xref: /freebsd/sys/kern/vfs_bio.c (revision b60a118dd88613da59c373d9a6eff5f8d35b53ea)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * Portions of this software were developed by Konstantin Belousov
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * this file contains a new buffer I/O scheme implementing a coherent
34  * VM object and buffer cache scheme.  Pains have been taken to make
35  * sure that the performance degradation associated with schemes such
36  * as this is not realized.
37  *
38  * Author:  John S. Dyson
39  * Significant help during the development and debugging phases
40  * had been provided by David Greenman, also of the FreeBSD core team.
41  *
42  * see man buf(9) for more info.
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/conf.h>
52 #include <sys/buf.h>
53 #include <sys/devicestat.h>
54 #include <sys/eventhandler.h>
55 #include <sys/fail.h>
56 #include <sys/limits.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/kernel.h>
62 #include <sys/kthread.h>
63 #include <sys/proc.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
66 #include <sys/sysctl.h>
67 #include <sys/sysproto.h>
68 #include <sys/vmem.h>
69 #include <sys/vmmeter.h>
70 #include <sys/vnode.h>
71 #include <sys/watchdog.h>
72 #include <geom/geom.h>
73 #include <vm/vm.h>
74 #include <vm/vm_param.h>
75 #include <vm/vm_kern.h>
76 #include <vm/vm_pageout.h>
77 #include <vm/vm_page.h>
78 #include <vm/vm_object.h>
79 #include <vm/vm_extern.h>
80 #include <vm/vm_map.h>
81 #include <vm/swap_pager.h>
82 #include "opt_compat.h"
83 #include "opt_swap.h"
84 
85 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
86 
87 struct	bio_ops bioops;		/* I/O operation notification */
88 
89 struct	buf_ops buf_ops_bio = {
90 	.bop_name	=	"buf_ops_bio",
91 	.bop_write	=	bufwrite,
92 	.bop_strategy	=	bufstrategy,
93 	.bop_sync	=	bufsync,
94 	.bop_bdflush	=	bufbdflush,
95 };
96 
97 static struct buf *buf;		/* buffer header pool */
98 extern struct buf *swbuf;	/* Swap buffer header pool. */
99 caddr_t unmapped_buf;
100 
101 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
102 struct proc *bufdaemonproc;
103 
104 static int inmem(struct vnode *vp, daddr_t blkno);
105 static void vm_hold_free_pages(struct buf *bp, int newbsize);
106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
107 		vm_offset_t to);
108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
110 		vm_page_t m);
111 static void vfs_clean_pages_dirty_buf(struct buf *bp);
112 static void vfs_setdirty_locked_object(struct buf *bp);
113 static void vfs_vmio_invalidate(struct buf *bp);
114 static void vfs_vmio_truncate(struct buf *bp, int npages);
115 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
116 static int vfs_bio_clcheck(struct vnode *vp, int size,
117 		daddr_t lblkno, daddr_t blkno);
118 static int buf_flush(struct vnode *vp, int);
119 static int flushbufqueues(struct vnode *, int, int);
120 static void buf_daemon(void);
121 static void bremfreel(struct buf *bp);
122 static __inline void bd_wakeup(void);
123 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
124 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
125     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
126 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
127 #endif
128 
129 int vmiodirenable = TRUE;
130 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
131     "Use the VM system for directory writes");
132 long runningbufspace;
133 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
134     "Amount of presently outstanding async buffer io");
135 static long bufspace;
136 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
137     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
138 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
139     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
140 #else
141 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
142     "Physical memory used for buffers");
143 #endif
144 static long bufkvaspace;
145 SYSCTL_LONG(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 0,
146     "Kernel virtual memory used for buffers");
147 static long maxbufspace;
148 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
149     "Maximum allowed value of bufspace (including buf_daemon)");
150 static long bufmallocspace;
151 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
152     "Amount of malloced memory for buffers");
153 static long maxbufmallocspace;
154 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
155     "Maximum amount of malloced memory for buffers");
156 static long lobufspace;
157 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
158     "Minimum amount of buffers we want to have");
159 long hibufspace;
160 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
161     "Maximum allowed value of bufspace (excluding buf_daemon)");
162 static int bufreusecnt;
163 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
164     "Number of times we have reused a buffer");
165 static int buffreekvacnt;
166 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
167     "Number of times we have freed the KVA space from some buffer");
168 static int bufdefragcnt;
169 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
170     "Number of times we have had to repeat buffer allocation to defragment");
171 static long lorunningspace;
172 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
173     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
174     "Minimum preferred space used for in-progress I/O");
175 static long hirunningspace;
176 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
177     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
178     "Maximum amount of space to use for in-progress I/O");
179 int dirtybufferflushes;
180 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
181     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
182 int bdwriteskip;
183 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
184     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
185 int altbufferflushes;
186 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
187     0, "Number of fsync flushes to limit dirty buffers");
188 static int recursiveflushes;
189 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
190     0, "Number of flushes skipped due to being recursive");
191 static int numdirtybuffers;
192 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
193     "Number of buffers that are dirty (has unwritten changes) at the moment");
194 static int lodirtybuffers;
195 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
196     "How many buffers we want to have free before bufdaemon can sleep");
197 static int hidirtybuffers;
198 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
199     "When the number of dirty buffers is considered severe");
200 int dirtybufthresh;
201 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
202     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
203 static int numfreebuffers;
204 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
205     "Number of free buffers");
206 static int lofreebuffers;
207 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
208    "XXX Unused");
209 static int hifreebuffers;
210 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
211    "XXX Complicatedly unused");
212 static int getnewbufcalls;
213 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
214    "Number of calls to getnewbuf");
215 static int getnewbufrestarts;
216 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
217     "Number of times getnewbuf has had to restart a buffer aquisition");
218 static int mappingrestarts;
219 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
220     "Number of times getblk has had to restart a buffer mapping for "
221     "unmapped buffer");
222 static int flushbufqtarget = 100;
223 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
224     "Amount of work to do in flushbufqueues when helping bufdaemon");
225 static long notbufdflushes;
226 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes, 0,
227     "Number of dirty buffer flushes done by the bufdaemon helpers");
228 static long barrierwrites;
229 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
230     "Number of barrier writes");
231 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
232     &unmapped_buf_allowed, 0,
233     "Permit the use of the unmapped i/o");
234 
235 /*
236  * Lock for the non-dirty bufqueues
237  */
238 static struct mtx_padalign bqclean;
239 
240 /*
241  * Lock for the dirty queue.
242  */
243 static struct mtx_padalign bqdirty;
244 
245 /*
246  * This lock synchronizes access to bd_request.
247  */
248 static struct mtx_padalign bdlock;
249 
250 /*
251  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
252  * waitrunningbufspace().
253  */
254 static struct mtx_padalign rbreqlock;
255 
256 /*
257  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
258  */
259 static struct rwlock_padalign nblock;
260 
261 /*
262  * Lock that protects bdirtywait.
263  */
264 static struct mtx_padalign bdirtylock;
265 
266 /*
267  * Wakeup point for bufdaemon, as well as indicator of whether it is already
268  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
269  * is idling.
270  */
271 static int bd_request;
272 
273 /*
274  * Request for the buf daemon to write more buffers than is indicated by
275  * lodirtybuf.  This may be necessary to push out excess dependencies or
276  * defragment the address space where a simple count of the number of dirty
277  * buffers is insufficient to characterize the demand for flushing them.
278  */
279 static int bd_speedupreq;
280 
281 /*
282  * bogus page -- for I/O to/from partially complete buffers
283  * this is a temporary solution to the problem, but it is not
284  * really that bad.  it would be better to split the buffer
285  * for input in the case of buffers partially already in memory,
286  * but the code is intricate enough already.
287  */
288 vm_page_t bogus_page;
289 
290 /*
291  * Synchronization (sleep/wakeup) variable for active buffer space requests.
292  * Set when wait starts, cleared prior to wakeup().
293  * Used in runningbufwakeup() and waitrunningbufspace().
294  */
295 static int runningbufreq;
296 
297 /*
298  * Synchronization (sleep/wakeup) variable for buffer requests.
299  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
300  * by and/or.
301  * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
302  * getnewbuf(), and getblk().
303  */
304 static volatile int needsbuffer;
305 
306 /*
307  * Synchronization for bwillwrite() waiters.
308  */
309 static int bdirtywait;
310 
311 /*
312  * Definitions for the buffer free lists.
313  */
314 #define BUFFER_QUEUES	4	/* number of free buffer queues */
315 
316 #define QUEUE_NONE	0	/* on no queue */
317 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
318 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
319 #define QUEUE_EMPTY	3	/* empty buffer headers */
320 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
321 
322 /* Queues for free buffers with various properties */
323 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
324 #ifdef INVARIANTS
325 static int bq_len[BUFFER_QUEUES];
326 #endif
327 
328 /*
329  * Single global constant for BUF_WMESG, to avoid getting multiple references.
330  * buf_wmesg is referred from macros.
331  */
332 const char *buf_wmesg = BUF_WMESG;
333 
334 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
335 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
336 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
337 
338 static int
339 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
340 {
341 	long value;
342 	int error;
343 
344 	value = *(long *)arg1;
345 	error = sysctl_handle_long(oidp, &value, 0, req);
346 	if (error != 0 || req->newptr == NULL)
347 		return (error);
348 	mtx_lock(&rbreqlock);
349 	if (arg1 == &hirunningspace) {
350 		if (value < lorunningspace)
351 			error = EINVAL;
352 		else
353 			hirunningspace = value;
354 	} else {
355 		KASSERT(arg1 == &lorunningspace,
356 		    ("%s: unknown arg1", __func__));
357 		if (value > hirunningspace)
358 			error = EINVAL;
359 		else
360 			lorunningspace = value;
361 	}
362 	mtx_unlock(&rbreqlock);
363 	return (error);
364 }
365 
366 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
367     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
368 static int
369 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
370 {
371 	long lvalue;
372 	int ivalue;
373 
374 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
375 		return (sysctl_handle_long(oidp, arg1, arg2, req));
376 	lvalue = *(long *)arg1;
377 	if (lvalue > INT_MAX)
378 		/* On overflow, still write out a long to trigger ENOMEM. */
379 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
380 	ivalue = lvalue;
381 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
382 }
383 #endif
384 
385 /*
386  *	bqlock:
387  *
388  *	Return the appropriate queue lock based on the index.
389  */
390 static inline struct mtx *
391 bqlock(int qindex)
392 {
393 
394 	if (qindex == QUEUE_DIRTY)
395 		return (struct mtx *)(&bqdirty);
396 	return (struct mtx *)(&bqclean);
397 }
398 
399 /*
400  *	bdirtywakeup:
401  *
402  *	Wakeup any bwillwrite() waiters.
403  */
404 static void
405 bdirtywakeup(void)
406 {
407 	mtx_lock(&bdirtylock);
408 	if (bdirtywait) {
409 		bdirtywait = 0;
410 		wakeup(&bdirtywait);
411 	}
412 	mtx_unlock(&bdirtylock);
413 }
414 
415 /*
416  *	bdirtysub:
417  *
418  *	Decrement the numdirtybuffers count by one and wakeup any
419  *	threads blocked in bwillwrite().
420  */
421 static void
422 bdirtysub(void)
423 {
424 
425 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
426 	    (lodirtybuffers + hidirtybuffers) / 2)
427 		bdirtywakeup();
428 }
429 
430 /*
431  *	bdirtyadd:
432  *
433  *	Increment the numdirtybuffers count by one and wakeup the buf
434  *	daemon if needed.
435  */
436 static void
437 bdirtyadd(void)
438 {
439 
440 	/*
441 	 * Only do the wakeup once as we cross the boundary.  The
442 	 * buf daemon will keep running until the condition clears.
443 	 */
444 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
445 	    (lodirtybuffers + hidirtybuffers) / 2)
446 		bd_wakeup();
447 }
448 
449 /*
450  *	bufspacewakeup:
451  *
452  *	Called when buffer space is potentially available for recovery.
453  *	getnewbuf() will block on this flag when it is unable to free
454  *	sufficient buffer space.  Buffer space becomes recoverable when
455  *	bp's get placed back in the queues.
456  */
457 static __inline void
458 bufspacewakeup(void)
459 {
460 	int need_wakeup, on;
461 
462 	/*
463 	 * If someone is waiting for bufspace, wake them up.  Even
464 	 * though we may not have freed the kva space yet, the waiting
465 	 * process will be able to now.
466 	 */
467 	rw_rlock(&nblock);
468 	for (;;) {
469 		need_wakeup = 0;
470 		on = needsbuffer;
471 		if ((on & VFS_BIO_NEED_BUFSPACE) == 0)
472 			break;
473 		need_wakeup = 1;
474 		if (atomic_cmpset_rel_int(&needsbuffer, on,
475 		    on & ~VFS_BIO_NEED_BUFSPACE))
476 			break;
477 	}
478 	if (need_wakeup)
479 		wakeup(__DEVOLATILE(void *, &needsbuffer));
480 	rw_runlock(&nblock);
481 }
482 
483 /*
484  *	bufspaceadjust:
485  *
486  *	Adjust the reported bufspace for a KVA managed buffer, possibly
487  * 	waking any waiters.
488  */
489 static void
490 bufspaceadjust(struct buf *bp, int bufsize)
491 {
492 	int diff;
493 
494 	KASSERT((bp->b_flags & B_MALLOC) == 0,
495 	    ("bufspaceadjust: malloc buf %p", bp));
496 	diff = bufsize - bp->b_bufsize;
497 	if (diff < 0) {
498 		atomic_subtract_long(&bufspace, -diff);
499 		bufspacewakeup();
500 	} else
501 		atomic_add_long(&bufspace, diff);
502 	bp->b_bufsize = bufsize;
503 }
504 
505 /*
506  *	bufmallocadjust:
507  *
508  *	Adjust the reported bufspace for a malloc managed buffer, possibly
509  *	waking any waiters.
510  */
511 static void
512 bufmallocadjust(struct buf *bp, int bufsize)
513 {
514 	int diff;
515 
516 	KASSERT((bp->b_flags & B_MALLOC) != 0,
517 	    ("bufmallocadjust: non-malloc buf %p", bp));
518 	diff = bufsize - bp->b_bufsize;
519 	if (diff < 0) {
520 		atomic_subtract_long(&bufmallocspace, -diff);
521 		bufspacewakeup();
522 	} else
523 		atomic_add_long(&bufmallocspace, diff);
524 	bp->b_bufsize = bufsize;
525 }
526 
527 /*
528  *	runningwakeup:
529  *
530  *	Wake up processes that are waiting on asynchronous writes to fall
531  *	below lorunningspace.
532  */
533 static void
534 runningwakeup(void)
535 {
536 
537 	mtx_lock(&rbreqlock);
538 	if (runningbufreq) {
539 		runningbufreq = 0;
540 		wakeup(&runningbufreq);
541 	}
542 	mtx_unlock(&rbreqlock);
543 }
544 
545 /*
546  *	runningbufwakeup:
547  *
548  *	Decrement the outstanding write count according.
549  */
550 void
551 runningbufwakeup(struct buf *bp)
552 {
553 	long space, bspace;
554 
555 	bspace = bp->b_runningbufspace;
556 	if (bspace == 0)
557 		return;
558 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
559 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
560 	    space, bspace));
561 	bp->b_runningbufspace = 0;
562 	/*
563 	 * Only acquire the lock and wakeup on the transition from exceeding
564 	 * the threshold to falling below it.
565 	 */
566 	if (space < lorunningspace)
567 		return;
568 	if (space - bspace > lorunningspace)
569 		return;
570 	runningwakeup();
571 }
572 
573 /*
574  *	bufcountadd:
575  *
576  *	Called when a buffer has been added to one of the free queues to
577  *	account for the buffer and to wakeup anyone waiting for free buffers.
578  *	This typically occurs when large amounts of metadata are being handled
579  *	by the buffer cache ( else buffer space runs out first, usually ).
580  */
581 static __inline void
582 bufcountadd(struct buf *bp)
583 {
584 	int mask, need_wakeup, old, on;
585 
586 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
587 	    ("buf %p already counted as free", bp));
588 	bp->b_flags |= B_INFREECNT;
589 	old = atomic_fetchadd_int(&numfreebuffers, 1);
590 	KASSERT(old >= 0 && old < nbuf,
591 	    ("numfreebuffers climbed to %d", old + 1));
592 	mask = VFS_BIO_NEED_ANY;
593 	if (numfreebuffers >= hifreebuffers)
594 		mask |= VFS_BIO_NEED_FREE;
595 	rw_rlock(&nblock);
596 	for (;;) {
597 		need_wakeup = 0;
598 		on = needsbuffer;
599 		if (on == 0)
600 			break;
601 		need_wakeup = 1;
602 		if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask))
603 			break;
604 	}
605 	if (need_wakeup)
606 		wakeup(__DEVOLATILE(void *, &needsbuffer));
607 	rw_runlock(&nblock);
608 }
609 
610 /*
611  *	bufcountsub:
612  *
613  *	Decrement the numfreebuffers count as needed.
614  */
615 static void
616 bufcountsub(struct buf *bp)
617 {
618 	int old;
619 
620 	/*
621 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
622 	 * delayed-write, the buffer was free and we must decrement
623 	 * numfreebuffers.
624 	 */
625 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
626 		KASSERT((bp->b_flags & B_INFREECNT) != 0,
627 		    ("buf %p not counted in numfreebuffers", bp));
628 		bp->b_flags &= ~B_INFREECNT;
629 		old = atomic_fetchadd_int(&numfreebuffers, -1);
630 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
631 	}
632 }
633 
634 /*
635  *	waitrunningbufspace()
636  *
637  *	runningbufspace is a measure of the amount of I/O currently
638  *	running.  This routine is used in async-write situations to
639  *	prevent creating huge backups of pending writes to a device.
640  *	Only asynchronous writes are governed by this function.
641  *
642  *	This does NOT turn an async write into a sync write.  It waits
643  *	for earlier writes to complete and generally returns before the
644  *	caller's write has reached the device.
645  */
646 void
647 waitrunningbufspace(void)
648 {
649 
650 	mtx_lock(&rbreqlock);
651 	while (runningbufspace > hirunningspace) {
652 		runningbufreq = 1;
653 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
654 	}
655 	mtx_unlock(&rbreqlock);
656 }
657 
658 
659 /*
660  *	vfs_buf_test_cache:
661  *
662  *	Called when a buffer is extended.  This function clears the B_CACHE
663  *	bit if the newly extended portion of the buffer does not contain
664  *	valid data.
665  */
666 static __inline void
667 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
668     vm_offset_t size, vm_page_t m)
669 {
670 
671 	VM_OBJECT_ASSERT_LOCKED(m->object);
672 	if (bp->b_flags & B_CACHE) {
673 		int base = (foff + off) & PAGE_MASK;
674 		if (vm_page_is_valid(m, base, size) == 0)
675 			bp->b_flags &= ~B_CACHE;
676 	}
677 }
678 
679 /* Wake up the buffer daemon if necessary */
680 static __inline void
681 bd_wakeup(void)
682 {
683 
684 	mtx_lock(&bdlock);
685 	if (bd_request == 0) {
686 		bd_request = 1;
687 		wakeup(&bd_request);
688 	}
689 	mtx_unlock(&bdlock);
690 }
691 
692 /*
693  * bd_speedup - speedup the buffer cache flushing code
694  */
695 void
696 bd_speedup(void)
697 {
698 	int needwake;
699 
700 	mtx_lock(&bdlock);
701 	needwake = 0;
702 	if (bd_speedupreq == 0 || bd_request == 0)
703 		needwake = 1;
704 	bd_speedupreq = 1;
705 	bd_request = 1;
706 	if (needwake)
707 		wakeup(&bd_request);
708 	mtx_unlock(&bdlock);
709 }
710 
711 #ifndef NSWBUF_MIN
712 #define	NSWBUF_MIN	16
713 #endif
714 
715 #ifdef __i386__
716 #define	TRANSIENT_DENOM	5
717 #else
718 #define	TRANSIENT_DENOM 10
719 #endif
720 
721 /*
722  * Calculating buffer cache scaling values and reserve space for buffer
723  * headers.  This is called during low level kernel initialization and
724  * may be called more then once.  We CANNOT write to the memory area
725  * being reserved at this time.
726  */
727 caddr_t
728 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
729 {
730 	int tuned_nbuf;
731 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
732 
733 	/*
734 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
735 	 * PAGE_SIZE is >= 1K)
736 	 */
737 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
738 
739 	/*
740 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
741 	 * For the first 64MB of ram nominally allocate sufficient buffers to
742 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
743 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
744 	 * the buffer cache we limit the eventual kva reservation to
745 	 * maxbcache bytes.
746 	 *
747 	 * factor represents the 1/4 x ram conversion.
748 	 */
749 	if (nbuf == 0) {
750 		int factor = 4 * BKVASIZE / 1024;
751 
752 		nbuf = 50;
753 		if (physmem_est > 4096)
754 			nbuf += min((physmem_est - 4096) / factor,
755 			    65536 / factor);
756 		if (physmem_est > 65536)
757 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
758 			    32 * 1024 * 1024 / (factor * 5));
759 
760 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
761 			nbuf = maxbcache / BKVASIZE;
762 		tuned_nbuf = 1;
763 	} else
764 		tuned_nbuf = 0;
765 
766 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
767 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
768 	if (nbuf > maxbuf) {
769 		if (!tuned_nbuf)
770 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
771 			    maxbuf);
772 		nbuf = maxbuf;
773 	}
774 
775 	/*
776 	 * Ideal allocation size for the transient bio submap is 10%
777 	 * of the maximal space buffer map.  This roughly corresponds
778 	 * to the amount of the buffer mapped for typical UFS load.
779 	 *
780 	 * Clip the buffer map to reserve space for the transient
781 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
782 	 * maximum buffer map extent on the platform.
783 	 *
784 	 * The fall-back to the maxbuf in case of maxbcache unset,
785 	 * allows to not trim the buffer KVA for the architectures
786 	 * with ample KVA space.
787 	 */
788 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
789 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
790 		buf_sz = (long)nbuf * BKVASIZE;
791 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
792 		    (TRANSIENT_DENOM - 1)) {
793 			/*
794 			 * There is more KVA than memory.  Do not
795 			 * adjust buffer map size, and assign the rest
796 			 * of maxbuf to transient map.
797 			 */
798 			biotmap_sz = maxbuf_sz - buf_sz;
799 		} else {
800 			/*
801 			 * Buffer map spans all KVA we could afford on
802 			 * this platform.  Give 10% (20% on i386) of
803 			 * the buffer map to the transient bio map.
804 			 */
805 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
806 			buf_sz -= biotmap_sz;
807 		}
808 		if (biotmap_sz / INT_MAX > MAXPHYS)
809 			bio_transient_maxcnt = INT_MAX;
810 		else
811 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
812 		/*
813 		 * Artifically limit to 1024 simultaneous in-flight I/Os
814 		 * using the transient mapping.
815 		 */
816 		if (bio_transient_maxcnt > 1024)
817 			bio_transient_maxcnt = 1024;
818 		if (tuned_nbuf)
819 			nbuf = buf_sz / BKVASIZE;
820 	}
821 
822 	/*
823 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
824 	 * We have no less then 16 and no more then 256.
825 	 */
826 	nswbuf = min(nbuf / 4, 256);
827 	TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
828 	if (nswbuf < NSWBUF_MIN)
829 		nswbuf = NSWBUF_MIN;
830 
831 	/*
832 	 * Reserve space for the buffer cache buffers
833 	 */
834 	swbuf = (void *)v;
835 	v = (caddr_t)(swbuf + nswbuf);
836 	buf = (void *)v;
837 	v = (caddr_t)(buf + nbuf);
838 
839 	return(v);
840 }
841 
842 /* Initialize the buffer subsystem.  Called before use of any buffers. */
843 void
844 bufinit(void)
845 {
846 	struct buf *bp;
847 	int i;
848 
849 	CTASSERT(MAXBCACHEBUF >= MAXBSIZE);
850 	mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
851 	mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
852 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
853 	rw_init(&nblock, "needsbuffer lock");
854 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
855 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
856 
857 	/* next, make a null set of free lists */
858 	for (i = 0; i < BUFFER_QUEUES; i++)
859 		TAILQ_INIT(&bufqueues[i]);
860 
861 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
862 
863 	/* finally, initialize each buffer header and stick on empty q */
864 	for (i = 0; i < nbuf; i++) {
865 		bp = &buf[i];
866 		bzero(bp, sizeof *bp);
867 		bp->b_flags = B_INVAL | B_INFREECNT;
868 		bp->b_rcred = NOCRED;
869 		bp->b_wcred = NOCRED;
870 		bp->b_qindex = QUEUE_EMPTY;
871 		bp->b_xflags = 0;
872 		bp->b_data = bp->b_kvabase = unmapped_buf;
873 		LIST_INIT(&bp->b_dep);
874 		BUF_LOCKINIT(bp);
875 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
876 #ifdef INVARIANTS
877 		bq_len[QUEUE_EMPTY]++;
878 #endif
879 	}
880 
881 	/*
882 	 * maxbufspace is the absolute maximum amount of buffer space we are
883 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
884 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
885 	 * used by most other processes.  The differential is required to
886 	 * ensure that buf_daemon is able to run when other processes might
887 	 * be blocked waiting for buffer space.
888 	 *
889 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
890 	 * this may result in KVM fragmentation which is not handled optimally
891 	 * by the system.
892 	 */
893 	maxbufspace = (long)nbuf * BKVASIZE;
894 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10);
895 	lobufspace = hibufspace - MAXBCACHEBUF;
896 
897 	/*
898 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
899 	 * arbitrarily and may need further tuning. It corresponds to
900 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
901 	 * which fits with many RAID controllers' tagged queuing limits.
902 	 * The lower 1 MiB limit is the historical upper limit for
903 	 * hirunningspace.
904 	 */
905 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF),
906 	    16 * 1024 * 1024), 1024 * 1024);
907 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF);
908 
909 /*
910  * Limit the amount of malloc memory since it is wired permanently into
911  * the kernel space.  Even though this is accounted for in the buffer
912  * allocation, we don't want the malloced region to grow uncontrolled.
913  * The malloc scheme improves memory utilization significantly on average
914  * (small) directories.
915  */
916 	maxbufmallocspace = hibufspace / 20;
917 
918 /*
919  * Reduce the chance of a deadlock occuring by limiting the number
920  * of delayed-write dirty buffers we allow to stack up.
921  */
922 	hidirtybuffers = nbuf / 4 + 20;
923 	dirtybufthresh = hidirtybuffers * 9 / 10;
924 	numdirtybuffers = 0;
925 /*
926  * To support extreme low-memory systems, make sure hidirtybuffers cannot
927  * eat up all available buffer space.  This occurs when our minimum cannot
928  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
929  * BKVASIZE'd buffers.
930  */
931 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
932 		hidirtybuffers >>= 1;
933 	}
934 	lodirtybuffers = hidirtybuffers / 2;
935 
936 /*
937  * Try to keep the number of free buffers in the specified range,
938  * and give special processes (e.g. like buf_daemon) access to an
939  * emergency reserve.
940  */
941 	lofreebuffers = nbuf / 18 + 5;
942 	hifreebuffers = 2 * lofreebuffers;
943 	numfreebuffers = nbuf;
944 
945 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
946 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
947 }
948 
949 #ifdef INVARIANTS
950 static inline void
951 vfs_buf_check_mapped(struct buf *bp)
952 {
953 
954 	KASSERT(bp->b_kvabase != unmapped_buf,
955 	    ("mapped buf: b_kvabase was not updated %p", bp));
956 	KASSERT(bp->b_data != unmapped_buf,
957 	    ("mapped buf: b_data was not updated %p", bp));
958 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
959 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
960 }
961 
962 static inline void
963 vfs_buf_check_unmapped(struct buf *bp)
964 {
965 
966 	KASSERT(bp->b_data == unmapped_buf,
967 	    ("unmapped buf: corrupted b_data %p", bp));
968 }
969 
970 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
971 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
972 #else
973 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
974 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
975 #endif
976 
977 static int
978 isbufbusy(struct buf *bp)
979 {
980 	if (((bp->b_flags & (B_INVAL | B_PERSISTENT)) == 0 &&
981 	    BUF_ISLOCKED(bp)) ||
982 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
983 		return (1);
984 	return (0);
985 }
986 
987 /*
988  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
989  */
990 void
991 bufshutdown(int show_busybufs)
992 {
993 	static int first_buf_printf = 1;
994 	struct buf *bp;
995 	int iter, nbusy, pbusy;
996 #ifndef PREEMPTION
997 	int subiter;
998 #endif
999 
1000 	/*
1001 	 * Sync filesystems for shutdown
1002 	 */
1003 	wdog_kern_pat(WD_LASTVAL);
1004 	sys_sync(curthread, NULL);
1005 
1006 	/*
1007 	 * With soft updates, some buffers that are
1008 	 * written will be remarked as dirty until other
1009 	 * buffers are written.
1010 	 */
1011 	for (iter = pbusy = 0; iter < 20; iter++) {
1012 		nbusy = 0;
1013 		for (bp = &buf[nbuf]; --bp >= buf; )
1014 			if (isbufbusy(bp))
1015 				nbusy++;
1016 		if (nbusy == 0) {
1017 			if (first_buf_printf)
1018 				printf("All buffers synced.");
1019 			break;
1020 		}
1021 		if (first_buf_printf) {
1022 			printf("Syncing disks, buffers remaining... ");
1023 			first_buf_printf = 0;
1024 		}
1025 		printf("%d ", nbusy);
1026 		if (nbusy < pbusy)
1027 			iter = 0;
1028 		pbusy = nbusy;
1029 
1030 		wdog_kern_pat(WD_LASTVAL);
1031 		sys_sync(curthread, NULL);
1032 
1033 #ifdef PREEMPTION
1034 		/*
1035 		 * Drop Giant and spin for a while to allow
1036 		 * interrupt threads to run.
1037 		 */
1038 		DROP_GIANT();
1039 		DELAY(50000 * iter);
1040 		PICKUP_GIANT();
1041 #else
1042 		/*
1043 		 * Drop Giant and context switch several times to
1044 		 * allow interrupt threads to run.
1045 		 */
1046 		DROP_GIANT();
1047 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1048 			thread_lock(curthread);
1049 			mi_switch(SW_VOL, NULL);
1050 			thread_unlock(curthread);
1051 			DELAY(1000);
1052 		}
1053 		PICKUP_GIANT();
1054 #endif
1055 	}
1056 	printf("\n");
1057 	/*
1058 	 * Count only busy local buffers to prevent forcing
1059 	 * a fsck if we're just a client of a wedged NFS server
1060 	 */
1061 	nbusy = 0;
1062 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1063 		if (isbufbusy(bp)) {
1064 #if 0
1065 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1066 			if (bp->b_dev == NULL) {
1067 				TAILQ_REMOVE(&mountlist,
1068 				    bp->b_vp->v_mount, mnt_list);
1069 				continue;
1070 			}
1071 #endif
1072 			nbusy++;
1073 			if (show_busybufs > 0) {
1074 				printf(
1075 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1076 				    nbusy, bp, bp->b_vp, bp->b_flags,
1077 				    (intmax_t)bp->b_blkno,
1078 				    (intmax_t)bp->b_lblkno);
1079 				BUF_LOCKPRINTINFO(bp);
1080 				if (show_busybufs > 1)
1081 					vn_printf(bp->b_vp,
1082 					    "vnode content: ");
1083 			}
1084 		}
1085 	}
1086 	if (nbusy) {
1087 		/*
1088 		 * Failed to sync all blocks. Indicate this and don't
1089 		 * unmount filesystems (thus forcing an fsck on reboot).
1090 		 */
1091 		printf("Giving up on %d buffers\n", nbusy);
1092 		DELAY(5000000);	/* 5 seconds */
1093 	} else {
1094 		if (!first_buf_printf)
1095 			printf("Final sync complete\n");
1096 		/*
1097 		 * Unmount filesystems
1098 		 */
1099 		if (panicstr == 0)
1100 			vfs_unmountall();
1101 	}
1102 	swapoff_all();
1103 	DELAY(100000);		/* wait for console output to finish */
1104 }
1105 
1106 static void
1107 bpmap_qenter(struct buf *bp)
1108 {
1109 
1110 	BUF_CHECK_MAPPED(bp);
1111 
1112 	/*
1113 	 * bp->b_data is relative to bp->b_offset, but
1114 	 * bp->b_offset may be offset into the first page.
1115 	 */
1116 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1117 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1118 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1119 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1120 }
1121 
1122 /*
1123  *	binsfree:
1124  *
1125  *	Insert the buffer into the appropriate free list.
1126  */
1127 static void
1128 binsfree(struct buf *bp, int qindex)
1129 {
1130 	struct mtx *olock, *nlock;
1131 
1132 	BUF_ASSERT_XLOCKED(bp);
1133 
1134 	nlock = bqlock(qindex);
1135 	/* Handle delayed bremfree() processing. */
1136 	if (bp->b_flags & B_REMFREE) {
1137 		olock = bqlock(bp->b_qindex);
1138 		mtx_lock(olock);
1139 		bremfreel(bp);
1140 		if (olock != nlock) {
1141 			mtx_unlock(olock);
1142 			mtx_lock(nlock);
1143 		}
1144 	} else
1145 		mtx_lock(nlock);
1146 
1147 	if (bp->b_qindex != QUEUE_NONE)
1148 		panic("binsfree: free buffer onto another queue???");
1149 
1150 	bp->b_qindex = qindex;
1151 	if (bp->b_flags & B_AGE)
1152 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1153 	else
1154 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1155 #ifdef INVARIANTS
1156 	bq_len[bp->b_qindex]++;
1157 #endif
1158 	mtx_unlock(nlock);
1159 
1160 	/*
1161 	 * Something we can maybe free or reuse.
1162 	 */
1163 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1164 		bufspacewakeup();
1165 
1166 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1167 		bufcountadd(bp);
1168 }
1169 
1170 /*
1171  *	bremfree:
1172  *
1173  *	Mark the buffer for removal from the appropriate free list.
1174  *
1175  */
1176 void
1177 bremfree(struct buf *bp)
1178 {
1179 
1180 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1181 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1182 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1183 	KASSERT(bp->b_qindex != QUEUE_NONE,
1184 	    ("bremfree: buffer %p not on a queue.", bp));
1185 	BUF_ASSERT_XLOCKED(bp);
1186 
1187 	bp->b_flags |= B_REMFREE;
1188 	bufcountsub(bp);
1189 }
1190 
1191 /*
1192  *	bremfreef:
1193  *
1194  *	Force an immediate removal from a free list.  Used only in nfs when
1195  *	it abuses the b_freelist pointer.
1196  */
1197 void
1198 bremfreef(struct buf *bp)
1199 {
1200 	struct mtx *qlock;
1201 
1202 	qlock = bqlock(bp->b_qindex);
1203 	mtx_lock(qlock);
1204 	bremfreel(bp);
1205 	mtx_unlock(qlock);
1206 }
1207 
1208 /*
1209  *	bremfreel:
1210  *
1211  *	Removes a buffer from the free list, must be called with the
1212  *	correct qlock held.
1213  */
1214 static void
1215 bremfreel(struct buf *bp)
1216 {
1217 
1218 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1219 	    bp, bp->b_vp, bp->b_flags);
1220 	KASSERT(bp->b_qindex != QUEUE_NONE,
1221 	    ("bremfreel: buffer %p not on a queue.", bp));
1222 	BUF_ASSERT_XLOCKED(bp);
1223 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1224 
1225 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1226 #ifdef INVARIANTS
1227 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1228 	    bp->b_qindex));
1229 	bq_len[bp->b_qindex]--;
1230 #endif
1231 	bp->b_qindex = QUEUE_NONE;
1232 	/*
1233 	 * If this was a delayed bremfree() we only need to remove the buffer
1234 	 * from the queue and return the stats are already done.
1235 	 */
1236 	if (bp->b_flags & B_REMFREE) {
1237 		bp->b_flags &= ~B_REMFREE;
1238 		return;
1239 	}
1240 	bufcountsub(bp);
1241 }
1242 
1243 /*
1244  *	bufkvafree:
1245  *
1246  *	Free the kva allocation for a buffer.
1247  *
1248  */
1249 static void
1250 bufkvafree(struct buf *bp)
1251 {
1252 
1253 #ifdef INVARIANTS
1254 	if (bp->b_kvasize == 0) {
1255 		KASSERT(bp->b_kvabase == unmapped_buf &&
1256 		    bp->b_data == unmapped_buf,
1257 		    ("Leaked KVA space on %p", bp));
1258 	} else if (buf_mapped(bp))
1259 		BUF_CHECK_MAPPED(bp);
1260 	else
1261 		BUF_CHECK_UNMAPPED(bp);
1262 #endif
1263 	if (bp->b_kvasize == 0)
1264 		return;
1265 
1266 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
1267 	atomic_subtract_long(&bufkvaspace, bp->b_kvasize);
1268 	atomic_add_int(&buffreekvacnt, 1);
1269 	bp->b_data = bp->b_kvabase = unmapped_buf;
1270 	bp->b_kvasize = 0;
1271 }
1272 
1273 /*
1274  *	bufkvaalloc:
1275  *
1276  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
1277  */
1278 static int
1279 bufkvaalloc(struct buf *bp, int maxsize, int gbflags)
1280 {
1281 	vm_offset_t addr;
1282 	int error;
1283 
1284 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
1285 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
1286 
1287 	bufkvafree(bp);
1288 
1289 	addr = 0;
1290 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
1291 	if (error != 0) {
1292 		/*
1293 		 * Buffer map is too fragmented.  Request the caller
1294 		 * to defragment the map.
1295 		 */
1296 		atomic_add_int(&bufdefragcnt, 1);
1297 		return (error);
1298 	}
1299 	bp->b_kvabase = (caddr_t)addr;
1300 	bp->b_kvasize = maxsize;
1301 	atomic_add_long(&bufkvaspace, bp->b_kvasize);
1302 	if ((gbflags & GB_UNMAPPED) != 0) {
1303 		bp->b_data = unmapped_buf;
1304 		BUF_CHECK_UNMAPPED(bp);
1305 	} else {
1306 		bp->b_data = bp->b_kvabase;
1307 		BUF_CHECK_MAPPED(bp);
1308 	}
1309 	return (0);
1310 }
1311 
1312 /*
1313  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1314  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1315  * the buffer is valid and we do not have to do anything.
1316  */
1317 void
1318 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1319     int cnt, struct ucred * cred)
1320 {
1321 	struct buf *rabp;
1322 	int i;
1323 
1324 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1325 		if (inmem(vp, *rablkno))
1326 			continue;
1327 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1328 
1329 		if ((rabp->b_flags & B_CACHE) == 0) {
1330 			if (!TD_IS_IDLETHREAD(curthread))
1331 				curthread->td_ru.ru_inblock++;
1332 			rabp->b_flags |= B_ASYNC;
1333 			rabp->b_flags &= ~B_INVAL;
1334 			rabp->b_ioflags &= ~BIO_ERROR;
1335 			rabp->b_iocmd = BIO_READ;
1336 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1337 				rabp->b_rcred = crhold(cred);
1338 			vfs_busy_pages(rabp, 0);
1339 			BUF_KERNPROC(rabp);
1340 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1341 			bstrategy(rabp);
1342 		} else {
1343 			brelse(rabp);
1344 		}
1345 	}
1346 }
1347 
1348 /*
1349  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1350  *
1351  * Get a buffer with the specified data.  Look in the cache first.  We
1352  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1353  * is set, the buffer is valid and we do not have to do anything, see
1354  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1355  */
1356 int
1357 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1358     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1359 {
1360 	struct buf *bp;
1361 	int rv = 0, readwait = 0;
1362 
1363 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1364 	/*
1365 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1366 	 */
1367 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1368 	if (bp == NULL)
1369 		return (EBUSY);
1370 
1371 	/* if not found in cache, do some I/O */
1372 	if ((bp->b_flags & B_CACHE) == 0) {
1373 		if (!TD_IS_IDLETHREAD(curthread))
1374 			curthread->td_ru.ru_inblock++;
1375 		bp->b_iocmd = BIO_READ;
1376 		bp->b_flags &= ~B_INVAL;
1377 		bp->b_ioflags &= ~BIO_ERROR;
1378 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1379 			bp->b_rcred = crhold(cred);
1380 		vfs_busy_pages(bp, 0);
1381 		bp->b_iooffset = dbtob(bp->b_blkno);
1382 		bstrategy(bp);
1383 		++readwait;
1384 	}
1385 
1386 	breada(vp, rablkno, rabsize, cnt, cred);
1387 
1388 	if (readwait) {
1389 		rv = bufwait(bp);
1390 	}
1391 	return (rv);
1392 }
1393 
1394 /*
1395  * Write, release buffer on completion.  (Done by iodone
1396  * if async).  Do not bother writing anything if the buffer
1397  * is invalid.
1398  *
1399  * Note that we set B_CACHE here, indicating that buffer is
1400  * fully valid and thus cacheable.  This is true even of NFS
1401  * now so we set it generally.  This could be set either here
1402  * or in biodone() since the I/O is synchronous.  We put it
1403  * here.
1404  */
1405 int
1406 bufwrite(struct buf *bp)
1407 {
1408 	int oldflags;
1409 	struct vnode *vp;
1410 	long space;
1411 	int vp_md;
1412 
1413 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1414 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
1415 		bp->b_flags |= B_INVAL | B_RELBUF;
1416 		bp->b_flags &= ~B_CACHE;
1417 		brelse(bp);
1418 		return (ENXIO);
1419 	}
1420 	if (bp->b_flags & B_INVAL) {
1421 		brelse(bp);
1422 		return (0);
1423 	}
1424 
1425 	if (bp->b_flags & B_BARRIER)
1426 		barrierwrites++;
1427 
1428 	oldflags = bp->b_flags;
1429 
1430 	BUF_ASSERT_HELD(bp);
1431 
1432 	if (bp->b_pin_count > 0)
1433 		bunpin_wait(bp);
1434 
1435 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1436 	    ("FFS background buffer should not get here %p", bp));
1437 
1438 	vp = bp->b_vp;
1439 	if (vp)
1440 		vp_md = vp->v_vflag & VV_MD;
1441 	else
1442 		vp_md = 0;
1443 
1444 	/*
1445 	 * Mark the buffer clean.  Increment the bufobj write count
1446 	 * before bundirty() call, to prevent other thread from seeing
1447 	 * empty dirty list and zero counter for writes in progress,
1448 	 * falsely indicating that the bufobj is clean.
1449 	 */
1450 	bufobj_wref(bp->b_bufobj);
1451 	bundirty(bp);
1452 
1453 	bp->b_flags &= ~B_DONE;
1454 	bp->b_ioflags &= ~BIO_ERROR;
1455 	bp->b_flags |= B_CACHE;
1456 	bp->b_iocmd = BIO_WRITE;
1457 
1458 	vfs_busy_pages(bp, 1);
1459 
1460 	/*
1461 	 * Normal bwrites pipeline writes
1462 	 */
1463 	bp->b_runningbufspace = bp->b_bufsize;
1464 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1465 
1466 	if (!TD_IS_IDLETHREAD(curthread))
1467 		curthread->td_ru.ru_oublock++;
1468 	if (oldflags & B_ASYNC)
1469 		BUF_KERNPROC(bp);
1470 	bp->b_iooffset = dbtob(bp->b_blkno);
1471 	bstrategy(bp);
1472 
1473 	if ((oldflags & B_ASYNC) == 0) {
1474 		int rtval = bufwait(bp);
1475 		brelse(bp);
1476 		return (rtval);
1477 	} else if (space > hirunningspace) {
1478 		/*
1479 		 * don't allow the async write to saturate the I/O
1480 		 * system.  We will not deadlock here because
1481 		 * we are blocking waiting for I/O that is already in-progress
1482 		 * to complete. We do not block here if it is the update
1483 		 * or syncer daemon trying to clean up as that can lead
1484 		 * to deadlock.
1485 		 */
1486 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1487 			waitrunningbufspace();
1488 	}
1489 
1490 	return (0);
1491 }
1492 
1493 void
1494 bufbdflush(struct bufobj *bo, struct buf *bp)
1495 {
1496 	struct buf *nbp;
1497 
1498 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1499 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1500 		altbufferflushes++;
1501 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1502 		BO_LOCK(bo);
1503 		/*
1504 		 * Try to find a buffer to flush.
1505 		 */
1506 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1507 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1508 			    BUF_LOCK(nbp,
1509 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1510 				continue;
1511 			if (bp == nbp)
1512 				panic("bdwrite: found ourselves");
1513 			BO_UNLOCK(bo);
1514 			/* Don't countdeps with the bo lock held. */
1515 			if (buf_countdeps(nbp, 0)) {
1516 				BO_LOCK(bo);
1517 				BUF_UNLOCK(nbp);
1518 				continue;
1519 			}
1520 			if (nbp->b_flags & B_CLUSTEROK) {
1521 				vfs_bio_awrite(nbp);
1522 			} else {
1523 				bremfree(nbp);
1524 				bawrite(nbp);
1525 			}
1526 			dirtybufferflushes++;
1527 			break;
1528 		}
1529 		if (nbp == NULL)
1530 			BO_UNLOCK(bo);
1531 	}
1532 }
1533 
1534 /*
1535  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1536  * anything if the buffer is marked invalid.
1537  *
1538  * Note that since the buffer must be completely valid, we can safely
1539  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1540  * biodone() in order to prevent getblk from writing the buffer
1541  * out synchronously.
1542  */
1543 void
1544 bdwrite(struct buf *bp)
1545 {
1546 	struct thread *td = curthread;
1547 	struct vnode *vp;
1548 	struct bufobj *bo;
1549 
1550 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1551 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1552 	KASSERT((bp->b_flags & B_BARRIER) == 0,
1553 	    ("Barrier request in delayed write %p", bp));
1554 	BUF_ASSERT_HELD(bp);
1555 
1556 	if (bp->b_flags & B_INVAL) {
1557 		brelse(bp);
1558 		return;
1559 	}
1560 
1561 	/*
1562 	 * If we have too many dirty buffers, don't create any more.
1563 	 * If we are wildly over our limit, then force a complete
1564 	 * cleanup. Otherwise, just keep the situation from getting
1565 	 * out of control. Note that we have to avoid a recursive
1566 	 * disaster and not try to clean up after our own cleanup!
1567 	 */
1568 	vp = bp->b_vp;
1569 	bo = bp->b_bufobj;
1570 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1571 		td->td_pflags |= TDP_INBDFLUSH;
1572 		BO_BDFLUSH(bo, bp);
1573 		td->td_pflags &= ~TDP_INBDFLUSH;
1574 	} else
1575 		recursiveflushes++;
1576 
1577 	bdirty(bp);
1578 	/*
1579 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1580 	 * true even of NFS now.
1581 	 */
1582 	bp->b_flags |= B_CACHE;
1583 
1584 	/*
1585 	 * This bmap keeps the system from needing to do the bmap later,
1586 	 * perhaps when the system is attempting to do a sync.  Since it
1587 	 * is likely that the indirect block -- or whatever other datastructure
1588 	 * that the filesystem needs is still in memory now, it is a good
1589 	 * thing to do this.  Note also, that if the pageout daemon is
1590 	 * requesting a sync -- there might not be enough memory to do
1591 	 * the bmap then...  So, this is important to do.
1592 	 */
1593 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1594 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1595 	}
1596 
1597 	/*
1598 	 * Set the *dirty* buffer range based upon the VM system dirty
1599 	 * pages.
1600 	 *
1601 	 * Mark the buffer pages as clean.  We need to do this here to
1602 	 * satisfy the vnode_pager and the pageout daemon, so that it
1603 	 * thinks that the pages have been "cleaned".  Note that since
1604 	 * the pages are in a delayed write buffer -- the VFS layer
1605 	 * "will" see that the pages get written out on the next sync,
1606 	 * or perhaps the cluster will be completed.
1607 	 */
1608 	vfs_clean_pages_dirty_buf(bp);
1609 	bqrelse(bp);
1610 
1611 	/*
1612 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1613 	 * due to the softdep code.
1614 	 */
1615 }
1616 
1617 /*
1618  *	bdirty:
1619  *
1620  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1621  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1622  *	itself to properly update it in the dirty/clean lists.  We mark it
1623  *	B_DONE to ensure that any asynchronization of the buffer properly
1624  *	clears B_DONE ( else a panic will occur later ).
1625  *
1626  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1627  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1628  *	should only be called if the buffer is known-good.
1629  *
1630  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1631  *	count.
1632  *
1633  *	The buffer must be on QUEUE_NONE.
1634  */
1635 void
1636 bdirty(struct buf *bp)
1637 {
1638 
1639 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1640 	    bp, bp->b_vp, bp->b_flags);
1641 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1642 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1643 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1644 	BUF_ASSERT_HELD(bp);
1645 	bp->b_flags &= ~(B_RELBUF);
1646 	bp->b_iocmd = BIO_WRITE;
1647 
1648 	if ((bp->b_flags & B_DELWRI) == 0) {
1649 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1650 		reassignbuf(bp);
1651 		bdirtyadd();
1652 	}
1653 }
1654 
1655 /*
1656  *	bundirty:
1657  *
1658  *	Clear B_DELWRI for buffer.
1659  *
1660  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1661  *	count.
1662  *
1663  *	The buffer must be on QUEUE_NONE.
1664  */
1665 
1666 void
1667 bundirty(struct buf *bp)
1668 {
1669 
1670 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1671 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1672 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1673 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1674 	BUF_ASSERT_HELD(bp);
1675 
1676 	if (bp->b_flags & B_DELWRI) {
1677 		bp->b_flags &= ~B_DELWRI;
1678 		reassignbuf(bp);
1679 		bdirtysub();
1680 	}
1681 	/*
1682 	 * Since it is now being written, we can clear its deferred write flag.
1683 	 */
1684 	bp->b_flags &= ~B_DEFERRED;
1685 }
1686 
1687 /*
1688  *	bawrite:
1689  *
1690  *	Asynchronous write.  Start output on a buffer, but do not wait for
1691  *	it to complete.  The buffer is released when the output completes.
1692  *
1693  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1694  *	B_INVAL buffers.  Not us.
1695  */
1696 void
1697 bawrite(struct buf *bp)
1698 {
1699 
1700 	bp->b_flags |= B_ASYNC;
1701 	(void) bwrite(bp);
1702 }
1703 
1704 /*
1705  *	babarrierwrite:
1706  *
1707  *	Asynchronous barrier write.  Start output on a buffer, but do not
1708  *	wait for it to complete.  Place a write barrier after this write so
1709  *	that this buffer and all buffers written before it are committed to
1710  *	the disk before any buffers written after this write are committed
1711  *	to the disk.  The buffer is released when the output completes.
1712  */
1713 void
1714 babarrierwrite(struct buf *bp)
1715 {
1716 
1717 	bp->b_flags |= B_ASYNC | B_BARRIER;
1718 	(void) bwrite(bp);
1719 }
1720 
1721 /*
1722  *	bbarrierwrite:
1723  *
1724  *	Synchronous barrier write.  Start output on a buffer and wait for
1725  *	it to complete.  Place a write barrier after this write so that
1726  *	this buffer and all buffers written before it are committed to
1727  *	the disk before any buffers written after this write are committed
1728  *	to the disk.  The buffer is released when the output completes.
1729  */
1730 int
1731 bbarrierwrite(struct buf *bp)
1732 {
1733 
1734 	bp->b_flags |= B_BARRIER;
1735 	return (bwrite(bp));
1736 }
1737 
1738 /*
1739  *	bwillwrite:
1740  *
1741  *	Called prior to the locking of any vnodes when we are expecting to
1742  *	write.  We do not want to starve the buffer cache with too many
1743  *	dirty buffers so we block here.  By blocking prior to the locking
1744  *	of any vnodes we attempt to avoid the situation where a locked vnode
1745  *	prevents the various system daemons from flushing related buffers.
1746  */
1747 void
1748 bwillwrite(void)
1749 {
1750 
1751 	if (numdirtybuffers >= hidirtybuffers) {
1752 		mtx_lock(&bdirtylock);
1753 		while (numdirtybuffers >= hidirtybuffers) {
1754 			bdirtywait = 1;
1755 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1756 			    "flswai", 0);
1757 		}
1758 		mtx_unlock(&bdirtylock);
1759 	}
1760 }
1761 
1762 /*
1763  * Return true if we have too many dirty buffers.
1764  */
1765 int
1766 buf_dirty_count_severe(void)
1767 {
1768 
1769 	return(numdirtybuffers >= hidirtybuffers);
1770 }
1771 
1772 /*
1773  *	brelse:
1774  *
1775  *	Release a busy buffer and, if requested, free its resources.  The
1776  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1777  *	to be accessed later as a cache entity or reused for other purposes.
1778  */
1779 void
1780 brelse(struct buf *bp)
1781 {
1782 	int qindex;
1783 
1784 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1785 	    bp, bp->b_vp, bp->b_flags);
1786 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1787 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1788 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
1789 	    ("brelse: non-VMIO buffer marked NOREUSE"));
1790 
1791 	if (BUF_LOCKRECURSED(bp)) {
1792 		/*
1793 		 * Do not process, in particular, do not handle the
1794 		 * B_INVAL/B_RELBUF and do not release to free list.
1795 		 */
1796 		BUF_UNLOCK(bp);
1797 		return;
1798 	}
1799 
1800 	if (bp->b_flags & B_MANAGED) {
1801 		bqrelse(bp);
1802 		return;
1803 	}
1804 
1805 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
1806 		BO_LOCK(bp->b_bufobj);
1807 		bp->b_vflags &= ~BV_BKGRDERR;
1808 		BO_UNLOCK(bp->b_bufobj);
1809 		bdirty(bp);
1810 	}
1811 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1812 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1813 		/*
1814 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1815 		 * pages from being scrapped.  If the error is anything
1816 		 * other than an I/O error (EIO), assume that retrying
1817 		 * is futile.
1818 		 */
1819 		bp->b_ioflags &= ~BIO_ERROR;
1820 		bdirty(bp);
1821 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1822 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1823 		/*
1824 		 * Either a failed I/O or we were asked to free or not
1825 		 * cache the buffer.
1826 		 */
1827 		bp->b_flags |= B_INVAL;
1828 		if (!LIST_EMPTY(&bp->b_dep))
1829 			buf_deallocate(bp);
1830 		if (bp->b_flags & B_DELWRI)
1831 			bdirtysub();
1832 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1833 		if ((bp->b_flags & B_VMIO) == 0) {
1834 			allocbuf(bp, 0);
1835 			if (bp->b_vp)
1836 				brelvp(bp);
1837 		}
1838 	}
1839 
1840 	/*
1841 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
1842 	 * is called with B_DELWRI set, the underlying pages may wind up
1843 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1844 	 * because pages associated with a B_DELWRI bp are marked clean.
1845 	 *
1846 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
1847 	 * if B_DELWRI is set.
1848 	 */
1849 	if (bp->b_flags & B_DELWRI)
1850 		bp->b_flags &= ~B_RELBUF;
1851 
1852 	/*
1853 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1854 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1855 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1856 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1857 	 *
1858 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1859 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1860 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1861 	 *
1862 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1863 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1864 	 * the commit state and we cannot afford to lose the buffer. If the
1865 	 * buffer has a background write in progress, we need to keep it
1866 	 * around to prevent it from being reconstituted and starting a second
1867 	 * background write.
1868 	 */
1869 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
1870 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
1871 	    !(bp->b_vp->v_mount != NULL &&
1872 	    (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1873 	    !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI))) {
1874 		vfs_vmio_invalidate(bp);
1875 		allocbuf(bp, 0);
1876 	}
1877 
1878 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
1879 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
1880 		allocbuf(bp, 0);
1881 		bp->b_flags &= ~B_NOREUSE;
1882 		if (bp->b_vp != NULL)
1883 			brelvp(bp);
1884 	}
1885 
1886 	/*
1887 	 * If the buffer has junk contents signal it and eventually
1888 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1889 	 * doesn't find it.
1890 	 */
1891 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1892 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1893 		bp->b_flags |= B_INVAL;
1894 	if (bp->b_flags & B_INVAL) {
1895 		if (bp->b_flags & B_DELWRI)
1896 			bundirty(bp);
1897 		if (bp->b_vp)
1898 			brelvp(bp);
1899 	}
1900 
1901 	/* buffers with no memory */
1902 	if (bp->b_bufsize == 0) {
1903 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1904 		if (bp->b_vflags & BV_BKGRDINPROG)
1905 			panic("losing buffer 1");
1906 		bufkvafree(bp);
1907 		qindex = QUEUE_EMPTY;
1908 		bp->b_flags |= B_AGE;
1909 	/* buffers with junk contents */
1910 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1911 	    (bp->b_ioflags & BIO_ERROR)) {
1912 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1913 		if (bp->b_vflags & BV_BKGRDINPROG)
1914 			panic("losing buffer 2");
1915 		qindex = QUEUE_CLEAN;
1916 		bp->b_flags |= B_AGE;
1917 	/* remaining buffers */
1918 	} else if (bp->b_flags & B_DELWRI)
1919 		qindex = QUEUE_DIRTY;
1920 	else
1921 		qindex = QUEUE_CLEAN;
1922 
1923 	binsfree(bp, qindex);
1924 
1925 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1926 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1927 		panic("brelse: not dirty");
1928 	/* unlock */
1929 	BUF_UNLOCK(bp);
1930 }
1931 
1932 /*
1933  * Release a buffer back to the appropriate queue but do not try to free
1934  * it.  The buffer is expected to be used again soon.
1935  *
1936  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1937  * biodone() to requeue an async I/O on completion.  It is also used when
1938  * known good buffers need to be requeued but we think we may need the data
1939  * again soon.
1940  *
1941  * XXX we should be able to leave the B_RELBUF hint set on completion.
1942  */
1943 void
1944 bqrelse(struct buf *bp)
1945 {
1946 	int qindex;
1947 
1948 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1949 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1950 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1951 
1952 	if (BUF_LOCKRECURSED(bp)) {
1953 		/* do not release to free list */
1954 		BUF_UNLOCK(bp);
1955 		return;
1956 	}
1957 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1958 
1959 	if (bp->b_flags & B_MANAGED) {
1960 		if (bp->b_flags & B_REMFREE)
1961 			bremfreef(bp);
1962 		goto out;
1963 	}
1964 
1965 	/* buffers with stale but valid contents */
1966 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
1967 	    BV_BKGRDERR)) == BV_BKGRDERR) {
1968 		BO_LOCK(bp->b_bufobj);
1969 		bp->b_vflags &= ~BV_BKGRDERR;
1970 		BO_UNLOCK(bp->b_bufobj);
1971 		qindex = QUEUE_DIRTY;
1972 	} else {
1973 		if ((bp->b_flags & B_DELWRI) == 0 &&
1974 		    (bp->b_xflags & BX_VNDIRTY))
1975 			panic("bqrelse: not dirty");
1976 		if ((bp->b_flags & B_NOREUSE) != 0) {
1977 			brelse(bp);
1978 			return;
1979 		}
1980 		qindex = QUEUE_CLEAN;
1981 	}
1982 	binsfree(bp, qindex);
1983 
1984 out:
1985 	/* unlock */
1986 	BUF_UNLOCK(bp);
1987 }
1988 
1989 /*
1990  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
1991  * restore bogus pages.
1992  */
1993 static void
1994 vfs_vmio_iodone(struct buf *bp)
1995 {
1996 	vm_ooffset_t foff;
1997 	vm_page_t m;
1998 	vm_object_t obj;
1999 	struct vnode *vp;
2000 	int bogus, i, iosize;
2001 
2002 	obj = bp->b_bufobj->bo_object;
2003 	KASSERT(obj->paging_in_progress >= bp->b_npages,
2004 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2005 	    obj->paging_in_progress, bp->b_npages));
2006 
2007 	vp = bp->b_vp;
2008 	KASSERT(vp->v_holdcnt > 0,
2009 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2010 	KASSERT(vp->v_object != NULL,
2011 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2012 
2013 	foff = bp->b_offset;
2014 	KASSERT(bp->b_offset != NOOFFSET,
2015 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2016 
2017 	bogus = 0;
2018 	iosize = bp->b_bcount - bp->b_resid;
2019 	VM_OBJECT_WLOCK(obj);
2020 	for (i = 0; i < bp->b_npages; i++) {
2021 		int resid;
2022 
2023 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2024 		if (resid > iosize)
2025 			resid = iosize;
2026 
2027 		/*
2028 		 * cleanup bogus pages, restoring the originals
2029 		 */
2030 		m = bp->b_pages[i];
2031 		if (m == bogus_page) {
2032 			bogus = 1;
2033 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2034 			if (m == NULL)
2035 				panic("biodone: page disappeared!");
2036 			bp->b_pages[i] = m;
2037 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2038 			/*
2039 			 * In the write case, the valid and clean bits are
2040 			 * already changed correctly ( see bdwrite() ), so we
2041 			 * only need to do this here in the read case.
2042 			 */
2043 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2044 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2045 			    "has unexpected dirty bits", m));
2046 			vfs_page_set_valid(bp, foff, m);
2047 		}
2048 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2049 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2050 		    (intmax_t)foff, (uintmax_t)m->pindex));
2051 
2052 		vm_page_sunbusy(m);
2053 		vm_object_pip_subtract(obj, 1);
2054 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2055 		iosize -= resid;
2056 	}
2057 	vm_object_pip_wakeupn(obj, 0);
2058 	VM_OBJECT_WUNLOCK(obj);
2059 	if (bogus && buf_mapped(bp)) {
2060 		BUF_CHECK_MAPPED(bp);
2061 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2062 		    bp->b_pages, bp->b_npages);
2063 	}
2064 }
2065 
2066 /*
2067  * Unwire a page held by a buf and place it on the appropriate vm queue.
2068  */
2069 static void
2070 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2071 {
2072 	bool freed;
2073 
2074 	vm_page_lock(m);
2075 	if (vm_page_unwire(m, PQ_NONE)) {
2076 		/*
2077 		 * Determine if the page should be freed before adding
2078 		 * it to the inactive queue.
2079 		 */
2080 		if (m->valid == 0) {
2081 			freed = !vm_page_busied(m);
2082 			if (freed)
2083 				vm_page_free(m);
2084 		} else if ((bp->b_flags & B_DIRECT) != 0)
2085 			freed = vm_page_try_to_free(m);
2086 		else
2087 			freed = false;
2088 		if (!freed) {
2089 			/*
2090 			 * If the page is unlikely to be reused, let the
2091 			 * VM know.  Otherwise, maintain LRU page
2092 			 * ordering and put the page at the tail of the
2093 			 * inactive queue.
2094 			 */
2095 			if ((bp->b_flags & B_NOREUSE) != 0)
2096 				vm_page_deactivate_noreuse(m);
2097 			else
2098 				vm_page_deactivate(m);
2099 		}
2100 	}
2101 	vm_page_unlock(m);
2102 }
2103 
2104 /*
2105  * Perform page invalidation when a buffer is released.  The fully invalid
2106  * pages will be reclaimed later in vfs_vmio_truncate().
2107  */
2108 static void
2109 vfs_vmio_invalidate(struct buf *bp)
2110 {
2111 	vm_object_t obj;
2112 	vm_page_t m;
2113 	int i, resid, poffset, presid;
2114 
2115 	if (buf_mapped(bp)) {
2116 		BUF_CHECK_MAPPED(bp);
2117 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2118 	} else
2119 		BUF_CHECK_UNMAPPED(bp);
2120 	/*
2121 	 * Get the base offset and length of the buffer.  Note that
2122 	 * in the VMIO case if the buffer block size is not
2123 	 * page-aligned then b_data pointer may not be page-aligned.
2124 	 * But our b_pages[] array *IS* page aligned.
2125 	 *
2126 	 * block sizes less then DEV_BSIZE (usually 512) are not
2127 	 * supported due to the page granularity bits (m->valid,
2128 	 * m->dirty, etc...).
2129 	 *
2130 	 * See man buf(9) for more information
2131 	 */
2132 	obj = bp->b_bufobj->bo_object;
2133 	resid = bp->b_bufsize;
2134 	poffset = bp->b_offset & PAGE_MASK;
2135 	VM_OBJECT_WLOCK(obj);
2136 	for (i = 0; i < bp->b_npages; i++) {
2137 		m = bp->b_pages[i];
2138 		if (m == bogus_page)
2139 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2140 		bp->b_pages[i] = NULL;
2141 
2142 		presid = resid > (PAGE_SIZE - poffset) ?
2143 		    (PAGE_SIZE - poffset) : resid;
2144 		KASSERT(presid >= 0, ("brelse: extra page"));
2145 		while (vm_page_xbusied(m)) {
2146 			vm_page_lock(m);
2147 			VM_OBJECT_WUNLOCK(obj);
2148 			vm_page_busy_sleep(m, "mbncsh");
2149 			VM_OBJECT_WLOCK(obj);
2150 		}
2151 		if (pmap_page_wired_mappings(m) == 0)
2152 			vm_page_set_invalid(m, poffset, presid);
2153 		vfs_vmio_unwire(bp, m);
2154 		resid -= presid;
2155 		poffset = 0;
2156 	}
2157 	VM_OBJECT_WUNLOCK(obj);
2158 	bp->b_npages = 0;
2159 }
2160 
2161 /*
2162  * Page-granular truncation of an existing VMIO buffer.
2163  */
2164 static void
2165 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2166 {
2167 	vm_object_t obj;
2168 	vm_page_t m;
2169 	int i;
2170 
2171 	if (bp->b_npages == desiredpages)
2172 		return;
2173 
2174 	if (buf_mapped(bp)) {
2175 		BUF_CHECK_MAPPED(bp);
2176 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2177 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2178 	} else
2179 		BUF_CHECK_UNMAPPED(bp);
2180 	obj = bp->b_bufobj->bo_object;
2181 	if (obj != NULL)
2182 		VM_OBJECT_WLOCK(obj);
2183 	for (i = desiredpages; i < bp->b_npages; i++) {
2184 		m = bp->b_pages[i];
2185 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
2186 		bp->b_pages[i] = NULL;
2187 		vfs_vmio_unwire(bp, m);
2188 	}
2189 	if (obj != NULL)
2190 		VM_OBJECT_WUNLOCK(obj);
2191 	bp->b_npages = desiredpages;
2192 }
2193 
2194 /*
2195  * Byte granular extension of VMIO buffers.
2196  */
2197 static void
2198 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
2199 {
2200 	/*
2201 	 * We are growing the buffer, possibly in a
2202 	 * byte-granular fashion.
2203 	 */
2204 	vm_object_t obj;
2205 	vm_offset_t toff;
2206 	vm_offset_t tinc;
2207 	vm_page_t m;
2208 
2209 	/*
2210 	 * Step 1, bring in the VM pages from the object, allocating
2211 	 * them if necessary.  We must clear B_CACHE if these pages
2212 	 * are not valid for the range covered by the buffer.
2213 	 */
2214 	obj = bp->b_bufobj->bo_object;
2215 	VM_OBJECT_WLOCK(obj);
2216 	while (bp->b_npages < desiredpages) {
2217 		/*
2218 		 * We must allocate system pages since blocking
2219 		 * here could interfere with paging I/O, no
2220 		 * matter which process we are.
2221 		 *
2222 		 * Only exclusive busy can be tested here.
2223 		 * Blocking on shared busy might lead to
2224 		 * deadlocks once allocbuf() is called after
2225 		 * pages are vfs_busy_pages().
2226 		 */
2227 		m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages,
2228 		    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2229 		    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY |
2230 		    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
2231 		if (m->valid == 0)
2232 			bp->b_flags &= ~B_CACHE;
2233 		bp->b_pages[bp->b_npages] = m;
2234 		++bp->b_npages;
2235 	}
2236 
2237 	/*
2238 	 * Step 2.  We've loaded the pages into the buffer,
2239 	 * we have to figure out if we can still have B_CACHE
2240 	 * set.  Note that B_CACHE is set according to the
2241 	 * byte-granular range ( bcount and size ), not the
2242 	 * aligned range ( newbsize ).
2243 	 *
2244 	 * The VM test is against m->valid, which is DEV_BSIZE
2245 	 * aligned.  Needless to say, the validity of the data
2246 	 * needs to also be DEV_BSIZE aligned.  Note that this
2247 	 * fails with NFS if the server or some other client
2248 	 * extends the file's EOF.  If our buffer is resized,
2249 	 * B_CACHE may remain set! XXX
2250 	 */
2251 	toff = bp->b_bcount;
2252 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2253 	while ((bp->b_flags & B_CACHE) && toff < size) {
2254 		vm_pindex_t pi;
2255 
2256 		if (tinc > (size - toff))
2257 			tinc = size - toff;
2258 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
2259 		m = bp->b_pages[pi];
2260 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
2261 		toff += tinc;
2262 		tinc = PAGE_SIZE;
2263 	}
2264 	VM_OBJECT_WUNLOCK(obj);
2265 
2266 	/*
2267 	 * Step 3, fixup the KVA pmap.
2268 	 */
2269 	if (buf_mapped(bp))
2270 		bpmap_qenter(bp);
2271 	else
2272 		BUF_CHECK_UNMAPPED(bp);
2273 }
2274 
2275 /*
2276  * Check to see if a block at a particular lbn is available for a clustered
2277  * write.
2278  */
2279 static int
2280 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
2281 {
2282 	struct buf *bpa;
2283 	int match;
2284 
2285 	match = 0;
2286 
2287 	/* If the buf isn't in core skip it */
2288 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
2289 		return (0);
2290 
2291 	/* If the buf is busy we don't want to wait for it */
2292 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2293 		return (0);
2294 
2295 	/* Only cluster with valid clusterable delayed write buffers */
2296 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
2297 	    (B_DELWRI | B_CLUSTEROK))
2298 		goto done;
2299 
2300 	if (bpa->b_bufsize != size)
2301 		goto done;
2302 
2303 	/*
2304 	 * Check to see if it is in the expected place on disk and that the
2305 	 * block has been mapped.
2306 	 */
2307 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
2308 		match = 1;
2309 done:
2310 	BUF_UNLOCK(bpa);
2311 	return (match);
2312 }
2313 
2314 /*
2315  *	vfs_bio_awrite:
2316  *
2317  *	Implement clustered async writes for clearing out B_DELWRI buffers.
2318  *	This is much better then the old way of writing only one buffer at
2319  *	a time.  Note that we may not be presented with the buffers in the
2320  *	correct order, so we search for the cluster in both directions.
2321  */
2322 int
2323 vfs_bio_awrite(struct buf *bp)
2324 {
2325 	struct bufobj *bo;
2326 	int i;
2327 	int j;
2328 	daddr_t lblkno = bp->b_lblkno;
2329 	struct vnode *vp = bp->b_vp;
2330 	int ncl;
2331 	int nwritten;
2332 	int size;
2333 	int maxcl;
2334 	int gbflags;
2335 
2336 	bo = &vp->v_bufobj;
2337 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
2338 	/*
2339 	 * right now we support clustered writing only to regular files.  If
2340 	 * we find a clusterable block we could be in the middle of a cluster
2341 	 * rather then at the beginning.
2342 	 */
2343 	if ((vp->v_type == VREG) &&
2344 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
2345 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
2346 
2347 		size = vp->v_mount->mnt_stat.f_iosize;
2348 		maxcl = MAXPHYS / size;
2349 
2350 		BO_RLOCK(bo);
2351 		for (i = 1; i < maxcl; i++)
2352 			if (vfs_bio_clcheck(vp, size, lblkno + i,
2353 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2354 				break;
2355 
2356 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2357 			if (vfs_bio_clcheck(vp, size, lblkno - j,
2358 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2359 				break;
2360 		BO_RUNLOCK(bo);
2361 		--j;
2362 		ncl = i + j;
2363 		/*
2364 		 * this is a possible cluster write
2365 		 */
2366 		if (ncl != 1) {
2367 			BUF_UNLOCK(bp);
2368 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2369 			    gbflags);
2370 			return (nwritten);
2371 		}
2372 	}
2373 	bremfree(bp);
2374 	bp->b_flags |= B_ASYNC;
2375 	/*
2376 	 * default (old) behavior, writing out only one block
2377 	 *
2378 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
2379 	 */
2380 	nwritten = bp->b_bufsize;
2381 	(void) bwrite(bp);
2382 
2383 	return (nwritten);
2384 }
2385 
2386 /*
2387  * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2388  * locked vnode is supplied.
2389  */
2390 static void
2391 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2392     int defrag)
2393 {
2394 	struct thread *td;
2395 	char *waitmsg;
2396 	int error, fl, flags, norunbuf;
2397 
2398 	mtx_assert(&bqclean, MA_OWNED);
2399 
2400 	if (defrag) {
2401 		flags = VFS_BIO_NEED_BUFSPACE;
2402 		waitmsg = "nbufkv";
2403 	} else if (bufspace >= hibufspace) {
2404 		waitmsg = "nbufbs";
2405 		flags = VFS_BIO_NEED_BUFSPACE;
2406 	} else {
2407 		waitmsg = "newbuf";
2408 		flags = VFS_BIO_NEED_ANY;
2409 	}
2410 	atomic_set_int(&needsbuffer, flags);
2411 	mtx_unlock(&bqclean);
2412 
2413 	bd_speedup();	/* heeeelp */
2414 	if ((gbflags & GB_NOWAIT_BD) != 0)
2415 		return;
2416 
2417 	td = curthread;
2418 	rw_wlock(&nblock);
2419 	while ((needsbuffer & flags) != 0) {
2420 		if (vp != NULL && vp->v_type != VCHR &&
2421 		    (td->td_pflags & TDP_BUFNEED) == 0) {
2422 			rw_wunlock(&nblock);
2423 			/*
2424 			 * getblk() is called with a vnode locked, and
2425 			 * some majority of the dirty buffers may as
2426 			 * well belong to the vnode.  Flushing the
2427 			 * buffers there would make a progress that
2428 			 * cannot be achieved by the buf_daemon, that
2429 			 * cannot lock the vnode.
2430 			 */
2431 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2432 			    (td->td_pflags & TDP_NORUNNINGBUF);
2433 
2434 			/*
2435 			 * Play bufdaemon.  The getnewbuf() function
2436 			 * may be called while the thread owns lock
2437 			 * for another dirty buffer for the same
2438 			 * vnode, which makes it impossible to use
2439 			 * VOP_FSYNC() there, due to the buffer lock
2440 			 * recursion.
2441 			 */
2442 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2443 			fl = buf_flush(vp, flushbufqtarget);
2444 			td->td_pflags &= norunbuf;
2445 			rw_wlock(&nblock);
2446 			if (fl != 0)
2447 				continue;
2448 			if ((needsbuffer & flags) == 0)
2449 				break;
2450 		}
2451 		error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
2452 		    (PRIBIO + 4) | slpflag, waitmsg, slptimeo);
2453 		if (error != 0)
2454 			break;
2455 	}
2456 	rw_wunlock(&nblock);
2457 }
2458 
2459 static void
2460 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2461 {
2462 
2463 	CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2464 	    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2465 	     bp->b_kvasize, bp->b_bufsize, qindex);
2466 	mtx_assert(&bqclean, MA_NOTOWNED);
2467 
2468 	/*
2469 	 * Note: we no longer distinguish between VMIO and non-VMIO
2470 	 * buffers.
2471 	 */
2472 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
2473 	    ("invalid buffer %p flags %#x found in queue %d", bp, bp->b_flags,
2474 	    qindex));
2475 
2476 	/*
2477 	 * When recycling a clean buffer we have to truncate it and
2478 	 * release the vnode.
2479 	 */
2480 	if (qindex == QUEUE_CLEAN) {
2481 		allocbuf(bp, 0);
2482 		if (bp->b_vp != NULL)
2483 			brelvp(bp);
2484 	}
2485 
2486 	/*
2487 	 * Get the rest of the buffer freed up.  b_kva* is still valid
2488 	 * after this operation.
2489 	 */
2490 	if (bp->b_rcred != NOCRED) {
2491 		crfree(bp->b_rcred);
2492 		bp->b_rcred = NOCRED;
2493 	}
2494 	if (bp->b_wcred != NOCRED) {
2495 		crfree(bp->b_wcred);
2496 		bp->b_wcred = NOCRED;
2497 	}
2498 	if (!LIST_EMPTY(&bp->b_dep))
2499 		buf_deallocate(bp);
2500 	if (bp->b_vflags & BV_BKGRDINPROG)
2501 		panic("losing buffer 3");
2502 	KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.  qindex: %d",
2503 	    bp, bp->b_vp, qindex));
2504 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2505 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2506 	KASSERT(bp->b_npages == 0,
2507 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
2508 
2509 	bp->b_flags = 0;
2510 	bp->b_ioflags = 0;
2511 	bp->b_xflags = 0;
2512 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
2513 	    ("buf %p still counted as free?", bp));
2514 	bp->b_vflags = 0;
2515 	bp->b_vp = NULL;
2516 	bp->b_blkno = bp->b_lblkno = 0;
2517 	bp->b_offset = NOOFFSET;
2518 	bp->b_iodone = 0;
2519 	bp->b_error = 0;
2520 	bp->b_resid = 0;
2521 	bp->b_bcount = 0;
2522 	bp->b_npages = 0;
2523 	bp->b_dirtyoff = bp->b_dirtyend = 0;
2524 	bp->b_bufobj = NULL;
2525 	bp->b_pin_count = 0;
2526 	bp->b_data = bp->b_kvabase;
2527 	bp->b_fsprivate1 = NULL;
2528 	bp->b_fsprivate2 = NULL;
2529 	bp->b_fsprivate3 = NULL;
2530 
2531 	LIST_INIT(&bp->b_dep);
2532 }
2533 
2534 static struct buf *
2535 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2536 {
2537 	struct buf *bp, *nbp;
2538 	int nqindex, qindex, pass;
2539 
2540 	KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2541 
2542 	pass = 0;
2543 restart:
2544 	if (pass != 0)
2545 		atomic_add_int(&getnewbufrestarts, 1);
2546 
2547 	nbp = NULL;
2548 	mtx_lock(&bqclean);
2549 	/*
2550 	 * If we're not defragging or low on bufspace attempt to make a new
2551 	 * buf from a header.
2552 	 */
2553 	if (defrag == 0 && bufspace + maxsize < hibufspace) {
2554 		nqindex = QUEUE_EMPTY;
2555 		nbp = TAILQ_FIRST(&bufqueues[nqindex]);
2556 	}
2557 	/*
2558 	 * All available buffers might be clean or we need to start recycling.
2559 	 */
2560 	if (nbp == NULL) {
2561 		nqindex = QUEUE_CLEAN;
2562 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2563 	}
2564 
2565 	/*
2566 	 * Run scan, possibly freeing data and/or kva mappings on the fly
2567 	 * depending.
2568 	 */
2569 	while ((bp = nbp) != NULL) {
2570 		qindex = nqindex;
2571 
2572 		/*
2573 		 * Calculate next bp (we can only use it if we do not
2574 		 * release the bqlock)
2575 		 */
2576 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2577 			switch (qindex) {
2578 			case QUEUE_EMPTY:
2579 				nqindex = QUEUE_CLEAN;
2580 				nbp = TAILQ_FIRST(&bufqueues[nqindex]);
2581 				if (nbp != NULL)
2582 					break;
2583 				/* FALLTHROUGH */
2584 			case QUEUE_CLEAN:
2585 				if (metadata && pass == 0) {
2586 					pass = 1;
2587 					nqindex = QUEUE_EMPTY;
2588 					nbp = TAILQ_FIRST(&bufqueues[nqindex]);
2589 				}
2590 				/*
2591 				 * nbp is NULL.
2592 				 */
2593 				break;
2594 			}
2595 		}
2596 		/*
2597 		 * If we are defragging then we need a buffer with
2598 		 * b_kvasize != 0.  This situation occurs when we
2599 		 * have many unmapped bufs.
2600 		 */
2601 		if (defrag && bp->b_kvasize == 0)
2602 			continue;
2603 
2604 		/*
2605 		 * Start freeing the bp.  This is somewhat involved.  nbp
2606 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2607 		 */
2608 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2609 			continue;
2610 		/*
2611 		 * BKGRDINPROG can only be set with the buf and bufobj
2612 		 * locks both held.  We tolerate a race to clear it here.
2613 		 */
2614 		if (bp->b_vflags & BV_BKGRDINPROG) {
2615 			BUF_UNLOCK(bp);
2616 			continue;
2617 		}
2618 
2619 		/*
2620 		 * Requeue the background write buffer with error.
2621 		 */
2622 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
2623 			bremfreel(bp);
2624 			mtx_unlock(&bqclean);
2625 			bqrelse(bp);
2626 			continue;
2627 		}
2628 
2629 		KASSERT(bp->b_qindex == qindex,
2630 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2631 
2632 		bremfreel(bp);
2633 		mtx_unlock(&bqclean);
2634 
2635 		/*
2636 		 * NOTE:  nbp is now entirely invalid.  We can only restart
2637 		 * the scan from this point on.
2638 		 */
2639 		getnewbuf_reuse_bp(bp, qindex);
2640 		mtx_assert(&bqclean, MA_NOTOWNED);
2641 
2642 		/*
2643 		 * If we are defragging then free the buffer.
2644 		 */
2645 		if (defrag) {
2646 			bp->b_flags |= B_INVAL;
2647 			brelse(bp);
2648 			defrag = 0;
2649 			goto restart;
2650 		}
2651 
2652 		/*
2653 		 * Notify any waiters for the buffer lock about
2654 		 * identity change by freeing the buffer.
2655 		 */
2656 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2657 			bp->b_flags |= B_INVAL;
2658 			brelse(bp);
2659 			goto restart;
2660 		}
2661 
2662 		if (metadata)
2663 			break;
2664 
2665 		/*
2666 		 * If we are overcomitted then recover the buffer and its
2667 		 * KVM space.  This occurs in rare situations when multiple
2668 		 * processes are blocked in getnewbuf() or allocbuf().
2669 		 */
2670 		if (bufspace >= hibufspace && bp->b_kvasize != 0) {
2671 			bp->b_flags |= B_INVAL;
2672 			brelse(bp);
2673 			goto restart;
2674 		}
2675 		break;
2676 	}
2677 	return (bp);
2678 }
2679 
2680 /*
2681  *	getnewbuf:
2682  *
2683  *	Find and initialize a new buffer header, freeing up existing buffers
2684  *	in the bufqueues as necessary.  The new buffer is returned locked.
2685  *
2686  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
2687  *	buffer away, the caller must set B_INVAL prior to calling brelse().
2688  *
2689  *	We block if:
2690  *		We have insufficient buffer headers
2691  *		We have insufficient buffer space
2692  *		buffer_arena is too fragmented ( space reservation fails )
2693  *		If we have to flush dirty buffers ( but we try to avoid this )
2694  */
2695 static struct buf *
2696 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2697     int gbflags)
2698 {
2699 	struct buf *bp;
2700 	int defrag, metadata;
2701 
2702 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2703 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2704 	if (!unmapped_buf_allowed)
2705 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2706 
2707 	defrag = 0;
2708 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2709 	    vp->v_type == VCHR)
2710 		metadata = 1;
2711 	else
2712 		metadata = 0;
2713 	/*
2714 	 * We can't afford to block since we might be holding a vnode lock,
2715 	 * which may prevent system daemons from running.  We deal with
2716 	 * low-memory situations by proactively returning memory and running
2717 	 * async I/O rather then sync I/O.
2718 	 */
2719 	atomic_add_int(&getnewbufcalls, 1);
2720 restart:
2721 	bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2722 	    GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2723 	if (bp != NULL)
2724 		defrag = 0;
2725 
2726 	/*
2727 	 * If we exhausted our list, sleep as appropriate.  We may have to
2728 	 * wakeup various daemons and write out some dirty buffers.
2729 	 *
2730 	 * Generally we are sleeping due to insufficient buffer space.
2731 	 */
2732 	if (bp == NULL) {
2733 		mtx_assert(&bqclean, MA_OWNED);
2734 		getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2735 		mtx_assert(&bqclean, MA_NOTOWNED);
2736 	} else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2737 		mtx_assert(&bqclean, MA_NOTOWNED);
2738 
2739 		bufkvafree(bp);
2740 		atomic_add_int(&bufreusecnt, 1);
2741 	} else {
2742 		mtx_assert(&bqclean, MA_NOTOWNED);
2743 
2744 		/*
2745 		 * We finally have a valid bp.  We aren't quite out of the
2746 		 * woods, we still have to reserve kva space. In order to
2747 		 * keep fragmentation sane we only allocate kva in BKVASIZE
2748 		 * chunks.
2749 		 */
2750 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2751 
2752 		if (maxsize != bp->b_kvasize &&
2753 		    bufkvaalloc(bp, maxsize, gbflags)) {
2754 			defrag = 1;
2755 			bp->b_flags |= B_INVAL;
2756 			brelse(bp);
2757 			goto restart;
2758 		} else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) ==
2759 		    (GB_UNMAPPED | GB_KVAALLOC)) {
2760 			bp->b_data = unmapped_buf;
2761 			BUF_CHECK_UNMAPPED(bp);
2762 		}
2763 		atomic_add_int(&bufreusecnt, 1);
2764 	}
2765 	return (bp);
2766 }
2767 
2768 /*
2769  *	buf_daemon:
2770  *
2771  *	buffer flushing daemon.  Buffers are normally flushed by the
2772  *	update daemon but if it cannot keep up this process starts to
2773  *	take the load in an attempt to prevent getnewbuf() from blocking.
2774  */
2775 
2776 static struct kproc_desc buf_kp = {
2777 	"bufdaemon",
2778 	buf_daemon,
2779 	&bufdaemonproc
2780 };
2781 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2782 
2783 static int
2784 buf_flush(struct vnode *vp, int target)
2785 {
2786 	int flushed;
2787 
2788 	flushed = flushbufqueues(vp, target, 0);
2789 	if (flushed == 0) {
2790 		/*
2791 		 * Could not find any buffers without rollback
2792 		 * dependencies, so just write the first one
2793 		 * in the hopes of eventually making progress.
2794 		 */
2795 		if (vp != NULL && target > 2)
2796 			target /= 2;
2797 		flushbufqueues(vp, target, 1);
2798 	}
2799 	return (flushed);
2800 }
2801 
2802 static void
2803 buf_daemon()
2804 {
2805 	int lodirty;
2806 
2807 	/*
2808 	 * This process needs to be suspended prior to shutdown sync.
2809 	 */
2810 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2811 	    SHUTDOWN_PRI_LAST);
2812 
2813 	/*
2814 	 * This process is allowed to take the buffer cache to the limit
2815 	 */
2816 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2817 	mtx_lock(&bdlock);
2818 	for (;;) {
2819 		bd_request = 0;
2820 		mtx_unlock(&bdlock);
2821 
2822 		kproc_suspend_check(bufdaemonproc);
2823 		lodirty = lodirtybuffers;
2824 		if (bd_speedupreq) {
2825 			lodirty = numdirtybuffers / 2;
2826 			bd_speedupreq = 0;
2827 		}
2828 		/*
2829 		 * Do the flush.  Limit the amount of in-transit I/O we
2830 		 * allow to build up, otherwise we would completely saturate
2831 		 * the I/O system.
2832 		 */
2833 		while (numdirtybuffers > lodirty) {
2834 			if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
2835 				break;
2836 			kern_yield(PRI_USER);
2837 		}
2838 
2839 		/*
2840 		 * Only clear bd_request if we have reached our low water
2841 		 * mark.  The buf_daemon normally waits 1 second and
2842 		 * then incrementally flushes any dirty buffers that have
2843 		 * built up, within reason.
2844 		 *
2845 		 * If we were unable to hit our low water mark and couldn't
2846 		 * find any flushable buffers, we sleep for a short period
2847 		 * to avoid endless loops on unlockable buffers.
2848 		 */
2849 		mtx_lock(&bdlock);
2850 		if (numdirtybuffers <= lodirtybuffers) {
2851 			/*
2852 			 * We reached our low water mark, reset the
2853 			 * request and sleep until we are needed again.
2854 			 * The sleep is just so the suspend code works.
2855 			 */
2856 			bd_request = 0;
2857 			/*
2858 			 * Do an extra wakeup in case dirty threshold
2859 			 * changed via sysctl and the explicit transition
2860 			 * out of shortfall was missed.
2861 			 */
2862 			bdirtywakeup();
2863 			if (runningbufspace <= lorunningspace)
2864 				runningwakeup();
2865 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2866 		} else {
2867 			/*
2868 			 * We couldn't find any flushable dirty buffers but
2869 			 * still have too many dirty buffers, we
2870 			 * have to sleep and try again.  (rare)
2871 			 */
2872 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2873 		}
2874 	}
2875 }
2876 
2877 /*
2878  *	flushbufqueues:
2879  *
2880  *	Try to flush a buffer in the dirty queue.  We must be careful to
2881  *	free up B_INVAL buffers instead of write them, which NFS is
2882  *	particularly sensitive to.
2883  */
2884 static int flushwithdeps = 0;
2885 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2886     0, "Number of buffers flushed with dependecies that require rollbacks");
2887 
2888 static int
2889 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
2890 {
2891 	struct buf *sentinel;
2892 	struct vnode *vp;
2893 	struct mount *mp;
2894 	struct buf *bp;
2895 	int hasdeps;
2896 	int flushed;
2897 	int queue;
2898 	int error;
2899 	bool unlock;
2900 
2901 	flushed = 0;
2902 	queue = QUEUE_DIRTY;
2903 	bp = NULL;
2904 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2905 	sentinel->b_qindex = QUEUE_SENTINEL;
2906 	mtx_lock(&bqdirty);
2907 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2908 	mtx_unlock(&bqdirty);
2909 	while (flushed != target) {
2910 		maybe_yield();
2911 		mtx_lock(&bqdirty);
2912 		bp = TAILQ_NEXT(sentinel, b_freelist);
2913 		if (bp != NULL) {
2914 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2915 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2916 			    b_freelist);
2917 		} else {
2918 			mtx_unlock(&bqdirty);
2919 			break;
2920 		}
2921 		/*
2922 		 * Skip sentinels inserted by other invocations of the
2923 		 * flushbufqueues(), taking care to not reorder them.
2924 		 *
2925 		 * Only flush the buffers that belong to the
2926 		 * vnode locked by the curthread.
2927 		 */
2928 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
2929 		    bp->b_vp != lvp)) {
2930 			mtx_unlock(&bqdirty);
2931  			continue;
2932 		}
2933 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2934 		mtx_unlock(&bqdirty);
2935 		if (error != 0)
2936 			continue;
2937 		if (bp->b_pin_count > 0) {
2938 			BUF_UNLOCK(bp);
2939 			continue;
2940 		}
2941 		/*
2942 		 * BKGRDINPROG can only be set with the buf and bufobj
2943 		 * locks both held.  We tolerate a race to clear it here.
2944 		 */
2945 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2946 		    (bp->b_flags & B_DELWRI) == 0) {
2947 			BUF_UNLOCK(bp);
2948 			continue;
2949 		}
2950 		if (bp->b_flags & B_INVAL) {
2951 			bremfreef(bp);
2952 			brelse(bp);
2953 			flushed++;
2954 			continue;
2955 		}
2956 
2957 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2958 			if (flushdeps == 0) {
2959 				BUF_UNLOCK(bp);
2960 				continue;
2961 			}
2962 			hasdeps = 1;
2963 		} else
2964 			hasdeps = 0;
2965 		/*
2966 		 * We must hold the lock on a vnode before writing
2967 		 * one of its buffers. Otherwise we may confuse, or
2968 		 * in the case of a snapshot vnode, deadlock the
2969 		 * system.
2970 		 *
2971 		 * The lock order here is the reverse of the normal
2972 		 * of vnode followed by buf lock.  This is ok because
2973 		 * the NOWAIT will prevent deadlock.
2974 		 */
2975 		vp = bp->b_vp;
2976 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2977 			BUF_UNLOCK(bp);
2978 			continue;
2979 		}
2980 		if (lvp == NULL) {
2981 			unlock = true;
2982 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2983 		} else {
2984 			ASSERT_VOP_LOCKED(vp, "getbuf");
2985 			unlock = false;
2986 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2987 			    vn_lock(vp, LK_TRYUPGRADE);
2988 		}
2989 		if (error == 0) {
2990 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2991 			    bp, bp->b_vp, bp->b_flags);
2992 			if (curproc == bufdaemonproc) {
2993 				vfs_bio_awrite(bp);
2994 			} else {
2995 				bremfree(bp);
2996 				bwrite(bp);
2997 				notbufdflushes++;
2998 			}
2999 			vn_finished_write(mp);
3000 			if (unlock)
3001 				VOP_UNLOCK(vp, 0);
3002 			flushwithdeps += hasdeps;
3003 			flushed++;
3004 
3005 			/*
3006 			 * Sleeping on runningbufspace while holding
3007 			 * vnode lock leads to deadlock.
3008 			 */
3009 			if (curproc == bufdaemonproc &&
3010 			    runningbufspace > hirunningspace)
3011 				waitrunningbufspace();
3012 			continue;
3013 		}
3014 		vn_finished_write(mp);
3015 		BUF_UNLOCK(bp);
3016 	}
3017 	mtx_lock(&bqdirty);
3018 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
3019 	mtx_unlock(&bqdirty);
3020 	free(sentinel, M_TEMP);
3021 	return (flushed);
3022 }
3023 
3024 /*
3025  * Check to see if a block is currently memory resident.
3026  */
3027 struct buf *
3028 incore(struct bufobj *bo, daddr_t blkno)
3029 {
3030 	struct buf *bp;
3031 
3032 	BO_RLOCK(bo);
3033 	bp = gbincore(bo, blkno);
3034 	BO_RUNLOCK(bo);
3035 	return (bp);
3036 }
3037 
3038 /*
3039  * Returns true if no I/O is needed to access the
3040  * associated VM object.  This is like incore except
3041  * it also hunts around in the VM system for the data.
3042  */
3043 
3044 static int
3045 inmem(struct vnode * vp, daddr_t blkno)
3046 {
3047 	vm_object_t obj;
3048 	vm_offset_t toff, tinc, size;
3049 	vm_page_t m;
3050 	vm_ooffset_t off;
3051 
3052 	ASSERT_VOP_LOCKED(vp, "inmem");
3053 
3054 	if (incore(&vp->v_bufobj, blkno))
3055 		return 1;
3056 	if (vp->v_mount == NULL)
3057 		return 0;
3058 	obj = vp->v_object;
3059 	if (obj == NULL)
3060 		return (0);
3061 
3062 	size = PAGE_SIZE;
3063 	if (size > vp->v_mount->mnt_stat.f_iosize)
3064 		size = vp->v_mount->mnt_stat.f_iosize;
3065 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3066 
3067 	VM_OBJECT_RLOCK(obj);
3068 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3069 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3070 		if (!m)
3071 			goto notinmem;
3072 		tinc = size;
3073 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3074 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3075 		if (vm_page_is_valid(m,
3076 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3077 			goto notinmem;
3078 	}
3079 	VM_OBJECT_RUNLOCK(obj);
3080 	return 1;
3081 
3082 notinmem:
3083 	VM_OBJECT_RUNLOCK(obj);
3084 	return (0);
3085 }
3086 
3087 /*
3088  * Set the dirty range for a buffer based on the status of the dirty
3089  * bits in the pages comprising the buffer.  The range is limited
3090  * to the size of the buffer.
3091  *
3092  * Tell the VM system that the pages associated with this buffer
3093  * are clean.  This is used for delayed writes where the data is
3094  * going to go to disk eventually without additional VM intevention.
3095  *
3096  * Note that while we only really need to clean through to b_bcount, we
3097  * just go ahead and clean through to b_bufsize.
3098  */
3099 static void
3100 vfs_clean_pages_dirty_buf(struct buf *bp)
3101 {
3102 	vm_ooffset_t foff, noff, eoff;
3103 	vm_page_t m;
3104 	int i;
3105 
3106 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3107 		return;
3108 
3109 	foff = bp->b_offset;
3110 	KASSERT(bp->b_offset != NOOFFSET,
3111 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3112 
3113 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3114 	vfs_drain_busy_pages(bp);
3115 	vfs_setdirty_locked_object(bp);
3116 	for (i = 0; i < bp->b_npages; i++) {
3117 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3118 		eoff = noff;
3119 		if (eoff > bp->b_offset + bp->b_bufsize)
3120 			eoff = bp->b_offset + bp->b_bufsize;
3121 		m = bp->b_pages[i];
3122 		vfs_page_set_validclean(bp, foff, m);
3123 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3124 		foff = noff;
3125 	}
3126 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3127 }
3128 
3129 static void
3130 vfs_setdirty_locked_object(struct buf *bp)
3131 {
3132 	vm_object_t object;
3133 	int i;
3134 
3135 	object = bp->b_bufobj->bo_object;
3136 	VM_OBJECT_ASSERT_WLOCKED(object);
3137 
3138 	/*
3139 	 * We qualify the scan for modified pages on whether the
3140 	 * object has been flushed yet.
3141 	 */
3142 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3143 		vm_offset_t boffset;
3144 		vm_offset_t eoffset;
3145 
3146 		/*
3147 		 * test the pages to see if they have been modified directly
3148 		 * by users through the VM system.
3149 		 */
3150 		for (i = 0; i < bp->b_npages; i++)
3151 			vm_page_test_dirty(bp->b_pages[i]);
3152 
3153 		/*
3154 		 * Calculate the encompassing dirty range, boffset and eoffset,
3155 		 * (eoffset - boffset) bytes.
3156 		 */
3157 
3158 		for (i = 0; i < bp->b_npages; i++) {
3159 			if (bp->b_pages[i]->dirty)
3160 				break;
3161 		}
3162 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3163 
3164 		for (i = bp->b_npages - 1; i >= 0; --i) {
3165 			if (bp->b_pages[i]->dirty) {
3166 				break;
3167 			}
3168 		}
3169 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3170 
3171 		/*
3172 		 * Fit it to the buffer.
3173 		 */
3174 
3175 		if (eoffset > bp->b_bcount)
3176 			eoffset = bp->b_bcount;
3177 
3178 		/*
3179 		 * If we have a good dirty range, merge with the existing
3180 		 * dirty range.
3181 		 */
3182 
3183 		if (boffset < eoffset) {
3184 			if (bp->b_dirtyoff > boffset)
3185 				bp->b_dirtyoff = boffset;
3186 			if (bp->b_dirtyend < eoffset)
3187 				bp->b_dirtyend = eoffset;
3188 		}
3189 	}
3190 }
3191 
3192 /*
3193  * Allocate the KVA mapping for an existing buffer.
3194  * If an unmapped buffer is provided but a mapped buffer is requested, take
3195  * also care to properly setup mappings between pages and KVA.
3196  */
3197 static void
3198 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3199 {
3200 	struct buf *scratch_bp;
3201 	int bsize, maxsize, need_mapping, need_kva;
3202 	off_t offset;
3203 
3204 	need_mapping = bp->b_data == unmapped_buf &&
3205 	    (gbflags & GB_UNMAPPED) == 0;
3206 	need_kva = bp->b_kvabase == unmapped_buf &&
3207 	    bp->b_data == unmapped_buf &&
3208 	    (gbflags & GB_KVAALLOC) != 0;
3209 	if (!need_mapping && !need_kva)
3210 		return;
3211 
3212 	BUF_CHECK_UNMAPPED(bp);
3213 
3214 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3215 		/*
3216 		 * Buffer is not mapped, but the KVA was already
3217 		 * reserved at the time of the instantiation.  Use the
3218 		 * allocated space.
3219 		 */
3220 		goto has_addr;
3221 	}
3222 
3223 	/*
3224 	 * Calculate the amount of the address space we would reserve
3225 	 * if the buffer was mapped.
3226 	 */
3227 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3228 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3229 	offset = blkno * bsize;
3230 	maxsize = size + (offset & PAGE_MASK);
3231 	maxsize = imax(maxsize, bsize);
3232 
3233 mapping_loop:
3234 	if (bufkvaalloc(bp, maxsize, gbflags)) {
3235 		/*
3236 		 * Request defragmentation. getnewbuf() returns us the
3237 		 * allocated space by the scratch buffer KVA.
3238 		 */
3239 		scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
3240 		    (GB_UNMAPPED | GB_KVAALLOC));
3241 		if (scratch_bp == NULL) {
3242 			if ((gbflags & GB_NOWAIT_BD) != 0) {
3243 				/*
3244 				 * XXXKIB: defragmentation cannot
3245 				 * succeed, not sure what else to do.
3246 				 */
3247 				panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3248 			}
3249 			atomic_add_int(&mappingrestarts, 1);
3250 			goto mapping_loop;
3251 		}
3252 		KASSERT(scratch_bp->b_kvabase != unmapped_buf,
3253 		    ("scratch bp has no KVA %p", scratch_bp));
3254 		/* Grab pointers. */
3255 		bp->b_kvabase = scratch_bp->b_kvabase;
3256 		bp->b_kvasize = scratch_bp->b_kvasize;
3257 		bp->b_data = scratch_bp->b_data;
3258 
3259 		/* Get rid of the scratch buffer. */
3260 		scratch_bp->b_kvasize = 0;
3261 		scratch_bp->b_flags |= B_INVAL;
3262 		scratch_bp->b_data = scratch_bp->b_kvabase = unmapped_buf;
3263 		brelse(scratch_bp);
3264 	}
3265 has_addr:
3266 	if (need_mapping) {
3267 		/* b_offset is handled by bpmap_qenter. */
3268 		bp->b_data = bp->b_kvabase;
3269 		BUF_CHECK_MAPPED(bp);
3270 		bpmap_qenter(bp);
3271 	}
3272 }
3273 
3274 /*
3275  *	getblk:
3276  *
3277  *	Get a block given a specified block and offset into a file/device.
3278  *	The buffers B_DONE bit will be cleared on return, making it almost
3279  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3280  *	return.  The caller should clear B_INVAL prior to initiating a
3281  *	READ.
3282  *
3283  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3284  *	an existing buffer.
3285  *
3286  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3287  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3288  *	and then cleared based on the backing VM.  If the previous buffer is
3289  *	non-0-sized but invalid, B_CACHE will be cleared.
3290  *
3291  *	If getblk() must create a new buffer, the new buffer is returned with
3292  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3293  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3294  *	backing VM.
3295  *
3296  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3297  *	B_CACHE bit is clear.
3298  *
3299  *	What this means, basically, is that the caller should use B_CACHE to
3300  *	determine whether the buffer is fully valid or not and should clear
3301  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3302  *	the buffer by loading its data area with something, the caller needs
3303  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3304  *	the caller should set B_CACHE ( as an optimization ), else the caller
3305  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3306  *	a write attempt or if it was a successfull read.  If the caller
3307  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3308  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3309  */
3310 struct buf *
3311 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3312     int flags)
3313 {
3314 	struct buf *bp;
3315 	struct bufobj *bo;
3316 	int bsize, error, maxsize, vmio;
3317 	off_t offset;
3318 
3319 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3320 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3321 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3322 	ASSERT_VOP_LOCKED(vp, "getblk");
3323 	if (size > MAXBCACHEBUF)
3324 		panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size,
3325 		    MAXBCACHEBUF);
3326 	if (!unmapped_buf_allowed)
3327 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3328 
3329 	bo = &vp->v_bufobj;
3330 loop:
3331 	BO_RLOCK(bo);
3332 	bp = gbincore(bo, blkno);
3333 	if (bp != NULL) {
3334 		int lockflags;
3335 		/*
3336 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3337 		 * it must be on a queue.
3338 		 */
3339 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3340 
3341 		if (flags & GB_LOCK_NOWAIT)
3342 			lockflags |= LK_NOWAIT;
3343 
3344 		error = BUF_TIMELOCK(bp, lockflags,
3345 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3346 
3347 		/*
3348 		 * If we slept and got the lock we have to restart in case
3349 		 * the buffer changed identities.
3350 		 */
3351 		if (error == ENOLCK)
3352 			goto loop;
3353 		/* We timed out or were interrupted. */
3354 		else if (error)
3355 			return (NULL);
3356 		/* If recursed, assume caller knows the rules. */
3357 		else if (BUF_LOCKRECURSED(bp))
3358 			goto end;
3359 
3360 		/*
3361 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3362 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3363 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3364 		 * backing VM cache.
3365 		 */
3366 		if (bp->b_flags & B_INVAL)
3367 			bp->b_flags &= ~B_CACHE;
3368 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3369 			bp->b_flags |= B_CACHE;
3370 		if (bp->b_flags & B_MANAGED)
3371 			MPASS(bp->b_qindex == QUEUE_NONE);
3372 		else
3373 			bremfree(bp);
3374 
3375 		/*
3376 		 * check for size inconsistencies for non-VMIO case.
3377 		 */
3378 		if (bp->b_bcount != size) {
3379 			if ((bp->b_flags & B_VMIO) == 0 ||
3380 			    (size > bp->b_kvasize)) {
3381 				if (bp->b_flags & B_DELWRI) {
3382 					/*
3383 					 * If buffer is pinned and caller does
3384 					 * not want sleep  waiting for it to be
3385 					 * unpinned, bail out
3386 					 * */
3387 					if (bp->b_pin_count > 0) {
3388 						if (flags & GB_LOCK_NOWAIT) {
3389 							bqrelse(bp);
3390 							return (NULL);
3391 						} else {
3392 							bunpin_wait(bp);
3393 						}
3394 					}
3395 					bp->b_flags |= B_NOCACHE;
3396 					bwrite(bp);
3397 				} else {
3398 					if (LIST_EMPTY(&bp->b_dep)) {
3399 						bp->b_flags |= B_RELBUF;
3400 						brelse(bp);
3401 					} else {
3402 						bp->b_flags |= B_NOCACHE;
3403 						bwrite(bp);
3404 					}
3405 				}
3406 				goto loop;
3407 			}
3408 		}
3409 
3410 		/*
3411 		 * Handle the case of unmapped buffer which should
3412 		 * become mapped, or the buffer for which KVA
3413 		 * reservation is requested.
3414 		 */
3415 		bp_unmapped_get_kva(bp, blkno, size, flags);
3416 
3417 		/*
3418 		 * If the size is inconsistant in the VMIO case, we can resize
3419 		 * the buffer.  This might lead to B_CACHE getting set or
3420 		 * cleared.  If the size has not changed, B_CACHE remains
3421 		 * unchanged from its previous state.
3422 		 */
3423 		allocbuf(bp, size);
3424 
3425 		KASSERT(bp->b_offset != NOOFFSET,
3426 		    ("getblk: no buffer offset"));
3427 
3428 		/*
3429 		 * A buffer with B_DELWRI set and B_CACHE clear must
3430 		 * be committed before we can return the buffer in
3431 		 * order to prevent the caller from issuing a read
3432 		 * ( due to B_CACHE not being set ) and overwriting
3433 		 * it.
3434 		 *
3435 		 * Most callers, including NFS and FFS, need this to
3436 		 * operate properly either because they assume they
3437 		 * can issue a read if B_CACHE is not set, or because
3438 		 * ( for example ) an uncached B_DELWRI might loop due
3439 		 * to softupdates re-dirtying the buffer.  In the latter
3440 		 * case, B_CACHE is set after the first write completes,
3441 		 * preventing further loops.
3442 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3443 		 * above while extending the buffer, we cannot allow the
3444 		 * buffer to remain with B_CACHE set after the write
3445 		 * completes or it will represent a corrupt state.  To
3446 		 * deal with this we set B_NOCACHE to scrap the buffer
3447 		 * after the write.
3448 		 *
3449 		 * We might be able to do something fancy, like setting
3450 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3451 		 * so the below call doesn't set B_CACHE, but that gets real
3452 		 * confusing.  This is much easier.
3453 		 */
3454 
3455 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3456 			bp->b_flags |= B_NOCACHE;
3457 			bwrite(bp);
3458 			goto loop;
3459 		}
3460 		bp->b_flags &= ~B_DONE;
3461 	} else {
3462 		/*
3463 		 * Buffer is not in-core, create new buffer.  The buffer
3464 		 * returned by getnewbuf() is locked.  Note that the returned
3465 		 * buffer is also considered valid (not marked B_INVAL).
3466 		 */
3467 		BO_RUNLOCK(bo);
3468 		/*
3469 		 * If the user does not want us to create the buffer, bail out
3470 		 * here.
3471 		 */
3472 		if (flags & GB_NOCREAT)
3473 			return NULL;
3474 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3475 			return NULL;
3476 
3477 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3478 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3479 		offset = blkno * bsize;
3480 		vmio = vp->v_object != NULL;
3481 		if (vmio) {
3482 			maxsize = size + (offset & PAGE_MASK);
3483 		} else {
3484 			maxsize = size;
3485 			/* Do not allow non-VMIO notmapped buffers. */
3486 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3487 		}
3488 		maxsize = imax(maxsize, bsize);
3489 
3490 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3491 		if (bp == NULL) {
3492 			if (slpflag || slptimeo)
3493 				return NULL;
3494 			goto loop;
3495 		}
3496 
3497 		/*
3498 		 * This code is used to make sure that a buffer is not
3499 		 * created while the getnewbuf routine is blocked.
3500 		 * This can be a problem whether the vnode is locked or not.
3501 		 * If the buffer is created out from under us, we have to
3502 		 * throw away the one we just created.
3503 		 *
3504 		 * Note: this must occur before we associate the buffer
3505 		 * with the vp especially considering limitations in
3506 		 * the splay tree implementation when dealing with duplicate
3507 		 * lblkno's.
3508 		 */
3509 		BO_LOCK(bo);
3510 		if (gbincore(bo, blkno)) {
3511 			BO_UNLOCK(bo);
3512 			bp->b_flags |= B_INVAL;
3513 			brelse(bp);
3514 			goto loop;
3515 		}
3516 
3517 		/*
3518 		 * Insert the buffer into the hash, so that it can
3519 		 * be found by incore.
3520 		 */
3521 		bp->b_blkno = bp->b_lblkno = blkno;
3522 		bp->b_offset = offset;
3523 		bgetvp(vp, bp);
3524 		BO_UNLOCK(bo);
3525 
3526 		/*
3527 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3528 		 * buffer size starts out as 0, B_CACHE will be set by
3529 		 * allocbuf() for the VMIO case prior to it testing the
3530 		 * backing store for validity.
3531 		 */
3532 
3533 		if (vmio) {
3534 			bp->b_flags |= B_VMIO;
3535 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3536 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3537 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3538 		} else {
3539 			bp->b_flags &= ~B_VMIO;
3540 			KASSERT(bp->b_bufobj->bo_object == NULL,
3541 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3542 			    bp, bp->b_bufobj->bo_object));
3543 			BUF_CHECK_MAPPED(bp);
3544 		}
3545 
3546 		allocbuf(bp, size);
3547 		bp->b_flags &= ~B_DONE;
3548 	}
3549 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3550 	BUF_ASSERT_HELD(bp);
3551 end:
3552 	KASSERT(bp->b_bufobj == bo,
3553 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3554 	return (bp);
3555 }
3556 
3557 /*
3558  * Get an empty, disassociated buffer of given size.  The buffer is initially
3559  * set to B_INVAL.
3560  */
3561 struct buf *
3562 geteblk(int size, int flags)
3563 {
3564 	struct buf *bp;
3565 	int maxsize;
3566 
3567 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3568 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3569 		if ((flags & GB_NOWAIT_BD) &&
3570 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3571 			return (NULL);
3572 	}
3573 	allocbuf(bp, size);
3574 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3575 	BUF_ASSERT_HELD(bp);
3576 	return (bp);
3577 }
3578 
3579 /*
3580  * Truncate the backing store for a non-vmio buffer.
3581  */
3582 static void
3583 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
3584 {
3585 
3586 	if (bp->b_flags & B_MALLOC) {
3587 		/*
3588 		 * malloced buffers are not shrunk
3589 		 */
3590 		if (newbsize == 0) {
3591 			bufmallocadjust(bp, 0);
3592 			free(bp->b_data, M_BIOBUF);
3593 			bp->b_data = bp->b_kvabase;
3594 			bp->b_flags &= ~B_MALLOC;
3595 		}
3596 		return;
3597 	}
3598 	vm_hold_free_pages(bp, newbsize);
3599 	bufspaceadjust(bp, newbsize);
3600 }
3601 
3602 /*
3603  * Extend the backing for a non-VMIO buffer.
3604  */
3605 static void
3606 vfs_nonvmio_extend(struct buf *bp, int newbsize)
3607 {
3608 	caddr_t origbuf;
3609 	int origbufsize;
3610 
3611 	/*
3612 	 * We only use malloced memory on the first allocation.
3613 	 * and revert to page-allocated memory when the buffer
3614 	 * grows.
3615 	 *
3616 	 * There is a potential smp race here that could lead
3617 	 * to bufmallocspace slightly passing the max.  It
3618 	 * is probably extremely rare and not worth worrying
3619 	 * over.
3620 	 */
3621 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
3622 	    bufmallocspace < maxbufmallocspace) {
3623 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
3624 		bp->b_flags |= B_MALLOC;
3625 		bufmallocadjust(bp, newbsize);
3626 		return;
3627 	}
3628 
3629 	/*
3630 	 * If the buffer is growing on its other-than-first
3631 	 * allocation then we revert to the page-allocation
3632 	 * scheme.
3633 	 */
3634 	origbuf = NULL;
3635 	origbufsize = 0;
3636 	if (bp->b_flags & B_MALLOC) {
3637 		origbuf = bp->b_data;
3638 		origbufsize = bp->b_bufsize;
3639 		bp->b_data = bp->b_kvabase;
3640 		bufmallocadjust(bp, 0);
3641 		bp->b_flags &= ~B_MALLOC;
3642 		newbsize = round_page(newbsize);
3643 	}
3644 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
3645 	    (vm_offset_t) bp->b_data + newbsize);
3646 	if (origbuf != NULL) {
3647 		bcopy(origbuf, bp->b_data, origbufsize);
3648 		free(origbuf, M_BIOBUF);
3649 	}
3650 	bufspaceadjust(bp, newbsize);
3651 }
3652 
3653 /*
3654  * This code constitutes the buffer memory from either anonymous system
3655  * memory (in the case of non-VMIO operations) or from an associated
3656  * VM object (in the case of VMIO operations).  This code is able to
3657  * resize a buffer up or down.
3658  *
3659  * Note that this code is tricky, and has many complications to resolve
3660  * deadlock or inconsistant data situations.  Tread lightly!!!
3661  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3662  * the caller.  Calling this code willy nilly can result in the loss of data.
3663  *
3664  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3665  * B_CACHE for the non-VMIO case.
3666  */
3667 int
3668 allocbuf(struct buf *bp, int size)
3669 {
3670 	int newbsize;
3671 
3672 	BUF_ASSERT_HELD(bp);
3673 
3674 	if (bp->b_bcount == size)
3675 		return (1);
3676 
3677 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
3678 		panic("allocbuf: buffer too small");
3679 
3680 	newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3681 	if ((bp->b_flags & B_VMIO) == 0) {
3682 		if ((bp->b_flags & B_MALLOC) == 0)
3683 			newbsize = round_page(newbsize);
3684 		/*
3685 		 * Just get anonymous memory from the kernel.  Don't
3686 		 * mess with B_CACHE.
3687 		 */
3688 		if (newbsize < bp->b_bufsize)
3689 			vfs_nonvmio_truncate(bp, newbsize);
3690 		else if (newbsize > bp->b_bufsize)
3691 			vfs_nonvmio_extend(bp, newbsize);
3692 	} else {
3693 		int desiredpages;
3694 
3695 		desiredpages = (size == 0) ? 0 :
3696 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3697 
3698 		if (bp->b_flags & B_MALLOC)
3699 			panic("allocbuf: VMIO buffer can't be malloced");
3700 		/*
3701 		 * Set B_CACHE initially if buffer is 0 length or will become
3702 		 * 0-length.
3703 		 */
3704 		if (size == 0 || bp->b_bufsize == 0)
3705 			bp->b_flags |= B_CACHE;
3706 
3707 		if (newbsize < bp->b_bufsize)
3708 			vfs_vmio_truncate(bp, desiredpages);
3709 		/* XXX This looks as if it should be newbsize > b_bufsize */
3710 		else if (size > bp->b_bcount)
3711 			vfs_vmio_extend(bp, desiredpages, size);
3712 		bufspaceadjust(bp, newbsize);
3713 	}
3714 	bp->b_bcount = size;		/* requested buffer size. */
3715 	return (1);
3716 }
3717 
3718 extern int inflight_transient_maps;
3719 
3720 void
3721 biodone(struct bio *bp)
3722 {
3723 	struct mtx *mtxp;
3724 	void (*done)(struct bio *);
3725 	vm_offset_t start, end;
3726 
3727 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3728 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3729 		bp->bio_flags |= BIO_UNMAPPED;
3730 		start = trunc_page((vm_offset_t)bp->bio_data);
3731 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3732 		bp->bio_data = unmapped_buf;
3733 		pmap_qremove(start, OFF_TO_IDX(end - start));
3734 		vmem_free(transient_arena, start, end - start);
3735 		atomic_add_int(&inflight_transient_maps, -1);
3736 	}
3737 	done = bp->bio_done;
3738 	if (done == NULL) {
3739 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
3740 		mtx_lock(mtxp);
3741 		bp->bio_flags |= BIO_DONE;
3742 		wakeup(bp);
3743 		mtx_unlock(mtxp);
3744 	} else {
3745 		bp->bio_flags |= BIO_DONE;
3746 		done(bp);
3747 	}
3748 }
3749 
3750 /*
3751  * Wait for a BIO to finish.
3752  */
3753 int
3754 biowait(struct bio *bp, const char *wchan)
3755 {
3756 	struct mtx *mtxp;
3757 
3758 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3759 	mtx_lock(mtxp);
3760 	while ((bp->bio_flags & BIO_DONE) == 0)
3761 		msleep(bp, mtxp, PRIBIO, wchan, 0);
3762 	mtx_unlock(mtxp);
3763 	if (bp->bio_error != 0)
3764 		return (bp->bio_error);
3765 	if (!(bp->bio_flags & BIO_ERROR))
3766 		return (0);
3767 	return (EIO);
3768 }
3769 
3770 void
3771 biofinish(struct bio *bp, struct devstat *stat, int error)
3772 {
3773 
3774 	if (error) {
3775 		bp->bio_error = error;
3776 		bp->bio_flags |= BIO_ERROR;
3777 	}
3778 	if (stat != NULL)
3779 		devstat_end_transaction_bio(stat, bp);
3780 	biodone(bp);
3781 }
3782 
3783 /*
3784  *	bufwait:
3785  *
3786  *	Wait for buffer I/O completion, returning error status.  The buffer
3787  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3788  *	error and cleared.
3789  */
3790 int
3791 bufwait(struct buf *bp)
3792 {
3793 	if (bp->b_iocmd == BIO_READ)
3794 		bwait(bp, PRIBIO, "biord");
3795 	else
3796 		bwait(bp, PRIBIO, "biowr");
3797 	if (bp->b_flags & B_EINTR) {
3798 		bp->b_flags &= ~B_EINTR;
3799 		return (EINTR);
3800 	}
3801 	if (bp->b_ioflags & BIO_ERROR) {
3802 		return (bp->b_error ? bp->b_error : EIO);
3803 	} else {
3804 		return (0);
3805 	}
3806 }
3807 
3808 /*
3809  *	bufdone:
3810  *
3811  *	Finish I/O on a buffer, optionally calling a completion function.
3812  *	This is usually called from an interrupt so process blocking is
3813  *	not allowed.
3814  *
3815  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3816  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3817  *	assuming B_INVAL is clear.
3818  *
3819  *	For the VMIO case, we set B_CACHE if the op was a read and no
3820  *	read error occured, or if the op was a write.  B_CACHE is never
3821  *	set if the buffer is invalid or otherwise uncacheable.
3822  *
3823  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3824  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3825  *	in the biodone routine.
3826  */
3827 void
3828 bufdone(struct buf *bp)
3829 {
3830 	struct bufobj *dropobj;
3831 	void    (*biodone)(struct buf *);
3832 
3833 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3834 	dropobj = NULL;
3835 
3836 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3837 	BUF_ASSERT_HELD(bp);
3838 
3839 	runningbufwakeup(bp);
3840 	if (bp->b_iocmd == BIO_WRITE)
3841 		dropobj = bp->b_bufobj;
3842 	/* call optional completion function if requested */
3843 	if (bp->b_iodone != NULL) {
3844 		biodone = bp->b_iodone;
3845 		bp->b_iodone = NULL;
3846 		(*biodone) (bp);
3847 		if (dropobj)
3848 			bufobj_wdrop(dropobj);
3849 		return;
3850 	}
3851 
3852 	bufdone_finish(bp);
3853 
3854 	if (dropobj)
3855 		bufobj_wdrop(dropobj);
3856 }
3857 
3858 void
3859 bufdone_finish(struct buf *bp)
3860 {
3861 	BUF_ASSERT_HELD(bp);
3862 
3863 	if (!LIST_EMPTY(&bp->b_dep))
3864 		buf_complete(bp);
3865 
3866 	if (bp->b_flags & B_VMIO) {
3867 		/*
3868 		 * Set B_CACHE if the op was a normal read and no error
3869 		 * occured.  B_CACHE is set for writes in the b*write()
3870 		 * routines.
3871 		 */
3872 		if (bp->b_iocmd == BIO_READ &&
3873 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3874 		    !(bp->b_ioflags & BIO_ERROR))
3875 			bp->b_flags |= B_CACHE;
3876 		vfs_vmio_iodone(bp);
3877 	}
3878 
3879 	/*
3880 	 * For asynchronous completions, release the buffer now. The brelse
3881 	 * will do a wakeup there if necessary - so no need to do a wakeup
3882 	 * here in the async case. The sync case always needs to do a wakeup.
3883 	 */
3884 	if (bp->b_flags & B_ASYNC) {
3885 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
3886 		    (bp->b_ioflags & BIO_ERROR))
3887 			brelse(bp);
3888 		else
3889 			bqrelse(bp);
3890 	} else
3891 		bdone(bp);
3892 }
3893 
3894 /*
3895  * This routine is called in lieu of iodone in the case of
3896  * incomplete I/O.  This keeps the busy status for pages
3897  * consistant.
3898  */
3899 void
3900 vfs_unbusy_pages(struct buf *bp)
3901 {
3902 	int i;
3903 	vm_object_t obj;
3904 	vm_page_t m;
3905 
3906 	runningbufwakeup(bp);
3907 	if (!(bp->b_flags & B_VMIO))
3908 		return;
3909 
3910 	obj = bp->b_bufobj->bo_object;
3911 	VM_OBJECT_WLOCK(obj);
3912 	for (i = 0; i < bp->b_npages; i++) {
3913 		m = bp->b_pages[i];
3914 		if (m == bogus_page) {
3915 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3916 			if (!m)
3917 				panic("vfs_unbusy_pages: page missing\n");
3918 			bp->b_pages[i] = m;
3919 			if (buf_mapped(bp)) {
3920 				BUF_CHECK_MAPPED(bp);
3921 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3922 				    bp->b_pages, bp->b_npages);
3923 			} else
3924 				BUF_CHECK_UNMAPPED(bp);
3925 		}
3926 		vm_object_pip_subtract(obj, 1);
3927 		vm_page_sunbusy(m);
3928 	}
3929 	vm_object_pip_wakeupn(obj, 0);
3930 	VM_OBJECT_WUNLOCK(obj);
3931 }
3932 
3933 /*
3934  * vfs_page_set_valid:
3935  *
3936  *	Set the valid bits in a page based on the supplied offset.   The
3937  *	range is restricted to the buffer's size.
3938  *
3939  *	This routine is typically called after a read completes.
3940  */
3941 static void
3942 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3943 {
3944 	vm_ooffset_t eoff;
3945 
3946 	/*
3947 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3948 	 * page boundary and eoff is not greater than the end of the buffer.
3949 	 * The end of the buffer, in this case, is our file EOF, not the
3950 	 * allocation size of the buffer.
3951 	 */
3952 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3953 	if (eoff > bp->b_offset + bp->b_bcount)
3954 		eoff = bp->b_offset + bp->b_bcount;
3955 
3956 	/*
3957 	 * Set valid range.  This is typically the entire buffer and thus the
3958 	 * entire page.
3959 	 */
3960 	if (eoff > off)
3961 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3962 }
3963 
3964 /*
3965  * vfs_page_set_validclean:
3966  *
3967  *	Set the valid bits and clear the dirty bits in a page based on the
3968  *	supplied offset.   The range is restricted to the buffer's size.
3969  */
3970 static void
3971 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3972 {
3973 	vm_ooffset_t soff, eoff;
3974 
3975 	/*
3976 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3977 	 * page boundry or cross the end of the buffer.  The end of the
3978 	 * buffer, in this case, is our file EOF, not the allocation size
3979 	 * of the buffer.
3980 	 */
3981 	soff = off;
3982 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3983 	if (eoff > bp->b_offset + bp->b_bcount)
3984 		eoff = bp->b_offset + bp->b_bcount;
3985 
3986 	/*
3987 	 * Set valid range.  This is typically the entire buffer and thus the
3988 	 * entire page.
3989 	 */
3990 	if (eoff > soff) {
3991 		vm_page_set_validclean(
3992 		    m,
3993 		   (vm_offset_t) (soff & PAGE_MASK),
3994 		   (vm_offset_t) (eoff - soff)
3995 		);
3996 	}
3997 }
3998 
3999 /*
4000  * Ensure that all buffer pages are not exclusive busied.  If any page is
4001  * exclusive busy, drain it.
4002  */
4003 void
4004 vfs_drain_busy_pages(struct buf *bp)
4005 {
4006 	vm_page_t m;
4007 	int i, last_busied;
4008 
4009 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4010 	last_busied = 0;
4011 	for (i = 0; i < bp->b_npages; i++) {
4012 		m = bp->b_pages[i];
4013 		if (vm_page_xbusied(m)) {
4014 			for (; last_busied < i; last_busied++)
4015 				vm_page_sbusy(bp->b_pages[last_busied]);
4016 			while (vm_page_xbusied(m)) {
4017 				vm_page_lock(m);
4018 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4019 				vm_page_busy_sleep(m, "vbpage");
4020 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4021 			}
4022 		}
4023 	}
4024 	for (i = 0; i < last_busied; i++)
4025 		vm_page_sunbusy(bp->b_pages[i]);
4026 }
4027 
4028 /*
4029  * This routine is called before a device strategy routine.
4030  * It is used to tell the VM system that paging I/O is in
4031  * progress, and treat the pages associated with the buffer
4032  * almost as being exclusive busy.  Also the object paging_in_progress
4033  * flag is handled to make sure that the object doesn't become
4034  * inconsistant.
4035  *
4036  * Since I/O has not been initiated yet, certain buffer flags
4037  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4038  * and should be ignored.
4039  */
4040 void
4041 vfs_busy_pages(struct buf *bp, int clear_modify)
4042 {
4043 	int i, bogus;
4044 	vm_object_t obj;
4045 	vm_ooffset_t foff;
4046 	vm_page_t m;
4047 
4048 	if (!(bp->b_flags & B_VMIO))
4049 		return;
4050 
4051 	obj = bp->b_bufobj->bo_object;
4052 	foff = bp->b_offset;
4053 	KASSERT(bp->b_offset != NOOFFSET,
4054 	    ("vfs_busy_pages: no buffer offset"));
4055 	VM_OBJECT_WLOCK(obj);
4056 	vfs_drain_busy_pages(bp);
4057 	if (bp->b_bufsize != 0)
4058 		vfs_setdirty_locked_object(bp);
4059 	bogus = 0;
4060 	for (i = 0; i < bp->b_npages; i++) {
4061 		m = bp->b_pages[i];
4062 
4063 		if ((bp->b_flags & B_CLUSTER) == 0) {
4064 			vm_object_pip_add(obj, 1);
4065 			vm_page_sbusy(m);
4066 		}
4067 		/*
4068 		 * When readying a buffer for a read ( i.e
4069 		 * clear_modify == 0 ), it is important to do
4070 		 * bogus_page replacement for valid pages in
4071 		 * partially instantiated buffers.  Partially
4072 		 * instantiated buffers can, in turn, occur when
4073 		 * reconstituting a buffer from its VM backing store
4074 		 * base.  We only have to do this if B_CACHE is
4075 		 * clear ( which causes the I/O to occur in the
4076 		 * first place ).  The replacement prevents the read
4077 		 * I/O from overwriting potentially dirty VM-backed
4078 		 * pages.  XXX bogus page replacement is, uh, bogus.
4079 		 * It may not work properly with small-block devices.
4080 		 * We need to find a better way.
4081 		 */
4082 		if (clear_modify) {
4083 			pmap_remove_write(m);
4084 			vfs_page_set_validclean(bp, foff, m);
4085 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4086 		    (bp->b_flags & B_CACHE) == 0) {
4087 			bp->b_pages[i] = bogus_page;
4088 			bogus++;
4089 		}
4090 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4091 	}
4092 	VM_OBJECT_WUNLOCK(obj);
4093 	if (bogus && buf_mapped(bp)) {
4094 		BUF_CHECK_MAPPED(bp);
4095 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4096 		    bp->b_pages, bp->b_npages);
4097 	}
4098 }
4099 
4100 /*
4101  *	vfs_bio_set_valid:
4102  *
4103  *	Set the range within the buffer to valid.  The range is
4104  *	relative to the beginning of the buffer, b_offset.  Note that
4105  *	b_offset itself may be offset from the beginning of the first
4106  *	page.
4107  */
4108 void
4109 vfs_bio_set_valid(struct buf *bp, int base, int size)
4110 {
4111 	int i, n;
4112 	vm_page_t m;
4113 
4114 	if (!(bp->b_flags & B_VMIO))
4115 		return;
4116 
4117 	/*
4118 	 * Fixup base to be relative to beginning of first page.
4119 	 * Set initial n to be the maximum number of bytes in the
4120 	 * first page that can be validated.
4121 	 */
4122 	base += (bp->b_offset & PAGE_MASK);
4123 	n = PAGE_SIZE - (base & PAGE_MASK);
4124 
4125 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4126 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4127 		m = bp->b_pages[i];
4128 		if (n > size)
4129 			n = size;
4130 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4131 		base += n;
4132 		size -= n;
4133 		n = PAGE_SIZE;
4134 	}
4135 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4136 }
4137 
4138 /*
4139  *	vfs_bio_clrbuf:
4140  *
4141  *	If the specified buffer is a non-VMIO buffer, clear the entire
4142  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4143  *	validate only the previously invalid portions of the buffer.
4144  *	This routine essentially fakes an I/O, so we need to clear
4145  *	BIO_ERROR and B_INVAL.
4146  *
4147  *	Note that while we only theoretically need to clear through b_bcount,
4148  *	we go ahead and clear through b_bufsize.
4149  */
4150 void
4151 vfs_bio_clrbuf(struct buf *bp)
4152 {
4153 	int i, j, mask, sa, ea, slide;
4154 
4155 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4156 		clrbuf(bp);
4157 		return;
4158 	}
4159 	bp->b_flags &= ~B_INVAL;
4160 	bp->b_ioflags &= ~BIO_ERROR;
4161 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4162 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4163 	    (bp->b_offset & PAGE_MASK) == 0) {
4164 		if (bp->b_pages[0] == bogus_page)
4165 			goto unlock;
4166 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4167 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4168 		if ((bp->b_pages[0]->valid & mask) == mask)
4169 			goto unlock;
4170 		if ((bp->b_pages[0]->valid & mask) == 0) {
4171 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4172 			bp->b_pages[0]->valid |= mask;
4173 			goto unlock;
4174 		}
4175 	}
4176 	sa = bp->b_offset & PAGE_MASK;
4177 	slide = 0;
4178 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4179 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4180 		ea = slide & PAGE_MASK;
4181 		if (ea == 0)
4182 			ea = PAGE_SIZE;
4183 		if (bp->b_pages[i] == bogus_page)
4184 			continue;
4185 		j = sa / DEV_BSIZE;
4186 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4187 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4188 		if ((bp->b_pages[i]->valid & mask) == mask)
4189 			continue;
4190 		if ((bp->b_pages[i]->valid & mask) == 0)
4191 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4192 		else {
4193 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4194 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4195 					pmap_zero_page_area(bp->b_pages[i],
4196 					    sa, DEV_BSIZE);
4197 				}
4198 			}
4199 		}
4200 		bp->b_pages[i]->valid |= mask;
4201 	}
4202 unlock:
4203 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4204 	bp->b_resid = 0;
4205 }
4206 
4207 void
4208 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4209 {
4210 	vm_page_t m;
4211 	int i, n;
4212 
4213 	if (buf_mapped(bp)) {
4214 		BUF_CHECK_MAPPED(bp);
4215 		bzero(bp->b_data + base, size);
4216 	} else {
4217 		BUF_CHECK_UNMAPPED(bp);
4218 		n = PAGE_SIZE - (base & PAGE_MASK);
4219 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4220 			m = bp->b_pages[i];
4221 			if (n > size)
4222 				n = size;
4223 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4224 			base += n;
4225 			size -= n;
4226 			n = PAGE_SIZE;
4227 		}
4228 	}
4229 }
4230 
4231 /*
4232  * vm_hold_load_pages and vm_hold_free_pages get pages into
4233  * a buffers address space.  The pages are anonymous and are
4234  * not associated with a file object.
4235  */
4236 static void
4237 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4238 {
4239 	vm_offset_t pg;
4240 	vm_page_t p;
4241 	int index;
4242 
4243 	BUF_CHECK_MAPPED(bp);
4244 
4245 	to = round_page(to);
4246 	from = round_page(from);
4247 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4248 
4249 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4250 tryagain:
4251 		/*
4252 		 * note: must allocate system pages since blocking here
4253 		 * could interfere with paging I/O, no matter which
4254 		 * process we are.
4255 		 */
4256 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4257 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4258 		if (p == NULL) {
4259 			VM_WAIT;
4260 			goto tryagain;
4261 		}
4262 		pmap_qenter(pg, &p, 1);
4263 		bp->b_pages[index] = p;
4264 	}
4265 	bp->b_npages = index;
4266 }
4267 
4268 /* Return pages associated with this buf to the vm system */
4269 static void
4270 vm_hold_free_pages(struct buf *bp, int newbsize)
4271 {
4272 	vm_offset_t from;
4273 	vm_page_t p;
4274 	int index, newnpages;
4275 
4276 	BUF_CHECK_MAPPED(bp);
4277 
4278 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4279 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4280 	if (bp->b_npages > newnpages)
4281 		pmap_qremove(from, bp->b_npages - newnpages);
4282 	for (index = newnpages; index < bp->b_npages; index++) {
4283 		p = bp->b_pages[index];
4284 		bp->b_pages[index] = NULL;
4285 		if (vm_page_sbusied(p))
4286 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4287 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4288 		p->wire_count--;
4289 		vm_page_free(p);
4290 		atomic_subtract_int(&vm_cnt.v_wire_count, 1);
4291 	}
4292 	bp->b_npages = newnpages;
4293 }
4294 
4295 /*
4296  * Map an IO request into kernel virtual address space.
4297  *
4298  * All requests are (re)mapped into kernel VA space.
4299  * Notice that we use b_bufsize for the size of the buffer
4300  * to be mapped.  b_bcount might be modified by the driver.
4301  *
4302  * Note that even if the caller determines that the address space should
4303  * be valid, a race or a smaller-file mapped into a larger space may
4304  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4305  * check the return value.
4306  *
4307  * This function only works with pager buffers.
4308  */
4309 int
4310 vmapbuf(struct buf *bp, int mapbuf)
4311 {
4312 	vm_prot_t prot;
4313 	int pidx;
4314 
4315 	if (bp->b_bufsize < 0)
4316 		return (-1);
4317 	prot = VM_PROT_READ;
4318 	if (bp->b_iocmd == BIO_READ)
4319 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4320 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4321 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4322 	    btoc(MAXPHYS))) < 0)
4323 		return (-1);
4324 	bp->b_npages = pidx;
4325 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4326 	if (mapbuf || !unmapped_buf_allowed) {
4327 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4328 		bp->b_data = bp->b_kvabase + bp->b_offset;
4329 	} else
4330 		bp->b_data = unmapped_buf;
4331 	return(0);
4332 }
4333 
4334 /*
4335  * Free the io map PTEs associated with this IO operation.
4336  * We also invalidate the TLB entries and restore the original b_addr.
4337  *
4338  * This function only works with pager buffers.
4339  */
4340 void
4341 vunmapbuf(struct buf *bp)
4342 {
4343 	int npages;
4344 
4345 	npages = bp->b_npages;
4346 	if (buf_mapped(bp))
4347 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4348 	vm_page_unhold_pages(bp->b_pages, npages);
4349 
4350 	bp->b_data = unmapped_buf;
4351 }
4352 
4353 void
4354 bdone(struct buf *bp)
4355 {
4356 	struct mtx *mtxp;
4357 
4358 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4359 	mtx_lock(mtxp);
4360 	bp->b_flags |= B_DONE;
4361 	wakeup(bp);
4362 	mtx_unlock(mtxp);
4363 }
4364 
4365 void
4366 bwait(struct buf *bp, u_char pri, const char *wchan)
4367 {
4368 	struct mtx *mtxp;
4369 
4370 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4371 	mtx_lock(mtxp);
4372 	while ((bp->b_flags & B_DONE) == 0)
4373 		msleep(bp, mtxp, pri, wchan, 0);
4374 	mtx_unlock(mtxp);
4375 }
4376 
4377 int
4378 bufsync(struct bufobj *bo, int waitfor)
4379 {
4380 
4381 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4382 }
4383 
4384 void
4385 bufstrategy(struct bufobj *bo, struct buf *bp)
4386 {
4387 	int i = 0;
4388 	struct vnode *vp;
4389 
4390 	vp = bp->b_vp;
4391 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4392 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4393 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4394 	i = VOP_STRATEGY(vp, bp);
4395 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4396 }
4397 
4398 void
4399 bufobj_wrefl(struct bufobj *bo)
4400 {
4401 
4402 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4403 	ASSERT_BO_WLOCKED(bo);
4404 	bo->bo_numoutput++;
4405 }
4406 
4407 void
4408 bufobj_wref(struct bufobj *bo)
4409 {
4410 
4411 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4412 	BO_LOCK(bo);
4413 	bo->bo_numoutput++;
4414 	BO_UNLOCK(bo);
4415 }
4416 
4417 void
4418 bufobj_wdrop(struct bufobj *bo)
4419 {
4420 
4421 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4422 	BO_LOCK(bo);
4423 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4424 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4425 		bo->bo_flag &= ~BO_WWAIT;
4426 		wakeup(&bo->bo_numoutput);
4427 	}
4428 	BO_UNLOCK(bo);
4429 }
4430 
4431 int
4432 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4433 {
4434 	int error;
4435 
4436 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4437 	ASSERT_BO_WLOCKED(bo);
4438 	error = 0;
4439 	while (bo->bo_numoutput) {
4440 		bo->bo_flag |= BO_WWAIT;
4441 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4442 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4443 		if (error)
4444 			break;
4445 	}
4446 	return (error);
4447 }
4448 
4449 void
4450 bpin(struct buf *bp)
4451 {
4452 	struct mtx *mtxp;
4453 
4454 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4455 	mtx_lock(mtxp);
4456 	bp->b_pin_count++;
4457 	mtx_unlock(mtxp);
4458 }
4459 
4460 void
4461 bunpin(struct buf *bp)
4462 {
4463 	struct mtx *mtxp;
4464 
4465 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4466 	mtx_lock(mtxp);
4467 	if (--bp->b_pin_count == 0)
4468 		wakeup(bp);
4469 	mtx_unlock(mtxp);
4470 }
4471 
4472 void
4473 bunpin_wait(struct buf *bp)
4474 {
4475 	struct mtx *mtxp;
4476 
4477 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4478 	mtx_lock(mtxp);
4479 	while (bp->b_pin_count > 0)
4480 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4481 	mtx_unlock(mtxp);
4482 }
4483 
4484 /*
4485  * Set bio_data or bio_ma for struct bio from the struct buf.
4486  */
4487 void
4488 bdata2bio(struct buf *bp, struct bio *bip)
4489 {
4490 
4491 	if (!buf_mapped(bp)) {
4492 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4493 		bip->bio_ma = bp->b_pages;
4494 		bip->bio_ma_n = bp->b_npages;
4495 		bip->bio_data = unmapped_buf;
4496 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4497 		bip->bio_flags |= BIO_UNMAPPED;
4498 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4499 		    PAGE_SIZE == bp->b_npages,
4500 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4501 		    (long long)bip->bio_length, bip->bio_ma_n));
4502 	} else {
4503 		bip->bio_data = bp->b_data;
4504 		bip->bio_ma = NULL;
4505 	}
4506 }
4507 
4508 #include "opt_ddb.h"
4509 #ifdef DDB
4510 #include <ddb/ddb.h>
4511 
4512 /* DDB command to show buffer data */
4513 DB_SHOW_COMMAND(buffer, db_show_buffer)
4514 {
4515 	/* get args */
4516 	struct buf *bp = (struct buf *)addr;
4517 
4518 	if (!have_addr) {
4519 		db_printf("usage: show buffer <addr>\n");
4520 		return;
4521 	}
4522 
4523 	db_printf("buf at %p\n", bp);
4524 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4525 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4526 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4527 	db_printf(
4528 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4529 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4530 	    "b_dep = %p\n",
4531 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4532 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4533 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4534 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
4535 	    bp->b_kvabase, bp->b_kvasize);
4536 	if (bp->b_npages) {
4537 		int i;
4538 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4539 		for (i = 0; i < bp->b_npages; i++) {
4540 			vm_page_t m;
4541 			m = bp->b_pages[i];
4542 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4543 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4544 			if ((i + 1) < bp->b_npages)
4545 				db_printf(",");
4546 		}
4547 		db_printf("\n");
4548 	}
4549 	db_printf(" ");
4550 	BUF_LOCKPRINTINFO(bp);
4551 }
4552 
4553 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4554 {
4555 	struct buf *bp;
4556 	int i;
4557 
4558 	for (i = 0; i < nbuf; i++) {
4559 		bp = &buf[i];
4560 		if (BUF_ISLOCKED(bp)) {
4561 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4562 			db_printf("\n");
4563 		}
4564 	}
4565 }
4566 
4567 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4568 {
4569 	struct vnode *vp;
4570 	struct buf *bp;
4571 
4572 	if (!have_addr) {
4573 		db_printf("usage: show vnodebufs <addr>\n");
4574 		return;
4575 	}
4576 	vp = (struct vnode *)addr;
4577 	db_printf("Clean buffers:\n");
4578 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4579 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4580 		db_printf("\n");
4581 	}
4582 	db_printf("Dirty buffers:\n");
4583 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4584 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4585 		db_printf("\n");
4586 	}
4587 }
4588 
4589 DB_COMMAND(countfreebufs, db_coundfreebufs)
4590 {
4591 	struct buf *bp;
4592 	int i, used = 0, nfree = 0;
4593 
4594 	if (have_addr) {
4595 		db_printf("usage: countfreebufs\n");
4596 		return;
4597 	}
4598 
4599 	for (i = 0; i < nbuf; i++) {
4600 		bp = &buf[i];
4601 		if ((bp->b_flags & B_INFREECNT) != 0)
4602 			nfree++;
4603 		else
4604 			used++;
4605 	}
4606 
4607 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4608 	    nfree + used);
4609 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4610 }
4611 #endif /* DDB */
4612