721 	caddr_t addr, kva;
3722 	vm_prot_t prot;
3723 	int pidx, i;
3724 	struct vm_page *m;
3725 	struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3726 
3727 	if (bp->b_bufsize < 0)
3728 		return (-1);
3729 	prot = VM_PROT_READ;
3730 	if (bp->b_iocmd == BIO_READ)
3731 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3732 	for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3733 	     addr < bp->b_data + bp->b_bufsize;
3734 	     addr += PAGE_SIZE, pidx++) {
3735 		/*
3736 		 * Do the vm_fault if needed; do the copy-on-write thing
3737 		 * when reading stuff off device into memory.
3738 		 *
3739 		 * NOTE! Must use pmap_extract() because addr may be in
3740 		 * the userland address space, and kextract is only guarenteed
3741 		 * to work for the kernland address space (see: sparc64 port).
3742 		 */
3743 retry:
3744 		if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3745 		    prot) < 0) {
3746 			vm_page_lock_queues();
3747 			for (i = 0; i < pidx; ++i) {
3748 				vm_page_unhold(bp->b_pages[i]);
3749 				bp->b_pages[i] = NULL;
3750 			}
3751 			vm_page_unlock_queues();
3752 			return(-1);
3753 		}
3754 		m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3755 		if (m == NULL)
3756 			goto retry;
3757 		bp->b_pages[pidx] = m;
3758 	}
3759 	if (pidx > btoc(MAXPHYS))
3760 		panic("vmapbuf: mapped more than MAXPHYS");
3761 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3762 
3763 	kva = bp->b_saveaddr;
3764 	bp->b_npages = pidx;
3765 	bp->b_saveaddr = bp->b_data;
3766 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3767 	return(0);
3768 }
3769 
3770 /*
3771  * Free the io map PTEs associated with this IO operation.
3772  * We also invalidate the TLB entries and restore the original b_addr.
3773  */
3774 void
3775 vunmapbuf(struct buf *bp)
3776 {
3777 	int pidx;
3778 	int npages;
3779 
3780 	npages = bp->b_npages;
3781 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3782 	vm_page_lock_queues();
3783 	for (pidx = 0; pidx < npages; pidx++)
3784 		vm_page_unhold(bp->b_pages[pidx]);
3785 	vm_page_unlock_queues();
3786 
3787 	bp->b_data = bp->b_saveaddr;
3788 }
3789 
3790 void
3791 bdone(struct buf *bp)
3792 {
3793 
3794 	mtx_lock(&bdonelock);
3795 	bp->b_flags |= B_DONE;
3796 	wakeup(bp);
3797 	mtx_unlock(&bdonelock);
3798 }
3799 
3800 void
3801 bwait(struct buf *bp, u_char pri, const char *wchan)
3802 {
3803 
3804 	mtx_lock(&bdonelock);
3805 	while ((bp->b_flags & B_DONE) == 0)
3806 		msleep(bp, &bdonelock, pri, wchan, 0);
3807 	mtx_unlock(&bdonelock);
3808 }
3809 
3810 int
3811 bufsync(struct bufobj *bo, int waitfor, struct thread *td)
3812 {
3813 
3814 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, td));
3815 }
3816 
3817 void
3818 bufstrategy(struct bufobj *bo, struct buf *bp)
3819 {
3820 	int i = 0;
3821 	struct vnode *vp;
3822 
3823 	vp = bp->b_vp;
3824 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3825 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3826 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3827 	i = VOP_STRATEGY(vp, bp);
3828 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3829 }
3830 
3831 void
3832 bufobj_wrefl(struct bufobj *bo)
3833 {
3834 
3835 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3836 	ASSERT_BO_LOCKED(bo);
3837 	bo->bo_numoutput++;
3838 }
3839 
3840 void
3841 bufobj_wref(struct bufobj *bo)
3842 {
3843 
3844 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3845 	BO_LOCK(bo);
3846 	bo->bo_numoutput++;
3847 	BO_UNLOCK(bo);
3848 }
3849 
3850 void
3851 bufobj_wdrop(struct bufobj *bo)
3852 {
3853 
3854 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3855 	BO_LOCK(bo);
3856 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3857 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3858 		bo->bo_flag &= ~BO_WWAIT;
3859 		wakeup(&bo->bo_numoutput);
3860 	}
3861 	BO_UNLOCK(bo);
3862 }
3863 
3864 int
3865 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3866 {
3867 	int error;
3868 
3869 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3870 	ASSERT_BO_LOCKED(bo);
3871 	error = 0;
3872 	while (bo->bo_numoutput) {
3873 		bo->bo_flag |= BO_WWAIT;
3874 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3875 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3876 		if (error)
3877 			break;
3878 	}
3879 	return (error);
3880 }
3881 
3882 void
3883 bpin(struct buf *bp)
3884 {
3885 	mtx_lock(&bpinlock);
3886 	bp->b_pin_count++;
3887 	mtx_unlock(&bpinlock);
3888 }
3889 
3890 void
3891 bunpin(struct buf *bp)
3892 {
3893 	mtx_lock(&bpinlock);
3894 	if (--bp->b_pin_count == 0)
3895 		wakeup(bp);
3896 	mtx_unlock(&bpinlock);
3897 }
3898 
3899 void
3900 bunpin_wait(struct buf *bp)
3901 {
3902 	mtx_lock(&bpinlock);
3903 	while (bp->b_pin_count > 0)
3904 		msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0);
3905 	mtx_unlock(&bpinlock);
3906 }
3907 
3908 #include "opt_ddb.h"
3909 #ifdef DDB
3910 #include <ddb/ddb.h>
3911 
3912 /* DDB command to show buffer data */
3913 DB_SHOW_COMMAND(buffer, db_show_buffer)
3914 {
3915 	/* get args */
3916 	struct buf *bp = (struct buf *)addr;
3917 
3918 	if (!have_addr) {
3919 		db_printf("usage: show buffer <addr>\n");
3920 		return;
3921 	}
3922 
3923 	db_printf("buf at %p\n", bp);
3924 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3925 	db_printf(
3926 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3927 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n",
3928 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3929 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno);
3930 	if (bp->b_npages) {
3931 		int i;
3932 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3933 		for (i = 0; i < bp->b_npages; i++) {
3934 			vm_page_t m;
3935 			m = bp->b_pages[i];
3936 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3937 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3938 			if ((i + 1) < bp->b_npages)
3939 				db_printf(",");
3940 		}
3941 		db_printf("\n");
3942 	}
3943 	lockmgr_printinfo(&bp->b_lock);
3944 }
3945 
3946 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
3947 {
3948 	struct buf *bp;
3949 	int i;
3950 
3951 	for (i = 0; i < nbuf; i++) {
3952 		bp = &buf[i];
3953 		if (lockcount(&bp->b_lock)) {
3954 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
3955 			db_printf("\n");
3956 		}
3957 	}
3958 }
3959 #endif /* DDB */
3960