xref: /freebsd/sys/kern/vfs_bio.c (revision bb15ca603fa442c72dde3f3cb8b46db6970e3950)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * this file contains a new buffer I/O scheme implementing a coherent
30  * VM object and buffer cache scheme.  Pains have been taken to make
31  * sure that the performance degradation associated with schemes such
32  * as this is not realized.
33  *
34  * Author:  John S. Dyson
35  * Significant help during the development and debugging phases
36  * had been provided by David Greenman, also of the FreeBSD core team.
37  *
38  * see man buf(9) for more info.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bio.h>
47 #include <sys/conf.h>
48 #include <sys/buf.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
51 #include <sys/fail.h>
52 #include <sys/limits.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mount.h>
56 #include <sys/mutex.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
59 #include <sys/proc.h>
60 #include <sys/resourcevar.h>
61 #include <sys/sysctl.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
64 #include <geom/geom.h>
65 #include <vm/vm.h>
66 #include <vm/vm_param.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_pageout.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_extern.h>
72 #include <vm/vm_map.h>
73 #include "opt_compat.h"
74 #include "opt_directio.h"
75 #include "opt_swap.h"
76 
77 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
78 
79 struct	bio_ops bioops;		/* I/O operation notification */
80 
81 struct	buf_ops buf_ops_bio = {
82 	.bop_name	=	"buf_ops_bio",
83 	.bop_write	=	bufwrite,
84 	.bop_strategy	=	bufstrategy,
85 	.bop_sync	=	bufsync,
86 	.bop_bdflush	=	bufbdflush,
87 };
88 
89 /*
90  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
91  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
92  */
93 struct buf *buf;		/* buffer header pool */
94 
95 static struct proc *bufdaemonproc;
96 
97 static int inmem(struct vnode *vp, daddr_t blkno);
98 static void vm_hold_free_pages(struct buf *bp, int newbsize);
99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
100 		vm_offset_t to);
101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
103 		vm_page_t m);
104 static void vfs_drain_busy_pages(struct buf *bp);
105 static void vfs_clean_pages_dirty_buf(struct buf *bp);
106 static void vfs_setdirty_locked_object(struct buf *bp);
107 static void vfs_vmio_release(struct buf *bp);
108 static int vfs_bio_clcheck(struct vnode *vp, int size,
109 		daddr_t lblkno, daddr_t blkno);
110 static int buf_do_flush(struct vnode *vp);
111 static int flushbufqueues(struct vnode *, int, int);
112 static void buf_daemon(void);
113 static void bremfreel(struct buf *bp);
114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
115     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
117 #endif
118 
119 int vmiodirenable = TRUE;
120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
121     "Use the VM system for directory writes");
122 long runningbufspace;
123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
124     "Amount of presently outstanding async buffer io");
125 static long bufspace;
126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
127     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
129     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
130 #else
131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
132     "Virtual memory used for buffers");
133 #endif
134 static long maxbufspace;
135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
136     "Maximum allowed value of bufspace (including buf_daemon)");
137 static long bufmallocspace;
138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
139     "Amount of malloced memory for buffers");
140 static long maxbufmallocspace;
141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
142     "Maximum amount of malloced memory for buffers");
143 static long lobufspace;
144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
145     "Minimum amount of buffers we want to have");
146 long hibufspace;
147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
148     "Maximum allowed value of bufspace (excluding buf_daemon)");
149 static int bufreusecnt;
150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
151     "Number of times we have reused a buffer");
152 static int buffreekvacnt;
153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
154     "Number of times we have freed the KVA space from some buffer");
155 static int bufdefragcnt;
156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
157     "Number of times we have had to repeat buffer allocation to defragment");
158 static long lorunningspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
160     "Minimum preferred space used for in-progress I/O");
161 static long hirunningspace;
162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
163     "Maximum amount of space to use for in-progress I/O");
164 int dirtybufferflushes;
165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
166     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
167 int bdwriteskip;
168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
169     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
170 int altbufferflushes;
171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
172     0, "Number of fsync flushes to limit dirty buffers");
173 static int recursiveflushes;
174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
175     0, "Number of flushes skipped due to being recursive");
176 static int numdirtybuffers;
177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
178     "Number of buffers that are dirty (has unwritten changes) at the moment");
179 static int lodirtybuffers;
180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
181     "How many buffers we want to have free before bufdaemon can sleep");
182 static int hidirtybuffers;
183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
184     "When the number of dirty buffers is considered severe");
185 int dirtybufthresh;
186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
187     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
188 static int numfreebuffers;
189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
190     "Number of free buffers");
191 static int lofreebuffers;
192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
193    "XXX Unused");
194 static int hifreebuffers;
195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
196    "XXX Complicatedly unused");
197 static int getnewbufcalls;
198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
199    "Number of calls to getnewbuf");
200 static int getnewbufrestarts;
201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
202     "Number of times getnewbuf has had to restart a buffer aquisition");
203 static int flushbufqtarget = 100;
204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
205     "Amount of work to do in flushbufqueues when helping bufdaemon");
206 static long notbufdflashes;
207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, &notbufdflashes, 0,
208     "Number of dirty buffer flushes done by the bufdaemon helpers");
209 
210 /*
211  * Wakeup point for bufdaemon, as well as indicator of whether it is already
212  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
213  * is idling.
214  */
215 static int bd_request;
216 
217 /*
218  * Request for the buf daemon to write more buffers than is indicated by
219  * lodirtybuf.  This may be necessary to push out excess dependencies or
220  * defragment the address space where a simple count of the number of dirty
221  * buffers is insufficient to characterize the demand for flushing them.
222  */
223 static int bd_speedupreq;
224 
225 /*
226  * This lock synchronizes access to bd_request.
227  */
228 static struct mtx bdlock;
229 
230 /*
231  * bogus page -- for I/O to/from partially complete buffers
232  * this is a temporary solution to the problem, but it is not
233  * really that bad.  it would be better to split the buffer
234  * for input in the case of buffers partially already in memory,
235  * but the code is intricate enough already.
236  */
237 vm_page_t bogus_page;
238 
239 /*
240  * Synchronization (sleep/wakeup) variable for active buffer space requests.
241  * Set when wait starts, cleared prior to wakeup().
242  * Used in runningbufwakeup() and waitrunningbufspace().
243  */
244 static int runningbufreq;
245 
246 /*
247  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
248  * waitrunningbufspace().
249  */
250 static struct mtx rbreqlock;
251 
252 /*
253  * Synchronization (sleep/wakeup) variable for buffer requests.
254  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
255  * by and/or.
256  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
257  * getnewbuf(), and getblk().
258  */
259 static int needsbuffer;
260 
261 /*
262  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
263  */
264 static struct mtx nblock;
265 
266 /*
267  * Definitions for the buffer free lists.
268  */
269 #define BUFFER_QUEUES	6	/* number of free buffer queues */
270 
271 #define QUEUE_NONE	0	/* on no queue */
272 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
273 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
274 #define QUEUE_DIRTY_GIANT 3	/* B_DELWRI buffers that need giant */
275 #define QUEUE_EMPTYKVA	4	/* empty buffer headers w/KVA assignment */
276 #define QUEUE_EMPTY	5	/* empty buffer headers */
277 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
278 
279 /* Queues for free buffers with various properties */
280 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
281 
282 /* Lock for the bufqueues */
283 static struct mtx bqlock;
284 
285 /*
286  * Single global constant for BUF_WMESG, to avoid getting multiple references.
287  * buf_wmesg is referred from macros.
288  */
289 const char *buf_wmesg = BUF_WMESG;
290 
291 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
292 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
293 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
294 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
295 
296 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
297     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
298 static int
299 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
300 {
301 	long lvalue;
302 	int ivalue;
303 
304 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
305 		return (sysctl_handle_long(oidp, arg1, arg2, req));
306 	lvalue = *(long *)arg1;
307 	if (lvalue > INT_MAX)
308 		/* On overflow, still write out a long to trigger ENOMEM. */
309 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
310 	ivalue = lvalue;
311 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
312 }
313 #endif
314 
315 #ifdef DIRECTIO
316 extern void ffs_rawread_setup(void);
317 #endif /* DIRECTIO */
318 /*
319  *	numdirtywakeup:
320  *
321  *	If someone is blocked due to there being too many dirty buffers,
322  *	and numdirtybuffers is now reasonable, wake them up.
323  */
324 
325 static __inline void
326 numdirtywakeup(int level)
327 {
328 
329 	if (numdirtybuffers <= level) {
330 		mtx_lock(&nblock);
331 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
332 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
333 			wakeup(&needsbuffer);
334 		}
335 		mtx_unlock(&nblock);
336 	}
337 }
338 
339 /*
340  *	bufspacewakeup:
341  *
342  *	Called when buffer space is potentially available for recovery.
343  *	getnewbuf() will block on this flag when it is unable to free
344  *	sufficient buffer space.  Buffer space becomes recoverable when
345  *	bp's get placed back in the queues.
346  */
347 
348 static __inline void
349 bufspacewakeup(void)
350 {
351 
352 	/*
353 	 * If someone is waiting for BUF space, wake them up.  Even
354 	 * though we haven't freed the kva space yet, the waiting
355 	 * process will be able to now.
356 	 */
357 	mtx_lock(&nblock);
358 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
359 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
360 		wakeup(&needsbuffer);
361 	}
362 	mtx_unlock(&nblock);
363 }
364 
365 /*
366  * runningbufwakeup() - in-progress I/O accounting.
367  *
368  */
369 void
370 runningbufwakeup(struct buf *bp)
371 {
372 
373 	if (bp->b_runningbufspace) {
374 		atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
375 		bp->b_runningbufspace = 0;
376 		mtx_lock(&rbreqlock);
377 		if (runningbufreq && runningbufspace <= lorunningspace) {
378 			runningbufreq = 0;
379 			wakeup(&runningbufreq);
380 		}
381 		mtx_unlock(&rbreqlock);
382 	}
383 }
384 
385 /*
386  *	bufcountwakeup:
387  *
388  *	Called when a buffer has been added to one of the free queues to
389  *	account for the buffer and to wakeup anyone waiting for free buffers.
390  *	This typically occurs when large amounts of metadata are being handled
391  *	by the buffer cache ( else buffer space runs out first, usually ).
392  */
393 
394 static __inline void
395 bufcountwakeup(struct buf *bp)
396 {
397 	int old;
398 
399 	KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
400 	    ("buf %p already counted as free", bp));
401 	if (bp->b_bufobj != NULL)
402 		mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
403 	bp->b_vflags |= BV_INFREECNT;
404 	old = atomic_fetchadd_int(&numfreebuffers, 1);
405 	KASSERT(old >= 0 && old < nbuf,
406 	    ("numfreebuffers climbed to %d", old + 1));
407 	mtx_lock(&nblock);
408 	if (needsbuffer) {
409 		needsbuffer &= ~VFS_BIO_NEED_ANY;
410 		if (numfreebuffers >= hifreebuffers)
411 			needsbuffer &= ~VFS_BIO_NEED_FREE;
412 		wakeup(&needsbuffer);
413 	}
414 	mtx_unlock(&nblock);
415 }
416 
417 /*
418  *	waitrunningbufspace()
419  *
420  *	runningbufspace is a measure of the amount of I/O currently
421  *	running.  This routine is used in async-write situations to
422  *	prevent creating huge backups of pending writes to a device.
423  *	Only asynchronous writes are governed by this function.
424  *
425  *	Reads will adjust runningbufspace, but will not block based on it.
426  *	The read load has a side effect of reducing the allowed write load.
427  *
428  *	This does NOT turn an async write into a sync write.  It waits
429  *	for earlier writes to complete and generally returns before the
430  *	caller's write has reached the device.
431  */
432 void
433 waitrunningbufspace(void)
434 {
435 
436 	mtx_lock(&rbreqlock);
437 	while (runningbufspace > hirunningspace) {
438 		++runningbufreq;
439 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
440 	}
441 	mtx_unlock(&rbreqlock);
442 }
443 
444 
445 /*
446  *	vfs_buf_test_cache:
447  *
448  *	Called when a buffer is extended.  This function clears the B_CACHE
449  *	bit if the newly extended portion of the buffer does not contain
450  *	valid data.
451  */
452 static __inline
453 void
454 vfs_buf_test_cache(struct buf *bp,
455 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
456 		  vm_page_t m)
457 {
458 
459 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
460 	if (bp->b_flags & B_CACHE) {
461 		int base = (foff + off) & PAGE_MASK;
462 		if (vm_page_is_valid(m, base, size) == 0)
463 			bp->b_flags &= ~B_CACHE;
464 	}
465 }
466 
467 /* Wake up the buffer daemon if necessary */
468 static __inline
469 void
470 bd_wakeup(int dirtybuflevel)
471 {
472 
473 	mtx_lock(&bdlock);
474 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
475 		bd_request = 1;
476 		wakeup(&bd_request);
477 	}
478 	mtx_unlock(&bdlock);
479 }
480 
481 /*
482  * bd_speedup - speedup the buffer cache flushing code
483  */
484 
485 void
486 bd_speedup(void)
487 {
488 	int needwake;
489 
490 	mtx_lock(&bdlock);
491 	needwake = 0;
492 	if (bd_speedupreq == 0 || bd_request == 0)
493 		needwake = 1;
494 	bd_speedupreq = 1;
495 	bd_request = 1;
496 	if (needwake)
497 		wakeup(&bd_request);
498 	mtx_unlock(&bdlock);
499 }
500 
501 /*
502  * Calculating buffer cache scaling values and reserve space for buffer
503  * headers.  This is called during low level kernel initialization and
504  * may be called more then once.  We CANNOT write to the memory area
505  * being reserved at this time.
506  */
507 caddr_t
508 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
509 {
510 	int tuned_nbuf;
511 	long maxbuf;
512 
513 	/*
514 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
515 	 * PAGE_SIZE is >= 1K)
516 	 */
517 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
518 
519 	/*
520 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
521 	 * For the first 64MB of ram nominally allocate sufficient buffers to
522 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
523 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
524 	 * the buffer cache we limit the eventual kva reservation to
525 	 * maxbcache bytes.
526 	 *
527 	 * factor represents the 1/4 x ram conversion.
528 	 */
529 	if (nbuf == 0) {
530 		int factor = 4 * BKVASIZE / 1024;
531 
532 		nbuf = 50;
533 		if (physmem_est > 4096)
534 			nbuf += min((physmem_est - 4096) / factor,
535 			    65536 / factor);
536 		if (physmem_est > 65536)
537 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
538 
539 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
540 			nbuf = maxbcache / BKVASIZE;
541 		tuned_nbuf = 1;
542 	} else
543 		tuned_nbuf = 0;
544 
545 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
546 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
547 	if (nbuf > maxbuf) {
548 		if (!tuned_nbuf)
549 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
550 			    maxbuf);
551 		nbuf = maxbuf;
552 	}
553 
554 	/*
555 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
556 	 * We have no less then 16 and no more then 256.
557 	 */
558 	nswbuf = max(min(nbuf/4, 256), 16);
559 #ifdef NSWBUF_MIN
560 	if (nswbuf < NSWBUF_MIN)
561 		nswbuf = NSWBUF_MIN;
562 #endif
563 #ifdef DIRECTIO
564 	ffs_rawread_setup();
565 #endif
566 
567 	/*
568 	 * Reserve space for the buffer cache buffers
569 	 */
570 	swbuf = (void *)v;
571 	v = (caddr_t)(swbuf + nswbuf);
572 	buf = (void *)v;
573 	v = (caddr_t)(buf + nbuf);
574 
575 	return(v);
576 }
577 
578 /* Initialize the buffer subsystem.  Called before use of any buffers. */
579 void
580 bufinit(void)
581 {
582 	struct buf *bp;
583 	int i;
584 
585 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
586 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
587 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
588 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
589 
590 	/* next, make a null set of free lists */
591 	for (i = 0; i < BUFFER_QUEUES; i++)
592 		TAILQ_INIT(&bufqueues[i]);
593 
594 	/* finally, initialize each buffer header and stick on empty q */
595 	for (i = 0; i < nbuf; i++) {
596 		bp = &buf[i];
597 		bzero(bp, sizeof *bp);
598 		bp->b_flags = B_INVAL;	/* we're just an empty header */
599 		bp->b_rcred = NOCRED;
600 		bp->b_wcred = NOCRED;
601 		bp->b_qindex = QUEUE_EMPTY;
602 		bp->b_vflags = BV_INFREECNT;	/* buf is counted as free */
603 		bp->b_xflags = 0;
604 		LIST_INIT(&bp->b_dep);
605 		BUF_LOCKINIT(bp);
606 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
607 	}
608 
609 	/*
610 	 * maxbufspace is the absolute maximum amount of buffer space we are
611 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
612 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
613 	 * used by most other processes.  The differential is required to
614 	 * ensure that buf_daemon is able to run when other processes might
615 	 * be blocked waiting for buffer space.
616 	 *
617 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
618 	 * this may result in KVM fragmentation which is not handled optimally
619 	 * by the system.
620 	 */
621 	maxbufspace = (long)nbuf * BKVASIZE;
622 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
623 	lobufspace = hibufspace - MAXBSIZE;
624 
625 	/*
626 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
627 	 * arbitrarily and may need further tuning. It corresponds to
628 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
629 	 * which fits with many RAID controllers' tagged queuing limits.
630 	 * The lower 1 MiB limit is the historical upper limit for
631 	 * hirunningspace.
632 	 */
633 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
634 	    16 * 1024 * 1024), 1024 * 1024);
635 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
636 
637 /*
638  * Limit the amount of malloc memory since it is wired permanently into
639  * the kernel space.  Even though this is accounted for in the buffer
640  * allocation, we don't want the malloced region to grow uncontrolled.
641  * The malloc scheme improves memory utilization significantly on average
642  * (small) directories.
643  */
644 	maxbufmallocspace = hibufspace / 20;
645 
646 /*
647  * Reduce the chance of a deadlock occuring by limiting the number
648  * of delayed-write dirty buffers we allow to stack up.
649  */
650 	hidirtybuffers = nbuf / 4 + 20;
651 	dirtybufthresh = hidirtybuffers * 9 / 10;
652 	numdirtybuffers = 0;
653 /*
654  * To support extreme low-memory systems, make sure hidirtybuffers cannot
655  * eat up all available buffer space.  This occurs when our minimum cannot
656  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
657  * BKVASIZE'd buffers.
658  */
659 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
660 		hidirtybuffers >>= 1;
661 	}
662 	lodirtybuffers = hidirtybuffers / 2;
663 
664 /*
665  * Try to keep the number of free buffers in the specified range,
666  * and give special processes (e.g. like buf_daemon) access to an
667  * emergency reserve.
668  */
669 	lofreebuffers = nbuf / 18 + 5;
670 	hifreebuffers = 2 * lofreebuffers;
671 	numfreebuffers = nbuf;
672 
673 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
674 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
675 }
676 
677 /*
678  * bfreekva() - free the kva allocation for a buffer.
679  *
680  *	Since this call frees up buffer space, we call bufspacewakeup().
681  */
682 static void
683 bfreekva(struct buf *bp)
684 {
685 
686 	if (bp->b_kvasize) {
687 		atomic_add_int(&buffreekvacnt, 1);
688 		atomic_subtract_long(&bufspace, bp->b_kvasize);
689 		vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
690 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
691 		bp->b_kvasize = 0;
692 		bufspacewakeup();
693 	}
694 }
695 
696 /*
697  *	bremfree:
698  *
699  *	Mark the buffer for removal from the appropriate free list in brelse.
700  *
701  */
702 void
703 bremfree(struct buf *bp)
704 {
705 	int old;
706 
707 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
708 	KASSERT((bp->b_flags & B_REMFREE) == 0,
709 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
710 	KASSERT(bp->b_qindex != QUEUE_NONE,
711 	    ("bremfree: buffer %p not on a queue.", bp));
712 	BUF_ASSERT_HELD(bp);
713 
714 	bp->b_flags |= B_REMFREE;
715 	/* Fixup numfreebuffers count.  */
716 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
717 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
718 		    ("buf %p not counted in numfreebuffers", bp));
719 		if (bp->b_bufobj != NULL)
720 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
721 		bp->b_vflags &= ~BV_INFREECNT;
722 		old = atomic_fetchadd_int(&numfreebuffers, -1);
723 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
724 	}
725 }
726 
727 /*
728  *	bremfreef:
729  *
730  *	Force an immediate removal from a free list.  Used only in nfs when
731  *	it abuses the b_freelist pointer.
732  */
733 void
734 bremfreef(struct buf *bp)
735 {
736 	mtx_lock(&bqlock);
737 	bremfreel(bp);
738 	mtx_unlock(&bqlock);
739 }
740 
741 /*
742  *	bremfreel:
743  *
744  *	Removes a buffer from the free list, must be called with the
745  *	bqlock held.
746  */
747 static void
748 bremfreel(struct buf *bp)
749 {
750 	int old;
751 
752 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
753 	    bp, bp->b_vp, bp->b_flags);
754 	KASSERT(bp->b_qindex != QUEUE_NONE,
755 	    ("bremfreel: buffer %p not on a queue.", bp));
756 	BUF_ASSERT_HELD(bp);
757 	mtx_assert(&bqlock, MA_OWNED);
758 
759 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
760 	bp->b_qindex = QUEUE_NONE;
761 	/*
762 	 * If this was a delayed bremfree() we only need to remove the buffer
763 	 * from the queue and return the stats are already done.
764 	 */
765 	if (bp->b_flags & B_REMFREE) {
766 		bp->b_flags &= ~B_REMFREE;
767 		return;
768 	}
769 	/*
770 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
771 	 * delayed-write, the buffer was free and we must decrement
772 	 * numfreebuffers.
773 	 */
774 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
775 		KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
776 		    ("buf %p not counted in numfreebuffers", bp));
777 		if (bp->b_bufobj != NULL)
778 			mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
779 		bp->b_vflags &= ~BV_INFREECNT;
780 		old = atomic_fetchadd_int(&numfreebuffers, -1);
781 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
782 	}
783 }
784 
785 
786 /*
787  * Get a buffer with the specified data.  Look in the cache first.  We
788  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
789  * is set, the buffer is valid and we do not have to do anything ( see
790  * getblk() ).  This is really just a special case of breadn().
791  */
792 int
793 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
794     struct buf **bpp)
795 {
796 
797 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
798 }
799 
800 /*
801  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
802  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
803  * the buffer is valid and we do not have to do anything.
804  */
805 void
806 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
807     int cnt, struct ucred * cred)
808 {
809 	struct buf *rabp;
810 	int i;
811 
812 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
813 		if (inmem(vp, *rablkno))
814 			continue;
815 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
816 
817 		if ((rabp->b_flags & B_CACHE) == 0) {
818 			if (!TD_IS_IDLETHREAD(curthread))
819 				curthread->td_ru.ru_inblock++;
820 			rabp->b_flags |= B_ASYNC;
821 			rabp->b_flags &= ~B_INVAL;
822 			rabp->b_ioflags &= ~BIO_ERROR;
823 			rabp->b_iocmd = BIO_READ;
824 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
825 				rabp->b_rcred = crhold(cred);
826 			vfs_busy_pages(rabp, 0);
827 			BUF_KERNPROC(rabp);
828 			rabp->b_iooffset = dbtob(rabp->b_blkno);
829 			bstrategy(rabp);
830 		} else {
831 			brelse(rabp);
832 		}
833 	}
834 }
835 
836 /*
837  * Operates like bread, but also starts asynchronous I/O on
838  * read-ahead blocks.
839  */
840 int
841 breadn(struct vnode * vp, daddr_t blkno, int size,
842     daddr_t * rablkno, int *rabsize,
843     int cnt, struct ucred * cred, struct buf **bpp)
844 {
845 	struct buf *bp;
846 	int rv = 0, readwait = 0;
847 
848 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
849 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
850 
851 	/* if not found in cache, do some I/O */
852 	if ((bp->b_flags & B_CACHE) == 0) {
853 		if (!TD_IS_IDLETHREAD(curthread))
854 			curthread->td_ru.ru_inblock++;
855 		bp->b_iocmd = BIO_READ;
856 		bp->b_flags &= ~B_INVAL;
857 		bp->b_ioflags &= ~BIO_ERROR;
858 		if (bp->b_rcred == NOCRED && cred != NOCRED)
859 			bp->b_rcred = crhold(cred);
860 		vfs_busy_pages(bp, 0);
861 		bp->b_iooffset = dbtob(bp->b_blkno);
862 		bstrategy(bp);
863 		++readwait;
864 	}
865 
866 	breada(vp, rablkno, rabsize, cnt, cred);
867 
868 	if (readwait) {
869 		rv = bufwait(bp);
870 	}
871 	return (rv);
872 }
873 
874 /*
875  * Write, release buffer on completion.  (Done by iodone
876  * if async).  Do not bother writing anything if the buffer
877  * is invalid.
878  *
879  * Note that we set B_CACHE here, indicating that buffer is
880  * fully valid and thus cacheable.  This is true even of NFS
881  * now so we set it generally.  This could be set either here
882  * or in biodone() since the I/O is synchronous.  We put it
883  * here.
884  */
885 int
886 bufwrite(struct buf *bp)
887 {
888 	int oldflags;
889 	struct vnode *vp;
890 	int vp_md;
891 
892 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
893 	if (bp->b_flags & B_INVAL) {
894 		brelse(bp);
895 		return (0);
896 	}
897 
898 	oldflags = bp->b_flags;
899 
900 	BUF_ASSERT_HELD(bp);
901 
902 	if (bp->b_pin_count > 0)
903 		bunpin_wait(bp);
904 
905 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
906 	    ("FFS background buffer should not get here %p", bp));
907 
908 	vp = bp->b_vp;
909 	if (vp)
910 		vp_md = vp->v_vflag & VV_MD;
911 	else
912 		vp_md = 0;
913 
914 	/* Mark the buffer clean */
915 	bundirty(bp);
916 
917 	bp->b_flags &= ~B_DONE;
918 	bp->b_ioflags &= ~BIO_ERROR;
919 	bp->b_flags |= B_CACHE;
920 	bp->b_iocmd = BIO_WRITE;
921 
922 	bufobj_wref(bp->b_bufobj);
923 	vfs_busy_pages(bp, 1);
924 
925 	/*
926 	 * Normal bwrites pipeline writes
927 	 */
928 	bp->b_runningbufspace = bp->b_bufsize;
929 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
930 
931 	if (!TD_IS_IDLETHREAD(curthread))
932 		curthread->td_ru.ru_oublock++;
933 	if (oldflags & B_ASYNC)
934 		BUF_KERNPROC(bp);
935 	bp->b_iooffset = dbtob(bp->b_blkno);
936 	bstrategy(bp);
937 
938 	if ((oldflags & B_ASYNC) == 0) {
939 		int rtval = bufwait(bp);
940 		brelse(bp);
941 		return (rtval);
942 	} else {
943 		/*
944 		 * don't allow the async write to saturate the I/O
945 		 * system.  We will not deadlock here because
946 		 * we are blocking waiting for I/O that is already in-progress
947 		 * to complete. We do not block here if it is the update
948 		 * or syncer daemon trying to clean up as that can lead
949 		 * to deadlock.
950 		 */
951 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
952 			waitrunningbufspace();
953 	}
954 
955 	return (0);
956 }
957 
958 void
959 bufbdflush(struct bufobj *bo, struct buf *bp)
960 {
961 	struct buf *nbp;
962 
963 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
964 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
965 		altbufferflushes++;
966 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
967 		BO_LOCK(bo);
968 		/*
969 		 * Try to find a buffer to flush.
970 		 */
971 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
972 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
973 			    BUF_LOCK(nbp,
974 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
975 				continue;
976 			if (bp == nbp)
977 				panic("bdwrite: found ourselves");
978 			BO_UNLOCK(bo);
979 			/* Don't countdeps with the bo lock held. */
980 			if (buf_countdeps(nbp, 0)) {
981 				BO_LOCK(bo);
982 				BUF_UNLOCK(nbp);
983 				continue;
984 			}
985 			if (nbp->b_flags & B_CLUSTEROK) {
986 				vfs_bio_awrite(nbp);
987 			} else {
988 				bremfree(nbp);
989 				bawrite(nbp);
990 			}
991 			dirtybufferflushes++;
992 			break;
993 		}
994 		if (nbp == NULL)
995 			BO_UNLOCK(bo);
996 	}
997 }
998 
999 /*
1000  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1001  * anything if the buffer is marked invalid.
1002  *
1003  * Note that since the buffer must be completely valid, we can safely
1004  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1005  * biodone() in order to prevent getblk from writing the buffer
1006  * out synchronously.
1007  */
1008 void
1009 bdwrite(struct buf *bp)
1010 {
1011 	struct thread *td = curthread;
1012 	struct vnode *vp;
1013 	struct bufobj *bo;
1014 
1015 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1016 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1017 	BUF_ASSERT_HELD(bp);
1018 
1019 	if (bp->b_flags & B_INVAL) {
1020 		brelse(bp);
1021 		return;
1022 	}
1023 
1024 	/*
1025 	 * If we have too many dirty buffers, don't create any more.
1026 	 * If we are wildly over our limit, then force a complete
1027 	 * cleanup. Otherwise, just keep the situation from getting
1028 	 * out of control. Note that we have to avoid a recursive
1029 	 * disaster and not try to clean up after our own cleanup!
1030 	 */
1031 	vp = bp->b_vp;
1032 	bo = bp->b_bufobj;
1033 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1034 		td->td_pflags |= TDP_INBDFLUSH;
1035 		BO_BDFLUSH(bo, bp);
1036 		td->td_pflags &= ~TDP_INBDFLUSH;
1037 	} else
1038 		recursiveflushes++;
1039 
1040 	bdirty(bp);
1041 	/*
1042 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1043 	 * true even of NFS now.
1044 	 */
1045 	bp->b_flags |= B_CACHE;
1046 
1047 	/*
1048 	 * This bmap keeps the system from needing to do the bmap later,
1049 	 * perhaps when the system is attempting to do a sync.  Since it
1050 	 * is likely that the indirect block -- or whatever other datastructure
1051 	 * that the filesystem needs is still in memory now, it is a good
1052 	 * thing to do this.  Note also, that if the pageout daemon is
1053 	 * requesting a sync -- there might not be enough memory to do
1054 	 * the bmap then...  So, this is important to do.
1055 	 */
1056 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1057 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1058 	}
1059 
1060 	/*
1061 	 * Set the *dirty* buffer range based upon the VM system dirty
1062 	 * pages.
1063 	 *
1064 	 * Mark the buffer pages as clean.  We need to do this here to
1065 	 * satisfy the vnode_pager and the pageout daemon, so that it
1066 	 * thinks that the pages have been "cleaned".  Note that since
1067 	 * the pages are in a delayed write buffer -- the VFS layer
1068 	 * "will" see that the pages get written out on the next sync,
1069 	 * or perhaps the cluster will be completed.
1070 	 */
1071 	vfs_clean_pages_dirty_buf(bp);
1072 	bqrelse(bp);
1073 
1074 	/*
1075 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
1076 	 * buffers (midpoint between our recovery point and our stall
1077 	 * point).
1078 	 */
1079 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1080 
1081 	/*
1082 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1083 	 * due to the softdep code.
1084 	 */
1085 }
1086 
1087 /*
1088  *	bdirty:
1089  *
1090  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1091  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1092  *	itself to properly update it in the dirty/clean lists.  We mark it
1093  *	B_DONE to ensure that any asynchronization of the buffer properly
1094  *	clears B_DONE ( else a panic will occur later ).
1095  *
1096  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1097  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1098  *	should only be called if the buffer is known-good.
1099  *
1100  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1101  *	count.
1102  *
1103  *	The buffer must be on QUEUE_NONE.
1104  */
1105 void
1106 bdirty(struct buf *bp)
1107 {
1108 
1109 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1110 	    bp, bp->b_vp, bp->b_flags);
1111 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1112 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1113 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1114 	BUF_ASSERT_HELD(bp);
1115 	bp->b_flags &= ~(B_RELBUF);
1116 	bp->b_iocmd = BIO_WRITE;
1117 
1118 	if ((bp->b_flags & B_DELWRI) == 0) {
1119 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1120 		reassignbuf(bp);
1121 		atomic_add_int(&numdirtybuffers, 1);
1122 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1123 	}
1124 }
1125 
1126 /*
1127  *	bundirty:
1128  *
1129  *	Clear B_DELWRI for buffer.
1130  *
1131  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1132  *	count.
1133  *
1134  *	The buffer must be on QUEUE_NONE.
1135  */
1136 
1137 void
1138 bundirty(struct buf *bp)
1139 {
1140 
1141 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1142 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1143 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1144 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1145 	BUF_ASSERT_HELD(bp);
1146 
1147 	if (bp->b_flags & B_DELWRI) {
1148 		bp->b_flags &= ~B_DELWRI;
1149 		reassignbuf(bp);
1150 		atomic_subtract_int(&numdirtybuffers, 1);
1151 		numdirtywakeup(lodirtybuffers);
1152 	}
1153 	/*
1154 	 * Since it is now being written, we can clear its deferred write flag.
1155 	 */
1156 	bp->b_flags &= ~B_DEFERRED;
1157 }
1158 
1159 /*
1160  *	bawrite:
1161  *
1162  *	Asynchronous write.  Start output on a buffer, but do not wait for
1163  *	it to complete.  The buffer is released when the output completes.
1164  *
1165  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1166  *	B_INVAL buffers.  Not us.
1167  */
1168 void
1169 bawrite(struct buf *bp)
1170 {
1171 
1172 	bp->b_flags |= B_ASYNC;
1173 	(void) bwrite(bp);
1174 }
1175 
1176 /*
1177  *	bwillwrite:
1178  *
1179  *	Called prior to the locking of any vnodes when we are expecting to
1180  *	write.  We do not want to starve the buffer cache with too many
1181  *	dirty buffers so we block here.  By blocking prior to the locking
1182  *	of any vnodes we attempt to avoid the situation where a locked vnode
1183  *	prevents the various system daemons from flushing related buffers.
1184  */
1185 
1186 void
1187 bwillwrite(void)
1188 {
1189 
1190 	if (numdirtybuffers >= hidirtybuffers) {
1191 		mtx_lock(&nblock);
1192 		while (numdirtybuffers >= hidirtybuffers) {
1193 			bd_wakeup(1);
1194 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1195 			msleep(&needsbuffer, &nblock,
1196 			    (PRIBIO + 4), "flswai", 0);
1197 		}
1198 		mtx_unlock(&nblock);
1199 	}
1200 }
1201 
1202 /*
1203  * Return true if we have too many dirty buffers.
1204  */
1205 int
1206 buf_dirty_count_severe(void)
1207 {
1208 
1209 	return(numdirtybuffers >= hidirtybuffers);
1210 }
1211 
1212 static __noinline int
1213 buf_vm_page_count_severe(void)
1214 {
1215 
1216 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1217 
1218 	return vm_page_count_severe();
1219 }
1220 
1221 /*
1222  *	brelse:
1223  *
1224  *	Release a busy buffer and, if requested, free its resources.  The
1225  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1226  *	to be accessed later as a cache entity or reused for other purposes.
1227  */
1228 void
1229 brelse(struct buf *bp)
1230 {
1231 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1232 	    bp, bp->b_vp, bp->b_flags);
1233 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1234 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1235 
1236 	if (bp->b_flags & B_MANAGED) {
1237 		bqrelse(bp);
1238 		return;
1239 	}
1240 
1241 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1242 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1243 		/*
1244 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1245 		 * pages from being scrapped.  If the error is anything
1246 		 * other than an I/O error (EIO), assume that retrying
1247 		 * is futile.
1248 		 */
1249 		bp->b_ioflags &= ~BIO_ERROR;
1250 		bdirty(bp);
1251 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1252 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1253 		/*
1254 		 * Either a failed I/O or we were asked to free or not
1255 		 * cache the buffer.
1256 		 */
1257 		bp->b_flags |= B_INVAL;
1258 		if (!LIST_EMPTY(&bp->b_dep))
1259 			buf_deallocate(bp);
1260 		if (bp->b_flags & B_DELWRI) {
1261 			atomic_subtract_int(&numdirtybuffers, 1);
1262 			numdirtywakeup(lodirtybuffers);
1263 		}
1264 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1265 		if ((bp->b_flags & B_VMIO) == 0) {
1266 			if (bp->b_bufsize)
1267 				allocbuf(bp, 0);
1268 			if (bp->b_vp)
1269 				brelvp(bp);
1270 		}
1271 	}
1272 
1273 	/*
1274 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1275 	 * is called with B_DELWRI set, the underlying pages may wind up
1276 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1277 	 * because pages associated with a B_DELWRI bp are marked clean.
1278 	 *
1279 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1280 	 * if B_DELWRI is set.
1281 	 *
1282 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1283 	 * on pages to return pages to the VM page queues.
1284 	 */
1285 	if (bp->b_flags & B_DELWRI)
1286 		bp->b_flags &= ~B_RELBUF;
1287 	else if (buf_vm_page_count_severe()) {
1288 		/*
1289 		 * The locking of the BO_LOCK is not necessary since
1290 		 * BKGRDINPROG cannot be set while we hold the buf
1291 		 * lock, it can only be cleared if it is already
1292 		 * pending.
1293 		 */
1294 		if (bp->b_vp) {
1295 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1296 				bp->b_flags |= B_RELBUF;
1297 		} else
1298 			bp->b_flags |= B_RELBUF;
1299 	}
1300 
1301 	/*
1302 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1303 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1304 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1305 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1306 	 *
1307 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1308 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1309 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1310 	 *
1311 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1312 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1313 	 * the commit state and we cannot afford to lose the buffer. If the
1314 	 * buffer has a background write in progress, we need to keep it
1315 	 * around to prevent it from being reconstituted and starting a second
1316 	 * background write.
1317 	 */
1318 	if ((bp->b_flags & B_VMIO)
1319 	    && !(bp->b_vp->v_mount != NULL &&
1320 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1321 		 !vn_isdisk(bp->b_vp, NULL) &&
1322 		 (bp->b_flags & B_DELWRI))
1323 	    ) {
1324 
1325 		int i, j, resid;
1326 		vm_page_t m;
1327 		off_t foff;
1328 		vm_pindex_t poff;
1329 		vm_object_t obj;
1330 
1331 		obj = bp->b_bufobj->bo_object;
1332 
1333 		/*
1334 		 * Get the base offset and length of the buffer.  Note that
1335 		 * in the VMIO case if the buffer block size is not
1336 		 * page-aligned then b_data pointer may not be page-aligned.
1337 		 * But our b_pages[] array *IS* page aligned.
1338 		 *
1339 		 * block sizes less then DEV_BSIZE (usually 512) are not
1340 		 * supported due to the page granularity bits (m->valid,
1341 		 * m->dirty, etc...).
1342 		 *
1343 		 * See man buf(9) for more information
1344 		 */
1345 		resid = bp->b_bufsize;
1346 		foff = bp->b_offset;
1347 		VM_OBJECT_LOCK(obj);
1348 		for (i = 0; i < bp->b_npages; i++) {
1349 			int had_bogus = 0;
1350 
1351 			m = bp->b_pages[i];
1352 
1353 			/*
1354 			 * If we hit a bogus page, fixup *all* the bogus pages
1355 			 * now.
1356 			 */
1357 			if (m == bogus_page) {
1358 				poff = OFF_TO_IDX(bp->b_offset);
1359 				had_bogus = 1;
1360 
1361 				for (j = i; j < bp->b_npages; j++) {
1362 					vm_page_t mtmp;
1363 					mtmp = bp->b_pages[j];
1364 					if (mtmp == bogus_page) {
1365 						mtmp = vm_page_lookup(obj, poff + j);
1366 						if (!mtmp) {
1367 							panic("brelse: page missing\n");
1368 						}
1369 						bp->b_pages[j] = mtmp;
1370 					}
1371 				}
1372 
1373 				if ((bp->b_flags & B_INVAL) == 0) {
1374 					pmap_qenter(
1375 					    trunc_page((vm_offset_t)bp->b_data),
1376 					    bp->b_pages, bp->b_npages);
1377 				}
1378 				m = bp->b_pages[i];
1379 			}
1380 			if ((bp->b_flags & B_NOCACHE) ||
1381 			    (bp->b_ioflags & BIO_ERROR &&
1382 			     bp->b_iocmd == BIO_READ)) {
1383 				int poffset = foff & PAGE_MASK;
1384 				int presid = resid > (PAGE_SIZE - poffset) ?
1385 					(PAGE_SIZE - poffset) : resid;
1386 
1387 				KASSERT(presid >= 0, ("brelse: extra page"));
1388 				vm_page_set_invalid(m, poffset, presid);
1389 				if (had_bogus)
1390 					printf("avoided corruption bug in bogus_page/brelse code\n");
1391 			}
1392 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1393 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1394 		}
1395 		VM_OBJECT_UNLOCK(obj);
1396 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1397 			vfs_vmio_release(bp);
1398 
1399 	} else if (bp->b_flags & B_VMIO) {
1400 
1401 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1402 			vfs_vmio_release(bp);
1403 		}
1404 
1405 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1406 		if (bp->b_bufsize != 0)
1407 			allocbuf(bp, 0);
1408 		if (bp->b_vp != NULL)
1409 			brelvp(bp);
1410 	}
1411 
1412 	if (BUF_LOCKRECURSED(bp)) {
1413 		/* do not release to free list */
1414 		BUF_UNLOCK(bp);
1415 		return;
1416 	}
1417 
1418 	/* enqueue */
1419 	mtx_lock(&bqlock);
1420 	/* Handle delayed bremfree() processing. */
1421 	if (bp->b_flags & B_REMFREE) {
1422 		struct bufobj *bo;
1423 
1424 		bo = bp->b_bufobj;
1425 		if (bo != NULL)
1426 			BO_LOCK(bo);
1427 		bremfreel(bp);
1428 		if (bo != NULL)
1429 			BO_UNLOCK(bo);
1430 	}
1431 	if (bp->b_qindex != QUEUE_NONE)
1432 		panic("brelse: free buffer onto another queue???");
1433 
1434 	/*
1435 	 * If the buffer has junk contents signal it and eventually
1436 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1437 	 * doesn't find it.
1438 	 */
1439 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1440 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1441 		bp->b_flags |= B_INVAL;
1442 	if (bp->b_flags & B_INVAL) {
1443 		if (bp->b_flags & B_DELWRI)
1444 			bundirty(bp);
1445 		if (bp->b_vp)
1446 			brelvp(bp);
1447 	}
1448 
1449 	/* buffers with no memory */
1450 	if (bp->b_bufsize == 0) {
1451 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1452 		if (bp->b_vflags & BV_BKGRDINPROG)
1453 			panic("losing buffer 1");
1454 		if (bp->b_kvasize) {
1455 			bp->b_qindex = QUEUE_EMPTYKVA;
1456 		} else {
1457 			bp->b_qindex = QUEUE_EMPTY;
1458 		}
1459 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1460 	/* buffers with junk contents */
1461 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1462 	    (bp->b_ioflags & BIO_ERROR)) {
1463 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1464 		if (bp->b_vflags & BV_BKGRDINPROG)
1465 			panic("losing buffer 2");
1466 		bp->b_qindex = QUEUE_CLEAN;
1467 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1468 	/* remaining buffers */
1469 	} else {
1470 		if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1471 		    (B_DELWRI|B_NEEDSGIANT))
1472 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1473 		else if (bp->b_flags & B_DELWRI)
1474 			bp->b_qindex = QUEUE_DIRTY;
1475 		else
1476 			bp->b_qindex = QUEUE_CLEAN;
1477 		if (bp->b_flags & B_AGE)
1478 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1479 		else
1480 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1481 	}
1482 	mtx_unlock(&bqlock);
1483 
1484 	/*
1485 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1486 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1487 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1488 	 * if B_INVAL is set ).
1489 	 */
1490 
1491 	if (!(bp->b_flags & B_DELWRI)) {
1492 		struct bufobj *bo;
1493 
1494 		bo = bp->b_bufobj;
1495 		if (bo != NULL)
1496 			BO_LOCK(bo);
1497 		bufcountwakeup(bp);
1498 		if (bo != NULL)
1499 			BO_UNLOCK(bo);
1500 	}
1501 
1502 	/*
1503 	 * Something we can maybe free or reuse
1504 	 */
1505 	if (bp->b_bufsize || bp->b_kvasize)
1506 		bufspacewakeup();
1507 
1508 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1509 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1510 		panic("brelse: not dirty");
1511 	/* unlock */
1512 	BUF_UNLOCK(bp);
1513 }
1514 
1515 /*
1516  * Release a buffer back to the appropriate queue but do not try to free
1517  * it.  The buffer is expected to be used again soon.
1518  *
1519  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1520  * biodone() to requeue an async I/O on completion.  It is also used when
1521  * known good buffers need to be requeued but we think we may need the data
1522  * again soon.
1523  *
1524  * XXX we should be able to leave the B_RELBUF hint set on completion.
1525  */
1526 void
1527 bqrelse(struct buf *bp)
1528 {
1529 	struct bufobj *bo;
1530 
1531 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1532 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1533 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1534 
1535 	if (BUF_LOCKRECURSED(bp)) {
1536 		/* do not release to free list */
1537 		BUF_UNLOCK(bp);
1538 		return;
1539 	}
1540 
1541 	bo = bp->b_bufobj;
1542 	if (bp->b_flags & B_MANAGED) {
1543 		if (bp->b_flags & B_REMFREE) {
1544 			mtx_lock(&bqlock);
1545 			if (bo != NULL)
1546 				BO_LOCK(bo);
1547 			bremfreel(bp);
1548 			if (bo != NULL)
1549 				BO_UNLOCK(bo);
1550 			mtx_unlock(&bqlock);
1551 		}
1552 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1553 		BUF_UNLOCK(bp);
1554 		return;
1555 	}
1556 
1557 	mtx_lock(&bqlock);
1558 	/* Handle delayed bremfree() processing. */
1559 	if (bp->b_flags & B_REMFREE) {
1560 		if (bo != NULL)
1561 			BO_LOCK(bo);
1562 		bremfreel(bp);
1563 		if (bo != NULL)
1564 			BO_UNLOCK(bo);
1565 	}
1566 	if (bp->b_qindex != QUEUE_NONE)
1567 		panic("bqrelse: free buffer onto another queue???");
1568 	/* buffers with stale but valid contents */
1569 	if (bp->b_flags & B_DELWRI) {
1570 		if (bp->b_flags & B_NEEDSGIANT)
1571 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1572 		else
1573 			bp->b_qindex = QUEUE_DIRTY;
1574 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1575 	} else {
1576 		/*
1577 		 * The locking of the BO_LOCK for checking of the
1578 		 * BV_BKGRDINPROG is not necessary since the
1579 		 * BV_BKGRDINPROG cannot be set while we hold the buf
1580 		 * lock, it can only be cleared if it is already
1581 		 * pending.
1582 		 */
1583 		if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1584 			bp->b_qindex = QUEUE_CLEAN;
1585 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1586 			    b_freelist);
1587 		} else {
1588 			/*
1589 			 * We are too low on memory, we have to try to free
1590 			 * the buffer (most importantly: the wired pages
1591 			 * making up its backing store) *now*.
1592 			 */
1593 			mtx_unlock(&bqlock);
1594 			brelse(bp);
1595 			return;
1596 		}
1597 	}
1598 	mtx_unlock(&bqlock);
1599 
1600 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) {
1601 		if (bo != NULL)
1602 			BO_LOCK(bo);
1603 		bufcountwakeup(bp);
1604 		if (bo != NULL)
1605 			BO_UNLOCK(bo);
1606 	}
1607 
1608 	/*
1609 	 * Something we can maybe free or reuse.
1610 	 */
1611 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1612 		bufspacewakeup();
1613 
1614 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1615 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1616 		panic("bqrelse: not dirty");
1617 	/* unlock */
1618 	BUF_UNLOCK(bp);
1619 }
1620 
1621 /* Give pages used by the bp back to the VM system (where possible) */
1622 static void
1623 vfs_vmio_release(struct buf *bp)
1624 {
1625 	int i;
1626 	vm_page_t m;
1627 
1628 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1629 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1630 	for (i = 0; i < bp->b_npages; i++) {
1631 		m = bp->b_pages[i];
1632 		bp->b_pages[i] = NULL;
1633 		/*
1634 		 * In order to keep page LRU ordering consistent, put
1635 		 * everything on the inactive queue.
1636 		 */
1637 		vm_page_lock(m);
1638 		vm_page_unwire(m, 0);
1639 		/*
1640 		 * We don't mess with busy pages, it is
1641 		 * the responsibility of the process that
1642 		 * busied the pages to deal with them.
1643 		 */
1644 		if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 &&
1645 		    m->wire_count == 0) {
1646 			/*
1647 			 * Might as well free the page if we can and it has
1648 			 * no valid data.  We also free the page if the
1649 			 * buffer was used for direct I/O
1650 			 */
1651 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1652 				vm_page_free(m);
1653 			} else if (bp->b_flags & B_DIRECT) {
1654 				vm_page_try_to_free(m);
1655 			} else if (buf_vm_page_count_severe()) {
1656 				vm_page_try_to_cache(m);
1657 			}
1658 		}
1659 		vm_page_unlock(m);
1660 	}
1661 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1662 
1663 	if (bp->b_bufsize) {
1664 		bufspacewakeup();
1665 		bp->b_bufsize = 0;
1666 	}
1667 	bp->b_npages = 0;
1668 	bp->b_flags &= ~B_VMIO;
1669 	if (bp->b_vp)
1670 		brelvp(bp);
1671 }
1672 
1673 /*
1674  * Check to see if a block at a particular lbn is available for a clustered
1675  * write.
1676  */
1677 static int
1678 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1679 {
1680 	struct buf *bpa;
1681 	int match;
1682 
1683 	match = 0;
1684 
1685 	/* If the buf isn't in core skip it */
1686 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1687 		return (0);
1688 
1689 	/* If the buf is busy we don't want to wait for it */
1690 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1691 		return (0);
1692 
1693 	/* Only cluster with valid clusterable delayed write buffers */
1694 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1695 	    (B_DELWRI | B_CLUSTEROK))
1696 		goto done;
1697 
1698 	if (bpa->b_bufsize != size)
1699 		goto done;
1700 
1701 	/*
1702 	 * Check to see if it is in the expected place on disk and that the
1703 	 * block has been mapped.
1704 	 */
1705 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1706 		match = 1;
1707 done:
1708 	BUF_UNLOCK(bpa);
1709 	return (match);
1710 }
1711 
1712 /*
1713  *	vfs_bio_awrite:
1714  *
1715  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1716  *	This is much better then the old way of writing only one buffer at
1717  *	a time.  Note that we may not be presented with the buffers in the
1718  *	correct order, so we search for the cluster in both directions.
1719  */
1720 int
1721 vfs_bio_awrite(struct buf *bp)
1722 {
1723 	struct bufobj *bo;
1724 	int i;
1725 	int j;
1726 	daddr_t lblkno = bp->b_lblkno;
1727 	struct vnode *vp = bp->b_vp;
1728 	int ncl;
1729 	int nwritten;
1730 	int size;
1731 	int maxcl;
1732 
1733 	bo = &vp->v_bufobj;
1734 	/*
1735 	 * right now we support clustered writing only to regular files.  If
1736 	 * we find a clusterable block we could be in the middle of a cluster
1737 	 * rather then at the beginning.
1738 	 */
1739 	if ((vp->v_type == VREG) &&
1740 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1741 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1742 
1743 		size = vp->v_mount->mnt_stat.f_iosize;
1744 		maxcl = MAXPHYS / size;
1745 
1746 		BO_LOCK(bo);
1747 		for (i = 1; i < maxcl; i++)
1748 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1749 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1750 				break;
1751 
1752 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1753 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1754 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1755 				break;
1756 		BO_UNLOCK(bo);
1757 		--j;
1758 		ncl = i + j;
1759 		/*
1760 		 * this is a possible cluster write
1761 		 */
1762 		if (ncl != 1) {
1763 			BUF_UNLOCK(bp);
1764 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1765 			return nwritten;
1766 		}
1767 	}
1768 	bremfree(bp);
1769 	bp->b_flags |= B_ASYNC;
1770 	/*
1771 	 * default (old) behavior, writing out only one block
1772 	 *
1773 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1774 	 */
1775 	nwritten = bp->b_bufsize;
1776 	(void) bwrite(bp);
1777 
1778 	return nwritten;
1779 }
1780 
1781 /*
1782  *	getnewbuf:
1783  *
1784  *	Find and initialize a new buffer header, freeing up existing buffers
1785  *	in the bufqueues as necessary.  The new buffer is returned locked.
1786  *
1787  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1788  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1789  *
1790  *	We block if:
1791  *		We have insufficient buffer headers
1792  *		We have insufficient buffer space
1793  *		buffer_map is too fragmented ( space reservation fails )
1794  *		If we have to flush dirty buffers ( but we try to avoid this )
1795  *
1796  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1797  *	Instead we ask the buf daemon to do it for us.  We attempt to
1798  *	avoid piecemeal wakeups of the pageout daemon.
1799  */
1800 
1801 static struct buf *
1802 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1803     int gbflags)
1804 {
1805 	struct thread *td;
1806 	struct buf *bp;
1807 	struct buf *nbp;
1808 	int defrag = 0;
1809 	int nqindex;
1810 	static int flushingbufs;
1811 
1812 	td = curthread;
1813 	/*
1814 	 * We can't afford to block since we might be holding a vnode lock,
1815 	 * which may prevent system daemons from running.  We deal with
1816 	 * low-memory situations by proactively returning memory and running
1817 	 * async I/O rather then sync I/O.
1818 	 */
1819 	atomic_add_int(&getnewbufcalls, 1);
1820 	atomic_subtract_int(&getnewbufrestarts, 1);
1821 restart:
1822 	atomic_add_int(&getnewbufrestarts, 1);
1823 
1824 	/*
1825 	 * Setup for scan.  If we do not have enough free buffers,
1826 	 * we setup a degenerate case that immediately fails.  Note
1827 	 * that if we are specially marked process, we are allowed to
1828 	 * dip into our reserves.
1829 	 *
1830 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1831 	 *
1832 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1833 	 * However, there are a number of cases (defragging, reusing, ...)
1834 	 * where we cannot backup.
1835 	 */
1836 	mtx_lock(&bqlock);
1837 	nqindex = QUEUE_EMPTYKVA;
1838 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1839 
1840 	if (nbp == NULL) {
1841 		/*
1842 		 * If no EMPTYKVA buffers and we are either
1843 		 * defragging or reusing, locate a CLEAN buffer
1844 		 * to free or reuse.  If bufspace useage is low
1845 		 * skip this step so we can allocate a new buffer.
1846 		 */
1847 		if (defrag || bufspace >= lobufspace) {
1848 			nqindex = QUEUE_CLEAN;
1849 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1850 		}
1851 
1852 		/*
1853 		 * If we could not find or were not allowed to reuse a
1854 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1855 		 * buffer.  We can only use an EMPTY buffer if allocating
1856 		 * its KVA would not otherwise run us out of buffer space.
1857 		 */
1858 		if (nbp == NULL && defrag == 0 &&
1859 		    bufspace + maxsize < hibufspace) {
1860 			nqindex = QUEUE_EMPTY;
1861 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1862 		}
1863 	}
1864 
1865 	/*
1866 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1867 	 * depending.
1868 	 */
1869 
1870 	while ((bp = nbp) != NULL) {
1871 		int qindex = nqindex;
1872 
1873 		/*
1874 		 * Calculate next bp ( we can only use it if we do not block
1875 		 * or do other fancy things ).
1876 		 */
1877 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1878 			switch(qindex) {
1879 			case QUEUE_EMPTY:
1880 				nqindex = QUEUE_EMPTYKVA;
1881 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1882 					break;
1883 				/* FALLTHROUGH */
1884 			case QUEUE_EMPTYKVA:
1885 				nqindex = QUEUE_CLEAN;
1886 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1887 					break;
1888 				/* FALLTHROUGH */
1889 			case QUEUE_CLEAN:
1890 				/*
1891 				 * nbp is NULL.
1892 				 */
1893 				break;
1894 			}
1895 		}
1896 		/*
1897 		 * If we are defragging then we need a buffer with
1898 		 * b_kvasize != 0.  XXX this situation should no longer
1899 		 * occur, if defrag is non-zero the buffer's b_kvasize
1900 		 * should also be non-zero at this point.  XXX
1901 		 */
1902 		if (defrag && bp->b_kvasize == 0) {
1903 			printf("Warning: defrag empty buffer %p\n", bp);
1904 			continue;
1905 		}
1906 
1907 		/*
1908 		 * Start freeing the bp.  This is somewhat involved.  nbp
1909 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1910 		 */
1911 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1912 			continue;
1913 		if (bp->b_vp) {
1914 			BO_LOCK(bp->b_bufobj);
1915 			if (bp->b_vflags & BV_BKGRDINPROG) {
1916 				BO_UNLOCK(bp->b_bufobj);
1917 				BUF_UNLOCK(bp);
1918 				continue;
1919 			}
1920 			BO_UNLOCK(bp->b_bufobj);
1921 		}
1922 		CTR6(KTR_BUF,
1923 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1924 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1925 		    bp->b_kvasize, bp->b_bufsize, qindex);
1926 
1927 		/*
1928 		 * Sanity Checks
1929 		 */
1930 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1931 
1932 		/*
1933 		 * Note: we no longer distinguish between VMIO and non-VMIO
1934 		 * buffers.
1935 		 */
1936 
1937 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1938 
1939 		if (bp->b_bufobj != NULL)
1940 			BO_LOCK(bp->b_bufobj);
1941 		bremfreel(bp);
1942 		if (bp->b_bufobj != NULL)
1943 			BO_UNLOCK(bp->b_bufobj);
1944 		mtx_unlock(&bqlock);
1945 
1946 		if (qindex == QUEUE_CLEAN) {
1947 			if (bp->b_flags & B_VMIO) {
1948 				bp->b_flags &= ~B_ASYNC;
1949 				vfs_vmio_release(bp);
1950 			}
1951 			if (bp->b_vp)
1952 				brelvp(bp);
1953 		}
1954 
1955 		/*
1956 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1957 		 * the scan from this point on.
1958 		 *
1959 		 * Get the rest of the buffer freed up.  b_kva* is still
1960 		 * valid after this operation.
1961 		 */
1962 
1963 		if (bp->b_rcred != NOCRED) {
1964 			crfree(bp->b_rcred);
1965 			bp->b_rcred = NOCRED;
1966 		}
1967 		if (bp->b_wcred != NOCRED) {
1968 			crfree(bp->b_wcred);
1969 			bp->b_wcred = NOCRED;
1970 		}
1971 		if (!LIST_EMPTY(&bp->b_dep))
1972 			buf_deallocate(bp);
1973 		if (bp->b_vflags & BV_BKGRDINPROG)
1974 			panic("losing buffer 3");
1975 		KASSERT(bp->b_vp == NULL,
1976 		    ("bp: %p still has vnode %p.  qindex: %d",
1977 		    bp, bp->b_vp, qindex));
1978 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1979 		   ("bp: %p still on a buffer list. xflags %X",
1980 		    bp, bp->b_xflags));
1981 
1982 		if (bp->b_bufsize)
1983 			allocbuf(bp, 0);
1984 
1985 		bp->b_flags = 0;
1986 		bp->b_ioflags = 0;
1987 		bp->b_xflags = 0;
1988 		KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
1989 		    ("buf %p still counted as free?", bp));
1990 		bp->b_vflags = 0;
1991 		bp->b_vp = NULL;
1992 		bp->b_blkno = bp->b_lblkno = 0;
1993 		bp->b_offset = NOOFFSET;
1994 		bp->b_iodone = 0;
1995 		bp->b_error = 0;
1996 		bp->b_resid = 0;
1997 		bp->b_bcount = 0;
1998 		bp->b_npages = 0;
1999 		bp->b_dirtyoff = bp->b_dirtyend = 0;
2000 		bp->b_bufobj = NULL;
2001 		bp->b_pin_count = 0;
2002 		bp->b_fsprivate1 = NULL;
2003 		bp->b_fsprivate2 = NULL;
2004 		bp->b_fsprivate3 = NULL;
2005 
2006 		LIST_INIT(&bp->b_dep);
2007 
2008 		/*
2009 		 * If we are defragging then free the buffer.
2010 		 */
2011 		if (defrag) {
2012 			bp->b_flags |= B_INVAL;
2013 			bfreekva(bp);
2014 			brelse(bp);
2015 			defrag = 0;
2016 			goto restart;
2017 		}
2018 
2019 		/*
2020 		 * Notify any waiters for the buffer lock about
2021 		 * identity change by freeing the buffer.
2022 		 */
2023 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2024 			bp->b_flags |= B_INVAL;
2025 			bfreekva(bp);
2026 			brelse(bp);
2027 			goto restart;
2028 		}
2029 
2030 		/*
2031 		 * If we are overcomitted then recover the buffer and its
2032 		 * KVM space.  This occurs in rare situations when multiple
2033 		 * processes are blocked in getnewbuf() or allocbuf().
2034 		 */
2035 		if (bufspace >= hibufspace)
2036 			flushingbufs = 1;
2037 		if (flushingbufs && bp->b_kvasize != 0) {
2038 			bp->b_flags |= B_INVAL;
2039 			bfreekva(bp);
2040 			brelse(bp);
2041 			goto restart;
2042 		}
2043 		if (bufspace < lobufspace)
2044 			flushingbufs = 0;
2045 		break;
2046 	}
2047 
2048 	/*
2049 	 * If we exhausted our list, sleep as appropriate.  We may have to
2050 	 * wakeup various daemons and write out some dirty buffers.
2051 	 *
2052 	 * Generally we are sleeping due to insufficient buffer space.
2053 	 */
2054 
2055 	if (bp == NULL) {
2056 		int flags, norunbuf;
2057 		char *waitmsg;
2058 		int fl;
2059 
2060 		if (defrag) {
2061 			flags = VFS_BIO_NEED_BUFSPACE;
2062 			waitmsg = "nbufkv";
2063 		} else if (bufspace >= hibufspace) {
2064 			waitmsg = "nbufbs";
2065 			flags = VFS_BIO_NEED_BUFSPACE;
2066 		} else {
2067 			waitmsg = "newbuf";
2068 			flags = VFS_BIO_NEED_ANY;
2069 		}
2070 		mtx_lock(&nblock);
2071 		needsbuffer |= flags;
2072 		mtx_unlock(&nblock);
2073 		mtx_unlock(&bqlock);
2074 
2075 		bd_speedup();	/* heeeelp */
2076 		if (gbflags & GB_NOWAIT_BD)
2077 			return (NULL);
2078 
2079 		mtx_lock(&nblock);
2080 		while (needsbuffer & flags) {
2081 			if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2082 				mtx_unlock(&nblock);
2083 				/*
2084 				 * getblk() is called with a vnode
2085 				 * locked, and some majority of the
2086 				 * dirty buffers may as well belong to
2087 				 * the vnode. Flushing the buffers
2088 				 * there would make a progress that
2089 				 * cannot be achieved by the
2090 				 * buf_daemon, that cannot lock the
2091 				 * vnode.
2092 				 */
2093 				norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2094 				    (td->td_pflags & TDP_NORUNNINGBUF);
2095 				/* play bufdaemon */
2096 				td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2097 				fl = buf_do_flush(vp);
2098 				td->td_pflags &= norunbuf;
2099 				mtx_lock(&nblock);
2100 				if (fl != 0)
2101 					continue;
2102 				if ((needsbuffer & flags) == 0)
2103 					break;
2104 			}
2105 			if (msleep(&needsbuffer, &nblock,
2106 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2107 				mtx_unlock(&nblock);
2108 				return (NULL);
2109 			}
2110 		}
2111 		mtx_unlock(&nblock);
2112 	} else {
2113 		/*
2114 		 * We finally have a valid bp.  We aren't quite out of the
2115 		 * woods, we still have to reserve kva space.  In order
2116 		 * to keep fragmentation sane we only allocate kva in
2117 		 * BKVASIZE chunks.
2118 		 */
2119 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2120 
2121 		if (maxsize != bp->b_kvasize) {
2122 			vm_offset_t addr = 0;
2123 
2124 			bfreekva(bp);
2125 
2126 			vm_map_lock(buffer_map);
2127 			if (vm_map_findspace(buffer_map,
2128 				vm_map_min(buffer_map), maxsize, &addr)) {
2129 				/*
2130 				 * Uh oh.  Buffer map is to fragmented.  We
2131 				 * must defragment the map.
2132 				 */
2133 				atomic_add_int(&bufdefragcnt, 1);
2134 				vm_map_unlock(buffer_map);
2135 				defrag = 1;
2136 				bp->b_flags |= B_INVAL;
2137 				brelse(bp);
2138 				goto restart;
2139 			}
2140 			if (addr) {
2141 				vm_map_insert(buffer_map, NULL, 0,
2142 					addr, addr + maxsize,
2143 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2144 
2145 				bp->b_kvabase = (caddr_t) addr;
2146 				bp->b_kvasize = maxsize;
2147 				atomic_add_long(&bufspace, bp->b_kvasize);
2148 				atomic_add_int(&bufreusecnt, 1);
2149 			}
2150 			vm_map_unlock(buffer_map);
2151 		}
2152 		bp->b_saveaddr = bp->b_kvabase;
2153 		bp->b_data = bp->b_saveaddr;
2154 	}
2155 	return(bp);
2156 }
2157 
2158 /*
2159  *	buf_daemon:
2160  *
2161  *	buffer flushing daemon.  Buffers are normally flushed by the
2162  *	update daemon but if it cannot keep up this process starts to
2163  *	take the load in an attempt to prevent getnewbuf() from blocking.
2164  */
2165 
2166 static struct kproc_desc buf_kp = {
2167 	"bufdaemon",
2168 	buf_daemon,
2169 	&bufdaemonproc
2170 };
2171 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2172 
2173 static int
2174 buf_do_flush(struct vnode *vp)
2175 {
2176 	int flushed;
2177 
2178 	flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2179 	/* The list empty check here is slightly racy */
2180 	if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2181 		mtx_lock(&Giant);
2182 		flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0);
2183 		mtx_unlock(&Giant);
2184 	}
2185 	if (flushed == 0) {
2186 		/*
2187 		 * Could not find any buffers without rollback
2188 		 * dependencies, so just write the first one
2189 		 * in the hopes of eventually making progress.
2190 		 */
2191 		flushbufqueues(vp, QUEUE_DIRTY, 1);
2192 		if (!TAILQ_EMPTY(
2193 			    &bufqueues[QUEUE_DIRTY_GIANT])) {
2194 			mtx_lock(&Giant);
2195 			flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1);
2196 			mtx_unlock(&Giant);
2197 		}
2198 	}
2199 	return (flushed);
2200 }
2201 
2202 static void
2203 buf_daemon()
2204 {
2205 	int lodirtysave;
2206 
2207 	/*
2208 	 * This process needs to be suspended prior to shutdown sync.
2209 	 */
2210 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2211 	    SHUTDOWN_PRI_LAST);
2212 
2213 	/*
2214 	 * This process is allowed to take the buffer cache to the limit
2215 	 */
2216 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2217 	mtx_lock(&bdlock);
2218 	for (;;) {
2219 		bd_request = 0;
2220 		mtx_unlock(&bdlock);
2221 
2222 		kproc_suspend_check(bufdaemonproc);
2223 		lodirtysave = lodirtybuffers;
2224 		if (bd_speedupreq) {
2225 			lodirtybuffers = numdirtybuffers / 2;
2226 			bd_speedupreq = 0;
2227 		}
2228 		/*
2229 		 * Do the flush.  Limit the amount of in-transit I/O we
2230 		 * allow to build up, otherwise we would completely saturate
2231 		 * the I/O system.  Wakeup any waiting processes before we
2232 		 * normally would so they can run in parallel with our drain.
2233 		 */
2234 		while (numdirtybuffers > lodirtybuffers) {
2235 			if (buf_do_flush(NULL) == 0)
2236 				break;
2237 			kern_yield(PRI_UNCHANGED);
2238 		}
2239 		lodirtybuffers = lodirtysave;
2240 
2241 		/*
2242 		 * Only clear bd_request if we have reached our low water
2243 		 * mark.  The buf_daemon normally waits 1 second and
2244 		 * then incrementally flushes any dirty buffers that have
2245 		 * built up, within reason.
2246 		 *
2247 		 * If we were unable to hit our low water mark and couldn't
2248 		 * find any flushable buffers, we sleep half a second.
2249 		 * Otherwise we loop immediately.
2250 		 */
2251 		mtx_lock(&bdlock);
2252 		if (numdirtybuffers <= lodirtybuffers) {
2253 			/*
2254 			 * We reached our low water mark, reset the
2255 			 * request and sleep until we are needed again.
2256 			 * The sleep is just so the suspend code works.
2257 			 */
2258 			bd_request = 0;
2259 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2260 		} else {
2261 			/*
2262 			 * We couldn't find any flushable dirty buffers but
2263 			 * still have too many dirty buffers, we
2264 			 * have to sleep and try again.  (rare)
2265 			 */
2266 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2267 		}
2268 	}
2269 }
2270 
2271 /*
2272  *	flushbufqueues:
2273  *
2274  *	Try to flush a buffer in the dirty queue.  We must be careful to
2275  *	free up B_INVAL buffers instead of write them, which NFS is
2276  *	particularly sensitive to.
2277  */
2278 static int flushwithdeps = 0;
2279 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2280     0, "Number of buffers flushed with dependecies that require rollbacks");
2281 
2282 static int
2283 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2284 {
2285 	struct buf *sentinel;
2286 	struct vnode *vp;
2287 	struct mount *mp;
2288 	struct buf *bp;
2289 	int hasdeps;
2290 	int flushed;
2291 	int target;
2292 
2293 	if (lvp == NULL) {
2294 		target = numdirtybuffers - lodirtybuffers;
2295 		if (flushdeps && target > 2)
2296 			target /= 2;
2297 	} else
2298 		target = flushbufqtarget;
2299 	flushed = 0;
2300 	bp = NULL;
2301 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2302 	sentinel->b_qindex = QUEUE_SENTINEL;
2303 	mtx_lock(&bqlock);
2304 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2305 	while (flushed != target) {
2306 		bp = TAILQ_NEXT(sentinel, b_freelist);
2307 		if (bp != NULL) {
2308 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2309 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2310 			    b_freelist);
2311 		} else
2312 			break;
2313 		/*
2314 		 * Skip sentinels inserted by other invocations of the
2315 		 * flushbufqueues(), taking care to not reorder them.
2316 		 */
2317 		if (bp->b_qindex == QUEUE_SENTINEL)
2318 			continue;
2319 		/*
2320 		 * Only flush the buffers that belong to the
2321 		 * vnode locked by the curthread.
2322 		 */
2323 		if (lvp != NULL && bp->b_vp != lvp)
2324 			continue;
2325 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2326 			continue;
2327 		if (bp->b_pin_count > 0) {
2328 			BUF_UNLOCK(bp);
2329 			continue;
2330 		}
2331 		BO_LOCK(bp->b_bufobj);
2332 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2333 		    (bp->b_flags & B_DELWRI) == 0) {
2334 			BO_UNLOCK(bp->b_bufobj);
2335 			BUF_UNLOCK(bp);
2336 			continue;
2337 		}
2338 		BO_UNLOCK(bp->b_bufobj);
2339 		if (bp->b_flags & B_INVAL) {
2340 			bremfreel(bp);
2341 			mtx_unlock(&bqlock);
2342 			brelse(bp);
2343 			flushed++;
2344 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2345 			mtx_lock(&bqlock);
2346 			continue;
2347 		}
2348 
2349 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2350 			if (flushdeps == 0) {
2351 				BUF_UNLOCK(bp);
2352 				continue;
2353 			}
2354 			hasdeps = 1;
2355 		} else
2356 			hasdeps = 0;
2357 		/*
2358 		 * We must hold the lock on a vnode before writing
2359 		 * one of its buffers. Otherwise we may confuse, or
2360 		 * in the case of a snapshot vnode, deadlock the
2361 		 * system.
2362 		 *
2363 		 * The lock order here is the reverse of the normal
2364 		 * of vnode followed by buf lock.  This is ok because
2365 		 * the NOWAIT will prevent deadlock.
2366 		 */
2367 		vp = bp->b_vp;
2368 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2369 			BUF_UNLOCK(bp);
2370 			continue;
2371 		}
2372 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2373 			mtx_unlock(&bqlock);
2374 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2375 			    bp, bp->b_vp, bp->b_flags);
2376 			if (curproc == bufdaemonproc)
2377 				vfs_bio_awrite(bp);
2378 			else {
2379 				bremfree(bp);
2380 				bwrite(bp);
2381 				notbufdflashes++;
2382 			}
2383 			vn_finished_write(mp);
2384 			VOP_UNLOCK(vp, 0);
2385 			flushwithdeps += hasdeps;
2386 			flushed++;
2387 
2388 			/*
2389 			 * Sleeping on runningbufspace while holding
2390 			 * vnode lock leads to deadlock.
2391 			 */
2392 			if (curproc == bufdaemonproc)
2393 				waitrunningbufspace();
2394 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2395 			mtx_lock(&bqlock);
2396 			continue;
2397 		}
2398 		vn_finished_write(mp);
2399 		BUF_UNLOCK(bp);
2400 	}
2401 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2402 	mtx_unlock(&bqlock);
2403 	free(sentinel, M_TEMP);
2404 	return (flushed);
2405 }
2406 
2407 /*
2408  * Check to see if a block is currently memory resident.
2409  */
2410 struct buf *
2411 incore(struct bufobj *bo, daddr_t blkno)
2412 {
2413 	struct buf *bp;
2414 
2415 	BO_LOCK(bo);
2416 	bp = gbincore(bo, blkno);
2417 	BO_UNLOCK(bo);
2418 	return (bp);
2419 }
2420 
2421 /*
2422  * Returns true if no I/O is needed to access the
2423  * associated VM object.  This is like incore except
2424  * it also hunts around in the VM system for the data.
2425  */
2426 
2427 static int
2428 inmem(struct vnode * vp, daddr_t blkno)
2429 {
2430 	vm_object_t obj;
2431 	vm_offset_t toff, tinc, size;
2432 	vm_page_t m;
2433 	vm_ooffset_t off;
2434 
2435 	ASSERT_VOP_LOCKED(vp, "inmem");
2436 
2437 	if (incore(&vp->v_bufobj, blkno))
2438 		return 1;
2439 	if (vp->v_mount == NULL)
2440 		return 0;
2441 	obj = vp->v_object;
2442 	if (obj == NULL)
2443 		return (0);
2444 
2445 	size = PAGE_SIZE;
2446 	if (size > vp->v_mount->mnt_stat.f_iosize)
2447 		size = vp->v_mount->mnt_stat.f_iosize;
2448 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2449 
2450 	VM_OBJECT_LOCK(obj);
2451 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2452 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2453 		if (!m)
2454 			goto notinmem;
2455 		tinc = size;
2456 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2457 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2458 		if (vm_page_is_valid(m,
2459 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2460 			goto notinmem;
2461 	}
2462 	VM_OBJECT_UNLOCK(obj);
2463 	return 1;
2464 
2465 notinmem:
2466 	VM_OBJECT_UNLOCK(obj);
2467 	return (0);
2468 }
2469 
2470 /*
2471  * Set the dirty range for a buffer based on the status of the dirty
2472  * bits in the pages comprising the buffer.  The range is limited
2473  * to the size of the buffer.
2474  *
2475  * Tell the VM system that the pages associated with this buffer
2476  * are clean.  This is used for delayed writes where the data is
2477  * going to go to disk eventually without additional VM intevention.
2478  *
2479  * Note that while we only really need to clean through to b_bcount, we
2480  * just go ahead and clean through to b_bufsize.
2481  */
2482 static void
2483 vfs_clean_pages_dirty_buf(struct buf *bp)
2484 {
2485 	vm_ooffset_t foff, noff, eoff;
2486 	vm_page_t m;
2487 	int i;
2488 
2489 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2490 		return;
2491 
2492 	foff = bp->b_offset;
2493 	KASSERT(bp->b_offset != NOOFFSET,
2494 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2495 
2496 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2497 	vfs_drain_busy_pages(bp);
2498 	vfs_setdirty_locked_object(bp);
2499 	for (i = 0; i < bp->b_npages; i++) {
2500 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2501 		eoff = noff;
2502 		if (eoff > bp->b_offset + bp->b_bufsize)
2503 			eoff = bp->b_offset + bp->b_bufsize;
2504 		m = bp->b_pages[i];
2505 		vfs_page_set_validclean(bp, foff, m);
2506 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2507 		foff = noff;
2508 	}
2509 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2510 }
2511 
2512 static void
2513 vfs_setdirty_locked_object(struct buf *bp)
2514 {
2515 	vm_object_t object;
2516 	int i;
2517 
2518 	object = bp->b_bufobj->bo_object;
2519 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2520 
2521 	/*
2522 	 * We qualify the scan for modified pages on whether the
2523 	 * object has been flushed yet.
2524 	 */
2525 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2526 		vm_offset_t boffset;
2527 		vm_offset_t eoffset;
2528 
2529 		/*
2530 		 * test the pages to see if they have been modified directly
2531 		 * by users through the VM system.
2532 		 */
2533 		for (i = 0; i < bp->b_npages; i++)
2534 			vm_page_test_dirty(bp->b_pages[i]);
2535 
2536 		/*
2537 		 * Calculate the encompassing dirty range, boffset and eoffset,
2538 		 * (eoffset - boffset) bytes.
2539 		 */
2540 
2541 		for (i = 0; i < bp->b_npages; i++) {
2542 			if (bp->b_pages[i]->dirty)
2543 				break;
2544 		}
2545 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2546 
2547 		for (i = bp->b_npages - 1; i >= 0; --i) {
2548 			if (bp->b_pages[i]->dirty) {
2549 				break;
2550 			}
2551 		}
2552 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2553 
2554 		/*
2555 		 * Fit it to the buffer.
2556 		 */
2557 
2558 		if (eoffset > bp->b_bcount)
2559 			eoffset = bp->b_bcount;
2560 
2561 		/*
2562 		 * If we have a good dirty range, merge with the existing
2563 		 * dirty range.
2564 		 */
2565 
2566 		if (boffset < eoffset) {
2567 			if (bp->b_dirtyoff > boffset)
2568 				bp->b_dirtyoff = boffset;
2569 			if (bp->b_dirtyend < eoffset)
2570 				bp->b_dirtyend = eoffset;
2571 		}
2572 	}
2573 }
2574 
2575 /*
2576  *	getblk:
2577  *
2578  *	Get a block given a specified block and offset into a file/device.
2579  *	The buffers B_DONE bit will be cleared on return, making it almost
2580  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2581  *	return.  The caller should clear B_INVAL prior to initiating a
2582  *	READ.
2583  *
2584  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2585  *	an existing buffer.
2586  *
2587  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2588  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2589  *	and then cleared based on the backing VM.  If the previous buffer is
2590  *	non-0-sized but invalid, B_CACHE will be cleared.
2591  *
2592  *	If getblk() must create a new buffer, the new buffer is returned with
2593  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2594  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2595  *	backing VM.
2596  *
2597  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2598  *	B_CACHE bit is clear.
2599  *
2600  *	What this means, basically, is that the caller should use B_CACHE to
2601  *	determine whether the buffer is fully valid or not and should clear
2602  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2603  *	the buffer by loading its data area with something, the caller needs
2604  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2605  *	the caller should set B_CACHE ( as an optimization ), else the caller
2606  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2607  *	a write attempt or if it was a successfull read.  If the caller
2608  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2609  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2610  */
2611 struct buf *
2612 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2613     int flags)
2614 {
2615 	struct buf *bp;
2616 	struct bufobj *bo;
2617 	int error;
2618 
2619 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2620 	ASSERT_VOP_LOCKED(vp, "getblk");
2621 	if (size > MAXBSIZE)
2622 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2623 
2624 	bo = &vp->v_bufobj;
2625 loop:
2626 	/*
2627 	 * Block if we are low on buffers.   Certain processes are allowed
2628 	 * to completely exhaust the buffer cache.
2629          *
2630          * If this check ever becomes a bottleneck it may be better to
2631          * move it into the else, when gbincore() fails.  At the moment
2632          * it isn't a problem.
2633 	 *
2634 	 * XXX remove if 0 sections (clean this up after its proven)
2635          */
2636 	if (numfreebuffers == 0) {
2637 		if (TD_IS_IDLETHREAD(curthread))
2638 			return NULL;
2639 		mtx_lock(&nblock);
2640 		needsbuffer |= VFS_BIO_NEED_ANY;
2641 		mtx_unlock(&nblock);
2642 	}
2643 
2644 	BO_LOCK(bo);
2645 	bp = gbincore(bo, blkno);
2646 	if (bp != NULL) {
2647 		int lockflags;
2648 		/*
2649 		 * Buffer is in-core.  If the buffer is not busy, it must
2650 		 * be on a queue.
2651 		 */
2652 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2653 
2654 		if (flags & GB_LOCK_NOWAIT)
2655 			lockflags |= LK_NOWAIT;
2656 
2657 		error = BUF_TIMELOCK(bp, lockflags,
2658 		    BO_MTX(bo), "getblk", slpflag, slptimeo);
2659 
2660 		/*
2661 		 * If we slept and got the lock we have to restart in case
2662 		 * the buffer changed identities.
2663 		 */
2664 		if (error == ENOLCK)
2665 			goto loop;
2666 		/* We timed out or were interrupted. */
2667 		else if (error)
2668 			return (NULL);
2669 
2670 		/*
2671 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2672 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2673 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2674 		 * backing VM cache.
2675 		 */
2676 		if (bp->b_flags & B_INVAL)
2677 			bp->b_flags &= ~B_CACHE;
2678 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2679 			bp->b_flags |= B_CACHE;
2680 		BO_LOCK(bo);
2681 		bremfree(bp);
2682 		BO_UNLOCK(bo);
2683 
2684 		/*
2685 		 * check for size inconsistancies for non-VMIO case.
2686 		 */
2687 
2688 		if (bp->b_bcount != size) {
2689 			if ((bp->b_flags & B_VMIO) == 0 ||
2690 			    (size > bp->b_kvasize)) {
2691 				if (bp->b_flags & B_DELWRI) {
2692 					/*
2693 					 * If buffer is pinned and caller does
2694 					 * not want sleep  waiting for it to be
2695 					 * unpinned, bail out
2696 					 * */
2697 					if (bp->b_pin_count > 0) {
2698 						if (flags & GB_LOCK_NOWAIT) {
2699 							bqrelse(bp);
2700 							return (NULL);
2701 						} else {
2702 							bunpin_wait(bp);
2703 						}
2704 					}
2705 					bp->b_flags |= B_NOCACHE;
2706 					bwrite(bp);
2707 				} else {
2708 					if (LIST_EMPTY(&bp->b_dep)) {
2709 						bp->b_flags |= B_RELBUF;
2710 						brelse(bp);
2711 					} else {
2712 						bp->b_flags |= B_NOCACHE;
2713 						bwrite(bp);
2714 					}
2715 				}
2716 				goto loop;
2717 			}
2718 		}
2719 
2720 		/*
2721 		 * If the size is inconsistant in the VMIO case, we can resize
2722 		 * the buffer.  This might lead to B_CACHE getting set or
2723 		 * cleared.  If the size has not changed, B_CACHE remains
2724 		 * unchanged from its previous state.
2725 		 */
2726 
2727 		if (bp->b_bcount != size)
2728 			allocbuf(bp, size);
2729 
2730 		KASSERT(bp->b_offset != NOOFFSET,
2731 		    ("getblk: no buffer offset"));
2732 
2733 		/*
2734 		 * A buffer with B_DELWRI set and B_CACHE clear must
2735 		 * be committed before we can return the buffer in
2736 		 * order to prevent the caller from issuing a read
2737 		 * ( due to B_CACHE not being set ) and overwriting
2738 		 * it.
2739 		 *
2740 		 * Most callers, including NFS and FFS, need this to
2741 		 * operate properly either because they assume they
2742 		 * can issue a read if B_CACHE is not set, or because
2743 		 * ( for example ) an uncached B_DELWRI might loop due
2744 		 * to softupdates re-dirtying the buffer.  In the latter
2745 		 * case, B_CACHE is set after the first write completes,
2746 		 * preventing further loops.
2747 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2748 		 * above while extending the buffer, we cannot allow the
2749 		 * buffer to remain with B_CACHE set after the write
2750 		 * completes or it will represent a corrupt state.  To
2751 		 * deal with this we set B_NOCACHE to scrap the buffer
2752 		 * after the write.
2753 		 *
2754 		 * We might be able to do something fancy, like setting
2755 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2756 		 * so the below call doesn't set B_CACHE, but that gets real
2757 		 * confusing.  This is much easier.
2758 		 */
2759 
2760 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2761 			bp->b_flags |= B_NOCACHE;
2762 			bwrite(bp);
2763 			goto loop;
2764 		}
2765 		bp->b_flags &= ~B_DONE;
2766 	} else {
2767 		int bsize, maxsize, vmio;
2768 		off_t offset;
2769 
2770 		/*
2771 		 * Buffer is not in-core, create new buffer.  The buffer
2772 		 * returned by getnewbuf() is locked.  Note that the returned
2773 		 * buffer is also considered valid (not marked B_INVAL).
2774 		 */
2775 		BO_UNLOCK(bo);
2776 		/*
2777 		 * If the user does not want us to create the buffer, bail out
2778 		 * here.
2779 		 */
2780 		if (flags & GB_NOCREAT)
2781 			return NULL;
2782 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
2783 		offset = blkno * bsize;
2784 		vmio = vp->v_object != NULL;
2785 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2786 		maxsize = imax(maxsize, bsize);
2787 
2788 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2789 		if (bp == NULL) {
2790 			if (slpflag || slptimeo)
2791 				return NULL;
2792 			goto loop;
2793 		}
2794 
2795 		/*
2796 		 * This code is used to make sure that a buffer is not
2797 		 * created while the getnewbuf routine is blocked.
2798 		 * This can be a problem whether the vnode is locked or not.
2799 		 * If the buffer is created out from under us, we have to
2800 		 * throw away the one we just created.
2801 		 *
2802 		 * Note: this must occur before we associate the buffer
2803 		 * with the vp especially considering limitations in
2804 		 * the splay tree implementation when dealing with duplicate
2805 		 * lblkno's.
2806 		 */
2807 		BO_LOCK(bo);
2808 		if (gbincore(bo, blkno)) {
2809 			BO_UNLOCK(bo);
2810 			bp->b_flags |= B_INVAL;
2811 			brelse(bp);
2812 			goto loop;
2813 		}
2814 
2815 		/*
2816 		 * Insert the buffer into the hash, so that it can
2817 		 * be found by incore.
2818 		 */
2819 		bp->b_blkno = bp->b_lblkno = blkno;
2820 		bp->b_offset = offset;
2821 		bgetvp(vp, bp);
2822 		BO_UNLOCK(bo);
2823 
2824 		/*
2825 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2826 		 * buffer size starts out as 0, B_CACHE will be set by
2827 		 * allocbuf() for the VMIO case prior to it testing the
2828 		 * backing store for validity.
2829 		 */
2830 
2831 		if (vmio) {
2832 			bp->b_flags |= B_VMIO;
2833 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2834 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2835 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2836 		} else {
2837 			bp->b_flags &= ~B_VMIO;
2838 			KASSERT(bp->b_bufobj->bo_object == NULL,
2839 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2840 			    bp, bp->b_bufobj->bo_object));
2841 		}
2842 
2843 		allocbuf(bp, size);
2844 		bp->b_flags &= ~B_DONE;
2845 	}
2846 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2847 	BUF_ASSERT_HELD(bp);
2848 	KASSERT(bp->b_bufobj == bo,
2849 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2850 	return (bp);
2851 }
2852 
2853 /*
2854  * Get an empty, disassociated buffer of given size.  The buffer is initially
2855  * set to B_INVAL.
2856  */
2857 struct buf *
2858 geteblk(int size, int flags)
2859 {
2860 	struct buf *bp;
2861 	int maxsize;
2862 
2863 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2864 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2865 		if ((flags & GB_NOWAIT_BD) &&
2866 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
2867 			return (NULL);
2868 	}
2869 	allocbuf(bp, size);
2870 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2871 	BUF_ASSERT_HELD(bp);
2872 	return (bp);
2873 }
2874 
2875 
2876 /*
2877  * This code constitutes the buffer memory from either anonymous system
2878  * memory (in the case of non-VMIO operations) or from an associated
2879  * VM object (in the case of VMIO operations).  This code is able to
2880  * resize a buffer up or down.
2881  *
2882  * Note that this code is tricky, and has many complications to resolve
2883  * deadlock or inconsistant data situations.  Tread lightly!!!
2884  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2885  * the caller.  Calling this code willy nilly can result in the loss of data.
2886  *
2887  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2888  * B_CACHE for the non-VMIO case.
2889  */
2890 
2891 int
2892 allocbuf(struct buf *bp, int size)
2893 {
2894 	int newbsize, mbsize;
2895 	int i;
2896 
2897 	BUF_ASSERT_HELD(bp);
2898 
2899 	if (bp->b_kvasize < size)
2900 		panic("allocbuf: buffer too small");
2901 
2902 	if ((bp->b_flags & B_VMIO) == 0) {
2903 		caddr_t origbuf;
2904 		int origbufsize;
2905 		/*
2906 		 * Just get anonymous memory from the kernel.  Don't
2907 		 * mess with B_CACHE.
2908 		 */
2909 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2910 		if (bp->b_flags & B_MALLOC)
2911 			newbsize = mbsize;
2912 		else
2913 			newbsize = round_page(size);
2914 
2915 		if (newbsize < bp->b_bufsize) {
2916 			/*
2917 			 * malloced buffers are not shrunk
2918 			 */
2919 			if (bp->b_flags & B_MALLOC) {
2920 				if (newbsize) {
2921 					bp->b_bcount = size;
2922 				} else {
2923 					free(bp->b_data, M_BIOBUF);
2924 					if (bp->b_bufsize) {
2925 						atomic_subtract_long(
2926 						    &bufmallocspace,
2927 						    bp->b_bufsize);
2928 						bufspacewakeup();
2929 						bp->b_bufsize = 0;
2930 					}
2931 					bp->b_saveaddr = bp->b_kvabase;
2932 					bp->b_data = bp->b_saveaddr;
2933 					bp->b_bcount = 0;
2934 					bp->b_flags &= ~B_MALLOC;
2935 				}
2936 				return 1;
2937 			}
2938 			vm_hold_free_pages(bp, newbsize);
2939 		} else if (newbsize > bp->b_bufsize) {
2940 			/*
2941 			 * We only use malloced memory on the first allocation.
2942 			 * and revert to page-allocated memory when the buffer
2943 			 * grows.
2944 			 */
2945 			/*
2946 			 * There is a potential smp race here that could lead
2947 			 * to bufmallocspace slightly passing the max.  It
2948 			 * is probably extremely rare and not worth worrying
2949 			 * over.
2950 			 */
2951 			if ( (bufmallocspace < maxbufmallocspace) &&
2952 				(bp->b_bufsize == 0) &&
2953 				(mbsize <= PAGE_SIZE/2)) {
2954 
2955 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2956 				bp->b_bufsize = mbsize;
2957 				bp->b_bcount = size;
2958 				bp->b_flags |= B_MALLOC;
2959 				atomic_add_long(&bufmallocspace, mbsize);
2960 				return 1;
2961 			}
2962 			origbuf = NULL;
2963 			origbufsize = 0;
2964 			/*
2965 			 * If the buffer is growing on its other-than-first allocation,
2966 			 * then we revert to the page-allocation scheme.
2967 			 */
2968 			if (bp->b_flags & B_MALLOC) {
2969 				origbuf = bp->b_data;
2970 				origbufsize = bp->b_bufsize;
2971 				bp->b_data = bp->b_kvabase;
2972 				if (bp->b_bufsize) {
2973 					atomic_subtract_long(&bufmallocspace,
2974 					    bp->b_bufsize);
2975 					bufspacewakeup();
2976 					bp->b_bufsize = 0;
2977 				}
2978 				bp->b_flags &= ~B_MALLOC;
2979 				newbsize = round_page(newbsize);
2980 			}
2981 			vm_hold_load_pages(
2982 			    bp,
2983 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2984 			    (vm_offset_t) bp->b_data + newbsize);
2985 			if (origbuf) {
2986 				bcopy(origbuf, bp->b_data, origbufsize);
2987 				free(origbuf, M_BIOBUF);
2988 			}
2989 		}
2990 	} else {
2991 		int desiredpages;
2992 
2993 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2994 		desiredpages = (size == 0) ? 0 :
2995 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2996 
2997 		if (bp->b_flags & B_MALLOC)
2998 			panic("allocbuf: VMIO buffer can't be malloced");
2999 		/*
3000 		 * Set B_CACHE initially if buffer is 0 length or will become
3001 		 * 0-length.
3002 		 */
3003 		if (size == 0 || bp->b_bufsize == 0)
3004 			bp->b_flags |= B_CACHE;
3005 
3006 		if (newbsize < bp->b_bufsize) {
3007 			/*
3008 			 * DEV_BSIZE aligned new buffer size is less then the
3009 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3010 			 * if we have to remove any pages.
3011 			 */
3012 			if (desiredpages < bp->b_npages) {
3013 				vm_page_t m;
3014 
3015 				pmap_qremove((vm_offset_t)trunc_page(
3016 				    (vm_offset_t)bp->b_data) +
3017 				    (desiredpages << PAGE_SHIFT),
3018 				    (bp->b_npages - desiredpages));
3019 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3020 				for (i = desiredpages; i < bp->b_npages; i++) {
3021 					/*
3022 					 * the page is not freed here -- it
3023 					 * is the responsibility of
3024 					 * vnode_pager_setsize
3025 					 */
3026 					m = bp->b_pages[i];
3027 					KASSERT(m != bogus_page,
3028 					    ("allocbuf: bogus page found"));
3029 					while (vm_page_sleep_if_busy(m, TRUE,
3030 					    "biodep"))
3031 						continue;
3032 
3033 					bp->b_pages[i] = NULL;
3034 					vm_page_lock(m);
3035 					vm_page_unwire(m, 0);
3036 					vm_page_unlock(m);
3037 				}
3038 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3039 				bp->b_npages = desiredpages;
3040 			}
3041 		} else if (size > bp->b_bcount) {
3042 			/*
3043 			 * We are growing the buffer, possibly in a
3044 			 * byte-granular fashion.
3045 			 */
3046 			vm_object_t obj;
3047 			vm_offset_t toff;
3048 			vm_offset_t tinc;
3049 
3050 			/*
3051 			 * Step 1, bring in the VM pages from the object,
3052 			 * allocating them if necessary.  We must clear
3053 			 * B_CACHE if these pages are not valid for the
3054 			 * range covered by the buffer.
3055 			 */
3056 
3057 			obj = bp->b_bufobj->bo_object;
3058 
3059 			VM_OBJECT_LOCK(obj);
3060 			while (bp->b_npages < desiredpages) {
3061 				vm_page_t m;
3062 
3063 				/*
3064 				 * We must allocate system pages since blocking
3065 				 * here could intefere with paging I/O, no
3066 				 * matter which process we are.
3067 				 *
3068 				 * We can only test VPO_BUSY here.  Blocking on
3069 				 * m->busy might lead to a deadlock:
3070 				 *  vm_fault->getpages->cluster_read->allocbuf
3071 				 * Thus, we specify VM_ALLOC_IGN_SBUSY.
3072 				 */
3073 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3074 				    bp->b_npages, VM_ALLOC_NOBUSY |
3075 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3076 				    VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY |
3077 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3078 				if (m->valid == 0)
3079 					bp->b_flags &= ~B_CACHE;
3080 				bp->b_pages[bp->b_npages] = m;
3081 				++bp->b_npages;
3082 			}
3083 
3084 			/*
3085 			 * Step 2.  We've loaded the pages into the buffer,
3086 			 * we have to figure out if we can still have B_CACHE
3087 			 * set.  Note that B_CACHE is set according to the
3088 			 * byte-granular range ( bcount and size ), new the
3089 			 * aligned range ( newbsize ).
3090 			 *
3091 			 * The VM test is against m->valid, which is DEV_BSIZE
3092 			 * aligned.  Needless to say, the validity of the data
3093 			 * needs to also be DEV_BSIZE aligned.  Note that this
3094 			 * fails with NFS if the server or some other client
3095 			 * extends the file's EOF.  If our buffer is resized,
3096 			 * B_CACHE may remain set! XXX
3097 			 */
3098 
3099 			toff = bp->b_bcount;
3100 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3101 
3102 			while ((bp->b_flags & B_CACHE) && toff < size) {
3103 				vm_pindex_t pi;
3104 
3105 				if (tinc > (size - toff))
3106 					tinc = size - toff;
3107 
3108 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3109 				    PAGE_SHIFT;
3110 
3111 				vfs_buf_test_cache(
3112 				    bp,
3113 				    bp->b_offset,
3114 				    toff,
3115 				    tinc,
3116 				    bp->b_pages[pi]
3117 				);
3118 				toff += tinc;
3119 				tinc = PAGE_SIZE;
3120 			}
3121 			VM_OBJECT_UNLOCK(obj);
3122 
3123 			/*
3124 			 * Step 3, fixup the KVM pmap.  Remember that
3125 			 * bp->b_data is relative to bp->b_offset, but
3126 			 * bp->b_offset may be offset into the first page.
3127 			 */
3128 
3129 			bp->b_data = (caddr_t)
3130 			    trunc_page((vm_offset_t)bp->b_data);
3131 			pmap_qenter(
3132 			    (vm_offset_t)bp->b_data,
3133 			    bp->b_pages,
3134 			    bp->b_npages
3135 			);
3136 
3137 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3138 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
3139 		}
3140 	}
3141 	if (newbsize < bp->b_bufsize)
3142 		bufspacewakeup();
3143 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3144 	bp->b_bcount = size;		/* requested buffer size	*/
3145 	return 1;
3146 }
3147 
3148 void
3149 biodone(struct bio *bp)
3150 {
3151 	struct mtx *mtxp;
3152 	void (*done)(struct bio *);
3153 
3154 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3155 	mtx_lock(mtxp);
3156 	bp->bio_flags |= BIO_DONE;
3157 	done = bp->bio_done;
3158 	if (done == NULL)
3159 		wakeup(bp);
3160 	mtx_unlock(mtxp);
3161 	if (done != NULL)
3162 		done(bp);
3163 }
3164 
3165 /*
3166  * Wait for a BIO to finish.
3167  *
3168  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3169  * case is not yet clear.
3170  */
3171 int
3172 biowait(struct bio *bp, const char *wchan)
3173 {
3174 	struct mtx *mtxp;
3175 
3176 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3177 	mtx_lock(mtxp);
3178 	while ((bp->bio_flags & BIO_DONE) == 0)
3179 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3180 	mtx_unlock(mtxp);
3181 	if (bp->bio_error != 0)
3182 		return (bp->bio_error);
3183 	if (!(bp->bio_flags & BIO_ERROR))
3184 		return (0);
3185 	return (EIO);
3186 }
3187 
3188 void
3189 biofinish(struct bio *bp, struct devstat *stat, int error)
3190 {
3191 
3192 	if (error) {
3193 		bp->bio_error = error;
3194 		bp->bio_flags |= BIO_ERROR;
3195 	}
3196 	if (stat != NULL)
3197 		devstat_end_transaction_bio(stat, bp);
3198 	biodone(bp);
3199 }
3200 
3201 /*
3202  *	bufwait:
3203  *
3204  *	Wait for buffer I/O completion, returning error status.  The buffer
3205  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3206  *	error and cleared.
3207  */
3208 int
3209 bufwait(struct buf *bp)
3210 {
3211 	if (bp->b_iocmd == BIO_READ)
3212 		bwait(bp, PRIBIO, "biord");
3213 	else
3214 		bwait(bp, PRIBIO, "biowr");
3215 	if (bp->b_flags & B_EINTR) {
3216 		bp->b_flags &= ~B_EINTR;
3217 		return (EINTR);
3218 	}
3219 	if (bp->b_ioflags & BIO_ERROR) {
3220 		return (bp->b_error ? bp->b_error : EIO);
3221 	} else {
3222 		return (0);
3223 	}
3224 }
3225 
3226  /*
3227   * Call back function from struct bio back up to struct buf.
3228   */
3229 static void
3230 bufdonebio(struct bio *bip)
3231 {
3232 	struct buf *bp;
3233 
3234 	bp = bip->bio_caller2;
3235 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3236 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3237 	bp->b_ioflags = bip->bio_flags;
3238 	bp->b_error = bip->bio_error;
3239 	if (bp->b_error)
3240 		bp->b_ioflags |= BIO_ERROR;
3241 	bufdone(bp);
3242 	g_destroy_bio(bip);
3243 }
3244 
3245 void
3246 dev_strategy(struct cdev *dev, struct buf *bp)
3247 {
3248 	struct cdevsw *csw;
3249 	struct bio *bip;
3250 	int ref;
3251 
3252 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3253 		panic("b_iocmd botch");
3254 	for (;;) {
3255 		bip = g_new_bio();
3256 		if (bip != NULL)
3257 			break;
3258 		/* Try again later */
3259 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3260 	}
3261 	bip->bio_cmd = bp->b_iocmd;
3262 	bip->bio_offset = bp->b_iooffset;
3263 	bip->bio_length = bp->b_bcount;
3264 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3265 	bip->bio_data = bp->b_data;
3266 	bip->bio_done = bufdonebio;
3267 	bip->bio_caller2 = bp;
3268 	bip->bio_dev = dev;
3269 	KASSERT(dev->si_refcount > 0,
3270 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3271 	    devtoname(dev)));
3272 	csw = dev_refthread(dev, &ref);
3273 	if (csw == NULL) {
3274 		g_destroy_bio(bip);
3275 		bp->b_error = ENXIO;
3276 		bp->b_ioflags = BIO_ERROR;
3277 		bufdone(bp);
3278 		return;
3279 	}
3280 	(*csw->d_strategy)(bip);
3281 	dev_relthread(dev, ref);
3282 }
3283 
3284 /*
3285  *	bufdone:
3286  *
3287  *	Finish I/O on a buffer, optionally calling a completion function.
3288  *	This is usually called from an interrupt so process blocking is
3289  *	not allowed.
3290  *
3291  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3292  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3293  *	assuming B_INVAL is clear.
3294  *
3295  *	For the VMIO case, we set B_CACHE if the op was a read and no
3296  *	read error occured, or if the op was a write.  B_CACHE is never
3297  *	set if the buffer is invalid or otherwise uncacheable.
3298  *
3299  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3300  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3301  *	in the biodone routine.
3302  */
3303 void
3304 bufdone(struct buf *bp)
3305 {
3306 	struct bufobj *dropobj;
3307 	void    (*biodone)(struct buf *);
3308 
3309 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3310 	dropobj = NULL;
3311 
3312 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3313 	BUF_ASSERT_HELD(bp);
3314 
3315 	runningbufwakeup(bp);
3316 	if (bp->b_iocmd == BIO_WRITE)
3317 		dropobj = bp->b_bufobj;
3318 	/* call optional completion function if requested */
3319 	if (bp->b_iodone != NULL) {
3320 		biodone = bp->b_iodone;
3321 		bp->b_iodone = NULL;
3322 		(*biodone) (bp);
3323 		if (dropobj)
3324 			bufobj_wdrop(dropobj);
3325 		return;
3326 	}
3327 
3328 	bufdone_finish(bp);
3329 
3330 	if (dropobj)
3331 		bufobj_wdrop(dropobj);
3332 }
3333 
3334 void
3335 bufdone_finish(struct buf *bp)
3336 {
3337 	BUF_ASSERT_HELD(bp);
3338 
3339 	if (!LIST_EMPTY(&bp->b_dep))
3340 		buf_complete(bp);
3341 
3342 	if (bp->b_flags & B_VMIO) {
3343 		vm_ooffset_t foff;
3344 		vm_page_t m;
3345 		vm_object_t obj;
3346 		struct vnode *vp;
3347 		int bogus, i, iosize;
3348 
3349 		obj = bp->b_bufobj->bo_object;
3350 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3351 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3352 		    obj->paging_in_progress, bp->b_npages));
3353 
3354 		vp = bp->b_vp;
3355 		KASSERT(vp->v_holdcnt > 0,
3356 		    ("biodone_finish: vnode %p has zero hold count", vp));
3357 		KASSERT(vp->v_object != NULL,
3358 		    ("biodone_finish: vnode %p has no vm_object", vp));
3359 
3360 		foff = bp->b_offset;
3361 		KASSERT(bp->b_offset != NOOFFSET,
3362 		    ("biodone_finish: bp %p has no buffer offset", bp));
3363 
3364 		/*
3365 		 * Set B_CACHE if the op was a normal read and no error
3366 		 * occured.  B_CACHE is set for writes in the b*write()
3367 		 * routines.
3368 		 */
3369 		iosize = bp->b_bcount - bp->b_resid;
3370 		if (bp->b_iocmd == BIO_READ &&
3371 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3372 		    !(bp->b_ioflags & BIO_ERROR)) {
3373 			bp->b_flags |= B_CACHE;
3374 		}
3375 		bogus = 0;
3376 		VM_OBJECT_LOCK(obj);
3377 		for (i = 0; i < bp->b_npages; i++) {
3378 			int bogusflag = 0;
3379 			int resid;
3380 
3381 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3382 			if (resid > iosize)
3383 				resid = iosize;
3384 
3385 			/*
3386 			 * cleanup bogus pages, restoring the originals
3387 			 */
3388 			m = bp->b_pages[i];
3389 			if (m == bogus_page) {
3390 				bogus = bogusflag = 1;
3391 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3392 				if (m == NULL)
3393 					panic("biodone: page disappeared!");
3394 				bp->b_pages[i] = m;
3395 			}
3396 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3397 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3398 			    (intmax_t)foff, (uintmax_t)m->pindex));
3399 
3400 			/*
3401 			 * In the write case, the valid and clean bits are
3402 			 * already changed correctly ( see bdwrite() ), so we
3403 			 * only need to do this here in the read case.
3404 			 */
3405 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3406 				KASSERT((m->dirty & vm_page_bits(foff &
3407 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3408 				    " page %p has unexpected dirty bits", m));
3409 				vfs_page_set_valid(bp, foff, m);
3410 			}
3411 
3412 			vm_page_io_finish(m);
3413 			vm_object_pip_subtract(obj, 1);
3414 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3415 			iosize -= resid;
3416 		}
3417 		vm_object_pip_wakeupn(obj, 0);
3418 		VM_OBJECT_UNLOCK(obj);
3419 		if (bogus)
3420 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3421 			    bp->b_pages, bp->b_npages);
3422 	}
3423 
3424 	/*
3425 	 * For asynchronous completions, release the buffer now. The brelse
3426 	 * will do a wakeup there if necessary - so no need to do a wakeup
3427 	 * here in the async case. The sync case always needs to do a wakeup.
3428 	 */
3429 
3430 	if (bp->b_flags & B_ASYNC) {
3431 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3432 			brelse(bp);
3433 		else
3434 			bqrelse(bp);
3435 	} else
3436 		bdone(bp);
3437 }
3438 
3439 /*
3440  * This routine is called in lieu of iodone in the case of
3441  * incomplete I/O.  This keeps the busy status for pages
3442  * consistant.
3443  */
3444 void
3445 vfs_unbusy_pages(struct buf *bp)
3446 {
3447 	int i;
3448 	vm_object_t obj;
3449 	vm_page_t m;
3450 
3451 	runningbufwakeup(bp);
3452 	if (!(bp->b_flags & B_VMIO))
3453 		return;
3454 
3455 	obj = bp->b_bufobj->bo_object;
3456 	VM_OBJECT_LOCK(obj);
3457 	for (i = 0; i < bp->b_npages; i++) {
3458 		m = bp->b_pages[i];
3459 		if (m == bogus_page) {
3460 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3461 			if (!m)
3462 				panic("vfs_unbusy_pages: page missing\n");
3463 			bp->b_pages[i] = m;
3464 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3465 			    bp->b_pages, bp->b_npages);
3466 		}
3467 		vm_object_pip_subtract(obj, 1);
3468 		vm_page_io_finish(m);
3469 	}
3470 	vm_object_pip_wakeupn(obj, 0);
3471 	VM_OBJECT_UNLOCK(obj);
3472 }
3473 
3474 /*
3475  * vfs_page_set_valid:
3476  *
3477  *	Set the valid bits in a page based on the supplied offset.   The
3478  *	range is restricted to the buffer's size.
3479  *
3480  *	This routine is typically called after a read completes.
3481  */
3482 static void
3483 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3484 {
3485 	vm_ooffset_t eoff;
3486 
3487 	/*
3488 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3489 	 * page boundary and eoff is not greater than the end of the buffer.
3490 	 * The end of the buffer, in this case, is our file EOF, not the
3491 	 * allocation size of the buffer.
3492 	 */
3493 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3494 	if (eoff > bp->b_offset + bp->b_bcount)
3495 		eoff = bp->b_offset + bp->b_bcount;
3496 
3497 	/*
3498 	 * Set valid range.  This is typically the entire buffer and thus the
3499 	 * entire page.
3500 	 */
3501 	if (eoff > off)
3502 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3503 }
3504 
3505 /*
3506  * vfs_page_set_validclean:
3507  *
3508  *	Set the valid bits and clear the dirty bits in a page based on the
3509  *	supplied offset.   The range is restricted to the buffer's size.
3510  */
3511 static void
3512 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3513 {
3514 	vm_ooffset_t soff, eoff;
3515 
3516 	/*
3517 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3518 	 * page boundry or cross the end of the buffer.  The end of the
3519 	 * buffer, in this case, is our file EOF, not the allocation size
3520 	 * of the buffer.
3521 	 */
3522 	soff = off;
3523 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3524 	if (eoff > bp->b_offset + bp->b_bcount)
3525 		eoff = bp->b_offset + bp->b_bcount;
3526 
3527 	/*
3528 	 * Set valid range.  This is typically the entire buffer and thus the
3529 	 * entire page.
3530 	 */
3531 	if (eoff > soff) {
3532 		vm_page_set_validclean(
3533 		    m,
3534 		   (vm_offset_t) (soff & PAGE_MASK),
3535 		   (vm_offset_t) (eoff - soff)
3536 		);
3537 	}
3538 }
3539 
3540 /*
3541  * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3542  * any page is busy, drain the flag.
3543  */
3544 static void
3545 vfs_drain_busy_pages(struct buf *bp)
3546 {
3547 	vm_page_t m;
3548 	int i, last_busied;
3549 
3550 	VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED);
3551 	last_busied = 0;
3552 	for (i = 0; i < bp->b_npages; i++) {
3553 		m = bp->b_pages[i];
3554 		if ((m->oflags & VPO_BUSY) != 0) {
3555 			for (; last_busied < i; last_busied++)
3556 				vm_page_busy(bp->b_pages[last_busied]);
3557 			while ((m->oflags & VPO_BUSY) != 0)
3558 				vm_page_sleep(m, "vbpage");
3559 		}
3560 	}
3561 	for (i = 0; i < last_busied; i++)
3562 		vm_page_wakeup(bp->b_pages[i]);
3563 }
3564 
3565 /*
3566  * This routine is called before a device strategy routine.
3567  * It is used to tell the VM system that paging I/O is in
3568  * progress, and treat the pages associated with the buffer
3569  * almost as being VPO_BUSY.  Also the object paging_in_progress
3570  * flag is handled to make sure that the object doesn't become
3571  * inconsistant.
3572  *
3573  * Since I/O has not been initiated yet, certain buffer flags
3574  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3575  * and should be ignored.
3576  */
3577 void
3578 vfs_busy_pages(struct buf *bp, int clear_modify)
3579 {
3580 	int i, bogus;
3581 	vm_object_t obj;
3582 	vm_ooffset_t foff;
3583 	vm_page_t m;
3584 
3585 	if (!(bp->b_flags & B_VMIO))
3586 		return;
3587 
3588 	obj = bp->b_bufobj->bo_object;
3589 	foff = bp->b_offset;
3590 	KASSERT(bp->b_offset != NOOFFSET,
3591 	    ("vfs_busy_pages: no buffer offset"));
3592 	VM_OBJECT_LOCK(obj);
3593 	vfs_drain_busy_pages(bp);
3594 	if (bp->b_bufsize != 0)
3595 		vfs_setdirty_locked_object(bp);
3596 	bogus = 0;
3597 	for (i = 0; i < bp->b_npages; i++) {
3598 		m = bp->b_pages[i];
3599 
3600 		if ((bp->b_flags & B_CLUSTER) == 0) {
3601 			vm_object_pip_add(obj, 1);
3602 			vm_page_io_start(m);
3603 		}
3604 		/*
3605 		 * When readying a buffer for a read ( i.e
3606 		 * clear_modify == 0 ), it is important to do
3607 		 * bogus_page replacement for valid pages in
3608 		 * partially instantiated buffers.  Partially
3609 		 * instantiated buffers can, in turn, occur when
3610 		 * reconstituting a buffer from its VM backing store
3611 		 * base.  We only have to do this if B_CACHE is
3612 		 * clear ( which causes the I/O to occur in the
3613 		 * first place ).  The replacement prevents the read
3614 		 * I/O from overwriting potentially dirty VM-backed
3615 		 * pages.  XXX bogus page replacement is, uh, bogus.
3616 		 * It may not work properly with small-block devices.
3617 		 * We need to find a better way.
3618 		 */
3619 		if (clear_modify) {
3620 			pmap_remove_write(m);
3621 			vfs_page_set_validclean(bp, foff, m);
3622 		} else if (m->valid == VM_PAGE_BITS_ALL &&
3623 		    (bp->b_flags & B_CACHE) == 0) {
3624 			bp->b_pages[i] = bogus_page;
3625 			bogus++;
3626 		}
3627 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3628 	}
3629 	VM_OBJECT_UNLOCK(obj);
3630 	if (bogus)
3631 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3632 		    bp->b_pages, bp->b_npages);
3633 }
3634 
3635 /*
3636  *	vfs_bio_set_valid:
3637  *
3638  *	Set the range within the buffer to valid.  The range is
3639  *	relative to the beginning of the buffer, b_offset.  Note that
3640  *	b_offset itself may be offset from the beginning of the first
3641  *	page.
3642  */
3643 void
3644 vfs_bio_set_valid(struct buf *bp, int base, int size)
3645 {
3646 	int i, n;
3647 	vm_page_t m;
3648 
3649 	if (!(bp->b_flags & B_VMIO))
3650 		return;
3651 
3652 	/*
3653 	 * Fixup base to be relative to beginning of first page.
3654 	 * Set initial n to be the maximum number of bytes in the
3655 	 * first page that can be validated.
3656 	 */
3657 	base += (bp->b_offset & PAGE_MASK);
3658 	n = PAGE_SIZE - (base & PAGE_MASK);
3659 
3660 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3661 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3662 		m = bp->b_pages[i];
3663 		if (n > size)
3664 			n = size;
3665 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
3666 		base += n;
3667 		size -= n;
3668 		n = PAGE_SIZE;
3669 	}
3670 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3671 }
3672 
3673 /*
3674  *	vfs_bio_clrbuf:
3675  *
3676  *	If the specified buffer is a non-VMIO buffer, clear the entire
3677  *	buffer.  If the specified buffer is a VMIO buffer, clear and
3678  *	validate only the previously invalid portions of the buffer.
3679  *	This routine essentially fakes an I/O, so we need to clear
3680  *	BIO_ERROR and B_INVAL.
3681  *
3682  *	Note that while we only theoretically need to clear through b_bcount,
3683  *	we go ahead and clear through b_bufsize.
3684  */
3685 void
3686 vfs_bio_clrbuf(struct buf *bp)
3687 {
3688 	int i, j, mask;
3689 	caddr_t sa, ea;
3690 
3691 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3692 		clrbuf(bp);
3693 		return;
3694 	}
3695 	bp->b_flags &= ~B_INVAL;
3696 	bp->b_ioflags &= ~BIO_ERROR;
3697 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3698 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3699 	    (bp->b_offset & PAGE_MASK) == 0) {
3700 		if (bp->b_pages[0] == bogus_page)
3701 			goto unlock;
3702 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3703 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3704 		if ((bp->b_pages[0]->valid & mask) == mask)
3705 			goto unlock;
3706 		if ((bp->b_pages[0]->valid & mask) == 0) {
3707 			bzero(bp->b_data, bp->b_bufsize);
3708 			bp->b_pages[0]->valid |= mask;
3709 			goto unlock;
3710 		}
3711 	}
3712 	ea = sa = bp->b_data;
3713 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3714 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3715 		ea = (caddr_t)(vm_offset_t)ulmin(
3716 		    (u_long)(vm_offset_t)ea,
3717 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3718 		if (bp->b_pages[i] == bogus_page)
3719 			continue;
3720 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3721 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3722 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3723 		if ((bp->b_pages[i]->valid & mask) == mask)
3724 			continue;
3725 		if ((bp->b_pages[i]->valid & mask) == 0)
3726 			bzero(sa, ea - sa);
3727 		else {
3728 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3729 				if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3730 					bzero(sa, DEV_BSIZE);
3731 			}
3732 		}
3733 		bp->b_pages[i]->valid |= mask;
3734 	}
3735 unlock:
3736 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3737 	bp->b_resid = 0;
3738 }
3739 
3740 /*
3741  * vm_hold_load_pages and vm_hold_free_pages get pages into
3742  * a buffers address space.  The pages are anonymous and are
3743  * not associated with a file object.
3744  */
3745 static void
3746 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3747 {
3748 	vm_offset_t pg;
3749 	vm_page_t p;
3750 	int index;
3751 
3752 	to = round_page(to);
3753 	from = round_page(from);
3754 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3755 
3756 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3757 tryagain:
3758 		/*
3759 		 * note: must allocate system pages since blocking here
3760 		 * could interfere with paging I/O, no matter which
3761 		 * process we are.
3762 		 */
3763 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
3764 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
3765 		if (p == NULL) {
3766 			VM_WAIT;
3767 			goto tryagain;
3768 		}
3769 		pmap_qenter(pg, &p, 1);
3770 		bp->b_pages[index] = p;
3771 	}
3772 	bp->b_npages = index;
3773 }
3774 
3775 /* Return pages associated with this buf to the vm system */
3776 static void
3777 vm_hold_free_pages(struct buf *bp, int newbsize)
3778 {
3779 	vm_offset_t from;
3780 	vm_page_t p;
3781 	int index, newnpages;
3782 
3783 	from = round_page((vm_offset_t)bp->b_data + newbsize);
3784 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3785 	if (bp->b_npages > newnpages)
3786 		pmap_qremove(from, bp->b_npages - newnpages);
3787 	for (index = newnpages; index < bp->b_npages; index++) {
3788 		p = bp->b_pages[index];
3789 		bp->b_pages[index] = NULL;
3790 		if (p->busy != 0)
3791 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3792 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
3793 		p->wire_count--;
3794 		vm_page_free(p);
3795 		atomic_subtract_int(&cnt.v_wire_count, 1);
3796 	}
3797 	bp->b_npages = newnpages;
3798 }
3799 
3800 /*
3801  * Map an IO request into kernel virtual address space.
3802  *
3803  * All requests are (re)mapped into kernel VA space.
3804  * Notice that we use b_bufsize for the size of the buffer
3805  * to be mapped.  b_bcount might be modified by the driver.
3806  *
3807  * Note that even if the caller determines that the address space should
3808  * be valid, a race or a smaller-file mapped into a larger space may
3809  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3810  * check the return value.
3811  */
3812 int
3813 vmapbuf(struct buf *bp)
3814 {
3815 	caddr_t kva;
3816 	vm_prot_t prot;
3817 	int pidx;
3818 
3819 	if (bp->b_bufsize < 0)
3820 		return (-1);
3821 	prot = VM_PROT_READ;
3822 	if (bp->b_iocmd == BIO_READ)
3823 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3824 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
3825 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
3826 	    btoc(MAXPHYS))) < 0)
3827 		return (-1);
3828 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3829 
3830 	kva = bp->b_saveaddr;
3831 	bp->b_npages = pidx;
3832 	bp->b_saveaddr = bp->b_data;
3833 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3834 	return(0);
3835 }
3836 
3837 /*
3838  * Free the io map PTEs associated with this IO operation.
3839  * We also invalidate the TLB entries and restore the original b_addr.
3840  */
3841 void
3842 vunmapbuf(struct buf *bp)
3843 {
3844 	int npages;
3845 
3846 	npages = bp->b_npages;
3847 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3848 	vm_page_unhold_pages(bp->b_pages, npages);
3849 
3850 	bp->b_data = bp->b_saveaddr;
3851 }
3852 
3853 void
3854 bdone(struct buf *bp)
3855 {
3856 	struct mtx *mtxp;
3857 
3858 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3859 	mtx_lock(mtxp);
3860 	bp->b_flags |= B_DONE;
3861 	wakeup(bp);
3862 	mtx_unlock(mtxp);
3863 }
3864 
3865 void
3866 bwait(struct buf *bp, u_char pri, const char *wchan)
3867 {
3868 	struct mtx *mtxp;
3869 
3870 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3871 	mtx_lock(mtxp);
3872 	while ((bp->b_flags & B_DONE) == 0)
3873 		msleep(bp, mtxp, pri, wchan, 0);
3874 	mtx_unlock(mtxp);
3875 }
3876 
3877 int
3878 bufsync(struct bufobj *bo, int waitfor)
3879 {
3880 
3881 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3882 }
3883 
3884 void
3885 bufstrategy(struct bufobj *bo, struct buf *bp)
3886 {
3887 	int i = 0;
3888 	struct vnode *vp;
3889 
3890 	vp = bp->b_vp;
3891 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3892 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3893 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3894 	i = VOP_STRATEGY(vp, bp);
3895 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3896 }
3897 
3898 void
3899 bufobj_wrefl(struct bufobj *bo)
3900 {
3901 
3902 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3903 	ASSERT_BO_LOCKED(bo);
3904 	bo->bo_numoutput++;
3905 }
3906 
3907 void
3908 bufobj_wref(struct bufobj *bo)
3909 {
3910 
3911 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3912 	BO_LOCK(bo);
3913 	bo->bo_numoutput++;
3914 	BO_UNLOCK(bo);
3915 }
3916 
3917 void
3918 bufobj_wdrop(struct bufobj *bo)
3919 {
3920 
3921 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3922 	BO_LOCK(bo);
3923 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3924 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3925 		bo->bo_flag &= ~BO_WWAIT;
3926 		wakeup(&bo->bo_numoutput);
3927 	}
3928 	BO_UNLOCK(bo);
3929 }
3930 
3931 int
3932 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3933 {
3934 	int error;
3935 
3936 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3937 	ASSERT_BO_LOCKED(bo);
3938 	error = 0;
3939 	while (bo->bo_numoutput) {
3940 		bo->bo_flag |= BO_WWAIT;
3941 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3942 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3943 		if (error)
3944 			break;
3945 	}
3946 	return (error);
3947 }
3948 
3949 void
3950 bpin(struct buf *bp)
3951 {
3952 	struct mtx *mtxp;
3953 
3954 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3955 	mtx_lock(mtxp);
3956 	bp->b_pin_count++;
3957 	mtx_unlock(mtxp);
3958 }
3959 
3960 void
3961 bunpin(struct buf *bp)
3962 {
3963 	struct mtx *mtxp;
3964 
3965 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3966 	mtx_lock(mtxp);
3967 	if (--bp->b_pin_count == 0)
3968 		wakeup(bp);
3969 	mtx_unlock(mtxp);
3970 }
3971 
3972 void
3973 bunpin_wait(struct buf *bp)
3974 {
3975 	struct mtx *mtxp;
3976 
3977 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3978 	mtx_lock(mtxp);
3979 	while (bp->b_pin_count > 0)
3980 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
3981 	mtx_unlock(mtxp);
3982 }
3983 
3984 #include "opt_ddb.h"
3985 #ifdef DDB
3986 #include <ddb/ddb.h>
3987 
3988 /* DDB command to show buffer data */
3989 DB_SHOW_COMMAND(buffer, db_show_buffer)
3990 {
3991 	/* get args */
3992 	struct buf *bp = (struct buf *)addr;
3993 
3994 	if (!have_addr) {
3995 		db_printf("usage: show buffer <addr>\n");
3996 		return;
3997 	}
3998 
3999 	db_printf("buf at %p\n", bp);
4000 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4001 	db_printf(
4002 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4003 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4004 	    "b_dep = %p\n",
4005 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4006 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4007 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4008 	if (bp->b_npages) {
4009 		int i;
4010 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4011 		for (i = 0; i < bp->b_npages; i++) {
4012 			vm_page_t m;
4013 			m = bp->b_pages[i];
4014 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4015 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4016 			if ((i + 1) < bp->b_npages)
4017 				db_printf(",");
4018 		}
4019 		db_printf("\n");
4020 	}
4021 	db_printf(" ");
4022 	BUF_LOCKPRINTINFO(bp);
4023 }
4024 
4025 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4026 {
4027 	struct buf *bp;
4028 	int i;
4029 
4030 	for (i = 0; i < nbuf; i++) {
4031 		bp = &buf[i];
4032 		if (BUF_ISLOCKED(bp)) {
4033 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4034 			db_printf("\n");
4035 		}
4036 	}
4037 }
4038 
4039 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4040 {
4041 	struct vnode *vp;
4042 	struct buf *bp;
4043 
4044 	if (!have_addr) {
4045 		db_printf("usage: show vnodebufs <addr>\n");
4046 		return;
4047 	}
4048 	vp = (struct vnode *)addr;
4049 	db_printf("Clean buffers:\n");
4050 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4051 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4052 		db_printf("\n");
4053 	}
4054 	db_printf("Dirty buffers:\n");
4055 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4056 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4057 		db_printf("\n");
4058 	}
4059 }
4060 
4061 DB_COMMAND(countfreebufs, db_coundfreebufs)
4062 {
4063 	struct buf *bp;
4064 	int i, used = 0, nfree = 0;
4065 
4066 	if (have_addr) {
4067 		db_printf("usage: countfreebufs\n");
4068 		return;
4069 	}
4070 
4071 	for (i = 0; i < nbuf; i++) {
4072 		bp = &buf[i];
4073 		if ((bp->b_vflags & BV_INFREECNT) != 0)
4074 			nfree++;
4075 		else
4076 			used++;
4077 	}
4078 
4079 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4080 	    nfree + used);
4081 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4082 }
4083 #endif /* DDB */
4084