xref: /freebsd/sys/kern/vfs_bio.c (revision 5861f9665471e98e544f6fa3ce73c4912229ff82)
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, vm_offset_t from,
99 		vm_offset_t to);
100 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
101 		vm_offset_t to);
102 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
103 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
104 		vm_page_t m);
105 static void vfs_clean_pages(struct buf *bp);
106 static void vfs_setdirty(struct buf *bp);
107 static void vfs_setdirty_locked_object(struct buf *bp);
108 static void vfs_vmio_release(struct buf *bp);
109 static int vfs_bio_clcheck(struct vnode *vp, int size,
110 		daddr_t lblkno, daddr_t blkno);
111 static int buf_do_flush(struct vnode *vp);
112 static int flushbufqueues(struct vnode *, int, int);
113 static void buf_daemon(void);
114 static void bremfreel(struct buf *bp);
115 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
116     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
117 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
118 #endif
119 
120 int vmiodirenable = TRUE;
121 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
122     "Use the VM system for directory writes");
123 long runningbufspace;
124 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
125     "Amount of presently outstanding async buffer io");
126 static long bufspace;
127 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
128     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
129 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
130     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
131 #else
132 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
133     "Virtual memory used for buffers");
134 #endif
135 static long maxbufspace;
136 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
137     "Maximum allowed value of bufspace (including buf_daemon)");
138 static long bufmallocspace;
139 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
140     "Amount of malloced memory for buffers");
141 static long maxbufmallocspace;
142 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
143     "Maximum amount of malloced memory for buffers");
144 static long lobufspace;
145 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
146     "Minimum amount of buffers we want to have");
147 long hibufspace;
148 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
149     "Maximum allowed value of bufspace (excluding buf_daemon)");
150 static int bufreusecnt;
151 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
152     "Number of times we have reused a buffer");
153 static int buffreekvacnt;
154 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
155     "Number of times we have freed the KVA space from some buffer");
156 static int bufdefragcnt;
157 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
158     "Number of times we have had to repeat buffer allocation to defragment");
159 static long lorunningspace;
160 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
161     "Minimum preferred space used for in-progress I/O");
162 static long hirunningspace;
163 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
164     "Maximum amount of space to use for in-progress I/O");
165 int dirtybufferflushes;
166 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
167     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
168 int bdwriteskip;
169 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
170     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
171 int altbufferflushes;
172 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
173     0, "Number of fsync flushes to limit dirty buffers");
174 static int recursiveflushes;
175 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
176     0, "Number of flushes skipped due to being recursive");
177 static int numdirtybuffers;
178 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
179     "Number of buffers that are dirty (has unwritten changes) at the moment");
180 static int lodirtybuffers;
181 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
182     "How many buffers we want to have free before bufdaemon can sleep");
183 static int hidirtybuffers;
184 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
185     "When the number of dirty buffers is considered severe");
186 int dirtybufthresh;
187 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
188     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
189 static int numfreebuffers;
190 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
191     "Number of free buffers");
192 static int lofreebuffers;
193 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
194    "XXX Unused");
195 static int hifreebuffers;
196 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
197    "XXX Complicatedly unused");
198 static int getnewbufcalls;
199 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
200    "Number of calls to getnewbuf");
201 static int getnewbufrestarts;
202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
203     "Number of times getnewbuf has had to restart a buffer aquisition");
204 static int flushbufqtarget = 100;
205 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
206     "Amount of work to do in flushbufqueues when helping bufdaemon");
207 static long notbufdflashes;
208 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, &notbufdflashes, 0,
209     "Number of dirty buffer flushes done by the bufdaemon helpers");
210 
211 /*
212  * Wakeup point for bufdaemon, as well as indicator of whether it is already
213  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
214  * is idling.
215  */
216 static int bd_request;
217 
218 /*
219  * This lock synchronizes access to bd_request.
220  */
221 static struct mtx bdlock;
222 
223 /*
224  * bogus page -- for I/O to/from partially complete buffers
225  * this is a temporary solution to the problem, but it is not
226  * really that bad.  it would be better to split the buffer
227  * for input in the case of buffers partially already in memory,
228  * but the code is intricate enough already.
229  */
230 vm_page_t bogus_page;
231 
232 /*
233  * Synchronization (sleep/wakeup) variable for active buffer space requests.
234  * Set when wait starts, cleared prior to wakeup().
235  * Used in runningbufwakeup() and waitrunningbufspace().
236  */
237 static int runningbufreq;
238 
239 /*
240  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
241  * waitrunningbufspace().
242  */
243 static struct mtx rbreqlock;
244 
245 /*
246  * Synchronization (sleep/wakeup) variable for buffer requests.
247  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
248  * by and/or.
249  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
250  * getnewbuf(), and getblk().
251  */
252 static int needsbuffer;
253 
254 /*
255  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
256  */
257 static struct mtx nblock;
258 
259 /*
260  * Definitions for the buffer free lists.
261  */
262 #define BUFFER_QUEUES	6	/* number of free buffer queues */
263 
264 #define QUEUE_NONE	0	/* on no queue */
265 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
266 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
267 #define QUEUE_DIRTY_GIANT 3	/* B_DELWRI buffers that need giant */
268 #define QUEUE_EMPTYKVA	4	/* empty buffer headers w/KVA assignment */
269 #define QUEUE_EMPTY	5	/* empty buffer headers */
270 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
271 
272 /* Queues for free buffers with various properties */
273 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
274 
275 /* Lock for the bufqueues */
276 static struct mtx bqlock;
277 
278 /*
279  * Single global constant for BUF_WMESG, to avoid getting multiple references.
280  * buf_wmesg is referred from macros.
281  */
282 const char *buf_wmesg = BUF_WMESG;
283 
284 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
285 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
286 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
287 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
288 
289 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
290     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
291 static int
292 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
293 {
294 	long lvalue;
295 	int ivalue;
296 
297 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
298 		return (sysctl_handle_long(oidp, arg1, arg2, req));
299 	lvalue = *(long *)arg1;
300 	if (lvalue > INT_MAX)
301 		/* On overflow, still write out a long to trigger ENOMEM. */
302 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
303 	ivalue = lvalue;
304 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
305 }
306 #endif
307 
308 #ifdef DIRECTIO
309 extern void ffs_rawread_setup(void);
310 #endif /* DIRECTIO */
311 /*
312  *	numdirtywakeup:
313  *
314  *	If someone is blocked due to there being too many dirty buffers,
315  *	and numdirtybuffers is now reasonable, wake them up.
316  */
317 
318 static __inline void
319 numdirtywakeup(int level)
320 {
321 
322 	if (numdirtybuffers <= level) {
323 		mtx_lock(&nblock);
324 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
325 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
326 			wakeup(&needsbuffer);
327 		}
328 		mtx_unlock(&nblock);
329 	}
330 }
331 
332 /*
333  *	bufspacewakeup:
334  *
335  *	Called when buffer space is potentially available for recovery.
336  *	getnewbuf() will block on this flag when it is unable to free
337  *	sufficient buffer space.  Buffer space becomes recoverable when
338  *	bp's get placed back in the queues.
339  */
340 
341 static __inline void
342 bufspacewakeup(void)
343 {
344 
345 	/*
346 	 * If someone is waiting for BUF space, wake them up.  Even
347 	 * though we haven't freed the kva space yet, the waiting
348 	 * process will be able to now.
349 	 */
350 	mtx_lock(&nblock);
351 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
352 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
353 		wakeup(&needsbuffer);
354 	}
355 	mtx_unlock(&nblock);
356 }
357 
358 /*
359  * runningbufwakeup() - in-progress I/O accounting.
360  *
361  */
362 void
363 runningbufwakeup(struct buf *bp)
364 {
365 
366 	if (bp->b_runningbufspace) {
367 		atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
368 		bp->b_runningbufspace = 0;
369 		mtx_lock(&rbreqlock);
370 		if (runningbufreq && runningbufspace <= lorunningspace) {
371 			runningbufreq = 0;
372 			wakeup(&runningbufreq);
373 		}
374 		mtx_unlock(&rbreqlock);
375 	}
376 }
377 
378 /*
379  *	bufcountwakeup:
380  *
381  *	Called when a buffer has been added to one of the free queues to
382  *	account for the buffer and to wakeup anyone waiting for free buffers.
383  *	This typically occurs when large amounts of metadata are being handled
384  *	by the buffer cache ( else buffer space runs out first, usually ).
385  */
386 
387 static __inline void
388 bufcountwakeup(void)
389 {
390 
391 	atomic_add_int(&numfreebuffers, 1);
392 	mtx_lock(&nblock);
393 	if (needsbuffer) {
394 		needsbuffer &= ~VFS_BIO_NEED_ANY;
395 		if (numfreebuffers >= hifreebuffers)
396 			needsbuffer &= ~VFS_BIO_NEED_FREE;
397 		wakeup(&needsbuffer);
398 	}
399 	mtx_unlock(&nblock);
400 }
401 
402 /*
403  *	waitrunningbufspace()
404  *
405  *	runningbufspace is a measure of the amount of I/O currently
406  *	running.  This routine is used in async-write situations to
407  *	prevent creating huge backups of pending writes to a device.
408  *	Only asynchronous writes are governed by this function.
409  *
410  *	Reads will adjust runningbufspace, but will not block based on it.
411  *	The read load has a side effect of reducing the allowed write load.
412  *
413  *	This does NOT turn an async write into a sync write.  It waits
414  *	for earlier writes to complete and generally returns before the
415  *	caller's write has reached the device.
416  */
417 void
418 waitrunningbufspace(void)
419 {
420 
421 	mtx_lock(&rbreqlock);
422 	while (runningbufspace > hirunningspace) {
423 		++runningbufreq;
424 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
425 	}
426 	mtx_unlock(&rbreqlock);
427 }
428 
429 
430 /*
431  *	vfs_buf_test_cache:
432  *
433  *	Called when a buffer is extended.  This function clears the B_CACHE
434  *	bit if the newly extended portion of the buffer does not contain
435  *	valid data.
436  */
437 static __inline
438 void
439 vfs_buf_test_cache(struct buf *bp,
440 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
441 		  vm_page_t m)
442 {
443 
444 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
445 	if (bp->b_flags & B_CACHE) {
446 		int base = (foff + off) & PAGE_MASK;
447 		if (vm_page_is_valid(m, base, size) == 0)
448 			bp->b_flags &= ~B_CACHE;
449 	}
450 }
451 
452 /* Wake up the buffer daemon if necessary */
453 static __inline
454 void
455 bd_wakeup(int dirtybuflevel)
456 {
457 
458 	mtx_lock(&bdlock);
459 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
460 		bd_request = 1;
461 		wakeup(&bd_request);
462 	}
463 	mtx_unlock(&bdlock);
464 }
465 
466 /*
467  * bd_speedup - speedup the buffer cache flushing code
468  */
469 
470 static __inline
471 void
472 bd_speedup(void)
473 {
474 
475 	bd_wakeup(1);
476 }
477 
478 /*
479  * Calculating buffer cache scaling values and reserve space for buffer
480  * headers.  This is called during low level kernel initialization and
481  * may be called more then once.  We CANNOT write to the memory area
482  * being reserved at this time.
483  */
484 caddr_t
485 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
486 {
487 	int tuned_nbuf;
488 	long maxbuf;
489 
490 	/*
491 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
492 	 * PAGE_SIZE is >= 1K)
493 	 */
494 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
495 
496 	/*
497 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
498 	 * For the first 64MB of ram nominally allocate sufficient buffers to
499 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
500 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
501 	 * the buffer cache we limit the eventual kva reservation to
502 	 * maxbcache bytes.
503 	 *
504 	 * factor represents the 1/4 x ram conversion.
505 	 */
506 	if (nbuf == 0) {
507 		int factor = 4 * BKVASIZE / 1024;
508 
509 		nbuf = 50;
510 		if (physmem_est > 4096)
511 			nbuf += min((physmem_est - 4096) / factor,
512 			    65536 / factor);
513 		if (physmem_est > 65536)
514 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
515 
516 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
517 			nbuf = maxbcache / BKVASIZE;
518 		tuned_nbuf = 1;
519 	} else
520 		tuned_nbuf = 0;
521 
522 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
523 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
524 	if (nbuf > maxbuf) {
525 		if (!tuned_nbuf)
526 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
527 			    maxbuf);
528 		nbuf = maxbuf;
529 	}
530 
531 	/*
532 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
533 	 * We have no less then 16 and no more then 256.
534 	 */
535 	nswbuf = max(min(nbuf/4, 256), 16);
536 #ifdef NSWBUF_MIN
537 	if (nswbuf < NSWBUF_MIN)
538 		nswbuf = NSWBUF_MIN;
539 #endif
540 #ifdef DIRECTIO
541 	ffs_rawread_setup();
542 #endif
543 
544 	/*
545 	 * Reserve space for the buffer cache buffers
546 	 */
547 	swbuf = (void *)v;
548 	v = (caddr_t)(swbuf + nswbuf);
549 	buf = (void *)v;
550 	v = (caddr_t)(buf + nbuf);
551 
552 	return(v);
553 }
554 
555 /* Initialize the buffer subsystem.  Called before use of any buffers. */
556 void
557 bufinit(void)
558 {
559 	struct buf *bp;
560 	int i;
561 
562 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
563 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
564 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
565 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
566 
567 	/* next, make a null set of free lists */
568 	for (i = 0; i < BUFFER_QUEUES; i++)
569 		TAILQ_INIT(&bufqueues[i]);
570 
571 	/* finally, initialize each buffer header and stick on empty q */
572 	for (i = 0; i < nbuf; i++) {
573 		bp = &buf[i];
574 		bzero(bp, sizeof *bp);
575 		bp->b_flags = B_INVAL;	/* we're just an empty header */
576 		bp->b_rcred = NOCRED;
577 		bp->b_wcred = NOCRED;
578 		bp->b_qindex = QUEUE_EMPTY;
579 		bp->b_vflags = 0;
580 		bp->b_xflags = 0;
581 		LIST_INIT(&bp->b_dep);
582 		BUF_LOCKINIT(bp);
583 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
584 	}
585 
586 	/*
587 	 * maxbufspace is the absolute maximum amount of buffer space we are
588 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
589 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
590 	 * used by most other processes.  The differential is required to
591 	 * ensure that buf_daemon is able to run when other processes might
592 	 * be blocked waiting for buffer space.
593 	 *
594 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
595 	 * this may result in KVM fragmentation which is not handled optimally
596 	 * by the system.
597 	 */
598 	maxbufspace = (long)nbuf * BKVASIZE;
599 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
600 	lobufspace = hibufspace - MAXBSIZE;
601 
602 	lorunningspace = 512 * 1024;
603 	hirunningspace = 1024 * 1024;
604 
605 /*
606  * Limit the amount of malloc memory since it is wired permanently into
607  * the kernel space.  Even though this is accounted for in the buffer
608  * allocation, we don't want the malloced region to grow uncontrolled.
609  * The malloc scheme improves memory utilization significantly on average
610  * (small) directories.
611  */
612 	maxbufmallocspace = hibufspace / 20;
613 
614 /*
615  * Reduce the chance of a deadlock occuring by limiting the number
616  * of delayed-write dirty buffers we allow to stack up.
617  */
618 	hidirtybuffers = nbuf / 4 + 20;
619 	dirtybufthresh = hidirtybuffers * 9 / 10;
620 	numdirtybuffers = 0;
621 /*
622  * To support extreme low-memory systems, make sure hidirtybuffers cannot
623  * eat up all available buffer space.  This occurs when our minimum cannot
624  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
625  * BKVASIZE'd (8K) buffers.
626  */
627 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
628 		hidirtybuffers >>= 1;
629 	}
630 	lodirtybuffers = hidirtybuffers / 2;
631 
632 /*
633  * Try to keep the number of free buffers in the specified range,
634  * and give special processes (e.g. like buf_daemon) access to an
635  * emergency reserve.
636  */
637 	lofreebuffers = nbuf / 18 + 5;
638 	hifreebuffers = 2 * lofreebuffers;
639 	numfreebuffers = nbuf;
640 
641 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
642 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
643 }
644 
645 /*
646  * bfreekva() - free the kva allocation for a buffer.
647  *
648  *	Since this call frees up buffer space, we call bufspacewakeup().
649  */
650 static void
651 bfreekva(struct buf *bp)
652 {
653 
654 	if (bp->b_kvasize) {
655 		atomic_add_int(&buffreekvacnt, 1);
656 		atomic_subtract_long(&bufspace, bp->b_kvasize);
657 		vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
658 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
659 		bp->b_kvasize = 0;
660 		bufspacewakeup();
661 	}
662 }
663 
664 /*
665  *	bremfree:
666  *
667  *	Mark the buffer for removal from the appropriate free list in brelse.
668  *
669  */
670 void
671 bremfree(struct buf *bp)
672 {
673 
674 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
675 	KASSERT((bp->b_flags & B_REMFREE) == 0,
676 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
677 	KASSERT(bp->b_qindex != QUEUE_NONE,
678 	    ("bremfree: buffer %p not on a queue.", bp));
679 	BUF_ASSERT_HELD(bp);
680 
681 	bp->b_flags |= B_REMFREE;
682 	/* Fixup numfreebuffers count.  */
683 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
684 		atomic_subtract_int(&numfreebuffers, 1);
685 }
686 
687 /*
688  *	bremfreef:
689  *
690  *	Force an immediate removal from a free list.  Used only in nfs when
691  *	it abuses the b_freelist pointer.
692  */
693 void
694 bremfreef(struct buf *bp)
695 {
696 	mtx_lock(&bqlock);
697 	bremfreel(bp);
698 	mtx_unlock(&bqlock);
699 }
700 
701 /*
702  *	bremfreel:
703  *
704  *	Removes a buffer from the free list, must be called with the
705  *	bqlock held.
706  */
707 static void
708 bremfreel(struct buf *bp)
709 {
710 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
711 	    bp, bp->b_vp, bp->b_flags);
712 	KASSERT(bp->b_qindex != QUEUE_NONE,
713 	    ("bremfreel: buffer %p not on a queue.", bp));
714 	BUF_ASSERT_HELD(bp);
715 	mtx_assert(&bqlock, MA_OWNED);
716 
717 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
718 	bp->b_qindex = QUEUE_NONE;
719 	/*
720 	 * If this was a delayed bremfree() we only need to remove the buffer
721 	 * from the queue and return the stats are already done.
722 	 */
723 	if (bp->b_flags & B_REMFREE) {
724 		bp->b_flags &= ~B_REMFREE;
725 		return;
726 	}
727 	/*
728 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
729 	 * delayed-write, the buffer was free and we must decrement
730 	 * numfreebuffers.
731 	 */
732 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
733 		atomic_subtract_int(&numfreebuffers, 1);
734 }
735 
736 
737 /*
738  * Get a buffer with the specified data.  Look in the cache first.  We
739  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
740  * is set, the buffer is valid and we do not have to do anything ( see
741  * getblk() ).  This is really just a special case of breadn().
742  */
743 int
744 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
745     struct buf **bpp)
746 {
747 
748 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
749 }
750 
751 /*
752  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
753  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
754  * the buffer is valid and we do not have to do anything.
755  */
756 void
757 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
758     int cnt, struct ucred * cred)
759 {
760 	struct buf *rabp;
761 	int i;
762 
763 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
764 		if (inmem(vp, *rablkno))
765 			continue;
766 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
767 
768 		if ((rabp->b_flags & B_CACHE) == 0) {
769 			if (!TD_IS_IDLETHREAD(curthread))
770 				curthread->td_ru.ru_inblock++;
771 			rabp->b_flags |= B_ASYNC;
772 			rabp->b_flags &= ~B_INVAL;
773 			rabp->b_ioflags &= ~BIO_ERROR;
774 			rabp->b_iocmd = BIO_READ;
775 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
776 				rabp->b_rcred = crhold(cred);
777 			vfs_busy_pages(rabp, 0);
778 			BUF_KERNPROC(rabp);
779 			rabp->b_iooffset = dbtob(rabp->b_blkno);
780 			bstrategy(rabp);
781 		} else {
782 			brelse(rabp);
783 		}
784 	}
785 }
786 
787 /*
788  * Operates like bread, but also starts asynchronous I/O on
789  * read-ahead blocks.
790  */
791 int
792 breadn(struct vnode * vp, daddr_t blkno, int size,
793     daddr_t * rablkno, int *rabsize,
794     int cnt, struct ucred * cred, struct buf **bpp)
795 {
796 	struct buf *bp;
797 	int rv = 0, readwait = 0;
798 
799 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
800 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
801 
802 	/* if not found in cache, do some I/O */
803 	if ((bp->b_flags & B_CACHE) == 0) {
804 		if (!TD_IS_IDLETHREAD(curthread))
805 			curthread->td_ru.ru_inblock++;
806 		bp->b_iocmd = BIO_READ;
807 		bp->b_flags &= ~B_INVAL;
808 		bp->b_ioflags &= ~BIO_ERROR;
809 		if (bp->b_rcred == NOCRED && cred != NOCRED)
810 			bp->b_rcred = crhold(cred);
811 		vfs_busy_pages(bp, 0);
812 		bp->b_iooffset = dbtob(bp->b_blkno);
813 		bstrategy(bp);
814 		++readwait;
815 	}
816 
817 	breada(vp, rablkno, rabsize, cnt, cred);
818 
819 	if (readwait) {
820 		rv = bufwait(bp);
821 	}
822 	return (rv);
823 }
824 
825 /*
826  * Write, release buffer on completion.  (Done by iodone
827  * if async).  Do not bother writing anything if the buffer
828  * is invalid.
829  *
830  * Note that we set B_CACHE here, indicating that buffer is
831  * fully valid and thus cacheable.  This is true even of NFS
832  * now so we set it generally.  This could be set either here
833  * or in biodone() since the I/O is synchronous.  We put it
834  * here.
835  */
836 int
837 bufwrite(struct buf *bp)
838 {
839 	int oldflags;
840 	struct vnode *vp;
841 	int vp_md;
842 
843 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
844 	if (bp->b_flags & B_INVAL) {
845 		brelse(bp);
846 		return (0);
847 	}
848 
849 	oldflags = bp->b_flags;
850 
851 	BUF_ASSERT_HELD(bp);
852 
853 	if (bp->b_pin_count > 0)
854 		bunpin_wait(bp);
855 
856 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
857 	    ("FFS background buffer should not get here %p", bp));
858 
859 	vp = bp->b_vp;
860 	if (vp)
861 		vp_md = vp->v_vflag & VV_MD;
862 	else
863 		vp_md = 0;
864 
865 	/* Mark the buffer clean */
866 	bundirty(bp);
867 
868 	bp->b_flags &= ~B_DONE;
869 	bp->b_ioflags &= ~BIO_ERROR;
870 	bp->b_flags |= B_CACHE;
871 	bp->b_iocmd = BIO_WRITE;
872 
873 	bufobj_wref(bp->b_bufobj);
874 	vfs_busy_pages(bp, 1);
875 
876 	/*
877 	 * Normal bwrites pipeline writes
878 	 */
879 	bp->b_runningbufspace = bp->b_bufsize;
880 	atomic_add_long(&runningbufspace, bp->b_runningbufspace);
881 
882 	if (!TD_IS_IDLETHREAD(curthread))
883 		curthread->td_ru.ru_oublock++;
884 	if (oldflags & B_ASYNC)
885 		BUF_KERNPROC(bp);
886 	bp->b_iooffset = dbtob(bp->b_blkno);
887 	bstrategy(bp);
888 
889 	if ((oldflags & B_ASYNC) == 0) {
890 		int rtval = bufwait(bp);
891 		brelse(bp);
892 		return (rtval);
893 	} else {
894 		/*
895 		 * don't allow the async write to saturate the I/O
896 		 * system.  We will not deadlock here because
897 		 * we are blocking waiting for I/O that is already in-progress
898 		 * to complete. We do not block here if it is the update
899 		 * or syncer daemon trying to clean up as that can lead
900 		 * to deadlock.
901 		 */
902 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
903 			waitrunningbufspace();
904 	}
905 
906 	return (0);
907 }
908 
909 void
910 bufbdflush(struct bufobj *bo, struct buf *bp)
911 {
912 	struct buf *nbp;
913 
914 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
915 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
916 		altbufferflushes++;
917 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
918 		BO_LOCK(bo);
919 		/*
920 		 * Try to find a buffer to flush.
921 		 */
922 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
923 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
924 			    BUF_LOCK(nbp,
925 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
926 				continue;
927 			if (bp == nbp)
928 				panic("bdwrite: found ourselves");
929 			BO_UNLOCK(bo);
930 			/* Don't countdeps with the bo lock held. */
931 			if (buf_countdeps(nbp, 0)) {
932 				BO_LOCK(bo);
933 				BUF_UNLOCK(nbp);
934 				continue;
935 			}
936 			if (nbp->b_flags & B_CLUSTEROK) {
937 				vfs_bio_awrite(nbp);
938 			} else {
939 				bremfree(nbp);
940 				bawrite(nbp);
941 			}
942 			dirtybufferflushes++;
943 			break;
944 		}
945 		if (nbp == NULL)
946 			BO_UNLOCK(bo);
947 	}
948 }
949 
950 /*
951  * Delayed write. (Buffer is marked dirty).  Do not bother writing
952  * anything if the buffer is marked invalid.
953  *
954  * Note that since the buffer must be completely valid, we can safely
955  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
956  * biodone() in order to prevent getblk from writing the buffer
957  * out synchronously.
958  */
959 void
960 bdwrite(struct buf *bp)
961 {
962 	struct thread *td = curthread;
963 	struct vnode *vp;
964 	struct bufobj *bo;
965 
966 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
967 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
968 	BUF_ASSERT_HELD(bp);
969 
970 	if (bp->b_flags & B_INVAL) {
971 		brelse(bp);
972 		return;
973 	}
974 
975 	/*
976 	 * If we have too many dirty buffers, don't create any more.
977 	 * If we are wildly over our limit, then force a complete
978 	 * cleanup. Otherwise, just keep the situation from getting
979 	 * out of control. Note that we have to avoid a recursive
980 	 * disaster and not try to clean up after our own cleanup!
981 	 */
982 	vp = bp->b_vp;
983 	bo = bp->b_bufobj;
984 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
985 		td->td_pflags |= TDP_INBDFLUSH;
986 		BO_BDFLUSH(bo, bp);
987 		td->td_pflags &= ~TDP_INBDFLUSH;
988 	} else
989 		recursiveflushes++;
990 
991 	bdirty(bp);
992 	/*
993 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
994 	 * true even of NFS now.
995 	 */
996 	bp->b_flags |= B_CACHE;
997 
998 	/*
999 	 * This bmap keeps the system from needing to do the bmap later,
1000 	 * perhaps when the system is attempting to do a sync.  Since it
1001 	 * is likely that the indirect block -- or whatever other datastructure
1002 	 * that the filesystem needs is still in memory now, it is a good
1003 	 * thing to do this.  Note also, that if the pageout daemon is
1004 	 * requesting a sync -- there might not be enough memory to do
1005 	 * the bmap then...  So, this is important to do.
1006 	 */
1007 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1008 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1009 	}
1010 
1011 	/*
1012 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
1013 	 */
1014 	vfs_setdirty(bp);
1015 
1016 	/*
1017 	 * We need to do this here to satisfy the vnode_pager and the
1018 	 * pageout daemon, so that it thinks that the pages have been
1019 	 * "cleaned".  Note that since the pages are in a delayed write
1020 	 * buffer -- the VFS layer "will" see that the pages get written
1021 	 * out on the next sync, or perhaps the cluster will be completed.
1022 	 */
1023 	vfs_clean_pages(bp);
1024 	bqrelse(bp);
1025 
1026 	/*
1027 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
1028 	 * buffers (midpoint between our recovery point and our stall
1029 	 * point).
1030 	 */
1031 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1032 
1033 	/*
1034 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1035 	 * due to the softdep code.
1036 	 */
1037 }
1038 
1039 /*
1040  *	bdirty:
1041  *
1042  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1043  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1044  *	itself to properly update it in the dirty/clean lists.  We mark it
1045  *	B_DONE to ensure that any asynchronization of the buffer properly
1046  *	clears B_DONE ( else a panic will occur later ).
1047  *
1048  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1049  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1050  *	should only be called if the buffer is known-good.
1051  *
1052  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1053  *	count.
1054  *
1055  *	The buffer must be on QUEUE_NONE.
1056  */
1057 void
1058 bdirty(struct buf *bp)
1059 {
1060 
1061 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1062 	    bp, bp->b_vp, bp->b_flags);
1063 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1064 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1065 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1066 	BUF_ASSERT_HELD(bp);
1067 	bp->b_flags &= ~(B_RELBUF);
1068 	bp->b_iocmd = BIO_WRITE;
1069 
1070 	if ((bp->b_flags & B_DELWRI) == 0) {
1071 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1072 		reassignbuf(bp);
1073 		atomic_add_int(&numdirtybuffers, 1);
1074 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1075 	}
1076 }
1077 
1078 /*
1079  *	bundirty:
1080  *
1081  *	Clear B_DELWRI for buffer.
1082  *
1083  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1084  *	count.
1085  *
1086  *	The buffer must be on QUEUE_NONE.
1087  */
1088 
1089 void
1090 bundirty(struct buf *bp)
1091 {
1092 
1093 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1094 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1095 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1096 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1097 	BUF_ASSERT_HELD(bp);
1098 
1099 	if (bp->b_flags & B_DELWRI) {
1100 		bp->b_flags &= ~B_DELWRI;
1101 		reassignbuf(bp);
1102 		atomic_subtract_int(&numdirtybuffers, 1);
1103 		numdirtywakeup(lodirtybuffers);
1104 	}
1105 	/*
1106 	 * Since it is now being written, we can clear its deferred write flag.
1107 	 */
1108 	bp->b_flags &= ~B_DEFERRED;
1109 }
1110 
1111 /*
1112  *	bawrite:
1113  *
1114  *	Asynchronous write.  Start output on a buffer, but do not wait for
1115  *	it to complete.  The buffer is released when the output completes.
1116  *
1117  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1118  *	B_INVAL buffers.  Not us.
1119  */
1120 void
1121 bawrite(struct buf *bp)
1122 {
1123 
1124 	bp->b_flags |= B_ASYNC;
1125 	(void) bwrite(bp);
1126 }
1127 
1128 /*
1129  *	bwillwrite:
1130  *
1131  *	Called prior to the locking of any vnodes when we are expecting to
1132  *	write.  We do not want to starve the buffer cache with too many
1133  *	dirty buffers so we block here.  By blocking prior to the locking
1134  *	of any vnodes we attempt to avoid the situation where a locked vnode
1135  *	prevents the various system daemons from flushing related buffers.
1136  */
1137 
1138 void
1139 bwillwrite(void)
1140 {
1141 
1142 	if (numdirtybuffers >= hidirtybuffers) {
1143 		mtx_lock(&nblock);
1144 		while (numdirtybuffers >= hidirtybuffers) {
1145 			bd_wakeup(1);
1146 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1147 			msleep(&needsbuffer, &nblock,
1148 			    (PRIBIO + 4), "flswai", 0);
1149 		}
1150 		mtx_unlock(&nblock);
1151 	}
1152 }
1153 
1154 /*
1155  * Return true if we have too many dirty buffers.
1156  */
1157 int
1158 buf_dirty_count_severe(void)
1159 {
1160 
1161 	return(numdirtybuffers >= hidirtybuffers);
1162 }
1163 
1164 static __noinline int
1165 buf_vm_page_count_severe(void)
1166 {
1167 
1168 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1169 
1170 	return vm_page_count_severe();
1171 }
1172 
1173 /*
1174  *	brelse:
1175  *
1176  *	Release a busy buffer and, if requested, free its resources.  The
1177  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1178  *	to be accessed later as a cache entity or reused for other purposes.
1179  */
1180 void
1181 brelse(struct buf *bp)
1182 {
1183 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1184 	    bp, bp->b_vp, bp->b_flags);
1185 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1186 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1187 
1188 	if (bp->b_flags & B_MANAGED) {
1189 		bqrelse(bp);
1190 		return;
1191 	}
1192 
1193 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1194 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1195 		/*
1196 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1197 		 * pages from being scrapped.  If the error is anything
1198 		 * other than an I/O error (EIO), assume that retrying
1199 		 * is futile.
1200 		 */
1201 		bp->b_ioflags &= ~BIO_ERROR;
1202 		bdirty(bp);
1203 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1204 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1205 		/*
1206 		 * Either a failed I/O or we were asked to free or not
1207 		 * cache the buffer.
1208 		 */
1209 		bp->b_flags |= B_INVAL;
1210 		if (!LIST_EMPTY(&bp->b_dep))
1211 			buf_deallocate(bp);
1212 		if (bp->b_flags & B_DELWRI) {
1213 			atomic_subtract_int(&numdirtybuffers, 1);
1214 			numdirtywakeup(lodirtybuffers);
1215 		}
1216 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1217 		if ((bp->b_flags & B_VMIO) == 0) {
1218 			if (bp->b_bufsize)
1219 				allocbuf(bp, 0);
1220 			if (bp->b_vp)
1221 				brelvp(bp);
1222 		}
1223 	}
1224 
1225 	/*
1226 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1227 	 * is called with B_DELWRI set, the underlying pages may wind up
1228 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1229 	 * because pages associated with a B_DELWRI bp are marked clean.
1230 	 *
1231 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1232 	 * if B_DELWRI is set.
1233 	 *
1234 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1235 	 * on pages to return pages to the VM page queues.
1236 	 */
1237 	if (bp->b_flags & B_DELWRI)
1238 		bp->b_flags &= ~B_RELBUF;
1239 	else if (buf_vm_page_count_severe()) {
1240 		/*
1241 		 * The locking of the BO_LOCK is not necessary since
1242 		 * BKGRDINPROG cannot be set while we hold the buf
1243 		 * lock, it can only be cleared if it is already
1244 		 * pending.
1245 		 */
1246 		if (bp->b_vp) {
1247 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1248 				bp->b_flags |= B_RELBUF;
1249 		} else
1250 			bp->b_flags |= B_RELBUF;
1251 	}
1252 
1253 	/*
1254 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1255 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1256 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1257 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1258 	 *
1259 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1260 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1261 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1262 	 *
1263 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1264 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1265 	 * the commit state and we cannot afford to lose the buffer. If the
1266 	 * buffer has a background write in progress, we need to keep it
1267 	 * around to prevent it from being reconstituted and starting a second
1268 	 * background write.
1269 	 */
1270 	if ((bp->b_flags & B_VMIO)
1271 	    && !(bp->b_vp->v_mount != NULL &&
1272 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1273 		 !vn_isdisk(bp->b_vp, NULL) &&
1274 		 (bp->b_flags & B_DELWRI))
1275 	    ) {
1276 
1277 		int i, j, resid;
1278 		vm_page_t m;
1279 		off_t foff;
1280 		vm_pindex_t poff;
1281 		vm_object_t obj;
1282 
1283 		obj = bp->b_bufobj->bo_object;
1284 
1285 		/*
1286 		 * Get the base offset and length of the buffer.  Note that
1287 		 * in the VMIO case if the buffer block size is not
1288 		 * page-aligned then b_data pointer may not be page-aligned.
1289 		 * But our b_pages[] array *IS* page aligned.
1290 		 *
1291 		 * block sizes less then DEV_BSIZE (usually 512) are not
1292 		 * supported due to the page granularity bits (m->valid,
1293 		 * m->dirty, etc...).
1294 		 *
1295 		 * See man buf(9) for more information
1296 		 */
1297 		resid = bp->b_bufsize;
1298 		foff = bp->b_offset;
1299 		VM_OBJECT_LOCK(obj);
1300 		for (i = 0; i < bp->b_npages; i++) {
1301 			int had_bogus = 0;
1302 
1303 			m = bp->b_pages[i];
1304 
1305 			/*
1306 			 * If we hit a bogus page, fixup *all* the bogus pages
1307 			 * now.
1308 			 */
1309 			if (m == bogus_page) {
1310 				poff = OFF_TO_IDX(bp->b_offset);
1311 				had_bogus = 1;
1312 
1313 				for (j = i; j < bp->b_npages; j++) {
1314 					vm_page_t mtmp;
1315 					mtmp = bp->b_pages[j];
1316 					if (mtmp == bogus_page) {
1317 						mtmp = vm_page_lookup(obj, poff + j);
1318 						if (!mtmp) {
1319 							panic("brelse: page missing\n");
1320 						}
1321 						bp->b_pages[j] = mtmp;
1322 					}
1323 				}
1324 
1325 				if ((bp->b_flags & B_INVAL) == 0) {
1326 					pmap_qenter(
1327 					    trunc_page((vm_offset_t)bp->b_data),
1328 					    bp->b_pages, bp->b_npages);
1329 				}
1330 				m = bp->b_pages[i];
1331 			}
1332 			if ((bp->b_flags & B_NOCACHE) ||
1333 			    (bp->b_ioflags & BIO_ERROR)) {
1334 				int poffset = foff & PAGE_MASK;
1335 				int presid = resid > (PAGE_SIZE - poffset) ?
1336 					(PAGE_SIZE - poffset) : resid;
1337 
1338 				KASSERT(presid >= 0, ("brelse: extra page"));
1339 				vm_page_lock_queues();
1340 				vm_page_set_invalid(m, poffset, presid);
1341 				vm_page_unlock_queues();
1342 				if (had_bogus)
1343 					printf("avoided corruption bug in bogus_page/brelse code\n");
1344 			}
1345 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1346 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1347 		}
1348 		VM_OBJECT_UNLOCK(obj);
1349 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1350 			vfs_vmio_release(bp);
1351 
1352 	} else if (bp->b_flags & B_VMIO) {
1353 
1354 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1355 			vfs_vmio_release(bp);
1356 		}
1357 
1358 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1359 		if (bp->b_bufsize != 0)
1360 			allocbuf(bp, 0);
1361 		if (bp->b_vp != NULL)
1362 			brelvp(bp);
1363 	}
1364 
1365 	if (BUF_LOCKRECURSED(bp)) {
1366 		/* do not release to free list */
1367 		BUF_UNLOCK(bp);
1368 		return;
1369 	}
1370 
1371 	/* enqueue */
1372 	mtx_lock(&bqlock);
1373 	/* Handle delayed bremfree() processing. */
1374 	if (bp->b_flags & B_REMFREE)
1375 		bremfreel(bp);
1376 	if (bp->b_qindex != QUEUE_NONE)
1377 		panic("brelse: free buffer onto another queue???");
1378 
1379 	/*
1380 	 * If the buffer has junk contents signal it and eventually
1381 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1382 	 * doesn't find it.
1383 	 */
1384 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1385 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1386 		bp->b_flags |= B_INVAL;
1387 	if (bp->b_flags & B_INVAL) {
1388 		if (bp->b_flags & B_DELWRI)
1389 			bundirty(bp);
1390 		if (bp->b_vp)
1391 			brelvp(bp);
1392 	}
1393 
1394 	/* buffers with no memory */
1395 	if (bp->b_bufsize == 0) {
1396 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1397 		if (bp->b_vflags & BV_BKGRDINPROG)
1398 			panic("losing buffer 1");
1399 		if (bp->b_kvasize) {
1400 			bp->b_qindex = QUEUE_EMPTYKVA;
1401 		} else {
1402 			bp->b_qindex = QUEUE_EMPTY;
1403 		}
1404 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1405 	/* buffers with junk contents */
1406 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1407 	    (bp->b_ioflags & BIO_ERROR)) {
1408 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1409 		if (bp->b_vflags & BV_BKGRDINPROG)
1410 			panic("losing buffer 2");
1411 		bp->b_qindex = QUEUE_CLEAN;
1412 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1413 	/* remaining buffers */
1414 	} else {
1415 		if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1416 		    (B_DELWRI|B_NEEDSGIANT))
1417 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1418 		else if (bp->b_flags & B_DELWRI)
1419 			bp->b_qindex = QUEUE_DIRTY;
1420 		else
1421 			bp->b_qindex = QUEUE_CLEAN;
1422 		if (bp->b_flags & B_AGE)
1423 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1424 		else
1425 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1426 	}
1427 	mtx_unlock(&bqlock);
1428 
1429 	/*
1430 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1431 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1432 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1433 	 * if B_INVAL is set ).
1434 	 */
1435 
1436 	if (!(bp->b_flags & B_DELWRI))
1437 		bufcountwakeup();
1438 
1439 	/*
1440 	 * Something we can maybe free or reuse
1441 	 */
1442 	if (bp->b_bufsize || bp->b_kvasize)
1443 		bufspacewakeup();
1444 
1445 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1446 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1447 		panic("brelse: not dirty");
1448 	/* unlock */
1449 	BUF_UNLOCK(bp);
1450 }
1451 
1452 /*
1453  * Release a buffer back to the appropriate queue but do not try to free
1454  * it.  The buffer is expected to be used again soon.
1455  *
1456  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1457  * biodone() to requeue an async I/O on completion.  It is also used when
1458  * known good buffers need to be requeued but we think we may need the data
1459  * again soon.
1460  *
1461  * XXX we should be able to leave the B_RELBUF hint set on completion.
1462  */
1463 void
1464 bqrelse(struct buf *bp)
1465 {
1466 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1467 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1468 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1469 
1470 	if (BUF_LOCKRECURSED(bp)) {
1471 		/* do not release to free list */
1472 		BUF_UNLOCK(bp);
1473 		return;
1474 	}
1475 
1476 	if (bp->b_flags & B_MANAGED) {
1477 		if (bp->b_flags & B_REMFREE) {
1478 			mtx_lock(&bqlock);
1479 			bremfreel(bp);
1480 			mtx_unlock(&bqlock);
1481 		}
1482 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1483 		BUF_UNLOCK(bp);
1484 		return;
1485 	}
1486 
1487 	mtx_lock(&bqlock);
1488 	/* Handle delayed bremfree() processing. */
1489 	if (bp->b_flags & B_REMFREE)
1490 		bremfreel(bp);
1491 	if (bp->b_qindex != QUEUE_NONE)
1492 		panic("bqrelse: free buffer onto another queue???");
1493 	/* buffers with stale but valid contents */
1494 	if (bp->b_flags & B_DELWRI) {
1495 		if (bp->b_flags & B_NEEDSGIANT)
1496 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1497 		else
1498 			bp->b_qindex = QUEUE_DIRTY;
1499 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1500 	} else {
1501 		/*
1502 		 * The locking of the BO_LOCK for checking of the
1503 		 * BV_BKGRDINPROG is not necessary since the
1504 		 * BV_BKGRDINPROG cannot be set while we hold the buf
1505 		 * lock, it can only be cleared if it is already
1506 		 * pending.
1507 		 */
1508 		if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1509 			bp->b_qindex = QUEUE_CLEAN;
1510 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1511 			    b_freelist);
1512 		} else {
1513 			/*
1514 			 * We are too low on memory, we have to try to free
1515 			 * the buffer (most importantly: the wired pages
1516 			 * making up its backing store) *now*.
1517 			 */
1518 			mtx_unlock(&bqlock);
1519 			brelse(bp);
1520 			return;
1521 		}
1522 	}
1523 	mtx_unlock(&bqlock);
1524 
1525 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1526 		bufcountwakeup();
1527 
1528 	/*
1529 	 * Something we can maybe free or reuse.
1530 	 */
1531 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1532 		bufspacewakeup();
1533 
1534 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1535 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1536 		panic("bqrelse: not dirty");
1537 	/* unlock */
1538 	BUF_UNLOCK(bp);
1539 }
1540 
1541 /* Give pages used by the bp back to the VM system (where possible) */
1542 static void
1543 vfs_vmio_release(struct buf *bp)
1544 {
1545 	int i;
1546 	vm_page_t m;
1547 
1548 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1549 	vm_page_lock_queues();
1550 	for (i = 0; i < bp->b_npages; i++) {
1551 		m = bp->b_pages[i];
1552 		bp->b_pages[i] = NULL;
1553 		/*
1554 		 * In order to keep page LRU ordering consistent, put
1555 		 * everything on the inactive queue.
1556 		 */
1557 		vm_page_unwire(m, 0);
1558 		/*
1559 		 * We don't mess with busy pages, it is
1560 		 * the responsibility of the process that
1561 		 * busied the pages to deal with them.
1562 		 */
1563 		if ((m->oflags & VPO_BUSY) || (m->busy != 0))
1564 			continue;
1565 
1566 		if (m->wire_count == 0) {
1567 			/*
1568 			 * Might as well free the page if we can and it has
1569 			 * no valid data.  We also free the page if the
1570 			 * buffer was used for direct I/O
1571 			 */
1572 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1573 			    m->hold_count == 0) {
1574 				vm_page_free(m);
1575 			} else if (bp->b_flags & B_DIRECT) {
1576 				vm_page_try_to_free(m);
1577 			} else if (buf_vm_page_count_severe()) {
1578 				vm_page_try_to_cache(m);
1579 			}
1580 		}
1581 	}
1582 	vm_page_unlock_queues();
1583 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1584 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1585 
1586 	if (bp->b_bufsize) {
1587 		bufspacewakeup();
1588 		bp->b_bufsize = 0;
1589 	}
1590 	bp->b_npages = 0;
1591 	bp->b_flags &= ~B_VMIO;
1592 	if (bp->b_vp)
1593 		brelvp(bp);
1594 }
1595 
1596 /*
1597  * Check to see if a block at a particular lbn is available for a clustered
1598  * write.
1599  */
1600 static int
1601 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1602 {
1603 	struct buf *bpa;
1604 	int match;
1605 
1606 	match = 0;
1607 
1608 	/* If the buf isn't in core skip it */
1609 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1610 		return (0);
1611 
1612 	/* If the buf is busy we don't want to wait for it */
1613 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1614 		return (0);
1615 
1616 	/* Only cluster with valid clusterable delayed write buffers */
1617 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1618 	    (B_DELWRI | B_CLUSTEROK))
1619 		goto done;
1620 
1621 	if (bpa->b_bufsize != size)
1622 		goto done;
1623 
1624 	/*
1625 	 * Check to see if it is in the expected place on disk and that the
1626 	 * block has been mapped.
1627 	 */
1628 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1629 		match = 1;
1630 done:
1631 	BUF_UNLOCK(bpa);
1632 	return (match);
1633 }
1634 
1635 /*
1636  *	vfs_bio_awrite:
1637  *
1638  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1639  *	This is much better then the old way of writing only one buffer at
1640  *	a time.  Note that we may not be presented with the buffers in the
1641  *	correct order, so we search for the cluster in both directions.
1642  */
1643 int
1644 vfs_bio_awrite(struct buf *bp)
1645 {
1646 	struct bufobj *bo;
1647 	int i;
1648 	int j;
1649 	daddr_t lblkno = bp->b_lblkno;
1650 	struct vnode *vp = bp->b_vp;
1651 	int ncl;
1652 	int nwritten;
1653 	int size;
1654 	int maxcl;
1655 
1656 	bo = &vp->v_bufobj;
1657 	/*
1658 	 * right now we support clustered writing only to regular files.  If
1659 	 * we find a clusterable block we could be in the middle of a cluster
1660 	 * rather then at the beginning.
1661 	 */
1662 	if ((vp->v_type == VREG) &&
1663 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1664 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1665 
1666 		size = vp->v_mount->mnt_stat.f_iosize;
1667 		maxcl = MAXPHYS / size;
1668 
1669 		BO_LOCK(bo);
1670 		for (i = 1; i < maxcl; i++)
1671 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1672 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1673 				break;
1674 
1675 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1676 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1677 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1678 				break;
1679 		BO_UNLOCK(bo);
1680 		--j;
1681 		ncl = i + j;
1682 		/*
1683 		 * this is a possible cluster write
1684 		 */
1685 		if (ncl != 1) {
1686 			BUF_UNLOCK(bp);
1687 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1688 			return nwritten;
1689 		}
1690 	}
1691 	bremfree(bp);
1692 	bp->b_flags |= B_ASYNC;
1693 	/*
1694 	 * default (old) behavior, writing out only one block
1695 	 *
1696 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1697 	 */
1698 	nwritten = bp->b_bufsize;
1699 	(void) bwrite(bp);
1700 
1701 	return nwritten;
1702 }
1703 
1704 /*
1705  *	getnewbuf:
1706  *
1707  *	Find and initialize a new buffer header, freeing up existing buffers
1708  *	in the bufqueues as necessary.  The new buffer is returned locked.
1709  *
1710  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1711  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1712  *
1713  *	We block if:
1714  *		We have insufficient buffer headers
1715  *		We have insufficient buffer space
1716  *		buffer_map is too fragmented ( space reservation fails )
1717  *		If we have to flush dirty buffers ( but we try to avoid this )
1718  *
1719  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1720  *	Instead we ask the buf daemon to do it for us.  We attempt to
1721  *	avoid piecemeal wakeups of the pageout daemon.
1722  */
1723 
1724 static struct buf *
1725 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1726     int gbflags)
1727 {
1728 	struct thread *td;
1729 	struct buf *bp;
1730 	struct buf *nbp;
1731 	int defrag = 0;
1732 	int nqindex;
1733 	static int flushingbufs;
1734 
1735 	td = curthread;
1736 	/*
1737 	 * We can't afford to block since we might be holding a vnode lock,
1738 	 * which may prevent system daemons from running.  We deal with
1739 	 * low-memory situations by proactively returning memory and running
1740 	 * async I/O rather then sync I/O.
1741 	 */
1742 	atomic_add_int(&getnewbufcalls, 1);
1743 	atomic_subtract_int(&getnewbufrestarts, 1);
1744 restart:
1745 	atomic_add_int(&getnewbufrestarts, 1);
1746 
1747 	/*
1748 	 * Setup for scan.  If we do not have enough free buffers,
1749 	 * we setup a degenerate case that immediately fails.  Note
1750 	 * that if we are specially marked process, we are allowed to
1751 	 * dip into our reserves.
1752 	 *
1753 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1754 	 *
1755 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1756 	 * However, there are a number of cases (defragging, reusing, ...)
1757 	 * where we cannot backup.
1758 	 */
1759 	mtx_lock(&bqlock);
1760 	nqindex = QUEUE_EMPTYKVA;
1761 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1762 
1763 	if (nbp == NULL) {
1764 		/*
1765 		 * If no EMPTYKVA buffers and we are either
1766 		 * defragging or reusing, locate a CLEAN buffer
1767 		 * to free or reuse.  If bufspace useage is low
1768 		 * skip this step so we can allocate a new buffer.
1769 		 */
1770 		if (defrag || bufspace >= lobufspace) {
1771 			nqindex = QUEUE_CLEAN;
1772 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1773 		}
1774 
1775 		/*
1776 		 * If we could not find or were not allowed to reuse a
1777 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1778 		 * buffer.  We can only use an EMPTY buffer if allocating
1779 		 * its KVA would not otherwise run us out of buffer space.
1780 		 */
1781 		if (nbp == NULL && defrag == 0 &&
1782 		    bufspace + maxsize < hibufspace) {
1783 			nqindex = QUEUE_EMPTY;
1784 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1785 		}
1786 	}
1787 
1788 	/*
1789 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1790 	 * depending.
1791 	 */
1792 
1793 	while ((bp = nbp) != NULL) {
1794 		int qindex = nqindex;
1795 
1796 		/*
1797 		 * Calculate next bp ( we can only use it if we do not block
1798 		 * or do other fancy things ).
1799 		 */
1800 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1801 			switch(qindex) {
1802 			case QUEUE_EMPTY:
1803 				nqindex = QUEUE_EMPTYKVA;
1804 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1805 					break;
1806 				/* FALLTHROUGH */
1807 			case QUEUE_EMPTYKVA:
1808 				nqindex = QUEUE_CLEAN;
1809 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1810 					break;
1811 				/* FALLTHROUGH */
1812 			case QUEUE_CLEAN:
1813 				/*
1814 				 * nbp is NULL.
1815 				 */
1816 				break;
1817 			}
1818 		}
1819 		/*
1820 		 * If we are defragging then we need a buffer with
1821 		 * b_kvasize != 0.  XXX this situation should no longer
1822 		 * occur, if defrag is non-zero the buffer's b_kvasize
1823 		 * should also be non-zero at this point.  XXX
1824 		 */
1825 		if (defrag && bp->b_kvasize == 0) {
1826 			printf("Warning: defrag empty buffer %p\n", bp);
1827 			continue;
1828 		}
1829 
1830 		/*
1831 		 * Start freeing the bp.  This is somewhat involved.  nbp
1832 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1833 		 */
1834 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1835 			continue;
1836 		if (bp->b_vp) {
1837 			BO_LOCK(bp->b_bufobj);
1838 			if (bp->b_vflags & BV_BKGRDINPROG) {
1839 				BO_UNLOCK(bp->b_bufobj);
1840 				BUF_UNLOCK(bp);
1841 				continue;
1842 			}
1843 			BO_UNLOCK(bp->b_bufobj);
1844 		}
1845 		CTR6(KTR_BUF,
1846 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1847 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1848 		    bp->b_kvasize, bp->b_bufsize, qindex);
1849 
1850 		/*
1851 		 * Sanity Checks
1852 		 */
1853 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1854 
1855 		/*
1856 		 * Note: we no longer distinguish between VMIO and non-VMIO
1857 		 * buffers.
1858 		 */
1859 
1860 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1861 
1862 		bremfreel(bp);
1863 		mtx_unlock(&bqlock);
1864 
1865 		if (qindex == QUEUE_CLEAN) {
1866 			if (bp->b_flags & B_VMIO) {
1867 				bp->b_flags &= ~B_ASYNC;
1868 				vfs_vmio_release(bp);
1869 			}
1870 			if (bp->b_vp)
1871 				brelvp(bp);
1872 		}
1873 
1874 		/*
1875 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1876 		 * the scan from this point on.
1877 		 *
1878 		 * Get the rest of the buffer freed up.  b_kva* is still
1879 		 * valid after this operation.
1880 		 */
1881 
1882 		if (bp->b_rcred != NOCRED) {
1883 			crfree(bp->b_rcred);
1884 			bp->b_rcred = NOCRED;
1885 		}
1886 		if (bp->b_wcred != NOCRED) {
1887 			crfree(bp->b_wcred);
1888 			bp->b_wcred = NOCRED;
1889 		}
1890 		if (!LIST_EMPTY(&bp->b_dep))
1891 			buf_deallocate(bp);
1892 		if (bp->b_vflags & BV_BKGRDINPROG)
1893 			panic("losing buffer 3");
1894 		KASSERT(bp->b_vp == NULL,
1895 		    ("bp: %p still has vnode %p.  qindex: %d",
1896 		    bp, bp->b_vp, qindex));
1897 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1898 		   ("bp: %p still on a buffer list. xflags %X",
1899 		    bp, bp->b_xflags));
1900 
1901 		if (bp->b_bufsize)
1902 			allocbuf(bp, 0);
1903 
1904 		bp->b_flags = 0;
1905 		bp->b_ioflags = 0;
1906 		bp->b_xflags = 0;
1907 		bp->b_vflags = 0;
1908 		bp->b_vp = NULL;
1909 		bp->b_blkno = bp->b_lblkno = 0;
1910 		bp->b_offset = NOOFFSET;
1911 		bp->b_iodone = 0;
1912 		bp->b_error = 0;
1913 		bp->b_resid = 0;
1914 		bp->b_bcount = 0;
1915 		bp->b_npages = 0;
1916 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1917 		bp->b_bufobj = NULL;
1918 		bp->b_pin_count = 0;
1919 		bp->b_fsprivate1 = NULL;
1920 		bp->b_fsprivate2 = NULL;
1921 		bp->b_fsprivate3 = NULL;
1922 
1923 		LIST_INIT(&bp->b_dep);
1924 
1925 		/*
1926 		 * If we are defragging then free the buffer.
1927 		 */
1928 		if (defrag) {
1929 			bp->b_flags |= B_INVAL;
1930 			bfreekva(bp);
1931 			brelse(bp);
1932 			defrag = 0;
1933 			goto restart;
1934 		}
1935 
1936 		/*
1937 		 * Notify any waiters for the buffer lock about
1938 		 * identity change by freeing the buffer.
1939 		 */
1940 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
1941 			bp->b_flags |= B_INVAL;
1942 			bfreekva(bp);
1943 			brelse(bp);
1944 			goto restart;
1945 		}
1946 
1947 		/*
1948 		 * If we are overcomitted then recover the buffer and its
1949 		 * KVM space.  This occurs in rare situations when multiple
1950 		 * processes are blocked in getnewbuf() or allocbuf().
1951 		 */
1952 		if (bufspace >= hibufspace)
1953 			flushingbufs = 1;
1954 		if (flushingbufs && bp->b_kvasize != 0) {
1955 			bp->b_flags |= B_INVAL;
1956 			bfreekva(bp);
1957 			brelse(bp);
1958 			goto restart;
1959 		}
1960 		if (bufspace < lobufspace)
1961 			flushingbufs = 0;
1962 		break;
1963 	}
1964 
1965 	/*
1966 	 * If we exhausted our list, sleep as appropriate.  We may have to
1967 	 * wakeup various daemons and write out some dirty buffers.
1968 	 *
1969 	 * Generally we are sleeping due to insufficient buffer space.
1970 	 */
1971 
1972 	if (bp == NULL) {
1973 		int flags, norunbuf;
1974 		char *waitmsg;
1975 		int fl;
1976 
1977 		if (defrag) {
1978 			flags = VFS_BIO_NEED_BUFSPACE;
1979 			waitmsg = "nbufkv";
1980 		} else if (bufspace >= hibufspace) {
1981 			waitmsg = "nbufbs";
1982 			flags = VFS_BIO_NEED_BUFSPACE;
1983 		} else {
1984 			waitmsg = "newbuf";
1985 			flags = VFS_BIO_NEED_ANY;
1986 		}
1987 		mtx_lock(&nblock);
1988 		needsbuffer |= flags;
1989 		mtx_unlock(&nblock);
1990 		mtx_unlock(&bqlock);
1991 
1992 		bd_speedup();	/* heeeelp */
1993 		if (gbflags & GB_NOWAIT_BD)
1994 			return (NULL);
1995 
1996 		mtx_lock(&nblock);
1997 		while (needsbuffer & flags) {
1998 			if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
1999 				mtx_unlock(&nblock);
2000 				/*
2001 				 * getblk() is called with a vnode
2002 				 * locked, and some majority of the
2003 				 * dirty buffers may as well belong to
2004 				 * the vnode. Flushing the buffers
2005 				 * there would make a progress that
2006 				 * cannot be achieved by the
2007 				 * buf_daemon, that cannot lock the
2008 				 * vnode.
2009 				 */
2010 				norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2011 				    (td->td_pflags & TDP_NORUNNINGBUF);
2012 				/* play bufdaemon */
2013 				td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2014 				fl = buf_do_flush(vp);
2015 				td->td_pflags &= norunbuf;
2016 				mtx_lock(&nblock);
2017 				if (fl != 0)
2018 					continue;
2019 				if ((needsbuffer & flags) == 0)
2020 					break;
2021 			}
2022 			if (msleep(&needsbuffer, &nblock,
2023 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2024 				mtx_unlock(&nblock);
2025 				return (NULL);
2026 			}
2027 		}
2028 		mtx_unlock(&nblock);
2029 	} else {
2030 		/*
2031 		 * We finally have a valid bp.  We aren't quite out of the
2032 		 * woods, we still have to reserve kva space.  In order
2033 		 * to keep fragmentation sane we only allocate kva in
2034 		 * BKVASIZE chunks.
2035 		 */
2036 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2037 
2038 		if (maxsize != bp->b_kvasize) {
2039 			vm_offset_t addr = 0;
2040 
2041 			bfreekva(bp);
2042 
2043 			vm_map_lock(buffer_map);
2044 			if (vm_map_findspace(buffer_map,
2045 				vm_map_min(buffer_map), maxsize, &addr)) {
2046 				/*
2047 				 * Uh oh.  Buffer map is to fragmented.  We
2048 				 * must defragment the map.
2049 				 */
2050 				atomic_add_int(&bufdefragcnt, 1);
2051 				vm_map_unlock(buffer_map);
2052 				defrag = 1;
2053 				bp->b_flags |= B_INVAL;
2054 				brelse(bp);
2055 				goto restart;
2056 			}
2057 			if (addr) {
2058 				vm_map_insert(buffer_map, NULL, 0,
2059 					addr, addr + maxsize,
2060 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2061 
2062 				bp->b_kvabase = (caddr_t) addr;
2063 				bp->b_kvasize = maxsize;
2064 				atomic_add_long(&bufspace, bp->b_kvasize);
2065 				atomic_add_int(&bufreusecnt, 1);
2066 			}
2067 			vm_map_unlock(buffer_map);
2068 		}
2069 		bp->b_saveaddr = bp->b_kvabase;
2070 		bp->b_data = bp->b_saveaddr;
2071 	}
2072 	return(bp);
2073 }
2074 
2075 /*
2076  *	buf_daemon:
2077  *
2078  *	buffer flushing daemon.  Buffers are normally flushed by the
2079  *	update daemon but if it cannot keep up this process starts to
2080  *	take the load in an attempt to prevent getnewbuf() from blocking.
2081  */
2082 
2083 static struct kproc_desc buf_kp = {
2084 	"bufdaemon",
2085 	buf_daemon,
2086 	&bufdaemonproc
2087 };
2088 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2089 
2090 static int
2091 buf_do_flush(struct vnode *vp)
2092 {
2093 	int flushed;
2094 
2095 	flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2096 	/* The list empty check here is slightly racy */
2097 	if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2098 		mtx_lock(&Giant);
2099 		flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0);
2100 		mtx_unlock(&Giant);
2101 	}
2102 	if (flushed == 0) {
2103 		/*
2104 		 * Could not find any buffers without rollback
2105 		 * dependencies, so just write the first one
2106 		 * in the hopes of eventually making progress.
2107 		 */
2108 		flushbufqueues(vp, QUEUE_DIRTY, 1);
2109 		if (!TAILQ_EMPTY(
2110 			    &bufqueues[QUEUE_DIRTY_GIANT])) {
2111 			mtx_lock(&Giant);
2112 			flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1);
2113 			mtx_unlock(&Giant);
2114 		}
2115 	}
2116 	return (flushed);
2117 }
2118 
2119 static void
2120 buf_daemon()
2121 {
2122 
2123 	/*
2124 	 * This process needs to be suspended prior to shutdown sync.
2125 	 */
2126 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2127 	    SHUTDOWN_PRI_LAST);
2128 
2129 	/*
2130 	 * This process is allowed to take the buffer cache to the limit
2131 	 */
2132 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2133 	mtx_lock(&bdlock);
2134 	for (;;) {
2135 		bd_request = 0;
2136 		mtx_unlock(&bdlock);
2137 
2138 		kproc_suspend_check(bufdaemonproc);
2139 
2140 		/*
2141 		 * Do the flush.  Limit the amount of in-transit I/O we
2142 		 * allow to build up, otherwise we would completely saturate
2143 		 * the I/O system.  Wakeup any waiting processes before we
2144 		 * normally would so they can run in parallel with our drain.
2145 		 */
2146 		while (numdirtybuffers > lodirtybuffers) {
2147 			if (buf_do_flush(NULL) == 0)
2148 				break;
2149 			uio_yield();
2150 		}
2151 
2152 		/*
2153 		 * Only clear bd_request if we have reached our low water
2154 		 * mark.  The buf_daemon normally waits 1 second and
2155 		 * then incrementally flushes any dirty buffers that have
2156 		 * built up, within reason.
2157 		 *
2158 		 * If we were unable to hit our low water mark and couldn't
2159 		 * find any flushable buffers, we sleep half a second.
2160 		 * Otherwise we loop immediately.
2161 		 */
2162 		mtx_lock(&bdlock);
2163 		if (numdirtybuffers <= lodirtybuffers) {
2164 			/*
2165 			 * We reached our low water mark, reset the
2166 			 * request and sleep until we are needed again.
2167 			 * The sleep is just so the suspend code works.
2168 			 */
2169 			bd_request = 0;
2170 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2171 		} else {
2172 			/*
2173 			 * We couldn't find any flushable dirty buffers but
2174 			 * still have too many dirty buffers, we
2175 			 * have to sleep and try again.  (rare)
2176 			 */
2177 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2178 		}
2179 	}
2180 }
2181 
2182 /*
2183  *	flushbufqueues:
2184  *
2185  *	Try to flush a buffer in the dirty queue.  We must be careful to
2186  *	free up B_INVAL buffers instead of write them, which NFS is
2187  *	particularly sensitive to.
2188  */
2189 static int flushwithdeps = 0;
2190 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2191     0, "Number of buffers flushed with dependecies that require rollbacks");
2192 
2193 static int
2194 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2195 {
2196 	struct buf *sentinel;
2197 	struct vnode *vp;
2198 	struct mount *mp;
2199 	struct buf *bp;
2200 	int hasdeps;
2201 	int flushed;
2202 	int target;
2203 
2204 	if (lvp == NULL) {
2205 		target = numdirtybuffers - lodirtybuffers;
2206 		if (flushdeps && target > 2)
2207 			target /= 2;
2208 	} else
2209 		target = flushbufqtarget;
2210 	flushed = 0;
2211 	bp = NULL;
2212 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2213 	sentinel->b_qindex = QUEUE_SENTINEL;
2214 	mtx_lock(&bqlock);
2215 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2216 	while (flushed != target) {
2217 		bp = TAILQ_NEXT(sentinel, b_freelist);
2218 		if (bp != NULL) {
2219 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2220 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2221 			    b_freelist);
2222 		} else
2223 			break;
2224 		/*
2225 		 * Skip sentinels inserted by other invocations of the
2226 		 * flushbufqueues(), taking care to not reorder them.
2227 		 */
2228 		if (bp->b_qindex == QUEUE_SENTINEL)
2229 			continue;
2230 		/*
2231 		 * Only flush the buffers that belong to the
2232 		 * vnode locked by the curthread.
2233 		 */
2234 		if (lvp != NULL && bp->b_vp != lvp)
2235 			continue;
2236 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2237 			continue;
2238 		if (bp->b_pin_count > 0) {
2239 			BUF_UNLOCK(bp);
2240 			continue;
2241 		}
2242 		BO_LOCK(bp->b_bufobj);
2243 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2244 		    (bp->b_flags & B_DELWRI) == 0) {
2245 			BO_UNLOCK(bp->b_bufobj);
2246 			BUF_UNLOCK(bp);
2247 			continue;
2248 		}
2249 		BO_UNLOCK(bp->b_bufobj);
2250 		if (bp->b_flags & B_INVAL) {
2251 			bremfreel(bp);
2252 			mtx_unlock(&bqlock);
2253 			brelse(bp);
2254 			flushed++;
2255 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2256 			mtx_lock(&bqlock);
2257 			continue;
2258 		}
2259 
2260 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2261 			if (flushdeps == 0) {
2262 				BUF_UNLOCK(bp);
2263 				continue;
2264 			}
2265 			hasdeps = 1;
2266 		} else
2267 			hasdeps = 0;
2268 		/*
2269 		 * We must hold the lock on a vnode before writing
2270 		 * one of its buffers. Otherwise we may confuse, or
2271 		 * in the case of a snapshot vnode, deadlock the
2272 		 * system.
2273 		 *
2274 		 * The lock order here is the reverse of the normal
2275 		 * of vnode followed by buf lock.  This is ok because
2276 		 * the NOWAIT will prevent deadlock.
2277 		 */
2278 		vp = bp->b_vp;
2279 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2280 			BUF_UNLOCK(bp);
2281 			continue;
2282 		}
2283 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2284 			mtx_unlock(&bqlock);
2285 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2286 			    bp, bp->b_vp, bp->b_flags);
2287 			if (curproc == bufdaemonproc)
2288 				vfs_bio_awrite(bp);
2289 			else {
2290 				bremfree(bp);
2291 				bwrite(bp);
2292 				notbufdflashes++;
2293 			}
2294 			vn_finished_write(mp);
2295 			VOP_UNLOCK(vp, 0);
2296 			flushwithdeps += hasdeps;
2297 			flushed++;
2298 
2299 			/*
2300 			 * Sleeping on runningbufspace while holding
2301 			 * vnode lock leads to deadlock.
2302 			 */
2303 			if (curproc == bufdaemonproc)
2304 				waitrunningbufspace();
2305 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2306 			mtx_lock(&bqlock);
2307 			continue;
2308 		}
2309 		vn_finished_write(mp);
2310 		BUF_UNLOCK(bp);
2311 	}
2312 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2313 	mtx_unlock(&bqlock);
2314 	free(sentinel, M_TEMP);
2315 	return (flushed);
2316 }
2317 
2318 /*
2319  * Check to see if a block is currently memory resident.
2320  */
2321 struct buf *
2322 incore(struct bufobj *bo, daddr_t blkno)
2323 {
2324 	struct buf *bp;
2325 
2326 	BO_LOCK(bo);
2327 	bp = gbincore(bo, blkno);
2328 	BO_UNLOCK(bo);
2329 	return (bp);
2330 }
2331 
2332 /*
2333  * Returns true if no I/O is needed to access the
2334  * associated VM object.  This is like incore except
2335  * it also hunts around in the VM system for the data.
2336  */
2337 
2338 static int
2339 inmem(struct vnode * vp, daddr_t blkno)
2340 {
2341 	vm_object_t obj;
2342 	vm_offset_t toff, tinc, size;
2343 	vm_page_t m;
2344 	vm_ooffset_t off;
2345 
2346 	ASSERT_VOP_LOCKED(vp, "inmem");
2347 
2348 	if (incore(&vp->v_bufobj, blkno))
2349 		return 1;
2350 	if (vp->v_mount == NULL)
2351 		return 0;
2352 	obj = vp->v_object;
2353 	if (obj == NULL)
2354 		return (0);
2355 
2356 	size = PAGE_SIZE;
2357 	if (size > vp->v_mount->mnt_stat.f_iosize)
2358 		size = vp->v_mount->mnt_stat.f_iosize;
2359 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2360 
2361 	VM_OBJECT_LOCK(obj);
2362 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2363 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2364 		if (!m)
2365 			goto notinmem;
2366 		tinc = size;
2367 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2368 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2369 		if (vm_page_is_valid(m,
2370 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2371 			goto notinmem;
2372 	}
2373 	VM_OBJECT_UNLOCK(obj);
2374 	return 1;
2375 
2376 notinmem:
2377 	VM_OBJECT_UNLOCK(obj);
2378 	return (0);
2379 }
2380 
2381 /*
2382  *	vfs_setdirty:
2383  *
2384  *	Sets the dirty range for a buffer based on the status of the dirty
2385  *	bits in the pages comprising the buffer.
2386  *
2387  *	The range is limited to the size of the buffer.
2388  *
2389  *	This routine is primarily used by NFS, but is generalized for the
2390  *	B_VMIO case.
2391  */
2392 static void
2393 vfs_setdirty(struct buf *bp)
2394 {
2395 
2396 	/*
2397 	 * Degenerate case - empty buffer
2398 	 */
2399 	if (bp->b_bufsize == 0)
2400 		return;
2401 
2402 	if ((bp->b_flags & B_VMIO) == 0)
2403 		return;
2404 
2405 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2406 	vfs_setdirty_locked_object(bp);
2407 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2408 }
2409 
2410 static void
2411 vfs_setdirty_locked_object(struct buf *bp)
2412 {
2413 	vm_object_t object;
2414 	int i;
2415 
2416 	object = bp->b_bufobj->bo_object;
2417 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2418 
2419 	/*
2420 	 * We qualify the scan for modified pages on whether the
2421 	 * object has been flushed yet.
2422 	 */
2423 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2424 		vm_offset_t boffset;
2425 		vm_offset_t eoffset;
2426 
2427 		vm_page_lock_queues();
2428 		/*
2429 		 * test the pages to see if they have been modified directly
2430 		 * by users through the VM system.
2431 		 */
2432 		for (i = 0; i < bp->b_npages; i++)
2433 			vm_page_test_dirty(bp->b_pages[i]);
2434 
2435 		/*
2436 		 * Calculate the encompassing dirty range, boffset and eoffset,
2437 		 * (eoffset - boffset) bytes.
2438 		 */
2439 
2440 		for (i = 0; i < bp->b_npages; i++) {
2441 			if (bp->b_pages[i]->dirty)
2442 				break;
2443 		}
2444 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2445 
2446 		for (i = bp->b_npages - 1; i >= 0; --i) {
2447 			if (bp->b_pages[i]->dirty) {
2448 				break;
2449 			}
2450 		}
2451 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2452 
2453 		vm_page_unlock_queues();
2454 		/*
2455 		 * Fit it to the buffer.
2456 		 */
2457 
2458 		if (eoffset > bp->b_bcount)
2459 			eoffset = bp->b_bcount;
2460 
2461 		/*
2462 		 * If we have a good dirty range, merge with the existing
2463 		 * dirty range.
2464 		 */
2465 
2466 		if (boffset < eoffset) {
2467 			if (bp->b_dirtyoff > boffset)
2468 				bp->b_dirtyoff = boffset;
2469 			if (bp->b_dirtyend < eoffset)
2470 				bp->b_dirtyend = eoffset;
2471 		}
2472 	}
2473 }
2474 
2475 /*
2476  *	getblk:
2477  *
2478  *	Get a block given a specified block and offset into a file/device.
2479  *	The buffers B_DONE bit will be cleared on return, making it almost
2480  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2481  *	return.  The caller should clear B_INVAL prior to initiating a
2482  *	READ.
2483  *
2484  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2485  *	an existing buffer.
2486  *
2487  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2488  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2489  *	and then cleared based on the backing VM.  If the previous buffer is
2490  *	non-0-sized but invalid, B_CACHE will be cleared.
2491  *
2492  *	If getblk() must create a new buffer, the new buffer is returned with
2493  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2494  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2495  *	backing VM.
2496  *
2497  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2498  *	B_CACHE bit is clear.
2499  *
2500  *	What this means, basically, is that the caller should use B_CACHE to
2501  *	determine whether the buffer is fully valid or not and should clear
2502  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2503  *	the buffer by loading its data area with something, the caller needs
2504  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2505  *	the caller should set B_CACHE ( as an optimization ), else the caller
2506  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2507  *	a write attempt or if it was a successfull read.  If the caller
2508  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2509  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2510  */
2511 struct buf *
2512 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2513     int flags)
2514 {
2515 	struct buf *bp;
2516 	struct bufobj *bo;
2517 	int error;
2518 
2519 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2520 	ASSERT_VOP_LOCKED(vp, "getblk");
2521 	if (size > MAXBSIZE)
2522 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2523 
2524 	bo = &vp->v_bufobj;
2525 loop:
2526 	/*
2527 	 * Block if we are low on buffers.   Certain processes are allowed
2528 	 * to completely exhaust the buffer cache.
2529          *
2530          * If this check ever becomes a bottleneck it may be better to
2531          * move it into the else, when gbincore() fails.  At the moment
2532          * it isn't a problem.
2533 	 *
2534 	 * XXX remove if 0 sections (clean this up after its proven)
2535          */
2536 	if (numfreebuffers == 0) {
2537 		if (TD_IS_IDLETHREAD(curthread))
2538 			return NULL;
2539 		mtx_lock(&nblock);
2540 		needsbuffer |= VFS_BIO_NEED_ANY;
2541 		mtx_unlock(&nblock);
2542 	}
2543 
2544 	BO_LOCK(bo);
2545 	bp = gbincore(bo, blkno);
2546 	if (bp != NULL) {
2547 		int lockflags;
2548 		/*
2549 		 * Buffer is in-core.  If the buffer is not busy, it must
2550 		 * be on a queue.
2551 		 */
2552 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2553 
2554 		if (flags & GB_LOCK_NOWAIT)
2555 			lockflags |= LK_NOWAIT;
2556 
2557 		error = BUF_TIMELOCK(bp, lockflags,
2558 		    BO_MTX(bo), "getblk", slpflag, slptimeo);
2559 
2560 		/*
2561 		 * If we slept and got the lock we have to restart in case
2562 		 * the buffer changed identities.
2563 		 */
2564 		if (error == ENOLCK)
2565 			goto loop;
2566 		/* We timed out or were interrupted. */
2567 		else if (error)
2568 			return (NULL);
2569 
2570 		/*
2571 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2572 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2573 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2574 		 * backing VM cache.
2575 		 */
2576 		if (bp->b_flags & B_INVAL)
2577 			bp->b_flags &= ~B_CACHE;
2578 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2579 			bp->b_flags |= B_CACHE;
2580 		bremfree(bp);
2581 
2582 		/*
2583 		 * check for size inconsistancies for non-VMIO case.
2584 		 */
2585 
2586 		if (bp->b_bcount != size) {
2587 			if ((bp->b_flags & B_VMIO) == 0 ||
2588 			    (size > bp->b_kvasize)) {
2589 				if (bp->b_flags & B_DELWRI) {
2590 					/*
2591 					 * If buffer is pinned and caller does
2592 					 * not want sleep  waiting for it to be
2593 					 * unpinned, bail out
2594 					 * */
2595 					if (bp->b_pin_count > 0) {
2596 						if (flags & GB_LOCK_NOWAIT) {
2597 							bqrelse(bp);
2598 							return (NULL);
2599 						} else {
2600 							bunpin_wait(bp);
2601 						}
2602 					}
2603 					bp->b_flags |= B_NOCACHE;
2604 					bwrite(bp);
2605 				} else {
2606 					if (LIST_EMPTY(&bp->b_dep)) {
2607 						bp->b_flags |= B_RELBUF;
2608 						brelse(bp);
2609 					} else {
2610 						bp->b_flags |= B_NOCACHE;
2611 						bwrite(bp);
2612 					}
2613 				}
2614 				goto loop;
2615 			}
2616 		}
2617 
2618 		/*
2619 		 * If the size is inconsistant in the VMIO case, we can resize
2620 		 * the buffer.  This might lead to B_CACHE getting set or
2621 		 * cleared.  If the size has not changed, B_CACHE remains
2622 		 * unchanged from its previous state.
2623 		 */
2624 
2625 		if (bp->b_bcount != size)
2626 			allocbuf(bp, size);
2627 
2628 		KASSERT(bp->b_offset != NOOFFSET,
2629 		    ("getblk: no buffer offset"));
2630 
2631 		/*
2632 		 * A buffer with B_DELWRI set and B_CACHE clear must
2633 		 * be committed before we can return the buffer in
2634 		 * order to prevent the caller from issuing a read
2635 		 * ( due to B_CACHE not being set ) and overwriting
2636 		 * it.
2637 		 *
2638 		 * Most callers, including NFS and FFS, need this to
2639 		 * operate properly either because they assume they
2640 		 * can issue a read if B_CACHE is not set, or because
2641 		 * ( for example ) an uncached B_DELWRI might loop due
2642 		 * to softupdates re-dirtying the buffer.  In the latter
2643 		 * case, B_CACHE is set after the first write completes,
2644 		 * preventing further loops.
2645 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2646 		 * above while extending the buffer, we cannot allow the
2647 		 * buffer to remain with B_CACHE set after the write
2648 		 * completes or it will represent a corrupt state.  To
2649 		 * deal with this we set B_NOCACHE to scrap the buffer
2650 		 * after the write.
2651 		 *
2652 		 * We might be able to do something fancy, like setting
2653 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2654 		 * so the below call doesn't set B_CACHE, but that gets real
2655 		 * confusing.  This is much easier.
2656 		 */
2657 
2658 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2659 			bp->b_flags |= B_NOCACHE;
2660 			bwrite(bp);
2661 			goto loop;
2662 		}
2663 		bp->b_flags &= ~B_DONE;
2664 	} else {
2665 		int bsize, maxsize, vmio;
2666 		off_t offset;
2667 
2668 		/*
2669 		 * Buffer is not in-core, create new buffer.  The buffer
2670 		 * returned by getnewbuf() is locked.  Note that the returned
2671 		 * buffer is also considered valid (not marked B_INVAL).
2672 		 */
2673 		BO_UNLOCK(bo);
2674 		/*
2675 		 * If the user does not want us to create the buffer, bail out
2676 		 * here.
2677 		 */
2678 		if (flags & GB_NOCREAT)
2679 			return NULL;
2680 		bsize = bo->bo_bsize;
2681 		offset = blkno * bsize;
2682 		vmio = vp->v_object != NULL;
2683 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2684 		maxsize = imax(maxsize, bsize);
2685 
2686 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2687 		if (bp == NULL) {
2688 			if (slpflag || slptimeo)
2689 				return NULL;
2690 			goto loop;
2691 		}
2692 
2693 		/*
2694 		 * This code is used to make sure that a buffer is not
2695 		 * created while the getnewbuf routine is blocked.
2696 		 * This can be a problem whether the vnode is locked or not.
2697 		 * If the buffer is created out from under us, we have to
2698 		 * throw away the one we just created.
2699 		 *
2700 		 * Note: this must occur before we associate the buffer
2701 		 * with the vp especially considering limitations in
2702 		 * the splay tree implementation when dealing with duplicate
2703 		 * lblkno's.
2704 		 */
2705 		BO_LOCK(bo);
2706 		if (gbincore(bo, blkno)) {
2707 			BO_UNLOCK(bo);
2708 			bp->b_flags |= B_INVAL;
2709 			brelse(bp);
2710 			goto loop;
2711 		}
2712 
2713 		/*
2714 		 * Insert the buffer into the hash, so that it can
2715 		 * be found by incore.
2716 		 */
2717 		bp->b_blkno = bp->b_lblkno = blkno;
2718 		bp->b_offset = offset;
2719 		bgetvp(vp, bp);
2720 		BO_UNLOCK(bo);
2721 
2722 		/*
2723 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2724 		 * buffer size starts out as 0, B_CACHE will be set by
2725 		 * allocbuf() for the VMIO case prior to it testing the
2726 		 * backing store for validity.
2727 		 */
2728 
2729 		if (vmio) {
2730 			bp->b_flags |= B_VMIO;
2731 #if defined(VFS_BIO_DEBUG)
2732 			if (vn_canvmio(vp) != TRUE)
2733 				printf("getblk: VMIO on vnode type %d\n",
2734 					vp->v_type);
2735 #endif
2736 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2737 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2738 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2739 		} else {
2740 			bp->b_flags &= ~B_VMIO;
2741 			KASSERT(bp->b_bufobj->bo_object == NULL,
2742 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2743 			    bp, bp->b_bufobj->bo_object));
2744 		}
2745 
2746 		allocbuf(bp, size);
2747 		bp->b_flags &= ~B_DONE;
2748 	}
2749 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2750 	BUF_ASSERT_HELD(bp);
2751 	KASSERT(bp->b_bufobj == bo,
2752 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2753 	return (bp);
2754 }
2755 
2756 /*
2757  * Get an empty, disassociated buffer of given size.  The buffer is initially
2758  * set to B_INVAL.
2759  */
2760 struct buf *
2761 geteblk(int size, int flags)
2762 {
2763 	struct buf *bp;
2764 	int maxsize;
2765 
2766 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2767 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2768 		if ((flags & GB_NOWAIT_BD) &&
2769 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
2770 			return (NULL);
2771 	}
2772 	allocbuf(bp, size);
2773 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2774 	BUF_ASSERT_HELD(bp);
2775 	return (bp);
2776 }
2777 
2778 
2779 /*
2780  * This code constitutes the buffer memory from either anonymous system
2781  * memory (in the case of non-VMIO operations) or from an associated
2782  * VM object (in the case of VMIO operations).  This code is able to
2783  * resize a buffer up or down.
2784  *
2785  * Note that this code is tricky, and has many complications to resolve
2786  * deadlock or inconsistant data situations.  Tread lightly!!!
2787  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2788  * the caller.  Calling this code willy nilly can result in the loss of data.
2789  *
2790  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2791  * B_CACHE for the non-VMIO case.
2792  */
2793 
2794 int
2795 allocbuf(struct buf *bp, int size)
2796 {
2797 	int newbsize, mbsize;
2798 	int i;
2799 
2800 	BUF_ASSERT_HELD(bp);
2801 
2802 	if (bp->b_kvasize < size)
2803 		panic("allocbuf: buffer too small");
2804 
2805 	if ((bp->b_flags & B_VMIO) == 0) {
2806 		caddr_t origbuf;
2807 		int origbufsize;
2808 		/*
2809 		 * Just get anonymous memory from the kernel.  Don't
2810 		 * mess with B_CACHE.
2811 		 */
2812 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2813 		if (bp->b_flags & B_MALLOC)
2814 			newbsize = mbsize;
2815 		else
2816 			newbsize = round_page(size);
2817 
2818 		if (newbsize < bp->b_bufsize) {
2819 			/*
2820 			 * malloced buffers are not shrunk
2821 			 */
2822 			if (bp->b_flags & B_MALLOC) {
2823 				if (newbsize) {
2824 					bp->b_bcount = size;
2825 				} else {
2826 					free(bp->b_data, M_BIOBUF);
2827 					if (bp->b_bufsize) {
2828 						atomic_subtract_long(
2829 						    &bufmallocspace,
2830 						    bp->b_bufsize);
2831 						bufspacewakeup();
2832 						bp->b_bufsize = 0;
2833 					}
2834 					bp->b_saveaddr = bp->b_kvabase;
2835 					bp->b_data = bp->b_saveaddr;
2836 					bp->b_bcount = 0;
2837 					bp->b_flags &= ~B_MALLOC;
2838 				}
2839 				return 1;
2840 			}
2841 			vm_hold_free_pages(
2842 			    bp,
2843 			    (vm_offset_t) bp->b_data + newbsize,
2844 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2845 		} else if (newbsize > bp->b_bufsize) {
2846 			/*
2847 			 * We only use malloced memory on the first allocation.
2848 			 * and revert to page-allocated memory when the buffer
2849 			 * grows.
2850 			 */
2851 			/*
2852 			 * There is a potential smp race here that could lead
2853 			 * to bufmallocspace slightly passing the max.  It
2854 			 * is probably extremely rare and not worth worrying
2855 			 * over.
2856 			 */
2857 			if ( (bufmallocspace < maxbufmallocspace) &&
2858 				(bp->b_bufsize == 0) &&
2859 				(mbsize <= PAGE_SIZE/2)) {
2860 
2861 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2862 				bp->b_bufsize = mbsize;
2863 				bp->b_bcount = size;
2864 				bp->b_flags |= B_MALLOC;
2865 				atomic_add_long(&bufmallocspace, mbsize);
2866 				return 1;
2867 			}
2868 			origbuf = NULL;
2869 			origbufsize = 0;
2870 			/*
2871 			 * If the buffer is growing on its other-than-first allocation,
2872 			 * then we revert to the page-allocation scheme.
2873 			 */
2874 			if (bp->b_flags & B_MALLOC) {
2875 				origbuf = bp->b_data;
2876 				origbufsize = bp->b_bufsize;
2877 				bp->b_data = bp->b_kvabase;
2878 				if (bp->b_bufsize) {
2879 					atomic_subtract_long(&bufmallocspace,
2880 					    bp->b_bufsize);
2881 					bufspacewakeup();
2882 					bp->b_bufsize = 0;
2883 				}
2884 				bp->b_flags &= ~B_MALLOC;
2885 				newbsize = round_page(newbsize);
2886 			}
2887 			vm_hold_load_pages(
2888 			    bp,
2889 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2890 			    (vm_offset_t) bp->b_data + newbsize);
2891 			if (origbuf) {
2892 				bcopy(origbuf, bp->b_data, origbufsize);
2893 				free(origbuf, M_BIOBUF);
2894 			}
2895 		}
2896 	} else {
2897 		int desiredpages;
2898 
2899 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2900 		desiredpages = (size == 0) ? 0 :
2901 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2902 
2903 		if (bp->b_flags & B_MALLOC)
2904 			panic("allocbuf: VMIO buffer can't be malloced");
2905 		/*
2906 		 * Set B_CACHE initially if buffer is 0 length or will become
2907 		 * 0-length.
2908 		 */
2909 		if (size == 0 || bp->b_bufsize == 0)
2910 			bp->b_flags |= B_CACHE;
2911 
2912 		if (newbsize < bp->b_bufsize) {
2913 			/*
2914 			 * DEV_BSIZE aligned new buffer size is less then the
2915 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2916 			 * if we have to remove any pages.
2917 			 */
2918 			if (desiredpages < bp->b_npages) {
2919 				vm_page_t m;
2920 
2921 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2922 				vm_page_lock_queues();
2923 				for (i = desiredpages; i < bp->b_npages; i++) {
2924 					/*
2925 					 * the page is not freed here -- it
2926 					 * is the responsibility of
2927 					 * vnode_pager_setsize
2928 					 */
2929 					m = bp->b_pages[i];
2930 					KASSERT(m != bogus_page,
2931 					    ("allocbuf: bogus page found"));
2932 					while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2933 						vm_page_lock_queues();
2934 
2935 					bp->b_pages[i] = NULL;
2936 					vm_page_unwire(m, 0);
2937 				}
2938 				vm_page_unlock_queues();
2939 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2940 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2941 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2942 				bp->b_npages = desiredpages;
2943 			}
2944 		} else if (size > bp->b_bcount) {
2945 			/*
2946 			 * We are growing the buffer, possibly in a
2947 			 * byte-granular fashion.
2948 			 */
2949 			vm_object_t obj;
2950 			vm_offset_t toff;
2951 			vm_offset_t tinc;
2952 
2953 			/*
2954 			 * Step 1, bring in the VM pages from the object,
2955 			 * allocating them if necessary.  We must clear
2956 			 * B_CACHE if these pages are not valid for the
2957 			 * range covered by the buffer.
2958 			 */
2959 
2960 			obj = bp->b_bufobj->bo_object;
2961 
2962 			VM_OBJECT_LOCK(obj);
2963 			while (bp->b_npages < desiredpages) {
2964 				vm_page_t m;
2965 				vm_pindex_t pi;
2966 
2967 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2968 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2969 					/*
2970 					 * note: must allocate system pages
2971 					 * since blocking here could intefere
2972 					 * with paging I/O, no matter which
2973 					 * process we are.
2974 					 */
2975 					m = vm_page_alloc(obj, pi,
2976 					    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2977 					    VM_ALLOC_WIRED);
2978 					if (m == NULL) {
2979 						atomic_add_int(&vm_pageout_deficit,
2980 						    desiredpages - bp->b_npages);
2981 						VM_OBJECT_UNLOCK(obj);
2982 						VM_WAIT;
2983 						VM_OBJECT_LOCK(obj);
2984 					} else {
2985 						if (m->valid == 0)
2986 							bp->b_flags &= ~B_CACHE;
2987 						bp->b_pages[bp->b_npages] = m;
2988 						++bp->b_npages;
2989 					}
2990 					continue;
2991 				}
2992 
2993 				/*
2994 				 * We found a page.  If we have to sleep on it,
2995 				 * retry because it might have gotten freed out
2996 				 * from under us.
2997 				 *
2998 				 * We can only test VPO_BUSY here.  Blocking on
2999 				 * m->busy might lead to a deadlock:
3000 				 *
3001 				 *  vm_fault->getpages->cluster_read->allocbuf
3002 				 *
3003 				 */
3004 				if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
3005 					continue;
3006 
3007 				/*
3008 				 * We have a good page.
3009 				 */
3010 				vm_page_lock_queues();
3011 				vm_page_wire(m);
3012 				vm_page_unlock_queues();
3013 				bp->b_pages[bp->b_npages] = m;
3014 				++bp->b_npages;
3015 			}
3016 
3017 			/*
3018 			 * Step 2.  We've loaded the pages into the buffer,
3019 			 * we have to figure out if we can still have B_CACHE
3020 			 * set.  Note that B_CACHE is set according to the
3021 			 * byte-granular range ( bcount and size ), new the
3022 			 * aligned range ( newbsize ).
3023 			 *
3024 			 * The VM test is against m->valid, which is DEV_BSIZE
3025 			 * aligned.  Needless to say, the validity of the data
3026 			 * needs to also be DEV_BSIZE aligned.  Note that this
3027 			 * fails with NFS if the server or some other client
3028 			 * extends the file's EOF.  If our buffer is resized,
3029 			 * B_CACHE may remain set! XXX
3030 			 */
3031 
3032 			toff = bp->b_bcount;
3033 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3034 
3035 			while ((bp->b_flags & B_CACHE) && toff < size) {
3036 				vm_pindex_t pi;
3037 
3038 				if (tinc > (size - toff))
3039 					tinc = size - toff;
3040 
3041 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3042 				    PAGE_SHIFT;
3043 
3044 				vfs_buf_test_cache(
3045 				    bp,
3046 				    bp->b_offset,
3047 				    toff,
3048 				    tinc,
3049 				    bp->b_pages[pi]
3050 				);
3051 				toff += tinc;
3052 				tinc = PAGE_SIZE;
3053 			}
3054 			VM_OBJECT_UNLOCK(obj);
3055 
3056 			/*
3057 			 * Step 3, fixup the KVM pmap.  Remember that
3058 			 * bp->b_data is relative to bp->b_offset, but
3059 			 * bp->b_offset may be offset into the first page.
3060 			 */
3061 
3062 			bp->b_data = (caddr_t)
3063 			    trunc_page((vm_offset_t)bp->b_data);
3064 			pmap_qenter(
3065 			    (vm_offset_t)bp->b_data,
3066 			    bp->b_pages,
3067 			    bp->b_npages
3068 			);
3069 
3070 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3071 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
3072 		}
3073 	}
3074 	if (newbsize < bp->b_bufsize)
3075 		bufspacewakeup();
3076 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3077 	bp->b_bcount = size;		/* requested buffer size	*/
3078 	return 1;
3079 }
3080 
3081 void
3082 biodone(struct bio *bp)
3083 {
3084 	struct mtx *mtxp;
3085 	void (*done)(struct bio *);
3086 
3087 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3088 	mtx_lock(mtxp);
3089 	bp->bio_flags |= BIO_DONE;
3090 	done = bp->bio_done;
3091 	if (done == NULL)
3092 		wakeup(bp);
3093 	mtx_unlock(mtxp);
3094 	if (done != NULL)
3095 		done(bp);
3096 }
3097 
3098 /*
3099  * Wait for a BIO to finish.
3100  *
3101  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3102  * case is not yet clear.
3103  */
3104 int
3105 biowait(struct bio *bp, const char *wchan)
3106 {
3107 	struct mtx *mtxp;
3108 
3109 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3110 	mtx_lock(mtxp);
3111 	while ((bp->bio_flags & BIO_DONE) == 0)
3112 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3113 	mtx_unlock(mtxp);
3114 	if (bp->bio_error != 0)
3115 		return (bp->bio_error);
3116 	if (!(bp->bio_flags & BIO_ERROR))
3117 		return (0);
3118 	return (EIO);
3119 }
3120 
3121 void
3122 biofinish(struct bio *bp, struct devstat *stat, int error)
3123 {
3124 
3125 	if (error) {
3126 		bp->bio_error = error;
3127 		bp->bio_flags |= BIO_ERROR;
3128 	}
3129 	if (stat != NULL)
3130 		devstat_end_transaction_bio(stat, bp);
3131 	biodone(bp);
3132 }
3133 
3134 /*
3135  *	bufwait:
3136  *
3137  *	Wait for buffer I/O completion, returning error status.  The buffer
3138  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3139  *	error and cleared.
3140  */
3141 int
3142 bufwait(struct buf *bp)
3143 {
3144 	if (bp->b_iocmd == BIO_READ)
3145 		bwait(bp, PRIBIO, "biord");
3146 	else
3147 		bwait(bp, PRIBIO, "biowr");
3148 	if (bp->b_flags & B_EINTR) {
3149 		bp->b_flags &= ~B_EINTR;
3150 		return (EINTR);
3151 	}
3152 	if (bp->b_ioflags & BIO_ERROR) {
3153 		return (bp->b_error ? bp->b_error : EIO);
3154 	} else {
3155 		return (0);
3156 	}
3157 }
3158 
3159  /*
3160   * Call back function from struct bio back up to struct buf.
3161   */
3162 static void
3163 bufdonebio(struct bio *bip)
3164 {
3165 	struct buf *bp;
3166 
3167 	bp = bip->bio_caller2;
3168 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3169 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3170 	bp->b_ioflags = bip->bio_flags;
3171 	bp->b_error = bip->bio_error;
3172 	if (bp->b_error)
3173 		bp->b_ioflags |= BIO_ERROR;
3174 	bufdone(bp);
3175 	g_destroy_bio(bip);
3176 }
3177 
3178 void
3179 dev_strategy(struct cdev *dev, struct buf *bp)
3180 {
3181 	struct cdevsw *csw;
3182 	struct bio *bip;
3183 
3184 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3185 		panic("b_iocmd botch");
3186 	for (;;) {
3187 		bip = g_new_bio();
3188 		if (bip != NULL)
3189 			break;
3190 		/* Try again later */
3191 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3192 	}
3193 	bip->bio_cmd = bp->b_iocmd;
3194 	bip->bio_offset = bp->b_iooffset;
3195 	bip->bio_length = bp->b_bcount;
3196 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3197 	bip->bio_data = bp->b_data;
3198 	bip->bio_done = bufdonebio;
3199 	bip->bio_caller2 = bp;
3200 	bip->bio_dev = dev;
3201 	KASSERT(dev->si_refcount > 0,
3202 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3203 	    devtoname(dev)));
3204 	csw = dev_refthread(dev);
3205 	if (csw == NULL) {
3206 		g_destroy_bio(bip);
3207 		bp->b_error = ENXIO;
3208 		bp->b_ioflags = BIO_ERROR;
3209 		bufdone(bp);
3210 		return;
3211 	}
3212 	(*csw->d_strategy)(bip);
3213 	dev_relthread(dev);
3214 }
3215 
3216 /*
3217  *	bufdone:
3218  *
3219  *	Finish I/O on a buffer, optionally calling a completion function.
3220  *	This is usually called from an interrupt so process blocking is
3221  *	not allowed.
3222  *
3223  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3224  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3225  *	assuming B_INVAL is clear.
3226  *
3227  *	For the VMIO case, we set B_CACHE if the op was a read and no
3228  *	read error occured, or if the op was a write.  B_CACHE is never
3229  *	set if the buffer is invalid or otherwise uncacheable.
3230  *
3231  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3232  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3233  *	in the biodone routine.
3234  */
3235 void
3236 bufdone(struct buf *bp)
3237 {
3238 	struct bufobj *dropobj;
3239 	void    (*biodone)(struct buf *);
3240 
3241 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3242 	dropobj = NULL;
3243 
3244 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3245 	BUF_ASSERT_HELD(bp);
3246 
3247 	runningbufwakeup(bp);
3248 	if (bp->b_iocmd == BIO_WRITE)
3249 		dropobj = bp->b_bufobj;
3250 	/* call optional completion function if requested */
3251 	if (bp->b_iodone != NULL) {
3252 		biodone = bp->b_iodone;
3253 		bp->b_iodone = NULL;
3254 		(*biodone) (bp);
3255 		if (dropobj)
3256 			bufobj_wdrop(dropobj);
3257 		return;
3258 	}
3259 
3260 	bufdone_finish(bp);
3261 
3262 	if (dropobj)
3263 		bufobj_wdrop(dropobj);
3264 }
3265 
3266 void
3267 bufdone_finish(struct buf *bp)
3268 {
3269 	BUF_ASSERT_HELD(bp);
3270 
3271 	if (!LIST_EMPTY(&bp->b_dep))
3272 		buf_complete(bp);
3273 
3274 	if (bp->b_flags & B_VMIO) {
3275 		int i;
3276 		vm_ooffset_t foff;
3277 		vm_page_t m;
3278 		vm_object_t obj;
3279 		int iosize;
3280 		struct vnode *vp = bp->b_vp;
3281 
3282 		obj = bp->b_bufobj->bo_object;
3283 
3284 #if defined(VFS_BIO_DEBUG)
3285 		mp_fixme("usecount and vflag accessed without locks.");
3286 		if (vp->v_usecount == 0) {
3287 			panic("biodone: zero vnode ref count");
3288 		}
3289 
3290 		KASSERT(vp->v_object != NULL,
3291 			("biodone: vnode %p has no vm_object", vp));
3292 #endif
3293 
3294 		foff = bp->b_offset;
3295 		KASSERT(bp->b_offset != NOOFFSET,
3296 		    ("biodone: no buffer offset"));
3297 
3298 		VM_OBJECT_LOCK(obj);
3299 #if defined(VFS_BIO_DEBUG)
3300 		if (obj->paging_in_progress < bp->b_npages) {
3301 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3302 			    obj->paging_in_progress, bp->b_npages);
3303 		}
3304 #endif
3305 
3306 		/*
3307 		 * Set B_CACHE if the op was a normal read and no error
3308 		 * occured.  B_CACHE is set for writes in the b*write()
3309 		 * routines.
3310 		 */
3311 		iosize = bp->b_bcount - bp->b_resid;
3312 		if (bp->b_iocmd == BIO_READ &&
3313 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3314 		    !(bp->b_ioflags & BIO_ERROR)) {
3315 			bp->b_flags |= B_CACHE;
3316 		}
3317 		for (i = 0; i < bp->b_npages; i++) {
3318 			int bogusflag = 0;
3319 			int resid;
3320 
3321 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3322 			if (resid > iosize)
3323 				resid = iosize;
3324 
3325 			/*
3326 			 * cleanup bogus pages, restoring the originals
3327 			 */
3328 			m = bp->b_pages[i];
3329 			if (m == bogus_page) {
3330 				bogusflag = 1;
3331 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3332 				if (m == NULL)
3333 					panic("biodone: page disappeared!");
3334 				bp->b_pages[i] = m;
3335 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3336 				    bp->b_pages, bp->b_npages);
3337 			}
3338 #if defined(VFS_BIO_DEBUG)
3339 			if (OFF_TO_IDX(foff) != m->pindex) {
3340 				printf(
3341 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3342 				    (intmax_t)foff, (uintmax_t)m->pindex);
3343 			}
3344 #endif
3345 
3346 			/*
3347 			 * In the write case, the valid and clean bits are
3348 			 * already changed correctly ( see bdwrite() ), so we
3349 			 * only need to do this here in the read case.
3350 			 */
3351 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3352 				KASSERT((m->dirty & vm_page_bits(foff &
3353 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3354 				    " page %p has unexpected dirty bits", m));
3355 				vfs_page_set_valid(bp, foff, m);
3356 			}
3357 
3358 			/*
3359 			 * when debugging new filesystems or buffer I/O methods, this
3360 			 * is the most common error that pops up.  if you see this, you
3361 			 * have not set the page busy flag correctly!!!
3362 			 */
3363 			if (m->busy == 0) {
3364 				printf("biodone: page busy < 0, "
3365 				    "pindex: %d, foff: 0x(%x,%x), "
3366 				    "resid: %d, index: %d\n",
3367 				    (int) m->pindex, (int)(foff >> 32),
3368 						(int) foff & 0xffffffff, resid, i);
3369 				if (!vn_isdisk(vp, NULL))
3370 					printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3371 					    (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3372 					    (intmax_t) bp->b_lblkno,
3373 					    bp->b_flags, bp->b_npages);
3374 				else
3375 					printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3376 					    (intmax_t) bp->b_lblkno,
3377 					    bp->b_flags, bp->b_npages);
3378 				printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3379 				    (u_long)m->valid, (u_long)m->dirty,
3380 				    m->wire_count);
3381 				panic("biodone: page busy < 0\n");
3382 			}
3383 			vm_page_io_finish(m);
3384 			vm_object_pip_subtract(obj, 1);
3385 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3386 			iosize -= resid;
3387 		}
3388 		vm_object_pip_wakeupn(obj, 0);
3389 		VM_OBJECT_UNLOCK(obj);
3390 	}
3391 
3392 	/*
3393 	 * For asynchronous completions, release the buffer now. The brelse
3394 	 * will do a wakeup there if necessary - so no need to do a wakeup
3395 	 * here in the async case. The sync case always needs to do a wakeup.
3396 	 */
3397 
3398 	if (bp->b_flags & B_ASYNC) {
3399 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3400 			brelse(bp);
3401 		else
3402 			bqrelse(bp);
3403 	} else
3404 		bdone(bp);
3405 }
3406 
3407 /*
3408  * This routine is called in lieu of iodone in the case of
3409  * incomplete I/O.  This keeps the busy status for pages
3410  * consistant.
3411  */
3412 void
3413 vfs_unbusy_pages(struct buf *bp)
3414 {
3415 	int i;
3416 	vm_object_t obj;
3417 	vm_page_t m;
3418 
3419 	runningbufwakeup(bp);
3420 	if (!(bp->b_flags & B_VMIO))
3421 		return;
3422 
3423 	obj = bp->b_bufobj->bo_object;
3424 	VM_OBJECT_LOCK(obj);
3425 	for (i = 0; i < bp->b_npages; i++) {
3426 		m = bp->b_pages[i];
3427 		if (m == bogus_page) {
3428 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3429 			if (!m)
3430 				panic("vfs_unbusy_pages: page missing\n");
3431 			bp->b_pages[i] = m;
3432 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3433 			    bp->b_pages, bp->b_npages);
3434 		}
3435 		vm_object_pip_subtract(obj, 1);
3436 		vm_page_io_finish(m);
3437 	}
3438 	vm_object_pip_wakeupn(obj, 0);
3439 	VM_OBJECT_UNLOCK(obj);
3440 }
3441 
3442 /*
3443  * vfs_page_set_valid:
3444  *
3445  *	Set the valid bits in a page based on the supplied offset.   The
3446  *	range is restricted to the buffer's size.
3447  *
3448  *	This routine is typically called after a read completes.
3449  */
3450 static void
3451 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3452 {
3453 	vm_ooffset_t eoff;
3454 
3455 	/*
3456 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3457 	 * page boundary and eoff is not greater than the end of the buffer.
3458 	 * The end of the buffer, in this case, is our file EOF, not the
3459 	 * allocation size of the buffer.
3460 	 */
3461 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3462 	if (eoff > bp->b_offset + bp->b_bcount)
3463 		eoff = bp->b_offset + bp->b_bcount;
3464 
3465 	/*
3466 	 * Set valid range.  This is typically the entire buffer and thus the
3467 	 * entire page.
3468 	 */
3469 	if (eoff > off)
3470 		vm_page_set_valid(m, off & PAGE_MASK, eoff - off);
3471 }
3472 
3473 /*
3474  * vfs_page_set_validclean:
3475  *
3476  *	Set the valid bits and clear the dirty bits in a page based on the
3477  *	supplied offset.   The range is restricted to the buffer's size.
3478  */
3479 static void
3480 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3481 {
3482 	vm_ooffset_t soff, eoff;
3483 
3484 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3485 	/*
3486 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3487 	 * page boundry or cross the end of the buffer.  The end of the
3488 	 * buffer, in this case, is our file EOF, not the allocation size
3489 	 * of the buffer.
3490 	 */
3491 	soff = off;
3492 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3493 	if (eoff > bp->b_offset + bp->b_bcount)
3494 		eoff = bp->b_offset + bp->b_bcount;
3495 
3496 	/*
3497 	 * Set valid range.  This is typically the entire buffer and thus the
3498 	 * entire page.
3499 	 */
3500 	if (eoff > soff) {
3501 		vm_page_set_validclean(
3502 		    m,
3503 		   (vm_offset_t) (soff & PAGE_MASK),
3504 		   (vm_offset_t) (eoff - soff)
3505 		);
3506 	}
3507 }
3508 
3509 /*
3510  * This routine is called before a device strategy routine.
3511  * It is used to tell the VM system that paging I/O is in
3512  * progress, and treat the pages associated with the buffer
3513  * almost as being VPO_BUSY.  Also the object paging_in_progress
3514  * flag is handled to make sure that the object doesn't become
3515  * inconsistant.
3516  *
3517  * Since I/O has not been initiated yet, certain buffer flags
3518  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3519  * and should be ignored.
3520  */
3521 void
3522 vfs_busy_pages(struct buf *bp, int clear_modify)
3523 {
3524 	int i, bogus;
3525 	vm_object_t obj;
3526 	vm_ooffset_t foff;
3527 	vm_page_t m;
3528 
3529 	if (!(bp->b_flags & B_VMIO))
3530 		return;
3531 
3532 	obj = bp->b_bufobj->bo_object;
3533 	foff = bp->b_offset;
3534 	KASSERT(bp->b_offset != NOOFFSET,
3535 	    ("vfs_busy_pages: no buffer offset"));
3536 	VM_OBJECT_LOCK(obj);
3537 	if (bp->b_bufsize != 0)
3538 		vfs_setdirty_locked_object(bp);
3539 retry:
3540 	for (i = 0; i < bp->b_npages; i++) {
3541 		m = bp->b_pages[i];
3542 
3543 		if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3544 			goto retry;
3545 	}
3546 	bogus = 0;
3547 	if (clear_modify)
3548 		vm_page_lock_queues();
3549 	for (i = 0; i < bp->b_npages; i++) {
3550 		m = bp->b_pages[i];
3551 
3552 		if ((bp->b_flags & B_CLUSTER) == 0) {
3553 			vm_object_pip_add(obj, 1);
3554 			vm_page_io_start(m);
3555 		}
3556 		/*
3557 		 * When readying a buffer for a read ( i.e
3558 		 * clear_modify == 0 ), it is important to do
3559 		 * bogus_page replacement for valid pages in
3560 		 * partially instantiated buffers.  Partially
3561 		 * instantiated buffers can, in turn, occur when
3562 		 * reconstituting a buffer from its VM backing store
3563 		 * base.  We only have to do this if B_CACHE is
3564 		 * clear ( which causes the I/O to occur in the
3565 		 * first place ).  The replacement prevents the read
3566 		 * I/O from overwriting potentially dirty VM-backed
3567 		 * pages.  XXX bogus page replacement is, uh, bogus.
3568 		 * It may not work properly with small-block devices.
3569 		 * We need to find a better way.
3570 		 */
3571 		if (clear_modify) {
3572 			pmap_remove_write(m);
3573 			vfs_page_set_validclean(bp, foff, m);
3574 		} else if (m->valid == VM_PAGE_BITS_ALL &&
3575 		    (bp->b_flags & B_CACHE) == 0) {
3576 			bp->b_pages[i] = bogus_page;
3577 			bogus++;
3578 		}
3579 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3580 	}
3581 	if (clear_modify)
3582 		vm_page_unlock_queues();
3583 	VM_OBJECT_UNLOCK(obj);
3584 	if (bogus)
3585 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3586 		    bp->b_pages, bp->b_npages);
3587 }
3588 
3589 /*
3590  * Tell the VM system that the pages associated with this buffer
3591  * are clean.  This is used for delayed writes where the data is
3592  * going to go to disk eventually without additional VM intevention.
3593  *
3594  * Note that while we only really need to clean through to b_bcount, we
3595  * just go ahead and clean through to b_bufsize.
3596  */
3597 static void
3598 vfs_clean_pages(struct buf *bp)
3599 {
3600 	int i;
3601 	vm_ooffset_t foff, noff, eoff;
3602 	vm_page_t m;
3603 
3604 	if (!(bp->b_flags & B_VMIO))
3605 		return;
3606 
3607 	foff = bp->b_offset;
3608 	KASSERT(bp->b_offset != NOOFFSET,
3609 	    ("vfs_clean_pages: no buffer offset"));
3610 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3611 	vm_page_lock_queues();
3612 	for (i = 0; i < bp->b_npages; i++) {
3613 		m = bp->b_pages[i];
3614 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3615 		eoff = noff;
3616 
3617 		if (eoff > bp->b_offset + bp->b_bufsize)
3618 			eoff = bp->b_offset + bp->b_bufsize;
3619 		vfs_page_set_validclean(bp, foff, m);
3620 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3621 		foff = noff;
3622 	}
3623 	vm_page_unlock_queues();
3624 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3625 }
3626 
3627 /*
3628  *	vfs_bio_set_valid:
3629  *
3630  *	Set the range within the buffer to valid.  The range is
3631  *	relative to the beginning of the buffer, b_offset.  Note that
3632  *	b_offset itself may be offset from the beginning of the first
3633  *	page.
3634  */
3635 void
3636 vfs_bio_set_valid(struct buf *bp, int base, int size)
3637 {
3638 	int i, n;
3639 	vm_page_t m;
3640 
3641 	if (!(bp->b_flags & B_VMIO))
3642 		return;
3643 
3644 	/*
3645 	 * Fixup base to be relative to beginning of first page.
3646 	 * Set initial n to be the maximum number of bytes in the
3647 	 * first page that can be validated.
3648 	 */
3649 	base += (bp->b_offset & PAGE_MASK);
3650 	n = PAGE_SIZE - (base & PAGE_MASK);
3651 
3652 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3653 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3654 		m = bp->b_pages[i];
3655 		if (n > size)
3656 			n = size;
3657 		vm_page_set_valid(m, base & PAGE_MASK, n);
3658 		base += n;
3659 		size -= n;
3660 		n = PAGE_SIZE;
3661 	}
3662 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3663 }
3664 
3665 /*
3666  *	vfs_bio_clrbuf:
3667  *
3668  *	If the specified buffer is a non-VMIO buffer, clear the entire
3669  *	buffer.  If the specified buffer is a VMIO buffer, clear and
3670  *	validate only the previously invalid portions of the buffer.
3671  *	This routine essentially fakes an I/O, so we need to clear
3672  *	BIO_ERROR and B_INVAL.
3673  *
3674  *	Note that while we only theoretically need to clear through b_bcount,
3675  *	we go ahead and clear through b_bufsize.
3676  */
3677 void
3678 vfs_bio_clrbuf(struct buf *bp)
3679 {
3680 	int i, j, mask;
3681 	caddr_t sa, ea;
3682 
3683 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3684 		clrbuf(bp);
3685 		return;
3686 	}
3687 	bp->b_flags &= ~B_INVAL;
3688 	bp->b_ioflags &= ~BIO_ERROR;
3689 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3690 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3691 	    (bp->b_offset & PAGE_MASK) == 0) {
3692 		if (bp->b_pages[0] == bogus_page)
3693 			goto unlock;
3694 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3695 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3696 		if ((bp->b_pages[0]->valid & mask) == mask)
3697 			goto unlock;
3698 		if ((bp->b_pages[0]->valid & mask) == 0) {
3699 			bzero(bp->b_data, bp->b_bufsize);
3700 			bp->b_pages[0]->valid |= mask;
3701 			goto unlock;
3702 		}
3703 	}
3704 	ea = sa = bp->b_data;
3705 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3706 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3707 		ea = (caddr_t)(vm_offset_t)ulmin(
3708 		    (u_long)(vm_offset_t)ea,
3709 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3710 		if (bp->b_pages[i] == bogus_page)
3711 			continue;
3712 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3713 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3714 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3715 		if ((bp->b_pages[i]->valid & mask) == mask)
3716 			continue;
3717 		if ((bp->b_pages[i]->valid & mask) == 0)
3718 			bzero(sa, ea - sa);
3719 		else {
3720 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3721 				if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3722 					bzero(sa, DEV_BSIZE);
3723 			}
3724 		}
3725 		bp->b_pages[i]->valid |= mask;
3726 	}
3727 unlock:
3728 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3729 	bp->b_resid = 0;
3730 }
3731 
3732 /*
3733  * vm_hold_load_pages and vm_hold_free_pages get pages into
3734  * a buffers address space.  The pages are anonymous and are
3735  * not associated with a file object.
3736  */
3737 static void
3738 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3739 {
3740 	vm_offset_t pg;
3741 	vm_page_t p;
3742 	int index;
3743 
3744 	to = round_page(to);
3745 	from = round_page(from);
3746 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3747 
3748 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3749 tryagain:
3750 		/*
3751 		 * note: must allocate system pages since blocking here
3752 		 * could interfere with paging I/O, no matter which
3753 		 * process we are.
3754 		 */
3755 		p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ |
3756 		    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3757 		if (!p) {
3758 			atomic_add_int(&vm_pageout_deficit,
3759 			    (to - pg) >> PAGE_SHIFT);
3760 			VM_WAIT;
3761 			goto tryagain;
3762 		}
3763 		pmap_qenter(pg, &p, 1);
3764 		bp->b_pages[index] = p;
3765 	}
3766 	bp->b_npages = index;
3767 }
3768 
3769 /* Return pages associated with this buf to the vm system */
3770 static void
3771 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3772 {
3773 	vm_offset_t pg;
3774 	vm_page_t p;
3775 	int index, newnpages;
3776 
3777 	from = round_page(from);
3778 	to = round_page(to);
3779 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3780 
3781 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3782 		p = bp->b_pages[index];
3783 		if (p && (index < bp->b_npages)) {
3784 			if (p->busy) {
3785 				printf(
3786 			    "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3787 				    (intmax_t)bp->b_blkno,
3788 				    (intmax_t)bp->b_lblkno);
3789 			}
3790 			bp->b_pages[index] = NULL;
3791 			pmap_qremove(pg, 1);
3792 			p->wire_count--;
3793 			vm_page_free(p);
3794 			atomic_subtract_int(&cnt.v_wire_count, 1);
3795 		}
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 addr, kva;
3816 	vm_prot_t prot;
3817 	int pidx, i;
3818 	struct vm_page *m;
3819 	struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3820 
3821 	if (bp->b_bufsize < 0)
3822 		return (-1);
3823 	prot = VM_PROT_READ;
3824 	if (bp->b_iocmd == BIO_READ)
3825 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3826 	for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3827 	     addr < bp->b_data + bp->b_bufsize;
3828 	     addr += PAGE_SIZE, pidx++) {
3829 		/*
3830 		 * Do the vm_fault if needed; do the copy-on-write thing
3831 		 * when reading stuff off device into memory.
3832 		 *
3833 		 * NOTE! Must use pmap_extract() because addr may be in
3834 		 * the userland address space, and kextract is only guarenteed
3835 		 * to work for the kernland address space (see: sparc64 port).
3836 		 */
3837 retry:
3838 		if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3839 		    prot) < 0) {
3840 			vm_page_lock_queues();
3841 			for (i = 0; i < pidx; ++i) {
3842 				vm_page_unhold(bp->b_pages[i]);
3843 				bp->b_pages[i] = NULL;
3844 			}
3845 			vm_page_unlock_queues();
3846 			return(-1);
3847 		}
3848 		m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3849 		if (m == NULL)
3850 			goto retry;
3851 		bp->b_pages[pidx] = m;
3852 	}
3853 	if (pidx > btoc(MAXPHYS))
3854 		panic("vmapbuf: mapped more than MAXPHYS");
3855 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3856 
3857 	kva = bp->b_saveaddr;
3858 	bp->b_npages = pidx;
3859 	bp->b_saveaddr = bp->b_data;
3860 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3861 	return(0);
3862 }
3863 
3864 /*
3865  * Free the io map PTEs associated with this IO operation.
3866  * We also invalidate the TLB entries and restore the original b_addr.
3867  */
3868 void
3869 vunmapbuf(struct buf *bp)
3870 {
3871 	int pidx;
3872 	int npages;
3873 
3874 	npages = bp->b_npages;
3875 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3876 	vm_page_lock_queues();
3877 	for (pidx = 0; pidx < npages; pidx++)
3878 		vm_page_unhold(bp->b_pages[pidx]);
3879 	vm_page_unlock_queues();
3880 
3881 	bp->b_data = bp->b_saveaddr;
3882 }
3883 
3884 void
3885 bdone(struct buf *bp)
3886 {
3887 	struct mtx *mtxp;
3888 
3889 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3890 	mtx_lock(mtxp);
3891 	bp->b_flags |= B_DONE;
3892 	wakeup(bp);
3893 	mtx_unlock(mtxp);
3894 }
3895 
3896 void
3897 bwait(struct buf *bp, u_char pri, const char *wchan)
3898 {
3899 	struct mtx *mtxp;
3900 
3901 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3902 	mtx_lock(mtxp);
3903 	while ((bp->b_flags & B_DONE) == 0)
3904 		msleep(bp, mtxp, pri, wchan, 0);
3905 	mtx_unlock(mtxp);
3906 }
3907 
3908 int
3909 bufsync(struct bufobj *bo, int waitfor)
3910 {
3911 
3912 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3913 }
3914 
3915 void
3916 bufstrategy(struct bufobj *bo, struct buf *bp)
3917 {
3918 	int i = 0;
3919 	struct vnode *vp;
3920 
3921 	vp = bp->b_vp;
3922 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3923 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3924 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3925 	i = VOP_STRATEGY(vp, bp);
3926 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3927 }
3928 
3929 void
3930 bufobj_wrefl(struct bufobj *bo)
3931 {
3932 
3933 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3934 	ASSERT_BO_LOCKED(bo);
3935 	bo->bo_numoutput++;
3936 }
3937 
3938 void
3939 bufobj_wref(struct bufobj *bo)
3940 {
3941 
3942 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3943 	BO_LOCK(bo);
3944 	bo->bo_numoutput++;
3945 	BO_UNLOCK(bo);
3946 }
3947 
3948 void
3949 bufobj_wdrop(struct bufobj *bo)
3950 {
3951 
3952 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3953 	BO_LOCK(bo);
3954 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3955 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3956 		bo->bo_flag &= ~BO_WWAIT;
3957 		wakeup(&bo->bo_numoutput);
3958 	}
3959 	BO_UNLOCK(bo);
3960 }
3961 
3962 int
3963 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3964 {
3965 	int error;
3966 
3967 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3968 	ASSERT_BO_LOCKED(bo);
3969 	error = 0;
3970 	while (bo->bo_numoutput) {
3971 		bo->bo_flag |= BO_WWAIT;
3972 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3973 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3974 		if (error)
3975 			break;
3976 	}
3977 	return (error);
3978 }
3979 
3980 void
3981 bpin(struct buf *bp)
3982 {
3983 	struct mtx *mtxp;
3984 
3985 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3986 	mtx_lock(mtxp);
3987 	bp->b_pin_count++;
3988 	mtx_unlock(mtxp);
3989 }
3990 
3991 void
3992 bunpin(struct buf *bp)
3993 {
3994 	struct mtx *mtxp;
3995 
3996 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3997 	mtx_lock(mtxp);
3998 	if (--bp->b_pin_count == 0)
3999 		wakeup(bp);
4000 	mtx_unlock(mtxp);
4001 }
4002 
4003 void
4004 bunpin_wait(struct buf *bp)
4005 {
4006 	struct mtx *mtxp;
4007 
4008 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4009 	mtx_lock(mtxp);
4010 	while (bp->b_pin_count > 0)
4011 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4012 	mtx_unlock(mtxp);
4013 }
4014 
4015 #include "opt_ddb.h"
4016 #ifdef DDB
4017 #include <ddb/ddb.h>
4018 
4019 /* DDB command to show buffer data */
4020 DB_SHOW_COMMAND(buffer, db_show_buffer)
4021 {
4022 	/* get args */
4023 	struct buf *bp = (struct buf *)addr;
4024 
4025 	if (!have_addr) {
4026 		db_printf("usage: show buffer <addr>\n");
4027 		return;
4028 	}
4029 
4030 	db_printf("buf at %p\n", bp);
4031 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4032 	db_printf(
4033 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4034 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n",
4035 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4036 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4037 	    bp->b_dep.lh_first);
4038 	if (bp->b_npages) {
4039 		int i;
4040 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4041 		for (i = 0; i < bp->b_npages; i++) {
4042 			vm_page_t m;
4043 			m = bp->b_pages[i];
4044 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4045 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4046 			if ((i + 1) < bp->b_npages)
4047 				db_printf(",");
4048 		}
4049 		db_printf("\n");
4050 	}
4051 	db_printf(" ");
4052 	lockmgr_printinfo(&bp->b_lock);
4053 }
4054 
4055 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4056 {
4057 	struct buf *bp;
4058 	int i;
4059 
4060 	for (i = 0; i < nbuf; i++) {
4061 		bp = &buf[i];
4062 		if (BUF_ISLOCKED(bp)) {
4063 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4064 			db_printf("\n");
4065 		}
4066 	}
4067 }
4068 
4069 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4070 {
4071 	struct vnode *vp;
4072 	struct buf *bp;
4073 
4074 	if (!have_addr) {
4075 		db_printf("usage: show vnodebufs <addr>\n");
4076 		return;
4077 	}
4078 	vp = (struct vnode *)addr;
4079 	db_printf("Clean buffers:\n");
4080 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4081 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4082 		db_printf("\n");
4083 	}
4084 	db_printf("Dirty buffers:\n");
4085 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4086 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4087 		db_printf("\n");
4088 	}
4089 }
4090 #endif /* DDB */
4091