xref: /freebsd/sys/kern/vfs_bio.c (revision 7f9dff23d3092aa33ad45b2b63e52469b3c13a6e)
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/racct.h>
65 #include <sys/resourcevar.h>
66 #include <sys/rwlock.h>
67 #include <sys/smp.h>
68 #include <sys/sysctl.h>
69 #include <sys/sysproto.h>
70 #include <sys/vmem.h>
71 #include <sys/vmmeter.h>
72 #include <sys/vnode.h>
73 #include <sys/watchdog.h>
74 #include <geom/geom.h>
75 #include <vm/vm.h>
76 #include <vm/vm_param.h>
77 #include <vm/vm_kern.h>
78 #include <vm/vm_object.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_pageout.h>
81 #include <vm/vm_pager.h>
82 #include <vm/vm_extern.h>
83 #include <vm/vm_map.h>
84 #include <vm/swap_pager.h>
85 #include "opt_compat.h"
86 #include "opt_swap.h"
87 
88 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
89 
90 struct	bio_ops bioops;		/* I/O operation notification */
91 
92 struct	buf_ops buf_ops_bio = {
93 	.bop_name	=	"buf_ops_bio",
94 	.bop_write	=	bufwrite,
95 	.bop_strategy	=	bufstrategy,
96 	.bop_sync	=	bufsync,
97 	.bop_bdflush	=	bufbdflush,
98 };
99 
100 static struct buf *buf;		/* buffer header pool */
101 extern struct buf *swbuf;	/* Swap buffer header pool. */
102 caddr_t unmapped_buf;
103 
104 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
105 struct proc *bufdaemonproc;
106 struct proc *bufspacedaemonproc;
107 
108 static int inmem(struct vnode *vp, daddr_t blkno);
109 static void vm_hold_free_pages(struct buf *bp, int newbsize);
110 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
111 		vm_offset_t to);
112 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
113 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
114 		vm_page_t m);
115 static void vfs_clean_pages_dirty_buf(struct buf *bp);
116 static void vfs_setdirty_locked_object(struct buf *bp);
117 static void vfs_vmio_invalidate(struct buf *bp);
118 static void vfs_vmio_truncate(struct buf *bp, int npages);
119 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
120 static int vfs_bio_clcheck(struct vnode *vp, int size,
121 		daddr_t lblkno, daddr_t blkno);
122 static int buf_flush(struct vnode *vp, int);
123 static int buf_recycle(bool);
124 static int buf_scan(bool);
125 static int flushbufqueues(struct vnode *, int, int);
126 static void buf_daemon(void);
127 static void bremfreel(struct buf *bp);
128 static __inline void bd_wakeup(void);
129 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
130 static void bufkva_reclaim(vmem_t *, int);
131 static void bufkva_free(struct buf *);
132 static int buf_import(void *, void **, int, int);
133 static void buf_release(void *, void **, int);
134 
135 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
136     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
137 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
138 #endif
139 
140 int vmiodirenable = TRUE;
141 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
142     "Use the VM system for directory writes");
143 long runningbufspace;
144 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
145     "Amount of presently outstanding async buffer io");
146 static long bufspace;
147 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
148     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
149 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
150     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
151 #else
152 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
153     "Physical memory used for buffers");
154 #endif
155 static long bufkvaspace;
156 SYSCTL_LONG(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 0,
157     "Kernel virtual memory used for buffers");
158 static long maxbufspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, &maxbufspace, 0,
160     "Maximum allowed value of bufspace (including metadata)");
161 static long bufmallocspace;
162 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
163     "Amount of malloced memory for buffers");
164 static long maxbufmallocspace;
165 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
166     0, "Maximum amount of malloced memory for buffers");
167 static long lobufspace;
168 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RW, &lobufspace, 0,
169     "Minimum amount of buffers we want to have");
170 long hibufspace;
171 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RW, &hibufspace, 0,
172     "Maximum allowed value of bufspace (excluding metadata)");
173 long bufspacethresh;
174 SYSCTL_LONG(_vfs, OID_AUTO, bufspacethresh, CTLFLAG_RW, &bufspacethresh,
175     0, "Bufspace consumed before waking the daemon to free some");
176 static int buffreekvacnt;
177 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
178     "Number of times we have freed the KVA space from some buffer");
179 static int bufdefragcnt;
180 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
181     "Number of times we have had to repeat buffer allocation to defragment");
182 static long lorunningspace;
183 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
184     CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
185     "Minimum preferred space used for in-progress I/O");
186 static long hirunningspace;
187 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
188     CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
189     "Maximum amount of space to use for in-progress I/O");
190 int dirtybufferflushes;
191 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
192     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
193 int bdwriteskip;
194 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
195     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
196 int altbufferflushes;
197 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
198     0, "Number of fsync flushes to limit dirty buffers");
199 static int recursiveflushes;
200 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
201     0, "Number of flushes skipped due to being recursive");
202 static int numdirtybuffers;
203 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
204     "Number of buffers that are dirty (has unwritten changes) at the moment");
205 static int lodirtybuffers;
206 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
207     "How many buffers we want to have free before bufdaemon can sleep");
208 static int hidirtybuffers;
209 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
210     "When the number of dirty buffers is considered severe");
211 int dirtybufthresh;
212 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
213     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
214 static int numfreebuffers;
215 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
216     "Number of free buffers");
217 static int lofreebuffers;
218 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
219    "Target number of free buffers");
220 static int hifreebuffers;
221 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
222    "Threshold for clean buffer recycling");
223 static int getnewbufcalls;
224 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
225    "Number of calls to getnewbuf");
226 static int getnewbufrestarts;
227 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
228     "Number of times getnewbuf has had to restart a buffer acquisition");
229 static int mappingrestarts;
230 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
231     "Number of times getblk has had to restart a buffer mapping for "
232     "unmapped buffer");
233 static int numbufallocfails;
234 SYSCTL_INT(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, 0,
235     "Number of times buffer allocations failed");
236 static int flushbufqtarget = 100;
237 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
238     "Amount of work to do in flushbufqueues when helping bufdaemon");
239 static long notbufdflushes;
240 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes, 0,
241     "Number of dirty buffer flushes done by the bufdaemon helpers");
242 static long barrierwrites;
243 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
244     "Number of barrier writes");
245 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
246     &unmapped_buf_allowed, 0,
247     "Permit the use of the unmapped i/o");
248 
249 /*
250  * This lock synchronizes access to bd_request.
251  */
252 static struct mtx_padalign bdlock;
253 
254 /*
255  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
256  * waitrunningbufspace().
257  */
258 static struct mtx_padalign rbreqlock;
259 
260 /*
261  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
262  */
263 static struct rwlock_padalign nblock;
264 
265 /*
266  * Lock that protects bdirtywait.
267  */
268 static struct mtx_padalign bdirtylock;
269 
270 /*
271  * Wakeup point for bufdaemon, as well as indicator of whether it is already
272  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
273  * is idling.
274  */
275 static int bd_request;
276 
277 /*
278  * Request/wakeup point for the bufspace daemon.
279  */
280 static int bufspace_request;
281 
282 /*
283  * Request for the buf daemon to write more buffers than is indicated by
284  * lodirtybuf.  This may be necessary to push out excess dependencies or
285  * defragment the address space where a simple count of the number of dirty
286  * buffers is insufficient to characterize the demand for flushing them.
287  */
288 static int bd_speedupreq;
289 
290 /*
291  * bogus page -- for I/O to/from partially complete buffers
292  * this is a temporary solution to the problem, but it is not
293  * really that bad.  it would be better to split the buffer
294  * for input in the case of buffers partially already in memory,
295  * but the code is intricate enough already.
296  */
297 vm_page_t bogus_page;
298 
299 /*
300  * Synchronization (sleep/wakeup) variable for active buffer space requests.
301  * Set when wait starts, cleared prior to wakeup().
302  * Used in runningbufwakeup() and waitrunningbufspace().
303  */
304 static int runningbufreq;
305 
306 /*
307  * Synchronization (sleep/wakeup) variable for buffer requests.
308  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
309  * by and/or.
310  * Used in numdirtywakeup(), bufspace_wakeup(), bwillwrite(),
311  * getnewbuf(), and getblk().
312  */
313 static volatile int needsbuffer;
314 
315 /*
316  * Synchronization for bwillwrite() waiters.
317  */
318 static int bdirtywait;
319 
320 /*
321  * Definitions for the buffer free lists.
322  */
323 #define QUEUE_NONE	0	/* on no queue */
324 #define QUEUE_EMPTY	1	/* empty buffer headers */
325 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
326 #define QUEUE_CLEAN	3	/* non-B_DELWRI buffers */
327 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
328 
329 /* Maximum number of clean buffer queues. */
330 #define	CLEAN_QUEUES	16
331 
332 /* Configured number of clean queues. */
333 static int clean_queues;
334 
335 /* Maximum number of buffer queues. */
336 #define BUFFER_QUEUES	(QUEUE_CLEAN + CLEAN_QUEUES)
337 
338 /* Queues for free buffers with various properties */
339 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
340 #ifdef INVARIANTS
341 static int bq_len[BUFFER_QUEUES];
342 #endif
343 
344 /*
345  * Lock for each bufqueue
346  */
347 static struct mtx_padalign bqlocks[BUFFER_QUEUES];
348 
349 /*
350  * per-cpu empty buffer cache.
351  */
352 uma_zone_t buf_zone;
353 
354 /*
355  * Single global constant for BUF_WMESG, to avoid getting multiple references.
356  * buf_wmesg is referred from macros.
357  */
358 const char *buf_wmesg = BUF_WMESG;
359 
360 static int
361 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
362 {
363 	long value;
364 	int error;
365 
366 	value = *(long *)arg1;
367 	error = sysctl_handle_long(oidp, &value, 0, req);
368 	if (error != 0 || req->newptr == NULL)
369 		return (error);
370 	mtx_lock(&rbreqlock);
371 	if (arg1 == &hirunningspace) {
372 		if (value < lorunningspace)
373 			error = EINVAL;
374 		else
375 			hirunningspace = value;
376 	} else {
377 		KASSERT(arg1 == &lorunningspace,
378 		    ("%s: unknown arg1", __func__));
379 		if (value > hirunningspace)
380 			error = EINVAL;
381 		else
382 			lorunningspace = value;
383 	}
384 	mtx_unlock(&rbreqlock);
385 	return (error);
386 }
387 
388 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
389     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
390 static int
391 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
392 {
393 	long lvalue;
394 	int ivalue;
395 
396 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
397 		return (sysctl_handle_long(oidp, arg1, arg2, req));
398 	lvalue = *(long *)arg1;
399 	if (lvalue > INT_MAX)
400 		/* On overflow, still write out a long to trigger ENOMEM. */
401 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
402 	ivalue = lvalue;
403 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
404 }
405 #endif
406 
407 static int
408 bqcleanq(void)
409 {
410 	static int nextq;
411 
412 	return ((atomic_fetchadd_int(&nextq, 1) % clean_queues) + QUEUE_CLEAN);
413 }
414 
415 static int
416 bqisclean(int qindex)
417 {
418 
419 	return (qindex >= QUEUE_CLEAN && qindex < QUEUE_CLEAN + CLEAN_QUEUES);
420 }
421 
422 /*
423  *	bqlock:
424  *
425  *	Return the appropriate queue lock based on the index.
426  */
427 static inline struct mtx *
428 bqlock(int qindex)
429 {
430 
431 	return (struct mtx *)&bqlocks[qindex];
432 }
433 
434 /*
435  *	bdirtywakeup:
436  *
437  *	Wakeup any bwillwrite() waiters.
438  */
439 static void
440 bdirtywakeup(void)
441 {
442 	mtx_lock(&bdirtylock);
443 	if (bdirtywait) {
444 		bdirtywait = 0;
445 		wakeup(&bdirtywait);
446 	}
447 	mtx_unlock(&bdirtylock);
448 }
449 
450 /*
451  *	bdirtysub:
452  *
453  *	Decrement the numdirtybuffers count by one and wakeup any
454  *	threads blocked in bwillwrite().
455  */
456 static void
457 bdirtysub(void)
458 {
459 
460 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
461 	    (lodirtybuffers + hidirtybuffers) / 2)
462 		bdirtywakeup();
463 }
464 
465 /*
466  *	bdirtyadd:
467  *
468  *	Increment the numdirtybuffers count by one and wakeup the buf
469  *	daemon if needed.
470  */
471 static void
472 bdirtyadd(void)
473 {
474 
475 	/*
476 	 * Only do the wakeup once as we cross the boundary.  The
477 	 * buf daemon will keep running until the condition clears.
478 	 */
479 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
480 	    (lodirtybuffers + hidirtybuffers) / 2)
481 		bd_wakeup();
482 }
483 
484 /*
485  *	bufspace_wakeup:
486  *
487  *	Called when buffer space is potentially available for recovery.
488  *	getnewbuf() will block on this flag when it is unable to free
489  *	sufficient buffer space.  Buffer space becomes recoverable when
490  *	bp's get placed back in the queues.
491  */
492 static void
493 bufspace_wakeup(void)
494 {
495 
496 	/*
497 	 * If someone is waiting for bufspace, wake them up.
498 	 *
499 	 * Since needsbuffer is set prior to doing an additional queue
500 	 * scan it is safe to check for the flag prior to acquiring the
501 	 * lock.  The thread that is preparing to scan again before
502 	 * blocking would discover the buf we released.
503 	 */
504 	if (needsbuffer) {
505 		rw_rlock(&nblock);
506 		if (atomic_cmpset_int(&needsbuffer, 1, 0) == 1)
507 			wakeup(__DEVOLATILE(void *, &needsbuffer));
508 		rw_runlock(&nblock);
509 	}
510 }
511 
512 /*
513  *	bufspace_daemonwakeup:
514  *
515  *	Wakeup the daemon responsible for freeing clean bufs.
516  */
517 static void
518 bufspace_daemonwakeup(void)
519 {
520 	rw_rlock(&nblock);
521 	if (bufspace_request == 0) {
522 		bufspace_request = 1;
523 		wakeup(&bufspace_request);
524 	}
525 	rw_runlock(&nblock);
526 }
527 
528 /*
529  *	bufspace_adjust:
530  *
531  *	Adjust the reported bufspace for a KVA managed buffer, possibly
532  * 	waking any waiters.
533  */
534 static void
535 bufspace_adjust(struct buf *bp, int bufsize)
536 {
537 	long space;
538 	int diff;
539 
540 	KASSERT((bp->b_flags & B_MALLOC) == 0,
541 	    ("bufspace_adjust: malloc buf %p", bp));
542 	diff = bufsize - bp->b_bufsize;
543 	if (diff < 0) {
544 		atomic_subtract_long(&bufspace, -diff);
545 		bufspace_wakeup();
546 	} else {
547 		space = atomic_fetchadd_long(&bufspace, diff);
548 		/* Wake up the daemon on the transition. */
549 		if (space < bufspacethresh && space + diff >= bufspacethresh)
550 			bufspace_daemonwakeup();
551 	}
552 	bp->b_bufsize = bufsize;
553 }
554 
555 /*
556  *	bufspace_reserve:
557  *
558  *	Reserve bufspace before calling allocbuf().  metadata has a
559  *	different space limit than data.
560  */
561 static int
562 bufspace_reserve(int size, bool metadata)
563 {
564 	long limit;
565 	long space;
566 
567 	if (metadata)
568 		limit = maxbufspace;
569 	else
570 		limit = hibufspace;
571 	do {
572 		space = bufspace;
573 		if (space + size > limit)
574 			return (ENOSPC);
575 	} while (atomic_cmpset_long(&bufspace, space, space + size) == 0);
576 
577 	/* Wake up the daemon on the transition. */
578 	if (space < bufspacethresh && space + size >= bufspacethresh)
579 		bufspace_daemonwakeup();
580 
581 	return (0);
582 }
583 
584 /*
585  *	bufspace_release:
586  *
587  *	Release reserved bufspace after bufspace_adjust() has consumed it.
588  */
589 static void
590 bufspace_release(int size)
591 {
592 	atomic_subtract_long(&bufspace, size);
593 	bufspace_wakeup();
594 }
595 
596 /*
597  *	bufspace_wait:
598  *
599  *	Wait for bufspace, acting as the buf daemon if a locked vnode is
600  *	supplied.  needsbuffer must be set in a safe fashion prior to
601  *	polling for space.  The operation must be re-tried on return.
602  */
603 static void
604 bufspace_wait(struct vnode *vp, int gbflags, int slpflag, int slptimeo)
605 {
606 	struct thread *td;
607 	int error, fl, norunbuf;
608 
609 	if ((gbflags & GB_NOWAIT_BD) != 0)
610 		return;
611 
612 	td = curthread;
613 	rw_wlock(&nblock);
614 	while (needsbuffer != 0) {
615 		if (vp != NULL && vp->v_type != VCHR &&
616 		    (td->td_pflags & TDP_BUFNEED) == 0) {
617 			rw_wunlock(&nblock);
618 			/*
619 			 * getblk() is called with a vnode locked, and
620 			 * some majority of the dirty buffers may as
621 			 * well belong to the vnode.  Flushing the
622 			 * buffers there would make a progress that
623 			 * cannot be achieved by the buf_daemon, that
624 			 * cannot lock the vnode.
625 			 */
626 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
627 			    (td->td_pflags & TDP_NORUNNINGBUF);
628 
629 			/*
630 			 * Play bufdaemon.  The getnewbuf() function
631 			 * may be called while the thread owns lock
632 			 * for another dirty buffer for the same
633 			 * vnode, which makes it impossible to use
634 			 * VOP_FSYNC() there, due to the buffer lock
635 			 * recursion.
636 			 */
637 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
638 			fl = buf_flush(vp, flushbufqtarget);
639 			td->td_pflags &= norunbuf;
640 			rw_wlock(&nblock);
641 			if (fl != 0)
642 				continue;
643 			if (needsbuffer == 0)
644 				break;
645 		}
646 		error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
647 		    (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
648 		if (error != 0)
649 			break;
650 	}
651 	rw_wunlock(&nblock);
652 }
653 
654 
655 /*
656  *	bufspace_daemon:
657  *
658  *	buffer space management daemon.  Tries to maintain some marginal
659  *	amount of free buffer space so that requesting processes neither
660  *	block nor work to reclaim buffers.
661  */
662 static void
663 bufspace_daemon(void)
664 {
665 	for (;;) {
666 		kproc_suspend_check(bufspacedaemonproc);
667 
668 		/*
669 		 * Free buffers from the clean queue until we meet our
670 		 * targets.
671 		 *
672 		 * Theory of operation:  The buffer cache is most efficient
673 		 * when some free buffer headers and space are always
674 		 * available to getnewbuf().  This daemon attempts to prevent
675 		 * the excessive blocking and synchronization associated
676 		 * with shortfall.  It goes through three phases according
677 		 * demand:
678 		 *
679 		 * 1)	The daemon wakes up voluntarily once per-second
680 		 *	during idle periods when the counters are below
681 		 *	the wakeup thresholds (bufspacethresh, lofreebuffers).
682 		 *
683 		 * 2)	The daemon wakes up as we cross the thresholds
684 		 *	ahead of any potential blocking.  This may bounce
685 		 *	slightly according to the rate of consumption and
686 		 *	release.
687 		 *
688 		 * 3)	The daemon and consumers are starved for working
689 		 *	clean buffers.  This is the 'bufspace' sleep below
690 		 *	which will inefficiently trade bufs with bqrelse
691 		 *	until we return to condition 2.
692 		 */
693 		while (bufspace > lobufspace ||
694 		    numfreebuffers < hifreebuffers) {
695 			if (buf_recycle(false) != 0) {
696 				atomic_set_int(&needsbuffer, 1);
697 				if (buf_recycle(false) != 0) {
698 					rw_wlock(&nblock);
699 					if (needsbuffer)
700 						rw_sleep(__DEVOLATILE(void *,
701 						    &needsbuffer), &nblock,
702 						    PRIBIO|PDROP, "bufspace",
703 						    hz/10);
704 					else
705 						rw_wunlock(&nblock);
706 				}
707 			}
708 			maybe_yield();
709 		}
710 
711 		/*
712 		 * Re-check our limits under the exclusive nblock.
713 		 */
714 		rw_wlock(&nblock);
715 		if (bufspace < bufspacethresh &&
716 		    numfreebuffers > lofreebuffers) {
717 			bufspace_request = 0;
718 			rw_sleep(&bufspace_request, &nblock, PRIBIO|PDROP,
719 			    "-", hz);
720 		} else
721 			rw_wunlock(&nblock);
722 	}
723 }
724 
725 static struct kproc_desc bufspace_kp = {
726 	"bufspacedaemon",
727 	bufspace_daemon,
728 	&bufspacedaemonproc
729 };
730 SYSINIT(bufspacedaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start,
731     &bufspace_kp);
732 
733 /*
734  *	bufmallocadjust:
735  *
736  *	Adjust the reported bufspace for a malloc managed buffer, possibly
737  *	waking any waiters.
738  */
739 static void
740 bufmallocadjust(struct buf *bp, int bufsize)
741 {
742 	int diff;
743 
744 	KASSERT((bp->b_flags & B_MALLOC) != 0,
745 	    ("bufmallocadjust: non-malloc buf %p", bp));
746 	diff = bufsize - bp->b_bufsize;
747 	if (diff < 0)
748 		atomic_subtract_long(&bufmallocspace, -diff);
749 	else
750 		atomic_add_long(&bufmallocspace, diff);
751 	bp->b_bufsize = bufsize;
752 }
753 
754 /*
755  *	runningwakeup:
756  *
757  *	Wake up processes that are waiting on asynchronous writes to fall
758  *	below lorunningspace.
759  */
760 static void
761 runningwakeup(void)
762 {
763 
764 	mtx_lock(&rbreqlock);
765 	if (runningbufreq) {
766 		runningbufreq = 0;
767 		wakeup(&runningbufreq);
768 	}
769 	mtx_unlock(&rbreqlock);
770 }
771 
772 /*
773  *	runningbufwakeup:
774  *
775  *	Decrement the outstanding write count according.
776  */
777 void
778 runningbufwakeup(struct buf *bp)
779 {
780 	long space, bspace;
781 
782 	bspace = bp->b_runningbufspace;
783 	if (bspace == 0)
784 		return;
785 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
786 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
787 	    space, bspace));
788 	bp->b_runningbufspace = 0;
789 	/*
790 	 * Only acquire the lock and wakeup on the transition from exceeding
791 	 * the threshold to falling below it.
792 	 */
793 	if (space < lorunningspace)
794 		return;
795 	if (space - bspace > lorunningspace)
796 		return;
797 	runningwakeup();
798 }
799 
800 /*
801  *	waitrunningbufspace()
802  *
803  *	runningbufspace is a measure of the amount of I/O currently
804  *	running.  This routine is used in async-write situations to
805  *	prevent creating huge backups of pending writes to a device.
806  *	Only asynchronous writes are governed by this function.
807  *
808  *	This does NOT turn an async write into a sync write.  It waits
809  *	for earlier writes to complete and generally returns before the
810  *	caller's write has reached the device.
811  */
812 void
813 waitrunningbufspace(void)
814 {
815 
816 	mtx_lock(&rbreqlock);
817 	while (runningbufspace > hirunningspace) {
818 		runningbufreq = 1;
819 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
820 	}
821 	mtx_unlock(&rbreqlock);
822 }
823 
824 
825 /*
826  *	vfs_buf_test_cache:
827  *
828  *	Called when a buffer is extended.  This function clears the B_CACHE
829  *	bit if the newly extended portion of the buffer does not contain
830  *	valid data.
831  */
832 static __inline void
833 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
834     vm_offset_t size, vm_page_t m)
835 {
836 
837 	VM_OBJECT_ASSERT_LOCKED(m->object);
838 	if (bp->b_flags & B_CACHE) {
839 		int base = (foff + off) & PAGE_MASK;
840 		if (vm_page_is_valid(m, base, size) == 0)
841 			bp->b_flags &= ~B_CACHE;
842 	}
843 }
844 
845 /* Wake up the buffer daemon if necessary */
846 static __inline void
847 bd_wakeup(void)
848 {
849 
850 	mtx_lock(&bdlock);
851 	if (bd_request == 0) {
852 		bd_request = 1;
853 		wakeup(&bd_request);
854 	}
855 	mtx_unlock(&bdlock);
856 }
857 
858 /*
859  * bd_speedup - speedup the buffer cache flushing code
860  */
861 void
862 bd_speedup(void)
863 {
864 	int needwake;
865 
866 	mtx_lock(&bdlock);
867 	needwake = 0;
868 	if (bd_speedupreq == 0 || bd_request == 0)
869 		needwake = 1;
870 	bd_speedupreq = 1;
871 	bd_request = 1;
872 	if (needwake)
873 		wakeup(&bd_request);
874 	mtx_unlock(&bdlock);
875 }
876 
877 #ifndef NSWBUF_MIN
878 #define	NSWBUF_MIN	16
879 #endif
880 
881 #ifdef __i386__
882 #define	TRANSIENT_DENOM	5
883 #else
884 #define	TRANSIENT_DENOM 10
885 #endif
886 
887 /*
888  * Calculating buffer cache scaling values and reserve space for buffer
889  * headers.  This is called during low level kernel initialization and
890  * may be called more then once.  We CANNOT write to the memory area
891  * being reserved at this time.
892  */
893 caddr_t
894 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
895 {
896 	int tuned_nbuf;
897 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
898 
899 	/*
900 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
901 	 * PAGE_SIZE is >= 1K)
902 	 */
903 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
904 
905 	/*
906 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
907 	 * For the first 64MB of ram nominally allocate sufficient buffers to
908 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
909 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
910 	 * the buffer cache we limit the eventual kva reservation to
911 	 * maxbcache bytes.
912 	 *
913 	 * factor represents the 1/4 x ram conversion.
914 	 */
915 	if (nbuf == 0) {
916 		int factor = 4 * BKVASIZE / 1024;
917 
918 		nbuf = 50;
919 		if (physmem_est > 4096)
920 			nbuf += min((physmem_est - 4096) / factor,
921 			    65536 / factor);
922 		if (physmem_est > 65536)
923 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
924 			    32 * 1024 * 1024 / (factor * 5));
925 
926 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
927 			nbuf = maxbcache / BKVASIZE;
928 		tuned_nbuf = 1;
929 	} else
930 		tuned_nbuf = 0;
931 
932 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
933 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
934 	if (nbuf > maxbuf) {
935 		if (!tuned_nbuf)
936 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
937 			    maxbuf);
938 		nbuf = maxbuf;
939 	}
940 
941 	/*
942 	 * Ideal allocation size for the transient bio submap is 10%
943 	 * of the maximal space buffer map.  This roughly corresponds
944 	 * to the amount of the buffer mapped for typical UFS load.
945 	 *
946 	 * Clip the buffer map to reserve space for the transient
947 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
948 	 * maximum buffer map extent on the platform.
949 	 *
950 	 * The fall-back to the maxbuf in case of maxbcache unset,
951 	 * allows to not trim the buffer KVA for the architectures
952 	 * with ample KVA space.
953 	 */
954 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
955 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
956 		buf_sz = (long)nbuf * BKVASIZE;
957 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
958 		    (TRANSIENT_DENOM - 1)) {
959 			/*
960 			 * There is more KVA than memory.  Do not
961 			 * adjust buffer map size, and assign the rest
962 			 * of maxbuf to transient map.
963 			 */
964 			biotmap_sz = maxbuf_sz - buf_sz;
965 		} else {
966 			/*
967 			 * Buffer map spans all KVA we could afford on
968 			 * this platform.  Give 10% (20% on i386) of
969 			 * the buffer map to the transient bio map.
970 			 */
971 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
972 			buf_sz -= biotmap_sz;
973 		}
974 		if (biotmap_sz / INT_MAX > MAXPHYS)
975 			bio_transient_maxcnt = INT_MAX;
976 		else
977 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
978 		/*
979 		 * Artificially limit to 1024 simultaneous in-flight I/Os
980 		 * using the transient mapping.
981 		 */
982 		if (bio_transient_maxcnt > 1024)
983 			bio_transient_maxcnt = 1024;
984 		if (tuned_nbuf)
985 			nbuf = buf_sz / BKVASIZE;
986 	}
987 
988 	/*
989 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
990 	 * We have no less then 16 and no more then 256.
991 	 */
992 	nswbuf = min(nbuf / 4, 256);
993 	TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
994 	if (nswbuf < NSWBUF_MIN)
995 		nswbuf = NSWBUF_MIN;
996 
997 	/*
998 	 * Reserve space for the buffer cache buffers
999 	 */
1000 	swbuf = (void *)v;
1001 	v = (caddr_t)(swbuf + nswbuf);
1002 	buf = (void *)v;
1003 	v = (caddr_t)(buf + nbuf);
1004 
1005 	return(v);
1006 }
1007 
1008 /* Initialize the buffer subsystem.  Called before use of any buffers. */
1009 void
1010 bufinit(void)
1011 {
1012 	struct buf *bp;
1013 	int i;
1014 
1015 	CTASSERT(MAXBCACHEBUF >= MAXBSIZE);
1016 	mtx_init(&bqlocks[QUEUE_DIRTY], "bufq dirty lock", NULL, MTX_DEF);
1017 	mtx_init(&bqlocks[QUEUE_EMPTY], "bufq empty lock", NULL, MTX_DEF);
1018 	for (i = QUEUE_CLEAN; i < QUEUE_CLEAN + CLEAN_QUEUES; i++)
1019 		mtx_init(&bqlocks[i], "bufq clean lock", NULL, MTX_DEF);
1020 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1021 	rw_init(&nblock, "needsbuffer lock");
1022 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1023 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1024 
1025 	/* next, make a null set of free lists */
1026 	for (i = 0; i < BUFFER_QUEUES; i++)
1027 		TAILQ_INIT(&bufqueues[i]);
1028 
1029 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1030 
1031 	/* finally, initialize each buffer header and stick on empty q */
1032 	for (i = 0; i < nbuf; i++) {
1033 		bp = &buf[i];
1034 		bzero(bp, sizeof *bp);
1035 		bp->b_flags = B_INVAL;
1036 		bp->b_rcred = NOCRED;
1037 		bp->b_wcred = NOCRED;
1038 		bp->b_qindex = QUEUE_EMPTY;
1039 		bp->b_xflags = 0;
1040 		bp->b_data = bp->b_kvabase = unmapped_buf;
1041 		LIST_INIT(&bp->b_dep);
1042 		BUF_LOCKINIT(bp);
1043 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
1044 #ifdef INVARIANTS
1045 		bq_len[QUEUE_EMPTY]++;
1046 #endif
1047 	}
1048 
1049 	/*
1050 	 * maxbufspace is the absolute maximum amount of buffer space we are
1051 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
1052 	 * is nominally used by metadata.  hibufspace is the nominal maximum
1053 	 * used by most other requests.  The differential is required to
1054 	 * ensure that metadata deadlocks don't occur.
1055 	 *
1056 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
1057 	 * this may result in KVM fragmentation which is not handled optimally
1058 	 * by the system. XXX This is less true with vmem.  We could use
1059 	 * PAGE_SIZE.
1060 	 */
1061 	maxbufspace = (long)nbuf * BKVASIZE;
1062 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10);
1063 	lobufspace = (hibufspace / 20) * 19; /* 95% */
1064 	bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1065 
1066 	/*
1067 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
1068 	 * arbitrarily and may need further tuning. It corresponds to
1069 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
1070 	 * which fits with many RAID controllers' tagged queuing limits.
1071 	 * The lower 1 MiB limit is the historical upper limit for
1072 	 * hirunningspace.
1073 	 */
1074 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF),
1075 	    16 * 1024 * 1024), 1024 * 1024);
1076 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF);
1077 
1078 	/*
1079 	 * Limit the amount of malloc memory since it is wired permanently into
1080 	 * the kernel space.  Even though this is accounted for in the buffer
1081 	 * allocation, we don't want the malloced region to grow uncontrolled.
1082 	 * The malloc scheme improves memory utilization significantly on
1083 	 * average (small) directories.
1084 	 */
1085 	maxbufmallocspace = hibufspace / 20;
1086 
1087 	/*
1088 	 * Reduce the chance of a deadlock occurring by limiting the number
1089 	 * of delayed-write dirty buffers we allow to stack up.
1090 	 */
1091 	hidirtybuffers = nbuf / 4 + 20;
1092 	dirtybufthresh = hidirtybuffers * 9 / 10;
1093 	numdirtybuffers = 0;
1094 	/*
1095 	 * To support extreme low-memory systems, make sure hidirtybuffers
1096 	 * cannot eat up all available buffer space.  This occurs when our
1097 	 * minimum cannot be met.  We try to size hidirtybuffers to 3/4 our
1098 	 * buffer space assuming BKVASIZE'd buffers.
1099 	 */
1100 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1101 		hidirtybuffers >>= 1;
1102 	}
1103 	lodirtybuffers = hidirtybuffers / 2;
1104 
1105 	/*
1106 	 * lofreebuffers should be sufficient to avoid stalling waiting on
1107 	 * buf headers under heavy utilization.  The bufs in per-cpu caches
1108 	 * are counted as free but will be unavailable to threads executing
1109 	 * on other cpus.
1110 	 *
1111 	 * hifreebuffers is the free target for the bufspace daemon.  This
1112 	 * should be set appropriately to limit work per-iteration.
1113 	 */
1114 	lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1115 	hifreebuffers = (3 * lofreebuffers) / 2;
1116 	numfreebuffers = nbuf;
1117 
1118 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
1119 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
1120 
1121 	/* Setup the kva and free list allocators. */
1122 	vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1123 	buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1124 	    NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1125 
1126 	/*
1127 	 * Size the clean queue according to the amount of buffer space.
1128 	 * One queue per-256mb up to the max.  More queues gives better
1129 	 * concurrency but less accurate LRU.
1130 	 */
1131 	clean_queues = MIN(howmany(maxbufspace, 256*1024*1024), CLEAN_QUEUES);
1132 
1133 }
1134 
1135 #ifdef INVARIANTS
1136 static inline void
1137 vfs_buf_check_mapped(struct buf *bp)
1138 {
1139 
1140 	KASSERT(bp->b_kvabase != unmapped_buf,
1141 	    ("mapped buf: b_kvabase was not updated %p", bp));
1142 	KASSERT(bp->b_data != unmapped_buf,
1143 	    ("mapped buf: b_data was not updated %p", bp));
1144 	KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1145 	    MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1146 }
1147 
1148 static inline void
1149 vfs_buf_check_unmapped(struct buf *bp)
1150 {
1151 
1152 	KASSERT(bp->b_data == unmapped_buf,
1153 	    ("unmapped buf: corrupted b_data %p", bp));
1154 }
1155 
1156 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1157 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1158 #else
1159 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
1160 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
1161 #endif
1162 
1163 static int
1164 isbufbusy(struct buf *bp)
1165 {
1166 	if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1167 	    ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1168 		return (1);
1169 	return (0);
1170 }
1171 
1172 /*
1173  * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1174  */
1175 void
1176 bufshutdown(int show_busybufs)
1177 {
1178 	static int first_buf_printf = 1;
1179 	struct buf *bp;
1180 	int iter, nbusy, pbusy;
1181 #ifndef PREEMPTION
1182 	int subiter;
1183 #endif
1184 
1185 	/*
1186 	 * Sync filesystems for shutdown
1187 	 */
1188 	wdog_kern_pat(WD_LASTVAL);
1189 	sys_sync(curthread, NULL);
1190 
1191 	/*
1192 	 * With soft updates, some buffers that are
1193 	 * written will be remarked as dirty until other
1194 	 * buffers are written.
1195 	 */
1196 	for (iter = pbusy = 0; iter < 20; iter++) {
1197 		nbusy = 0;
1198 		for (bp = &buf[nbuf]; --bp >= buf; )
1199 			if (isbufbusy(bp))
1200 				nbusy++;
1201 		if (nbusy == 0) {
1202 			if (first_buf_printf)
1203 				printf("All buffers synced.");
1204 			break;
1205 		}
1206 		if (first_buf_printf) {
1207 			printf("Syncing disks, buffers remaining... ");
1208 			first_buf_printf = 0;
1209 		}
1210 		printf("%d ", nbusy);
1211 		if (nbusy < pbusy)
1212 			iter = 0;
1213 		pbusy = nbusy;
1214 
1215 		wdog_kern_pat(WD_LASTVAL);
1216 		sys_sync(curthread, NULL);
1217 
1218 #ifdef PREEMPTION
1219 		/*
1220 		 * Drop Giant and spin for a while to allow
1221 		 * interrupt threads to run.
1222 		 */
1223 		DROP_GIANT();
1224 		DELAY(50000 * iter);
1225 		PICKUP_GIANT();
1226 #else
1227 		/*
1228 		 * Drop Giant and context switch several times to
1229 		 * allow interrupt threads to run.
1230 		 */
1231 		DROP_GIANT();
1232 		for (subiter = 0; subiter < 50 * iter; subiter++) {
1233 			thread_lock(curthread);
1234 			mi_switch(SW_VOL, NULL);
1235 			thread_unlock(curthread);
1236 			DELAY(1000);
1237 		}
1238 		PICKUP_GIANT();
1239 #endif
1240 	}
1241 	printf("\n");
1242 	/*
1243 	 * Count only busy local buffers to prevent forcing
1244 	 * a fsck if we're just a client of a wedged NFS server
1245 	 */
1246 	nbusy = 0;
1247 	for (bp = &buf[nbuf]; --bp >= buf; ) {
1248 		if (isbufbusy(bp)) {
1249 #if 0
1250 /* XXX: This is bogus.  We should probably have a BO_REMOTE flag instead */
1251 			if (bp->b_dev == NULL) {
1252 				TAILQ_REMOVE(&mountlist,
1253 				    bp->b_vp->v_mount, mnt_list);
1254 				continue;
1255 			}
1256 #endif
1257 			nbusy++;
1258 			if (show_busybufs > 0) {
1259 				printf(
1260 	    "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1261 				    nbusy, bp, bp->b_vp, bp->b_flags,
1262 				    (intmax_t)bp->b_blkno,
1263 				    (intmax_t)bp->b_lblkno);
1264 				BUF_LOCKPRINTINFO(bp);
1265 				if (show_busybufs > 1)
1266 					vn_printf(bp->b_vp,
1267 					    "vnode content: ");
1268 			}
1269 		}
1270 	}
1271 	if (nbusy) {
1272 		/*
1273 		 * Failed to sync all blocks. Indicate this and don't
1274 		 * unmount filesystems (thus forcing an fsck on reboot).
1275 		 */
1276 		printf("Giving up on %d buffers\n", nbusy);
1277 		DELAY(5000000);	/* 5 seconds */
1278 	} else {
1279 		if (!first_buf_printf)
1280 			printf("Final sync complete\n");
1281 		/*
1282 		 * Unmount filesystems
1283 		 */
1284 		if (panicstr == NULL)
1285 			vfs_unmountall();
1286 	}
1287 	swapoff_all();
1288 	DELAY(100000);		/* wait for console output to finish */
1289 }
1290 
1291 static void
1292 bpmap_qenter(struct buf *bp)
1293 {
1294 
1295 	BUF_CHECK_MAPPED(bp);
1296 
1297 	/*
1298 	 * bp->b_data is relative to bp->b_offset, but
1299 	 * bp->b_offset may be offset into the first page.
1300 	 */
1301 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1302 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1303 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1304 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
1305 }
1306 
1307 /*
1308  *	binsfree:
1309  *
1310  *	Insert the buffer into the appropriate free list.
1311  */
1312 static void
1313 binsfree(struct buf *bp, int qindex)
1314 {
1315 	struct mtx *olock, *nlock;
1316 
1317 	if (qindex != QUEUE_EMPTY) {
1318 		BUF_ASSERT_XLOCKED(bp);
1319 	}
1320 
1321 	/*
1322 	 * Stick to the same clean queue for the lifetime of the buf to
1323 	 * limit locking below.  Otherwise pick ont sequentially.
1324 	 */
1325 	if (qindex == QUEUE_CLEAN) {
1326 		if (bqisclean(bp->b_qindex))
1327 			qindex = bp->b_qindex;
1328 		else
1329 			qindex = bqcleanq();
1330 	}
1331 
1332 	/*
1333 	 * Handle delayed bremfree() processing.
1334 	 */
1335 	nlock = bqlock(qindex);
1336 	if (bp->b_flags & B_REMFREE) {
1337 		olock = bqlock(bp->b_qindex);
1338 		mtx_lock(olock);
1339 		bremfreel(bp);
1340 		if (olock != nlock) {
1341 			mtx_unlock(olock);
1342 			mtx_lock(nlock);
1343 		}
1344 	} else
1345 		mtx_lock(nlock);
1346 
1347 	if (bp->b_qindex != QUEUE_NONE)
1348 		panic("binsfree: free buffer onto another queue???");
1349 
1350 	bp->b_qindex = qindex;
1351 	if (bp->b_flags & B_AGE)
1352 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1353 	else
1354 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1355 #ifdef INVARIANTS
1356 	bq_len[bp->b_qindex]++;
1357 #endif
1358 	mtx_unlock(nlock);
1359 }
1360 
1361 /*
1362  * buf_free:
1363  *
1364  *	Free a buffer to the buf zone once it no longer has valid contents.
1365  */
1366 static void
1367 buf_free(struct buf *bp)
1368 {
1369 
1370 	if (bp->b_flags & B_REMFREE)
1371 		bremfreef(bp);
1372 	if (bp->b_vflags & BV_BKGRDINPROG)
1373 		panic("losing buffer 1");
1374 	if (bp->b_rcred != NOCRED) {
1375 		crfree(bp->b_rcred);
1376 		bp->b_rcred = NOCRED;
1377 	}
1378 	if (bp->b_wcred != NOCRED) {
1379 		crfree(bp->b_wcred);
1380 		bp->b_wcred = NOCRED;
1381 	}
1382 	if (!LIST_EMPTY(&bp->b_dep))
1383 		buf_deallocate(bp);
1384 	bufkva_free(bp);
1385 	BUF_UNLOCK(bp);
1386 	uma_zfree(buf_zone, bp);
1387 	atomic_add_int(&numfreebuffers, 1);
1388 	bufspace_wakeup();
1389 }
1390 
1391 /*
1392  * buf_import:
1393  *
1394  *	Import bufs into the uma cache from the buf list.  The system still
1395  *	expects a static array of bufs and much of the synchronization
1396  *	around bufs assumes type stable storage.  As a result, UMA is used
1397  *	only as a per-cpu cache of bufs still maintained on a global list.
1398  */
1399 static int
1400 buf_import(void *arg, void **store, int cnt, int flags)
1401 {
1402 	struct buf *bp;
1403 	int i;
1404 
1405 	mtx_lock(&bqlocks[QUEUE_EMPTY]);
1406 	for (i = 0; i < cnt; i++) {
1407 		bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1408 		if (bp == NULL)
1409 			break;
1410 		bremfreel(bp);
1411 		store[i] = bp;
1412 	}
1413 	mtx_unlock(&bqlocks[QUEUE_EMPTY]);
1414 
1415 	return (i);
1416 }
1417 
1418 /*
1419  * buf_release:
1420  *
1421  *	Release bufs from the uma cache back to the buffer queues.
1422  */
1423 static void
1424 buf_release(void *arg, void **store, int cnt)
1425 {
1426         int i;
1427 
1428         for (i = 0; i < cnt; i++)
1429 		binsfree(store[i], QUEUE_EMPTY);
1430 }
1431 
1432 /*
1433  * buf_alloc:
1434  *
1435  *	Allocate an empty buffer header.
1436  */
1437 static struct buf *
1438 buf_alloc(void)
1439 {
1440 	struct buf *bp;
1441 
1442 	bp = uma_zalloc(buf_zone, M_NOWAIT);
1443 	if (bp == NULL) {
1444 		bufspace_daemonwakeup();
1445 		atomic_add_int(&numbufallocfails, 1);
1446 		return (NULL);
1447 	}
1448 
1449 	/*
1450 	 * Wake-up the bufspace daemon on transition.
1451 	 */
1452 	if (atomic_fetchadd_int(&numfreebuffers, -1) == lofreebuffers)
1453 		bufspace_daemonwakeup();
1454 
1455 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1456 		panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1457 
1458 	KASSERT(bp->b_vp == NULL,
1459 	    ("bp: %p still has vnode %p.", bp, bp->b_vp));
1460 	KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1461 	    ("invalid buffer %p flags %#x", bp, bp->b_flags));
1462 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1463 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1464 	KASSERT(bp->b_npages == 0,
1465 	    ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1466 	KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1467 	KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1468 
1469 	bp->b_flags = 0;
1470 	bp->b_ioflags = 0;
1471 	bp->b_xflags = 0;
1472 	bp->b_vflags = 0;
1473 	bp->b_vp = NULL;
1474 	bp->b_blkno = bp->b_lblkno = 0;
1475 	bp->b_offset = NOOFFSET;
1476 	bp->b_iodone = 0;
1477 	bp->b_error = 0;
1478 	bp->b_resid = 0;
1479 	bp->b_bcount = 0;
1480 	bp->b_npages = 0;
1481 	bp->b_dirtyoff = bp->b_dirtyend = 0;
1482 	bp->b_bufobj = NULL;
1483 	bp->b_data = bp->b_kvabase = unmapped_buf;
1484 	bp->b_fsprivate1 = NULL;
1485 	bp->b_fsprivate2 = NULL;
1486 	bp->b_fsprivate3 = NULL;
1487 	LIST_INIT(&bp->b_dep);
1488 
1489 	return (bp);
1490 }
1491 
1492 /*
1493  *	buf_qrecycle:
1494  *
1495  *	Free a buffer from the given bufqueue.  kva controls whether the
1496  *	freed buf must own some kva resources.  This is used for
1497  *	defragmenting.
1498  */
1499 static int
1500 buf_qrecycle(int qindex, bool kva)
1501 {
1502 	struct buf *bp, *nbp;
1503 
1504 	if (kva)
1505 		atomic_add_int(&bufdefragcnt, 1);
1506 	nbp = NULL;
1507 	mtx_lock(&bqlocks[qindex]);
1508 	nbp = TAILQ_FIRST(&bufqueues[qindex]);
1509 
1510 	/*
1511 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1512 	 * depending.
1513 	 */
1514 	while ((bp = nbp) != NULL) {
1515 		/*
1516 		 * Calculate next bp (we can only use it if we do not
1517 		 * release the bqlock).
1518 		 */
1519 		nbp = TAILQ_NEXT(bp, b_freelist);
1520 
1521 		/*
1522 		 * If we are defragging then we need a buffer with
1523 		 * some kva to reclaim.
1524 		 */
1525 		if (kva && bp->b_kvasize == 0)
1526 			continue;
1527 
1528 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1529 			continue;
1530 
1531 		/*
1532 		 * Skip buffers with background writes in progress.
1533 		 */
1534 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1535 			BUF_UNLOCK(bp);
1536 			continue;
1537 		}
1538 
1539 		KASSERT(bp->b_qindex == qindex,
1540 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1541 		/*
1542 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1543 		 * the scan from this point on.
1544 		 */
1545 		bremfreel(bp);
1546 		mtx_unlock(&bqlocks[qindex]);
1547 
1548 		/*
1549 		 * Requeue the background write buffer with error and
1550 		 * restart the scan.
1551 		 */
1552 		if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1553 			bqrelse(bp);
1554 			mtx_lock(&bqlocks[qindex]);
1555 			nbp = TAILQ_FIRST(&bufqueues[qindex]);
1556 			continue;
1557 		}
1558 		bp->b_flags |= B_INVAL;
1559 		brelse(bp);
1560 		return (0);
1561 	}
1562 	mtx_unlock(&bqlocks[qindex]);
1563 
1564 	return (ENOBUFS);
1565 }
1566 
1567 /*
1568  *	buf_recycle:
1569  *
1570  *	Iterate through all clean queues until we find a buf to recycle or
1571  *	exhaust the search.
1572  */
1573 static int
1574 buf_recycle(bool kva)
1575 {
1576 	int qindex, first_qindex;
1577 
1578 	qindex = first_qindex = bqcleanq();
1579 	do {
1580 		if (buf_qrecycle(qindex, kva) == 0)
1581 			return (0);
1582 		if (++qindex == QUEUE_CLEAN + clean_queues)
1583 			qindex = QUEUE_CLEAN;
1584 	} while (qindex != first_qindex);
1585 
1586 	return (ENOBUFS);
1587 }
1588 
1589 /*
1590  *	buf_scan:
1591  *
1592  *	Scan the clean queues looking for a buffer to recycle.  needsbuffer
1593  *	is set on failure so that the caller may optionally bufspace_wait()
1594  *	in a race-free fashion.
1595  */
1596 static int
1597 buf_scan(bool defrag)
1598 {
1599 	int error;
1600 
1601 	/*
1602 	 * To avoid heavy synchronization and wakeup races we set
1603 	 * needsbuffer and re-poll before failing.  This ensures that
1604 	 * no frees can be missed between an unsuccessful poll and
1605 	 * going to sleep in a synchronized fashion.
1606 	 */
1607 	if ((error = buf_recycle(defrag)) != 0) {
1608 		atomic_set_int(&needsbuffer, 1);
1609 		bufspace_daemonwakeup();
1610 		error = buf_recycle(defrag);
1611 	}
1612 	if (error == 0)
1613 		atomic_add_int(&getnewbufrestarts, 1);
1614 	return (error);
1615 }
1616 
1617 /*
1618  *	bremfree:
1619  *
1620  *	Mark the buffer for removal from the appropriate free list.
1621  *
1622  */
1623 void
1624 bremfree(struct buf *bp)
1625 {
1626 
1627 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1628 	KASSERT((bp->b_flags & B_REMFREE) == 0,
1629 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
1630 	KASSERT(bp->b_qindex != QUEUE_NONE,
1631 	    ("bremfree: buffer %p not on a queue.", bp));
1632 	BUF_ASSERT_XLOCKED(bp);
1633 
1634 	bp->b_flags |= B_REMFREE;
1635 }
1636 
1637 /*
1638  *	bremfreef:
1639  *
1640  *	Force an immediate removal from a free list.  Used only in nfs when
1641  *	it abuses the b_freelist pointer.
1642  */
1643 void
1644 bremfreef(struct buf *bp)
1645 {
1646 	struct mtx *qlock;
1647 
1648 	qlock = bqlock(bp->b_qindex);
1649 	mtx_lock(qlock);
1650 	bremfreel(bp);
1651 	mtx_unlock(qlock);
1652 }
1653 
1654 /*
1655  *	bremfreel:
1656  *
1657  *	Removes a buffer from the free list, must be called with the
1658  *	correct qlock held.
1659  */
1660 static void
1661 bremfreel(struct buf *bp)
1662 {
1663 
1664 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1665 	    bp, bp->b_vp, bp->b_flags);
1666 	KASSERT(bp->b_qindex != QUEUE_NONE,
1667 	    ("bremfreel: buffer %p not on a queue.", bp));
1668 	if (bp->b_qindex != QUEUE_EMPTY) {
1669 		BUF_ASSERT_XLOCKED(bp);
1670 	}
1671 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1672 
1673 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1674 #ifdef INVARIANTS
1675 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1676 	    bp->b_qindex));
1677 	bq_len[bp->b_qindex]--;
1678 #endif
1679 	bp->b_qindex = QUEUE_NONE;
1680 	bp->b_flags &= ~B_REMFREE;
1681 }
1682 
1683 /*
1684  *	bufkva_free:
1685  *
1686  *	Free the kva allocation for a buffer.
1687  *
1688  */
1689 static void
1690 bufkva_free(struct buf *bp)
1691 {
1692 
1693 #ifdef INVARIANTS
1694 	if (bp->b_kvasize == 0) {
1695 		KASSERT(bp->b_kvabase == unmapped_buf &&
1696 		    bp->b_data == unmapped_buf,
1697 		    ("Leaked KVA space on %p", bp));
1698 	} else if (buf_mapped(bp))
1699 		BUF_CHECK_MAPPED(bp);
1700 	else
1701 		BUF_CHECK_UNMAPPED(bp);
1702 #endif
1703 	if (bp->b_kvasize == 0)
1704 		return;
1705 
1706 	vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
1707 	atomic_subtract_long(&bufkvaspace, bp->b_kvasize);
1708 	atomic_add_int(&buffreekvacnt, 1);
1709 	bp->b_data = bp->b_kvabase = unmapped_buf;
1710 	bp->b_kvasize = 0;
1711 }
1712 
1713 /*
1714  *	bufkva_alloc:
1715  *
1716  *	Allocate the buffer KVA and set b_kvasize and b_kvabase.
1717  */
1718 static int
1719 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
1720 {
1721 	vm_offset_t addr;
1722 	int error;
1723 
1724 	KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
1725 	    ("Invalid gbflags 0x%x in %s", gbflags, __func__));
1726 
1727 	bufkva_free(bp);
1728 
1729 	addr = 0;
1730 	error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
1731 	if (error != 0) {
1732 		/*
1733 		 * Buffer map is too fragmented.  Request the caller
1734 		 * to defragment the map.
1735 		 */
1736 		return (error);
1737 	}
1738 	bp->b_kvabase = (caddr_t)addr;
1739 	bp->b_kvasize = maxsize;
1740 	atomic_add_long(&bufkvaspace, bp->b_kvasize);
1741 	if ((gbflags & GB_UNMAPPED) != 0) {
1742 		bp->b_data = unmapped_buf;
1743 		BUF_CHECK_UNMAPPED(bp);
1744 	} else {
1745 		bp->b_data = bp->b_kvabase;
1746 		BUF_CHECK_MAPPED(bp);
1747 	}
1748 	return (0);
1749 }
1750 
1751 /*
1752  *	bufkva_reclaim:
1753  *
1754  *	Reclaim buffer kva by freeing buffers holding kva.  This is a vmem
1755  *	callback that fires to avoid returning failure.
1756  */
1757 static void
1758 bufkva_reclaim(vmem_t *vmem, int flags)
1759 {
1760 	int i;
1761 
1762 	for (i = 0; i < 5; i++)
1763 		if (buf_scan(true) != 0)
1764 			break;
1765 	return;
1766 }
1767 
1768 
1769 /*
1770  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1771  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1772  * the buffer is valid and we do not have to do anything.
1773  */
1774 void
1775 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1776     int cnt, struct ucred * cred)
1777 {
1778 	struct buf *rabp;
1779 	int i;
1780 
1781 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1782 		if (inmem(vp, *rablkno))
1783 			continue;
1784 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1785 
1786 		if ((rabp->b_flags & B_CACHE) == 0) {
1787 			if (!TD_IS_IDLETHREAD(curthread)) {
1788 #ifdef RACCT
1789 				if (racct_enable) {
1790 					PROC_LOCK(curproc);
1791 					racct_add_buf(curproc, rabp, 0);
1792 					PROC_UNLOCK(curproc);
1793 				}
1794 #endif /* RACCT */
1795 				curthread->td_ru.ru_inblock++;
1796 			}
1797 			rabp->b_flags |= B_ASYNC;
1798 			rabp->b_flags &= ~B_INVAL;
1799 			rabp->b_ioflags &= ~BIO_ERROR;
1800 			rabp->b_iocmd = BIO_READ;
1801 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1802 				rabp->b_rcred = crhold(cred);
1803 			vfs_busy_pages(rabp, 0);
1804 			BUF_KERNPROC(rabp);
1805 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1806 			bstrategy(rabp);
1807 		} else {
1808 			brelse(rabp);
1809 		}
1810 	}
1811 }
1812 
1813 /*
1814  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1815  *
1816  * Get a buffer with the specified data.  Look in the cache first.  We
1817  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1818  * is set, the buffer is valid and we do not have to do anything, see
1819  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1820  *
1821  * Always return a NULL buffer pointer (in bpp) when returning an error.
1822  */
1823 int
1824 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1825     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1826 {
1827 	struct buf *bp;
1828 	int rv = 0, readwait = 0;
1829 
1830 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1831 	/*
1832 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1833 	 */
1834 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1835 	if (bp == NULL)
1836 		return (EBUSY);
1837 
1838 	/* if not found in cache, do some I/O */
1839 	if ((bp->b_flags & B_CACHE) == 0) {
1840 		if (!TD_IS_IDLETHREAD(curthread)) {
1841 #ifdef RACCT
1842 			if (racct_enable) {
1843 				PROC_LOCK(curproc);
1844 				racct_add_buf(curproc, bp, 0);
1845 				PROC_UNLOCK(curproc);
1846 			}
1847 #endif /* RACCT */
1848 			curthread->td_ru.ru_inblock++;
1849 		}
1850 		bp->b_iocmd = BIO_READ;
1851 		bp->b_flags &= ~B_INVAL;
1852 		bp->b_ioflags &= ~BIO_ERROR;
1853 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1854 			bp->b_rcred = crhold(cred);
1855 		vfs_busy_pages(bp, 0);
1856 		bp->b_iooffset = dbtob(bp->b_blkno);
1857 		bstrategy(bp);
1858 		++readwait;
1859 	}
1860 
1861 	breada(vp, rablkno, rabsize, cnt, cred);
1862 
1863 	if (readwait) {
1864 		rv = bufwait(bp);
1865 		if (rv != 0) {
1866 			brelse(bp);
1867 			*bpp = NULL;
1868 		}
1869 	}
1870 	return (rv);
1871 }
1872 
1873 /*
1874  * Write, release buffer on completion.  (Done by iodone
1875  * if async).  Do not bother writing anything if the buffer
1876  * is invalid.
1877  *
1878  * Note that we set B_CACHE here, indicating that buffer is
1879  * fully valid and thus cacheable.  This is true even of NFS
1880  * now so we set it generally.  This could be set either here
1881  * or in biodone() since the I/O is synchronous.  We put it
1882  * here.
1883  */
1884 int
1885 bufwrite(struct buf *bp)
1886 {
1887 	int oldflags;
1888 	struct vnode *vp;
1889 	long space;
1890 	int vp_md;
1891 
1892 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1893 	if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
1894 		bp->b_flags |= B_INVAL | B_RELBUF;
1895 		bp->b_flags &= ~B_CACHE;
1896 		brelse(bp);
1897 		return (ENXIO);
1898 	}
1899 	if (bp->b_flags & B_INVAL) {
1900 		brelse(bp);
1901 		return (0);
1902 	}
1903 
1904 	if (bp->b_flags & B_BARRIER)
1905 		barrierwrites++;
1906 
1907 	oldflags = bp->b_flags;
1908 
1909 	BUF_ASSERT_HELD(bp);
1910 
1911 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1912 	    ("FFS background buffer should not get here %p", bp));
1913 
1914 	vp = bp->b_vp;
1915 	if (vp)
1916 		vp_md = vp->v_vflag & VV_MD;
1917 	else
1918 		vp_md = 0;
1919 
1920 	/*
1921 	 * Mark the buffer clean.  Increment the bufobj write count
1922 	 * before bundirty() call, to prevent other thread from seeing
1923 	 * empty dirty list and zero counter for writes in progress,
1924 	 * falsely indicating that the bufobj is clean.
1925 	 */
1926 	bufobj_wref(bp->b_bufobj);
1927 	bundirty(bp);
1928 
1929 	bp->b_flags &= ~B_DONE;
1930 	bp->b_ioflags &= ~BIO_ERROR;
1931 	bp->b_flags |= B_CACHE;
1932 	bp->b_iocmd = BIO_WRITE;
1933 
1934 	vfs_busy_pages(bp, 1);
1935 
1936 	/*
1937 	 * Normal bwrites pipeline writes
1938 	 */
1939 	bp->b_runningbufspace = bp->b_bufsize;
1940 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1941 
1942 	if (!TD_IS_IDLETHREAD(curthread)) {
1943 #ifdef RACCT
1944 		if (racct_enable) {
1945 			PROC_LOCK(curproc);
1946 			racct_add_buf(curproc, bp, 1);
1947 			PROC_UNLOCK(curproc);
1948 		}
1949 #endif /* RACCT */
1950 		curthread->td_ru.ru_oublock++;
1951 	}
1952 	if (oldflags & B_ASYNC)
1953 		BUF_KERNPROC(bp);
1954 	bp->b_iooffset = dbtob(bp->b_blkno);
1955 	buf_track(bp, __func__);
1956 	bstrategy(bp);
1957 
1958 	if ((oldflags & B_ASYNC) == 0) {
1959 		int rtval = bufwait(bp);
1960 		brelse(bp);
1961 		return (rtval);
1962 	} else if (space > hirunningspace) {
1963 		/*
1964 		 * don't allow the async write to saturate the I/O
1965 		 * system.  We will not deadlock here because
1966 		 * we are blocking waiting for I/O that is already in-progress
1967 		 * to complete. We do not block here if it is the update
1968 		 * or syncer daemon trying to clean up as that can lead
1969 		 * to deadlock.
1970 		 */
1971 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1972 			waitrunningbufspace();
1973 	}
1974 
1975 	return (0);
1976 }
1977 
1978 void
1979 bufbdflush(struct bufobj *bo, struct buf *bp)
1980 {
1981 	struct buf *nbp;
1982 
1983 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1984 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1985 		altbufferflushes++;
1986 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1987 		BO_LOCK(bo);
1988 		/*
1989 		 * Try to find a buffer to flush.
1990 		 */
1991 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1992 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1993 			    BUF_LOCK(nbp,
1994 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1995 				continue;
1996 			if (bp == nbp)
1997 				panic("bdwrite: found ourselves");
1998 			BO_UNLOCK(bo);
1999 			/* Don't countdeps with the bo lock held. */
2000 			if (buf_countdeps(nbp, 0)) {
2001 				BO_LOCK(bo);
2002 				BUF_UNLOCK(nbp);
2003 				continue;
2004 			}
2005 			if (nbp->b_flags & B_CLUSTEROK) {
2006 				vfs_bio_awrite(nbp);
2007 			} else {
2008 				bremfree(nbp);
2009 				bawrite(nbp);
2010 			}
2011 			dirtybufferflushes++;
2012 			break;
2013 		}
2014 		if (nbp == NULL)
2015 			BO_UNLOCK(bo);
2016 	}
2017 }
2018 
2019 /*
2020  * Delayed write. (Buffer is marked dirty).  Do not bother writing
2021  * anything if the buffer is marked invalid.
2022  *
2023  * Note that since the buffer must be completely valid, we can safely
2024  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
2025  * biodone() in order to prevent getblk from writing the buffer
2026  * out synchronously.
2027  */
2028 void
2029 bdwrite(struct buf *bp)
2030 {
2031 	struct thread *td = curthread;
2032 	struct vnode *vp;
2033 	struct bufobj *bo;
2034 
2035 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2036 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2037 	KASSERT((bp->b_flags & B_BARRIER) == 0,
2038 	    ("Barrier request in delayed write %p", bp));
2039 	BUF_ASSERT_HELD(bp);
2040 
2041 	if (bp->b_flags & B_INVAL) {
2042 		brelse(bp);
2043 		return;
2044 	}
2045 
2046 	/*
2047 	 * If we have too many dirty buffers, don't create any more.
2048 	 * If we are wildly over our limit, then force a complete
2049 	 * cleanup. Otherwise, just keep the situation from getting
2050 	 * out of control. Note that we have to avoid a recursive
2051 	 * disaster and not try to clean up after our own cleanup!
2052 	 */
2053 	vp = bp->b_vp;
2054 	bo = bp->b_bufobj;
2055 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2056 		td->td_pflags |= TDP_INBDFLUSH;
2057 		BO_BDFLUSH(bo, bp);
2058 		td->td_pflags &= ~TDP_INBDFLUSH;
2059 	} else
2060 		recursiveflushes++;
2061 
2062 	bdirty(bp);
2063 	/*
2064 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
2065 	 * true even of NFS now.
2066 	 */
2067 	bp->b_flags |= B_CACHE;
2068 
2069 	/*
2070 	 * This bmap keeps the system from needing to do the bmap later,
2071 	 * perhaps when the system is attempting to do a sync.  Since it
2072 	 * is likely that the indirect block -- or whatever other datastructure
2073 	 * that the filesystem needs is still in memory now, it is a good
2074 	 * thing to do this.  Note also, that if the pageout daemon is
2075 	 * requesting a sync -- there might not be enough memory to do
2076 	 * the bmap then...  So, this is important to do.
2077 	 */
2078 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2079 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2080 	}
2081 
2082 	buf_track(bp, __func__);
2083 
2084 	/*
2085 	 * Set the *dirty* buffer range based upon the VM system dirty
2086 	 * pages.
2087 	 *
2088 	 * Mark the buffer pages as clean.  We need to do this here to
2089 	 * satisfy the vnode_pager and the pageout daemon, so that it
2090 	 * thinks that the pages have been "cleaned".  Note that since
2091 	 * the pages are in a delayed write buffer -- the VFS layer
2092 	 * "will" see that the pages get written out on the next sync,
2093 	 * or perhaps the cluster will be completed.
2094 	 */
2095 	vfs_clean_pages_dirty_buf(bp);
2096 	bqrelse(bp);
2097 
2098 	/*
2099 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2100 	 * due to the softdep code.
2101 	 */
2102 }
2103 
2104 /*
2105  *	bdirty:
2106  *
2107  *	Turn buffer into delayed write request.  We must clear BIO_READ and
2108  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
2109  *	itself to properly update it in the dirty/clean lists.  We mark it
2110  *	B_DONE to ensure that any asynchronization of the buffer properly
2111  *	clears B_DONE ( else a panic will occur later ).
2112  *
2113  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2114  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
2115  *	should only be called if the buffer is known-good.
2116  *
2117  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2118  *	count.
2119  *
2120  *	The buffer must be on QUEUE_NONE.
2121  */
2122 void
2123 bdirty(struct buf *bp)
2124 {
2125 
2126 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2127 	    bp, bp->b_vp, bp->b_flags);
2128 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2129 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2130 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2131 	BUF_ASSERT_HELD(bp);
2132 	bp->b_flags &= ~(B_RELBUF);
2133 	bp->b_iocmd = BIO_WRITE;
2134 
2135 	if ((bp->b_flags & B_DELWRI) == 0) {
2136 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2137 		reassignbuf(bp);
2138 		bdirtyadd();
2139 	}
2140 }
2141 
2142 /*
2143  *	bundirty:
2144  *
2145  *	Clear B_DELWRI for buffer.
2146  *
2147  *	Since the buffer is not on a queue, we do not update the numfreebuffers
2148  *	count.
2149  *
2150  *	The buffer must be on QUEUE_NONE.
2151  */
2152 
2153 void
2154 bundirty(struct buf *bp)
2155 {
2156 
2157 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2158 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2159 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2160 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2161 	BUF_ASSERT_HELD(bp);
2162 
2163 	if (bp->b_flags & B_DELWRI) {
2164 		bp->b_flags &= ~B_DELWRI;
2165 		reassignbuf(bp);
2166 		bdirtysub();
2167 	}
2168 	/*
2169 	 * Since it is now being written, we can clear its deferred write flag.
2170 	 */
2171 	bp->b_flags &= ~B_DEFERRED;
2172 }
2173 
2174 /*
2175  *	bawrite:
2176  *
2177  *	Asynchronous write.  Start output on a buffer, but do not wait for
2178  *	it to complete.  The buffer is released when the output completes.
2179  *
2180  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
2181  *	B_INVAL buffers.  Not us.
2182  */
2183 void
2184 bawrite(struct buf *bp)
2185 {
2186 
2187 	bp->b_flags |= B_ASYNC;
2188 	(void) bwrite(bp);
2189 }
2190 
2191 /*
2192  *	babarrierwrite:
2193  *
2194  *	Asynchronous barrier write.  Start output on a buffer, but do not
2195  *	wait for it to complete.  Place a write barrier after this write so
2196  *	that this buffer and all buffers written before it are committed to
2197  *	the disk before any buffers written after this write are committed
2198  *	to the disk.  The buffer is released when the output completes.
2199  */
2200 void
2201 babarrierwrite(struct buf *bp)
2202 {
2203 
2204 	bp->b_flags |= B_ASYNC | B_BARRIER;
2205 	(void) bwrite(bp);
2206 }
2207 
2208 /*
2209  *	bbarrierwrite:
2210  *
2211  *	Synchronous barrier write.  Start output on a buffer and wait for
2212  *	it to complete.  Place a write barrier after this write so that
2213  *	this buffer and all buffers written before it are committed to
2214  *	the disk before any buffers written after this write are committed
2215  *	to the disk.  The buffer is released when the output completes.
2216  */
2217 int
2218 bbarrierwrite(struct buf *bp)
2219 {
2220 
2221 	bp->b_flags |= B_BARRIER;
2222 	return (bwrite(bp));
2223 }
2224 
2225 /*
2226  *	bwillwrite:
2227  *
2228  *	Called prior to the locking of any vnodes when we are expecting to
2229  *	write.  We do not want to starve the buffer cache with too many
2230  *	dirty buffers so we block here.  By blocking prior to the locking
2231  *	of any vnodes we attempt to avoid the situation where a locked vnode
2232  *	prevents the various system daemons from flushing related buffers.
2233  */
2234 void
2235 bwillwrite(void)
2236 {
2237 
2238 	if (numdirtybuffers >= hidirtybuffers) {
2239 		mtx_lock(&bdirtylock);
2240 		while (numdirtybuffers >= hidirtybuffers) {
2241 			bdirtywait = 1;
2242 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2243 			    "flswai", 0);
2244 		}
2245 		mtx_unlock(&bdirtylock);
2246 	}
2247 }
2248 
2249 /*
2250  * Return true if we have too many dirty buffers.
2251  */
2252 int
2253 buf_dirty_count_severe(void)
2254 {
2255 
2256 	return(numdirtybuffers >= hidirtybuffers);
2257 }
2258 
2259 /*
2260  *	brelse:
2261  *
2262  *	Release a busy buffer and, if requested, free its resources.  The
2263  *	buffer will be stashed in the appropriate bufqueue[] allowing it
2264  *	to be accessed later as a cache entity or reused for other purposes.
2265  */
2266 void
2267 brelse(struct buf *bp)
2268 {
2269 	int qindex;
2270 
2271 	/*
2272 	 * Many functions erroneously call brelse with a NULL bp under rare
2273 	 * error conditions. Simply return when called with a NULL bp.
2274 	 */
2275 	if (bp == NULL)
2276 		return;
2277 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2278 	    bp, bp->b_vp, bp->b_flags);
2279 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2280 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2281 	KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2282 	    ("brelse: non-VMIO buffer marked NOREUSE"));
2283 
2284 	if (BUF_LOCKRECURSED(bp)) {
2285 		/*
2286 		 * Do not process, in particular, do not handle the
2287 		 * B_INVAL/B_RELBUF and do not release to free list.
2288 		 */
2289 		BUF_UNLOCK(bp);
2290 		return;
2291 	}
2292 
2293 	if (bp->b_flags & B_MANAGED) {
2294 		bqrelse(bp);
2295 		return;
2296 	}
2297 
2298 	if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2299 		BO_LOCK(bp->b_bufobj);
2300 		bp->b_vflags &= ~BV_BKGRDERR;
2301 		BO_UNLOCK(bp->b_bufobj);
2302 		bdirty(bp);
2303 	}
2304 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2305 	    !(bp->b_flags & B_INVAL)) {
2306 		/*
2307 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
2308 		 * pages from being scrapped.
2309 		 */
2310 		bp->b_ioflags &= ~BIO_ERROR;
2311 		bdirty(bp);
2312 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2313 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2314 		/*
2315 		 * Either a failed read I/O or we were asked to free or not
2316 		 * cache the buffer.
2317 		 */
2318 		bp->b_flags |= B_INVAL;
2319 		if (!LIST_EMPTY(&bp->b_dep))
2320 			buf_deallocate(bp);
2321 		if (bp->b_flags & B_DELWRI)
2322 			bdirtysub();
2323 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
2324 		if ((bp->b_flags & B_VMIO) == 0) {
2325 			allocbuf(bp, 0);
2326 			if (bp->b_vp)
2327 				brelvp(bp);
2328 		}
2329 	}
2330 
2331 	/*
2332 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_truncate()
2333 	 * is called with B_DELWRI set, the underlying pages may wind up
2334 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
2335 	 * because pages associated with a B_DELWRI bp are marked clean.
2336 	 *
2337 	 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2338 	 * if B_DELWRI is set.
2339 	 */
2340 	if (bp->b_flags & B_DELWRI)
2341 		bp->b_flags &= ~B_RELBUF;
2342 
2343 	/*
2344 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
2345 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
2346 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
2347 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2348 	 *
2349 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2350 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
2351 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
2352 	 *
2353 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
2354 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2355 	 * the commit state and we cannot afford to lose the buffer. If the
2356 	 * buffer has a background write in progress, we need to keep it
2357 	 * around to prevent it from being reconstituted and starting a second
2358 	 * background write.
2359 	 */
2360 	if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2361 	    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2362 	    !(bp->b_vp->v_mount != NULL &&
2363 	    (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2364 	    !vn_isdisk(bp->b_vp, NULL) && (bp->b_flags & B_DELWRI))) {
2365 		vfs_vmio_invalidate(bp);
2366 		allocbuf(bp, 0);
2367 	}
2368 
2369 	if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2370 	    (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2371 		allocbuf(bp, 0);
2372 		bp->b_flags &= ~B_NOREUSE;
2373 		if (bp->b_vp != NULL)
2374 			brelvp(bp);
2375 	}
2376 
2377 	/*
2378 	 * If the buffer has junk contents signal it and eventually
2379 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2380 	 * doesn't find it.
2381 	 */
2382 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2383 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2384 		bp->b_flags |= B_INVAL;
2385 	if (bp->b_flags & B_INVAL) {
2386 		if (bp->b_flags & B_DELWRI)
2387 			bundirty(bp);
2388 		if (bp->b_vp)
2389 			brelvp(bp);
2390 	}
2391 
2392 	buf_track(bp, __func__);
2393 
2394 	/* buffers with no memory */
2395 	if (bp->b_bufsize == 0) {
2396 		buf_free(bp);
2397 		return;
2398 	}
2399 	/* buffers with junk contents */
2400 	if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2401 	    (bp->b_ioflags & BIO_ERROR)) {
2402 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2403 		if (bp->b_vflags & BV_BKGRDINPROG)
2404 			panic("losing buffer 2");
2405 		qindex = QUEUE_CLEAN;
2406 		bp->b_flags |= B_AGE;
2407 	/* remaining buffers */
2408 	} else if (bp->b_flags & B_DELWRI)
2409 		qindex = QUEUE_DIRTY;
2410 	else
2411 		qindex = QUEUE_CLEAN;
2412 
2413 	binsfree(bp, qindex);
2414 
2415 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
2416 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2417 		panic("brelse: not dirty");
2418 	/* unlock */
2419 	BUF_UNLOCK(bp);
2420 	if (qindex == QUEUE_CLEAN)
2421 		bufspace_wakeup();
2422 }
2423 
2424 /*
2425  * Release a buffer back to the appropriate queue but do not try to free
2426  * it.  The buffer is expected to be used again soon.
2427  *
2428  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2429  * biodone() to requeue an async I/O on completion.  It is also used when
2430  * known good buffers need to be requeued but we think we may need the data
2431  * again soon.
2432  *
2433  * XXX we should be able to leave the B_RELBUF hint set on completion.
2434  */
2435 void
2436 bqrelse(struct buf *bp)
2437 {
2438 	int qindex;
2439 
2440 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2441 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2442 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2443 
2444 	qindex = QUEUE_NONE;
2445 	if (BUF_LOCKRECURSED(bp)) {
2446 		/* do not release to free list */
2447 		BUF_UNLOCK(bp);
2448 		return;
2449 	}
2450 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2451 
2452 	if (bp->b_flags & B_MANAGED) {
2453 		if (bp->b_flags & B_REMFREE)
2454 			bremfreef(bp);
2455 		goto out;
2456 	}
2457 
2458 	/* buffers with stale but valid contents */
2459 	if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2460 	    BV_BKGRDERR)) == BV_BKGRDERR) {
2461 		BO_LOCK(bp->b_bufobj);
2462 		bp->b_vflags &= ~BV_BKGRDERR;
2463 		BO_UNLOCK(bp->b_bufobj);
2464 		qindex = QUEUE_DIRTY;
2465 	} else {
2466 		if ((bp->b_flags & B_DELWRI) == 0 &&
2467 		    (bp->b_xflags & BX_VNDIRTY))
2468 			panic("bqrelse: not dirty");
2469 		if ((bp->b_flags & B_NOREUSE) != 0) {
2470 			brelse(bp);
2471 			return;
2472 		}
2473 		qindex = QUEUE_CLEAN;
2474 	}
2475 	binsfree(bp, qindex);
2476 
2477 out:
2478 	buf_track(bp, __func__);
2479 	/* unlock */
2480 	BUF_UNLOCK(bp);
2481 	if (qindex == QUEUE_CLEAN)
2482 		bufspace_wakeup();
2483 }
2484 
2485 /*
2486  * Complete I/O to a VMIO backed page.  Validate the pages as appropriate,
2487  * restore bogus pages.
2488  */
2489 static void
2490 vfs_vmio_iodone(struct buf *bp)
2491 {
2492 	vm_ooffset_t foff;
2493 	vm_page_t m;
2494 	vm_object_t obj;
2495 	struct vnode *vp;
2496 	int bogus, i, iosize;
2497 
2498 	obj = bp->b_bufobj->bo_object;
2499 	KASSERT(obj->paging_in_progress >= bp->b_npages,
2500 	    ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2501 	    obj->paging_in_progress, bp->b_npages));
2502 
2503 	vp = bp->b_vp;
2504 	KASSERT(vp->v_holdcnt > 0,
2505 	    ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2506 	KASSERT(vp->v_object != NULL,
2507 	    ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2508 
2509 	foff = bp->b_offset;
2510 	KASSERT(bp->b_offset != NOOFFSET,
2511 	    ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2512 
2513 	bogus = 0;
2514 	iosize = bp->b_bcount - bp->b_resid;
2515 	VM_OBJECT_WLOCK(obj);
2516 	for (i = 0; i < bp->b_npages; i++) {
2517 		int resid;
2518 
2519 		resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2520 		if (resid > iosize)
2521 			resid = iosize;
2522 
2523 		/*
2524 		 * cleanup bogus pages, restoring the originals
2525 		 */
2526 		m = bp->b_pages[i];
2527 		if (m == bogus_page) {
2528 			bogus = 1;
2529 			m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2530 			if (m == NULL)
2531 				panic("biodone: page disappeared!");
2532 			bp->b_pages[i] = m;
2533 		} else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2534 			/*
2535 			 * In the write case, the valid and clean bits are
2536 			 * already changed correctly ( see bdwrite() ), so we
2537 			 * only need to do this here in the read case.
2538 			 */
2539 			KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2540 			    resid)) == 0, ("vfs_vmio_iodone: page %p "
2541 			    "has unexpected dirty bits", m));
2542 			vfs_page_set_valid(bp, foff, m);
2543 		}
2544 		KASSERT(OFF_TO_IDX(foff) == m->pindex,
2545 		    ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2546 		    (intmax_t)foff, (uintmax_t)m->pindex));
2547 
2548 		vm_page_sunbusy(m);
2549 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2550 		iosize -= resid;
2551 	}
2552 	vm_object_pip_wakeupn(obj, bp->b_npages);
2553 	VM_OBJECT_WUNLOCK(obj);
2554 	if (bogus && buf_mapped(bp)) {
2555 		BUF_CHECK_MAPPED(bp);
2556 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2557 		    bp->b_pages, bp->b_npages);
2558 	}
2559 }
2560 
2561 /*
2562  * Unwire a page held by a buf and place it on the appropriate vm queue.
2563  */
2564 static void
2565 vfs_vmio_unwire(struct buf *bp, vm_page_t m)
2566 {
2567 	bool freed;
2568 
2569 	vm_page_lock(m);
2570 	if (vm_page_unwire(m, PQ_NONE)) {
2571 		/*
2572 		 * Determine if the page should be freed before adding
2573 		 * it to the inactive queue.
2574 		 */
2575 		if (m->valid == 0) {
2576 			freed = !vm_page_busied(m);
2577 			if (freed)
2578 				vm_page_free(m);
2579 		} else if ((bp->b_flags & B_DIRECT) != 0)
2580 			freed = vm_page_try_to_free(m);
2581 		else
2582 			freed = false;
2583 		if (!freed) {
2584 			/*
2585 			 * If the page is unlikely to be reused, let the
2586 			 * VM know.  Otherwise, maintain LRU page
2587 			 * ordering and put the page at the tail of the
2588 			 * inactive queue.
2589 			 */
2590 			if ((bp->b_flags & B_NOREUSE) != 0)
2591 				vm_page_deactivate_noreuse(m);
2592 			else
2593 				vm_page_deactivate(m);
2594 		}
2595 	}
2596 	vm_page_unlock(m);
2597 }
2598 
2599 /*
2600  * Perform page invalidation when a buffer is released.  The fully invalid
2601  * pages will be reclaimed later in vfs_vmio_truncate().
2602  */
2603 static void
2604 vfs_vmio_invalidate(struct buf *bp)
2605 {
2606 	vm_object_t obj;
2607 	vm_page_t m;
2608 	int i, resid, poffset, presid;
2609 
2610 	if (buf_mapped(bp)) {
2611 		BUF_CHECK_MAPPED(bp);
2612 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2613 	} else
2614 		BUF_CHECK_UNMAPPED(bp);
2615 	/*
2616 	 * Get the base offset and length of the buffer.  Note that
2617 	 * in the VMIO case if the buffer block size is not
2618 	 * page-aligned then b_data pointer may not be page-aligned.
2619 	 * But our b_pages[] array *IS* page aligned.
2620 	 *
2621 	 * block sizes less then DEV_BSIZE (usually 512) are not
2622 	 * supported due to the page granularity bits (m->valid,
2623 	 * m->dirty, etc...).
2624 	 *
2625 	 * See man buf(9) for more information
2626 	 */
2627 	obj = bp->b_bufobj->bo_object;
2628 	resid = bp->b_bufsize;
2629 	poffset = bp->b_offset & PAGE_MASK;
2630 	VM_OBJECT_WLOCK(obj);
2631 	for (i = 0; i < bp->b_npages; i++) {
2632 		m = bp->b_pages[i];
2633 		if (m == bogus_page)
2634 			panic("vfs_vmio_invalidate: Unexpected bogus page.");
2635 		bp->b_pages[i] = NULL;
2636 
2637 		presid = resid > (PAGE_SIZE - poffset) ?
2638 		    (PAGE_SIZE - poffset) : resid;
2639 		KASSERT(presid >= 0, ("brelse: extra page"));
2640 		while (vm_page_xbusied(m)) {
2641 			vm_page_lock(m);
2642 			VM_OBJECT_WUNLOCK(obj);
2643 			vm_page_busy_sleep(m, "mbncsh", true);
2644 			VM_OBJECT_WLOCK(obj);
2645 		}
2646 		if (pmap_page_wired_mappings(m) == 0)
2647 			vm_page_set_invalid(m, poffset, presid);
2648 		vfs_vmio_unwire(bp, m);
2649 		resid -= presid;
2650 		poffset = 0;
2651 	}
2652 	VM_OBJECT_WUNLOCK(obj);
2653 	bp->b_npages = 0;
2654 }
2655 
2656 /*
2657  * Page-granular truncation of an existing VMIO buffer.
2658  */
2659 static void
2660 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2661 {
2662 	vm_object_t obj;
2663 	vm_page_t m;
2664 	int i;
2665 
2666 	if (bp->b_npages == desiredpages)
2667 		return;
2668 
2669 	if (buf_mapped(bp)) {
2670 		BUF_CHECK_MAPPED(bp);
2671 		pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2672 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2673 	} else
2674 		BUF_CHECK_UNMAPPED(bp);
2675 	obj = bp->b_bufobj->bo_object;
2676 	if (obj != NULL)
2677 		VM_OBJECT_WLOCK(obj);
2678 	for (i = desiredpages; i < bp->b_npages; i++) {
2679 		m = bp->b_pages[i];
2680 		KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
2681 		bp->b_pages[i] = NULL;
2682 		vfs_vmio_unwire(bp, m);
2683 	}
2684 	if (obj != NULL)
2685 		VM_OBJECT_WUNLOCK(obj);
2686 	bp->b_npages = desiredpages;
2687 }
2688 
2689 /*
2690  * Byte granular extension of VMIO buffers.
2691  */
2692 static void
2693 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
2694 {
2695 	/*
2696 	 * We are growing the buffer, possibly in a
2697 	 * byte-granular fashion.
2698 	 */
2699 	vm_object_t obj;
2700 	vm_offset_t toff;
2701 	vm_offset_t tinc;
2702 	vm_page_t m;
2703 
2704 	/*
2705 	 * Step 1, bring in the VM pages from the object, allocating
2706 	 * them if necessary.  We must clear B_CACHE if these pages
2707 	 * are not valid for the range covered by the buffer.
2708 	 */
2709 	obj = bp->b_bufobj->bo_object;
2710 	VM_OBJECT_WLOCK(obj);
2711 	while (bp->b_npages < desiredpages) {
2712 		/*
2713 		 * We must allocate system pages since blocking
2714 		 * here could interfere with paging I/O, no
2715 		 * matter which process we are.
2716 		 *
2717 		 * Only exclusive busy can be tested here.
2718 		 * Blocking on shared busy might lead to
2719 		 * deadlocks once allocbuf() is called after
2720 		 * pages are vfs_busy_pages().
2721 		 */
2722 		m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages,
2723 		    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2724 		    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY |
2725 		    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
2726 		if (m->valid == 0)
2727 			bp->b_flags &= ~B_CACHE;
2728 		bp->b_pages[bp->b_npages] = m;
2729 		++bp->b_npages;
2730 	}
2731 
2732 	/*
2733 	 * Step 2.  We've loaded the pages into the buffer,
2734 	 * we have to figure out if we can still have B_CACHE
2735 	 * set.  Note that B_CACHE is set according to the
2736 	 * byte-granular range ( bcount and size ), not the
2737 	 * aligned range ( newbsize ).
2738 	 *
2739 	 * The VM test is against m->valid, which is DEV_BSIZE
2740 	 * aligned.  Needless to say, the validity of the data
2741 	 * needs to also be DEV_BSIZE aligned.  Note that this
2742 	 * fails with NFS if the server or some other client
2743 	 * extends the file's EOF.  If our buffer is resized,
2744 	 * B_CACHE may remain set! XXX
2745 	 */
2746 	toff = bp->b_bcount;
2747 	tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2748 	while ((bp->b_flags & B_CACHE) && toff < size) {
2749 		vm_pindex_t pi;
2750 
2751 		if (tinc > (size - toff))
2752 			tinc = size - toff;
2753 		pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
2754 		m = bp->b_pages[pi];
2755 		vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
2756 		toff += tinc;
2757 		tinc = PAGE_SIZE;
2758 	}
2759 	VM_OBJECT_WUNLOCK(obj);
2760 
2761 	/*
2762 	 * Step 3, fixup the KVA pmap.
2763 	 */
2764 	if (buf_mapped(bp))
2765 		bpmap_qenter(bp);
2766 	else
2767 		BUF_CHECK_UNMAPPED(bp);
2768 }
2769 
2770 /*
2771  * Check to see if a block at a particular lbn is available for a clustered
2772  * write.
2773  */
2774 static int
2775 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
2776 {
2777 	struct buf *bpa;
2778 	int match;
2779 
2780 	match = 0;
2781 
2782 	/* If the buf isn't in core skip it */
2783 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
2784 		return (0);
2785 
2786 	/* If the buf is busy we don't want to wait for it */
2787 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2788 		return (0);
2789 
2790 	/* Only cluster with valid clusterable delayed write buffers */
2791 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
2792 	    (B_DELWRI | B_CLUSTEROK))
2793 		goto done;
2794 
2795 	if (bpa->b_bufsize != size)
2796 		goto done;
2797 
2798 	/*
2799 	 * Check to see if it is in the expected place on disk and that the
2800 	 * block has been mapped.
2801 	 */
2802 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
2803 		match = 1;
2804 done:
2805 	BUF_UNLOCK(bpa);
2806 	return (match);
2807 }
2808 
2809 /*
2810  *	vfs_bio_awrite:
2811  *
2812  *	Implement clustered async writes for clearing out B_DELWRI buffers.
2813  *	This is much better then the old way of writing only one buffer at
2814  *	a time.  Note that we may not be presented with the buffers in the
2815  *	correct order, so we search for the cluster in both directions.
2816  */
2817 int
2818 vfs_bio_awrite(struct buf *bp)
2819 {
2820 	struct bufobj *bo;
2821 	int i;
2822 	int j;
2823 	daddr_t lblkno = bp->b_lblkno;
2824 	struct vnode *vp = bp->b_vp;
2825 	int ncl;
2826 	int nwritten;
2827 	int size;
2828 	int maxcl;
2829 	int gbflags;
2830 
2831 	bo = &vp->v_bufobj;
2832 	gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
2833 	/*
2834 	 * right now we support clustered writing only to regular files.  If
2835 	 * we find a clusterable block we could be in the middle of a cluster
2836 	 * rather then at the beginning.
2837 	 */
2838 	if ((vp->v_type == VREG) &&
2839 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
2840 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
2841 
2842 		size = vp->v_mount->mnt_stat.f_iosize;
2843 		maxcl = MAXPHYS / size;
2844 
2845 		BO_RLOCK(bo);
2846 		for (i = 1; i < maxcl; i++)
2847 			if (vfs_bio_clcheck(vp, size, lblkno + i,
2848 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2849 				break;
2850 
2851 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2852 			if (vfs_bio_clcheck(vp, size, lblkno - j,
2853 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2854 				break;
2855 		BO_RUNLOCK(bo);
2856 		--j;
2857 		ncl = i + j;
2858 		/*
2859 		 * this is a possible cluster write
2860 		 */
2861 		if (ncl != 1) {
2862 			BUF_UNLOCK(bp);
2863 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2864 			    gbflags);
2865 			return (nwritten);
2866 		}
2867 	}
2868 	bremfree(bp);
2869 	bp->b_flags |= B_ASYNC;
2870 	/*
2871 	 * default (old) behavior, writing out only one block
2872 	 *
2873 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
2874 	 */
2875 	nwritten = bp->b_bufsize;
2876 	(void) bwrite(bp);
2877 
2878 	return (nwritten);
2879 }
2880 
2881 /*
2882  *	getnewbuf_kva:
2883  *
2884  *	Allocate KVA for an empty buf header according to gbflags.
2885  */
2886 static int
2887 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
2888 {
2889 
2890 	if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
2891 		/*
2892 		 * In order to keep fragmentation sane we only allocate kva
2893 		 * in BKVASIZE chunks.  XXX with vmem we can do page size.
2894 		 */
2895 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2896 
2897 		if (maxsize != bp->b_kvasize &&
2898 		    bufkva_alloc(bp, maxsize, gbflags))
2899 			return (ENOSPC);
2900 	}
2901 	return (0);
2902 }
2903 
2904 /*
2905  *	getnewbuf:
2906  *
2907  *	Find and initialize a new buffer header, freeing up existing buffers
2908  *	in the bufqueues as necessary.  The new buffer is returned locked.
2909  *
2910  *	We block if:
2911  *		We have insufficient buffer headers
2912  *		We have insufficient buffer space
2913  *		buffer_arena is too fragmented ( space reservation fails )
2914  *		If we have to flush dirty buffers ( but we try to avoid this )
2915  *
2916  *	The caller is responsible for releasing the reserved bufspace after
2917  *	allocbuf() is called.
2918  */
2919 static struct buf *
2920 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
2921 {
2922 	struct buf *bp;
2923 	bool metadata, reserved;
2924 
2925 	bp = NULL;
2926 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2927 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2928 	if (!unmapped_buf_allowed)
2929 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2930 
2931 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2932 	    vp->v_type == VCHR)
2933 		metadata = true;
2934 	else
2935 		metadata = false;
2936 	atomic_add_int(&getnewbufcalls, 1);
2937 	reserved = false;
2938 	do {
2939 		if (reserved == false &&
2940 		    bufspace_reserve(maxsize, metadata) != 0)
2941 			continue;
2942 		reserved = true;
2943 		if ((bp = buf_alloc()) == NULL)
2944 			continue;
2945 		if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
2946 			return (bp);
2947 		break;
2948 	} while(buf_scan(false) == 0);
2949 
2950 	if (reserved)
2951 		atomic_subtract_long(&bufspace, maxsize);
2952 	if (bp != NULL) {
2953 		bp->b_flags |= B_INVAL;
2954 		brelse(bp);
2955 	}
2956 	bufspace_wait(vp, gbflags, slpflag, slptimeo);
2957 
2958 	return (NULL);
2959 }
2960 
2961 /*
2962  *	buf_daemon:
2963  *
2964  *	buffer flushing daemon.  Buffers are normally flushed by the
2965  *	update daemon but if it cannot keep up this process starts to
2966  *	take the load in an attempt to prevent getnewbuf() from blocking.
2967  */
2968 static struct kproc_desc buf_kp = {
2969 	"bufdaemon",
2970 	buf_daemon,
2971 	&bufdaemonproc
2972 };
2973 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2974 
2975 static int
2976 buf_flush(struct vnode *vp, int target)
2977 {
2978 	int flushed;
2979 
2980 	flushed = flushbufqueues(vp, target, 0);
2981 	if (flushed == 0) {
2982 		/*
2983 		 * Could not find any buffers without rollback
2984 		 * dependencies, so just write the first one
2985 		 * in the hopes of eventually making progress.
2986 		 */
2987 		if (vp != NULL && target > 2)
2988 			target /= 2;
2989 		flushbufqueues(vp, target, 1);
2990 	}
2991 	return (flushed);
2992 }
2993 
2994 static void
2995 buf_daemon()
2996 {
2997 	int lodirty;
2998 
2999 	/*
3000 	 * This process needs to be suspended prior to shutdown sync.
3001 	 */
3002 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
3003 	    SHUTDOWN_PRI_LAST);
3004 
3005 	/*
3006 	 * This process is allowed to take the buffer cache to the limit
3007 	 */
3008 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3009 	mtx_lock(&bdlock);
3010 	for (;;) {
3011 		bd_request = 0;
3012 		mtx_unlock(&bdlock);
3013 
3014 		kproc_suspend_check(bufdaemonproc);
3015 		lodirty = lodirtybuffers;
3016 		if (bd_speedupreq) {
3017 			lodirty = numdirtybuffers / 2;
3018 			bd_speedupreq = 0;
3019 		}
3020 		/*
3021 		 * Do the flush.  Limit the amount of in-transit I/O we
3022 		 * allow to build up, otherwise we would completely saturate
3023 		 * the I/O system.
3024 		 */
3025 		while (numdirtybuffers > lodirty) {
3026 			if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
3027 				break;
3028 			kern_yield(PRI_USER);
3029 		}
3030 
3031 		/*
3032 		 * Only clear bd_request if we have reached our low water
3033 		 * mark.  The buf_daemon normally waits 1 second and
3034 		 * then incrementally flushes any dirty buffers that have
3035 		 * built up, within reason.
3036 		 *
3037 		 * If we were unable to hit our low water mark and couldn't
3038 		 * find any flushable buffers, we sleep for a short period
3039 		 * to avoid endless loops on unlockable buffers.
3040 		 */
3041 		mtx_lock(&bdlock);
3042 		if (numdirtybuffers <= lodirtybuffers) {
3043 			/*
3044 			 * We reached our low water mark, reset the
3045 			 * request and sleep until we are needed again.
3046 			 * The sleep is just so the suspend code works.
3047 			 */
3048 			bd_request = 0;
3049 			/*
3050 			 * Do an extra wakeup in case dirty threshold
3051 			 * changed via sysctl and the explicit transition
3052 			 * out of shortfall was missed.
3053 			 */
3054 			bdirtywakeup();
3055 			if (runningbufspace <= lorunningspace)
3056 				runningwakeup();
3057 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3058 		} else {
3059 			/*
3060 			 * We couldn't find any flushable dirty buffers but
3061 			 * still have too many dirty buffers, we
3062 			 * have to sleep and try again.  (rare)
3063 			 */
3064 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3065 		}
3066 	}
3067 }
3068 
3069 /*
3070  *	flushbufqueues:
3071  *
3072  *	Try to flush a buffer in the dirty queue.  We must be careful to
3073  *	free up B_INVAL buffers instead of write them, which NFS is
3074  *	particularly sensitive to.
3075  */
3076 static int flushwithdeps = 0;
3077 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3078     0, "Number of buffers flushed with dependecies that require rollbacks");
3079 
3080 static int
3081 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
3082 {
3083 	struct buf *sentinel;
3084 	struct vnode *vp;
3085 	struct mount *mp;
3086 	struct buf *bp;
3087 	int hasdeps;
3088 	int flushed;
3089 	int queue;
3090 	int error;
3091 	bool unlock;
3092 
3093 	flushed = 0;
3094 	queue = QUEUE_DIRTY;
3095 	bp = NULL;
3096 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3097 	sentinel->b_qindex = QUEUE_SENTINEL;
3098 	mtx_lock(&bqlocks[queue]);
3099 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
3100 	mtx_unlock(&bqlocks[queue]);
3101 	while (flushed != target) {
3102 		maybe_yield();
3103 		mtx_lock(&bqlocks[queue]);
3104 		bp = TAILQ_NEXT(sentinel, b_freelist);
3105 		if (bp != NULL) {
3106 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
3107 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
3108 			    b_freelist);
3109 		} else {
3110 			mtx_unlock(&bqlocks[queue]);
3111 			break;
3112 		}
3113 		/*
3114 		 * Skip sentinels inserted by other invocations of the
3115 		 * flushbufqueues(), taking care to not reorder them.
3116 		 *
3117 		 * Only flush the buffers that belong to the
3118 		 * vnode locked by the curthread.
3119 		 */
3120 		if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3121 		    bp->b_vp != lvp)) {
3122 			mtx_unlock(&bqlocks[queue]);
3123 			continue;
3124 		}
3125 		error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3126 		mtx_unlock(&bqlocks[queue]);
3127 		if (error != 0)
3128 			continue;
3129 
3130 		/*
3131 		 * BKGRDINPROG can only be set with the buf and bufobj
3132 		 * locks both held.  We tolerate a race to clear it here.
3133 		 */
3134 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3135 		    (bp->b_flags & B_DELWRI) == 0) {
3136 			BUF_UNLOCK(bp);
3137 			continue;
3138 		}
3139 		if (bp->b_flags & B_INVAL) {
3140 			bremfreef(bp);
3141 			brelse(bp);
3142 			flushed++;
3143 			continue;
3144 		}
3145 
3146 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3147 			if (flushdeps == 0) {
3148 				BUF_UNLOCK(bp);
3149 				continue;
3150 			}
3151 			hasdeps = 1;
3152 		} else
3153 			hasdeps = 0;
3154 		/*
3155 		 * We must hold the lock on a vnode before writing
3156 		 * one of its buffers. Otherwise we may confuse, or
3157 		 * in the case of a snapshot vnode, deadlock the
3158 		 * system.
3159 		 *
3160 		 * The lock order here is the reverse of the normal
3161 		 * of vnode followed by buf lock.  This is ok because
3162 		 * the NOWAIT will prevent deadlock.
3163 		 */
3164 		vp = bp->b_vp;
3165 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3166 			BUF_UNLOCK(bp);
3167 			continue;
3168 		}
3169 		if (lvp == NULL) {
3170 			unlock = true;
3171 			error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3172 		} else {
3173 			ASSERT_VOP_LOCKED(vp, "getbuf");
3174 			unlock = false;
3175 			error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3176 			    vn_lock(vp, LK_TRYUPGRADE);
3177 		}
3178 		if (error == 0) {
3179 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3180 			    bp, bp->b_vp, bp->b_flags);
3181 			if (curproc == bufdaemonproc) {
3182 				vfs_bio_awrite(bp);
3183 			} else {
3184 				bremfree(bp);
3185 				bwrite(bp);
3186 				notbufdflushes++;
3187 			}
3188 			vn_finished_write(mp);
3189 			if (unlock)
3190 				VOP_UNLOCK(vp, 0);
3191 			flushwithdeps += hasdeps;
3192 			flushed++;
3193 
3194 			/*
3195 			 * Sleeping on runningbufspace while holding
3196 			 * vnode lock leads to deadlock.
3197 			 */
3198 			if (curproc == bufdaemonproc &&
3199 			    runningbufspace > hirunningspace)
3200 				waitrunningbufspace();
3201 			continue;
3202 		}
3203 		vn_finished_write(mp);
3204 		BUF_UNLOCK(bp);
3205 	}
3206 	mtx_lock(&bqlocks[queue]);
3207 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
3208 	mtx_unlock(&bqlocks[queue]);
3209 	free(sentinel, M_TEMP);
3210 	return (flushed);
3211 }
3212 
3213 /*
3214  * Check to see if a block is currently memory resident.
3215  */
3216 struct buf *
3217 incore(struct bufobj *bo, daddr_t blkno)
3218 {
3219 	struct buf *bp;
3220 
3221 	BO_RLOCK(bo);
3222 	bp = gbincore(bo, blkno);
3223 	BO_RUNLOCK(bo);
3224 	return (bp);
3225 }
3226 
3227 /*
3228  * Returns true if no I/O is needed to access the
3229  * associated VM object.  This is like incore except
3230  * it also hunts around in the VM system for the data.
3231  */
3232 
3233 static int
3234 inmem(struct vnode * vp, daddr_t blkno)
3235 {
3236 	vm_object_t obj;
3237 	vm_offset_t toff, tinc, size;
3238 	vm_page_t m;
3239 	vm_ooffset_t off;
3240 
3241 	ASSERT_VOP_LOCKED(vp, "inmem");
3242 
3243 	if (incore(&vp->v_bufobj, blkno))
3244 		return 1;
3245 	if (vp->v_mount == NULL)
3246 		return 0;
3247 	obj = vp->v_object;
3248 	if (obj == NULL)
3249 		return (0);
3250 
3251 	size = PAGE_SIZE;
3252 	if (size > vp->v_mount->mnt_stat.f_iosize)
3253 		size = vp->v_mount->mnt_stat.f_iosize;
3254 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3255 
3256 	VM_OBJECT_RLOCK(obj);
3257 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3258 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3259 		if (!m)
3260 			goto notinmem;
3261 		tinc = size;
3262 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3263 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3264 		if (vm_page_is_valid(m,
3265 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3266 			goto notinmem;
3267 	}
3268 	VM_OBJECT_RUNLOCK(obj);
3269 	return 1;
3270 
3271 notinmem:
3272 	VM_OBJECT_RUNLOCK(obj);
3273 	return (0);
3274 }
3275 
3276 /*
3277  * Set the dirty range for a buffer based on the status of the dirty
3278  * bits in the pages comprising the buffer.  The range is limited
3279  * to the size of the buffer.
3280  *
3281  * Tell the VM system that the pages associated with this buffer
3282  * are clean.  This is used for delayed writes where the data is
3283  * going to go to disk eventually without additional VM intevention.
3284  *
3285  * Note that while we only really need to clean through to b_bcount, we
3286  * just go ahead and clean through to b_bufsize.
3287  */
3288 static void
3289 vfs_clean_pages_dirty_buf(struct buf *bp)
3290 {
3291 	vm_ooffset_t foff, noff, eoff;
3292 	vm_page_t m;
3293 	int i;
3294 
3295 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3296 		return;
3297 
3298 	foff = bp->b_offset;
3299 	KASSERT(bp->b_offset != NOOFFSET,
3300 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
3301 
3302 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3303 	vfs_drain_busy_pages(bp);
3304 	vfs_setdirty_locked_object(bp);
3305 	for (i = 0; i < bp->b_npages; i++) {
3306 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3307 		eoff = noff;
3308 		if (eoff > bp->b_offset + bp->b_bufsize)
3309 			eoff = bp->b_offset + bp->b_bufsize;
3310 		m = bp->b_pages[i];
3311 		vfs_page_set_validclean(bp, foff, m);
3312 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3313 		foff = noff;
3314 	}
3315 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3316 }
3317 
3318 static void
3319 vfs_setdirty_locked_object(struct buf *bp)
3320 {
3321 	vm_object_t object;
3322 	int i;
3323 
3324 	object = bp->b_bufobj->bo_object;
3325 	VM_OBJECT_ASSERT_WLOCKED(object);
3326 
3327 	/*
3328 	 * We qualify the scan for modified pages on whether the
3329 	 * object has been flushed yet.
3330 	 */
3331 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3332 		vm_offset_t boffset;
3333 		vm_offset_t eoffset;
3334 
3335 		/*
3336 		 * test the pages to see if they have been modified directly
3337 		 * by users through the VM system.
3338 		 */
3339 		for (i = 0; i < bp->b_npages; i++)
3340 			vm_page_test_dirty(bp->b_pages[i]);
3341 
3342 		/*
3343 		 * Calculate the encompassing dirty range, boffset and eoffset,
3344 		 * (eoffset - boffset) bytes.
3345 		 */
3346 
3347 		for (i = 0; i < bp->b_npages; i++) {
3348 			if (bp->b_pages[i]->dirty)
3349 				break;
3350 		}
3351 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3352 
3353 		for (i = bp->b_npages - 1; i >= 0; --i) {
3354 			if (bp->b_pages[i]->dirty) {
3355 				break;
3356 			}
3357 		}
3358 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3359 
3360 		/*
3361 		 * Fit it to the buffer.
3362 		 */
3363 
3364 		if (eoffset > bp->b_bcount)
3365 			eoffset = bp->b_bcount;
3366 
3367 		/*
3368 		 * If we have a good dirty range, merge with the existing
3369 		 * dirty range.
3370 		 */
3371 
3372 		if (boffset < eoffset) {
3373 			if (bp->b_dirtyoff > boffset)
3374 				bp->b_dirtyoff = boffset;
3375 			if (bp->b_dirtyend < eoffset)
3376 				bp->b_dirtyend = eoffset;
3377 		}
3378 	}
3379 }
3380 
3381 /*
3382  * Allocate the KVA mapping for an existing buffer.
3383  * If an unmapped buffer is provided but a mapped buffer is requested, take
3384  * also care to properly setup mappings between pages and KVA.
3385  */
3386 static void
3387 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3388 {
3389 	int bsize, maxsize, need_mapping, need_kva;
3390 	off_t offset;
3391 
3392 	need_mapping = bp->b_data == unmapped_buf &&
3393 	    (gbflags & GB_UNMAPPED) == 0;
3394 	need_kva = bp->b_kvabase == unmapped_buf &&
3395 	    bp->b_data == unmapped_buf &&
3396 	    (gbflags & GB_KVAALLOC) != 0;
3397 	if (!need_mapping && !need_kva)
3398 		return;
3399 
3400 	BUF_CHECK_UNMAPPED(bp);
3401 
3402 	if (need_mapping && bp->b_kvabase != unmapped_buf) {
3403 		/*
3404 		 * Buffer is not mapped, but the KVA was already
3405 		 * reserved at the time of the instantiation.  Use the
3406 		 * allocated space.
3407 		 */
3408 		goto has_addr;
3409 	}
3410 
3411 	/*
3412 	 * Calculate the amount of the address space we would reserve
3413 	 * if the buffer was mapped.
3414 	 */
3415 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3416 	KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3417 	offset = blkno * bsize;
3418 	maxsize = size + (offset & PAGE_MASK);
3419 	maxsize = imax(maxsize, bsize);
3420 
3421 	while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3422 		if ((gbflags & GB_NOWAIT_BD) != 0) {
3423 			/*
3424 			 * XXXKIB: defragmentation cannot
3425 			 * succeed, not sure what else to do.
3426 			 */
3427 			panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3428 		}
3429 		atomic_add_int(&mappingrestarts, 1);
3430 		bufspace_wait(bp->b_vp, gbflags, 0, 0);
3431 	}
3432 has_addr:
3433 	if (need_mapping) {
3434 		/* b_offset is handled by bpmap_qenter. */
3435 		bp->b_data = bp->b_kvabase;
3436 		BUF_CHECK_MAPPED(bp);
3437 		bpmap_qenter(bp);
3438 	}
3439 }
3440 
3441 /*
3442  *	getblk:
3443  *
3444  *	Get a block given a specified block and offset into a file/device.
3445  *	The buffers B_DONE bit will be cleared on return, making it almost
3446  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
3447  *	return.  The caller should clear B_INVAL prior to initiating a
3448  *	READ.
3449  *
3450  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3451  *	an existing buffer.
3452  *
3453  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3454  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3455  *	and then cleared based on the backing VM.  If the previous buffer is
3456  *	non-0-sized but invalid, B_CACHE will be cleared.
3457  *
3458  *	If getblk() must create a new buffer, the new buffer is returned with
3459  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3460  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3461  *	backing VM.
3462  *
3463  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3464  *	B_CACHE bit is clear.
3465  *
3466  *	What this means, basically, is that the caller should use B_CACHE to
3467  *	determine whether the buffer is fully valid or not and should clear
3468  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3469  *	the buffer by loading its data area with something, the caller needs
3470  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3471  *	the caller should set B_CACHE ( as an optimization ), else the caller
3472  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3473  *	a write attempt or if it was a successful read.  If the caller
3474  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3475  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3476  */
3477 struct buf *
3478 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3479     int flags)
3480 {
3481 	struct buf *bp;
3482 	struct bufobj *bo;
3483 	int bsize, error, maxsize, vmio;
3484 	off_t offset;
3485 
3486 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3487 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3488 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3489 	ASSERT_VOP_LOCKED(vp, "getblk");
3490 	if (size > MAXBCACHEBUF)
3491 		panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size,
3492 		    MAXBCACHEBUF);
3493 	if (!unmapped_buf_allowed)
3494 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3495 
3496 	bo = &vp->v_bufobj;
3497 loop:
3498 	BO_RLOCK(bo);
3499 	bp = gbincore(bo, blkno);
3500 	if (bp != NULL) {
3501 		int lockflags;
3502 		/*
3503 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3504 		 * it must be on a queue.
3505 		 */
3506 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3507 
3508 		if (flags & GB_LOCK_NOWAIT)
3509 			lockflags |= LK_NOWAIT;
3510 
3511 		error = BUF_TIMELOCK(bp, lockflags,
3512 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3513 
3514 		/*
3515 		 * If we slept and got the lock we have to restart in case
3516 		 * the buffer changed identities.
3517 		 */
3518 		if (error == ENOLCK)
3519 			goto loop;
3520 		/* We timed out or were interrupted. */
3521 		else if (error)
3522 			return (NULL);
3523 		/* If recursed, assume caller knows the rules. */
3524 		else if (BUF_LOCKRECURSED(bp))
3525 			goto end;
3526 
3527 		/*
3528 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3529 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3530 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3531 		 * backing VM cache.
3532 		 */
3533 		if (bp->b_flags & B_INVAL)
3534 			bp->b_flags &= ~B_CACHE;
3535 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3536 			bp->b_flags |= B_CACHE;
3537 		if (bp->b_flags & B_MANAGED)
3538 			MPASS(bp->b_qindex == QUEUE_NONE);
3539 		else
3540 			bremfree(bp);
3541 
3542 		/*
3543 		 * check for size inconsistencies for non-VMIO case.
3544 		 */
3545 		if (bp->b_bcount != size) {
3546 			if ((bp->b_flags & B_VMIO) == 0 ||
3547 			    (size > bp->b_kvasize)) {
3548 				if (bp->b_flags & B_DELWRI) {
3549 					bp->b_flags |= B_NOCACHE;
3550 					bwrite(bp);
3551 				} else {
3552 					if (LIST_EMPTY(&bp->b_dep)) {
3553 						bp->b_flags |= B_RELBUF;
3554 						brelse(bp);
3555 					} else {
3556 						bp->b_flags |= B_NOCACHE;
3557 						bwrite(bp);
3558 					}
3559 				}
3560 				goto loop;
3561 			}
3562 		}
3563 
3564 		/*
3565 		 * Handle the case of unmapped buffer which should
3566 		 * become mapped, or the buffer for which KVA
3567 		 * reservation is requested.
3568 		 */
3569 		bp_unmapped_get_kva(bp, blkno, size, flags);
3570 
3571 		/*
3572 		 * If the size is inconsistent in the VMIO case, we can resize
3573 		 * the buffer.  This might lead to B_CACHE getting set or
3574 		 * cleared.  If the size has not changed, B_CACHE remains
3575 		 * unchanged from its previous state.
3576 		 */
3577 		allocbuf(bp, size);
3578 
3579 		KASSERT(bp->b_offset != NOOFFSET,
3580 		    ("getblk: no buffer offset"));
3581 
3582 		/*
3583 		 * A buffer with B_DELWRI set and B_CACHE clear must
3584 		 * be committed before we can return the buffer in
3585 		 * order to prevent the caller from issuing a read
3586 		 * ( due to B_CACHE not being set ) and overwriting
3587 		 * it.
3588 		 *
3589 		 * Most callers, including NFS and FFS, need this to
3590 		 * operate properly either because they assume they
3591 		 * can issue a read if B_CACHE is not set, or because
3592 		 * ( for example ) an uncached B_DELWRI might loop due
3593 		 * to softupdates re-dirtying the buffer.  In the latter
3594 		 * case, B_CACHE is set after the first write completes,
3595 		 * preventing further loops.
3596 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3597 		 * above while extending the buffer, we cannot allow the
3598 		 * buffer to remain with B_CACHE set after the write
3599 		 * completes or it will represent a corrupt state.  To
3600 		 * deal with this we set B_NOCACHE to scrap the buffer
3601 		 * after the write.
3602 		 *
3603 		 * We might be able to do something fancy, like setting
3604 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3605 		 * so the below call doesn't set B_CACHE, but that gets real
3606 		 * confusing.  This is much easier.
3607 		 */
3608 
3609 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3610 			bp->b_flags |= B_NOCACHE;
3611 			bwrite(bp);
3612 			goto loop;
3613 		}
3614 		bp->b_flags &= ~B_DONE;
3615 	} else {
3616 		/*
3617 		 * Buffer is not in-core, create new buffer.  The buffer
3618 		 * returned by getnewbuf() is locked.  Note that the returned
3619 		 * buffer is also considered valid (not marked B_INVAL).
3620 		 */
3621 		BO_RUNLOCK(bo);
3622 		/*
3623 		 * If the user does not want us to create the buffer, bail out
3624 		 * here.
3625 		 */
3626 		if (flags & GB_NOCREAT)
3627 			return NULL;
3628 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3629 			return NULL;
3630 
3631 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3632 		KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3633 		offset = blkno * bsize;
3634 		vmio = vp->v_object != NULL;
3635 		if (vmio) {
3636 			maxsize = size + (offset & PAGE_MASK);
3637 		} else {
3638 			maxsize = size;
3639 			/* Do not allow non-VMIO notmapped buffers. */
3640 			flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3641 		}
3642 		maxsize = imax(maxsize, bsize);
3643 
3644 		bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
3645 		if (bp == NULL) {
3646 			if (slpflag || slptimeo)
3647 				return NULL;
3648 			/*
3649 			 * XXX This is here until the sleep path is diagnosed
3650 			 * enough to work under very low memory conditions.
3651 			 *
3652 			 * There's an issue on low memory, 4BSD+non-preempt
3653 			 * systems (eg MIPS routers with 32MB RAM) where buffer
3654 			 * exhaustion occurs without sleeping for buffer
3655 			 * reclaimation.  This just sticks in a loop and
3656 			 * constantly attempts to allocate a buffer, which
3657 			 * hits exhaustion and tries to wakeup bufdaemon.
3658 			 * This never happens because we never yield.
3659 			 *
3660 			 * The real solution is to identify and fix these cases
3661 			 * so we aren't effectively busy-waiting in a loop
3662 			 * until the reclaimation path has cycles to run.
3663 			 */
3664 			kern_yield(PRI_USER);
3665 			goto loop;
3666 		}
3667 
3668 		/*
3669 		 * This code is used to make sure that a buffer is not
3670 		 * created while the getnewbuf routine is blocked.
3671 		 * This can be a problem whether the vnode is locked or not.
3672 		 * If the buffer is created out from under us, we have to
3673 		 * throw away the one we just created.
3674 		 *
3675 		 * Note: this must occur before we associate the buffer
3676 		 * with the vp especially considering limitations in
3677 		 * the splay tree implementation when dealing with duplicate
3678 		 * lblkno's.
3679 		 */
3680 		BO_LOCK(bo);
3681 		if (gbincore(bo, blkno)) {
3682 			BO_UNLOCK(bo);
3683 			bp->b_flags |= B_INVAL;
3684 			brelse(bp);
3685 			bufspace_release(maxsize);
3686 			goto loop;
3687 		}
3688 
3689 		/*
3690 		 * Insert the buffer into the hash, so that it can
3691 		 * be found by incore.
3692 		 */
3693 		bp->b_blkno = bp->b_lblkno = blkno;
3694 		bp->b_offset = offset;
3695 		bgetvp(vp, bp);
3696 		BO_UNLOCK(bo);
3697 
3698 		/*
3699 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3700 		 * buffer size starts out as 0, B_CACHE will be set by
3701 		 * allocbuf() for the VMIO case prior to it testing the
3702 		 * backing store for validity.
3703 		 */
3704 
3705 		if (vmio) {
3706 			bp->b_flags |= B_VMIO;
3707 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3708 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3709 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3710 		} else {
3711 			bp->b_flags &= ~B_VMIO;
3712 			KASSERT(bp->b_bufobj->bo_object == NULL,
3713 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3714 			    bp, bp->b_bufobj->bo_object));
3715 			BUF_CHECK_MAPPED(bp);
3716 		}
3717 
3718 		allocbuf(bp, size);
3719 		bufspace_release(maxsize);
3720 		bp->b_flags &= ~B_DONE;
3721 	}
3722 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3723 	BUF_ASSERT_HELD(bp);
3724 end:
3725 	buf_track(bp, __func__);
3726 	KASSERT(bp->b_bufobj == bo,
3727 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3728 	return (bp);
3729 }
3730 
3731 /*
3732  * Get an empty, disassociated buffer of given size.  The buffer is initially
3733  * set to B_INVAL.
3734  */
3735 struct buf *
3736 geteblk(int size, int flags)
3737 {
3738 	struct buf *bp;
3739 	int maxsize;
3740 
3741 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3742 	while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
3743 		if ((flags & GB_NOWAIT_BD) &&
3744 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3745 			return (NULL);
3746 	}
3747 	allocbuf(bp, size);
3748 	bufspace_release(maxsize);
3749 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3750 	BUF_ASSERT_HELD(bp);
3751 	return (bp);
3752 }
3753 
3754 /*
3755  * Truncate the backing store for a non-vmio buffer.
3756  */
3757 static void
3758 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
3759 {
3760 
3761 	if (bp->b_flags & B_MALLOC) {
3762 		/*
3763 		 * malloced buffers are not shrunk
3764 		 */
3765 		if (newbsize == 0) {
3766 			bufmallocadjust(bp, 0);
3767 			free(bp->b_data, M_BIOBUF);
3768 			bp->b_data = bp->b_kvabase;
3769 			bp->b_flags &= ~B_MALLOC;
3770 		}
3771 		return;
3772 	}
3773 	vm_hold_free_pages(bp, newbsize);
3774 	bufspace_adjust(bp, newbsize);
3775 }
3776 
3777 /*
3778  * Extend the backing for a non-VMIO buffer.
3779  */
3780 static void
3781 vfs_nonvmio_extend(struct buf *bp, int newbsize)
3782 {
3783 	caddr_t origbuf;
3784 	int origbufsize;
3785 
3786 	/*
3787 	 * We only use malloced memory on the first allocation.
3788 	 * and revert to page-allocated memory when the buffer
3789 	 * grows.
3790 	 *
3791 	 * There is a potential smp race here that could lead
3792 	 * to bufmallocspace slightly passing the max.  It
3793 	 * is probably extremely rare and not worth worrying
3794 	 * over.
3795 	 */
3796 	if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
3797 	    bufmallocspace < maxbufmallocspace) {
3798 		bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
3799 		bp->b_flags |= B_MALLOC;
3800 		bufmallocadjust(bp, newbsize);
3801 		return;
3802 	}
3803 
3804 	/*
3805 	 * If the buffer is growing on its other-than-first
3806 	 * allocation then we revert to the page-allocation
3807 	 * scheme.
3808 	 */
3809 	origbuf = NULL;
3810 	origbufsize = 0;
3811 	if (bp->b_flags & B_MALLOC) {
3812 		origbuf = bp->b_data;
3813 		origbufsize = bp->b_bufsize;
3814 		bp->b_data = bp->b_kvabase;
3815 		bufmallocadjust(bp, 0);
3816 		bp->b_flags &= ~B_MALLOC;
3817 		newbsize = round_page(newbsize);
3818 	}
3819 	vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
3820 	    (vm_offset_t) bp->b_data + newbsize);
3821 	if (origbuf != NULL) {
3822 		bcopy(origbuf, bp->b_data, origbufsize);
3823 		free(origbuf, M_BIOBUF);
3824 	}
3825 	bufspace_adjust(bp, newbsize);
3826 }
3827 
3828 /*
3829  * This code constitutes the buffer memory from either anonymous system
3830  * memory (in the case of non-VMIO operations) or from an associated
3831  * VM object (in the case of VMIO operations).  This code is able to
3832  * resize a buffer up or down.
3833  *
3834  * Note that this code is tricky, and has many complications to resolve
3835  * deadlock or inconsistent data situations.  Tread lightly!!!
3836  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3837  * the caller.  Calling this code willy nilly can result in the loss of data.
3838  *
3839  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3840  * B_CACHE for the non-VMIO case.
3841  */
3842 int
3843 allocbuf(struct buf *bp, int size)
3844 {
3845 	int newbsize;
3846 
3847 	BUF_ASSERT_HELD(bp);
3848 
3849 	if (bp->b_bcount == size)
3850 		return (1);
3851 
3852 	if (bp->b_kvasize != 0 && bp->b_kvasize < size)
3853 		panic("allocbuf: buffer too small");
3854 
3855 	newbsize = roundup2(size, DEV_BSIZE);
3856 	if ((bp->b_flags & B_VMIO) == 0) {
3857 		if ((bp->b_flags & B_MALLOC) == 0)
3858 			newbsize = round_page(newbsize);
3859 		/*
3860 		 * Just get anonymous memory from the kernel.  Don't
3861 		 * mess with B_CACHE.
3862 		 */
3863 		if (newbsize < bp->b_bufsize)
3864 			vfs_nonvmio_truncate(bp, newbsize);
3865 		else if (newbsize > bp->b_bufsize)
3866 			vfs_nonvmio_extend(bp, newbsize);
3867 	} else {
3868 		int desiredpages;
3869 
3870 		desiredpages = (size == 0) ? 0 :
3871 		    num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3872 
3873 		if (bp->b_flags & B_MALLOC)
3874 			panic("allocbuf: VMIO buffer can't be malloced");
3875 		/*
3876 		 * Set B_CACHE initially if buffer is 0 length or will become
3877 		 * 0-length.
3878 		 */
3879 		if (size == 0 || bp->b_bufsize == 0)
3880 			bp->b_flags |= B_CACHE;
3881 
3882 		if (newbsize < bp->b_bufsize)
3883 			vfs_vmio_truncate(bp, desiredpages);
3884 		/* XXX This looks as if it should be newbsize > b_bufsize */
3885 		else if (size > bp->b_bcount)
3886 			vfs_vmio_extend(bp, desiredpages, size);
3887 		bufspace_adjust(bp, newbsize);
3888 	}
3889 	bp->b_bcount = size;		/* requested buffer size. */
3890 	return (1);
3891 }
3892 
3893 extern int inflight_transient_maps;
3894 
3895 void
3896 biodone(struct bio *bp)
3897 {
3898 	struct mtx *mtxp;
3899 	void (*done)(struct bio *);
3900 	vm_offset_t start, end;
3901 
3902 	biotrack(bp, __func__);
3903 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3904 		bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3905 		bp->bio_flags |= BIO_UNMAPPED;
3906 		start = trunc_page((vm_offset_t)bp->bio_data);
3907 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3908 		bp->bio_data = unmapped_buf;
3909 		pmap_qremove(start, OFF_TO_IDX(end - start));
3910 		vmem_free(transient_arena, start, end - start);
3911 		atomic_add_int(&inflight_transient_maps, -1);
3912 	}
3913 	done = bp->bio_done;
3914 	if (done == NULL) {
3915 		mtxp = mtx_pool_find(mtxpool_sleep, bp);
3916 		mtx_lock(mtxp);
3917 		bp->bio_flags |= BIO_DONE;
3918 		wakeup(bp);
3919 		mtx_unlock(mtxp);
3920 	} else {
3921 		bp->bio_flags |= BIO_DONE;
3922 		done(bp);
3923 	}
3924 }
3925 
3926 /*
3927  * Wait for a BIO to finish.
3928  */
3929 int
3930 biowait(struct bio *bp, const char *wchan)
3931 {
3932 	struct mtx *mtxp;
3933 
3934 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3935 	mtx_lock(mtxp);
3936 	while ((bp->bio_flags & BIO_DONE) == 0)
3937 		msleep(bp, mtxp, PRIBIO, wchan, 0);
3938 	mtx_unlock(mtxp);
3939 	if (bp->bio_error != 0)
3940 		return (bp->bio_error);
3941 	if (!(bp->bio_flags & BIO_ERROR))
3942 		return (0);
3943 	return (EIO);
3944 }
3945 
3946 void
3947 biofinish(struct bio *bp, struct devstat *stat, int error)
3948 {
3949 
3950 	if (error) {
3951 		bp->bio_error = error;
3952 		bp->bio_flags |= BIO_ERROR;
3953 	}
3954 	if (stat != NULL)
3955 		devstat_end_transaction_bio(stat, bp);
3956 	biodone(bp);
3957 }
3958 
3959 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
3960 void
3961 biotrack_buf(struct bio *bp, const char *location)
3962 {
3963 
3964 	buf_track(bp->bio_track_bp, location);
3965 }
3966 #endif
3967 
3968 /*
3969  *	bufwait:
3970  *
3971  *	Wait for buffer I/O completion, returning error status.  The buffer
3972  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3973  *	error and cleared.
3974  */
3975 int
3976 bufwait(struct buf *bp)
3977 {
3978 	if (bp->b_iocmd == BIO_READ)
3979 		bwait(bp, PRIBIO, "biord");
3980 	else
3981 		bwait(bp, PRIBIO, "biowr");
3982 	if (bp->b_flags & B_EINTR) {
3983 		bp->b_flags &= ~B_EINTR;
3984 		return (EINTR);
3985 	}
3986 	if (bp->b_ioflags & BIO_ERROR) {
3987 		return (bp->b_error ? bp->b_error : EIO);
3988 	} else {
3989 		return (0);
3990 	}
3991 }
3992 
3993 /*
3994  *	bufdone:
3995  *
3996  *	Finish I/O on a buffer, optionally calling a completion function.
3997  *	This is usually called from an interrupt so process blocking is
3998  *	not allowed.
3999  *
4000  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4001  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
4002  *	assuming B_INVAL is clear.
4003  *
4004  *	For the VMIO case, we set B_CACHE if the op was a read and no
4005  *	read error occurred, or if the op was a write.  B_CACHE is never
4006  *	set if the buffer is invalid or otherwise uncacheable.
4007  *
4008  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
4009  *	initiator to leave B_INVAL set to brelse the buffer out of existence
4010  *	in the biodone routine.
4011  */
4012 void
4013 bufdone(struct buf *bp)
4014 {
4015 	struct bufobj *dropobj;
4016 	void    (*biodone)(struct buf *);
4017 
4018 	buf_track(bp, __func__);
4019 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4020 	dropobj = NULL;
4021 
4022 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4023 	BUF_ASSERT_HELD(bp);
4024 
4025 	runningbufwakeup(bp);
4026 	if (bp->b_iocmd == BIO_WRITE)
4027 		dropobj = bp->b_bufobj;
4028 	/* call optional completion function if requested */
4029 	if (bp->b_iodone != NULL) {
4030 		biodone = bp->b_iodone;
4031 		bp->b_iodone = NULL;
4032 		(*biodone) (bp);
4033 		if (dropobj)
4034 			bufobj_wdrop(dropobj);
4035 		return;
4036 	}
4037 
4038 	bufdone_finish(bp);
4039 
4040 	if (dropobj)
4041 		bufobj_wdrop(dropobj);
4042 }
4043 
4044 void
4045 bufdone_finish(struct buf *bp)
4046 {
4047 	BUF_ASSERT_HELD(bp);
4048 
4049 	if (!LIST_EMPTY(&bp->b_dep))
4050 		buf_complete(bp);
4051 
4052 	if (bp->b_flags & B_VMIO) {
4053 		/*
4054 		 * Set B_CACHE if the op was a normal read and no error
4055 		 * occurred.  B_CACHE is set for writes in the b*write()
4056 		 * routines.
4057 		 */
4058 		if (bp->b_iocmd == BIO_READ &&
4059 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4060 		    !(bp->b_ioflags & BIO_ERROR))
4061 			bp->b_flags |= B_CACHE;
4062 		vfs_vmio_iodone(bp);
4063 	}
4064 
4065 	/*
4066 	 * For asynchronous completions, release the buffer now. The brelse
4067 	 * will do a wakeup there if necessary - so no need to do a wakeup
4068 	 * here in the async case. The sync case always needs to do a wakeup.
4069 	 */
4070 	if (bp->b_flags & B_ASYNC) {
4071 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4072 		    (bp->b_ioflags & BIO_ERROR))
4073 			brelse(bp);
4074 		else
4075 			bqrelse(bp);
4076 	} else
4077 		bdone(bp);
4078 }
4079 
4080 /*
4081  * This routine is called in lieu of iodone in the case of
4082  * incomplete I/O.  This keeps the busy status for pages
4083  * consistent.
4084  */
4085 void
4086 vfs_unbusy_pages(struct buf *bp)
4087 {
4088 	int i;
4089 	vm_object_t obj;
4090 	vm_page_t m;
4091 
4092 	runningbufwakeup(bp);
4093 	if (!(bp->b_flags & B_VMIO))
4094 		return;
4095 
4096 	obj = bp->b_bufobj->bo_object;
4097 	VM_OBJECT_WLOCK(obj);
4098 	for (i = 0; i < bp->b_npages; i++) {
4099 		m = bp->b_pages[i];
4100 		if (m == bogus_page) {
4101 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4102 			if (!m)
4103 				panic("vfs_unbusy_pages: page missing\n");
4104 			bp->b_pages[i] = m;
4105 			if (buf_mapped(bp)) {
4106 				BUF_CHECK_MAPPED(bp);
4107 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4108 				    bp->b_pages, bp->b_npages);
4109 			} else
4110 				BUF_CHECK_UNMAPPED(bp);
4111 		}
4112 		vm_page_sunbusy(m);
4113 	}
4114 	vm_object_pip_wakeupn(obj, bp->b_npages);
4115 	VM_OBJECT_WUNLOCK(obj);
4116 }
4117 
4118 /*
4119  * vfs_page_set_valid:
4120  *
4121  *	Set the valid bits in a page based on the supplied offset.   The
4122  *	range is restricted to the buffer's size.
4123  *
4124  *	This routine is typically called after a read completes.
4125  */
4126 static void
4127 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4128 {
4129 	vm_ooffset_t eoff;
4130 
4131 	/*
4132 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
4133 	 * page boundary and eoff is not greater than the end of the buffer.
4134 	 * The end of the buffer, in this case, is our file EOF, not the
4135 	 * allocation size of the buffer.
4136 	 */
4137 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4138 	if (eoff > bp->b_offset + bp->b_bcount)
4139 		eoff = bp->b_offset + bp->b_bcount;
4140 
4141 	/*
4142 	 * Set valid range.  This is typically the entire buffer and thus the
4143 	 * entire page.
4144 	 */
4145 	if (eoff > off)
4146 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4147 }
4148 
4149 /*
4150  * vfs_page_set_validclean:
4151  *
4152  *	Set the valid bits and clear the dirty bits in a page based on the
4153  *	supplied offset.   The range is restricted to the buffer's size.
4154  */
4155 static void
4156 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4157 {
4158 	vm_ooffset_t soff, eoff;
4159 
4160 	/*
4161 	 * Start and end offsets in buffer.  eoff - soff may not cross a
4162 	 * page boundary or cross the end of the buffer.  The end of the
4163 	 * buffer, in this case, is our file EOF, not the allocation size
4164 	 * of the buffer.
4165 	 */
4166 	soff = off;
4167 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4168 	if (eoff > bp->b_offset + bp->b_bcount)
4169 		eoff = bp->b_offset + bp->b_bcount;
4170 
4171 	/*
4172 	 * Set valid range.  This is typically the entire buffer and thus the
4173 	 * entire page.
4174 	 */
4175 	if (eoff > soff) {
4176 		vm_page_set_validclean(
4177 		    m,
4178 		   (vm_offset_t) (soff & PAGE_MASK),
4179 		   (vm_offset_t) (eoff - soff)
4180 		);
4181 	}
4182 }
4183 
4184 /*
4185  * Ensure that all buffer pages are not exclusive busied.  If any page is
4186  * exclusive busy, drain it.
4187  */
4188 void
4189 vfs_drain_busy_pages(struct buf *bp)
4190 {
4191 	vm_page_t m;
4192 	int i, last_busied;
4193 
4194 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4195 	last_busied = 0;
4196 	for (i = 0; i < bp->b_npages; i++) {
4197 		m = bp->b_pages[i];
4198 		if (vm_page_xbusied(m)) {
4199 			for (; last_busied < i; last_busied++)
4200 				vm_page_sbusy(bp->b_pages[last_busied]);
4201 			while (vm_page_xbusied(m)) {
4202 				vm_page_lock(m);
4203 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4204 				vm_page_busy_sleep(m, "vbpage", true);
4205 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4206 			}
4207 		}
4208 	}
4209 	for (i = 0; i < last_busied; i++)
4210 		vm_page_sunbusy(bp->b_pages[i]);
4211 }
4212 
4213 /*
4214  * This routine is called before a device strategy routine.
4215  * It is used to tell the VM system that paging I/O is in
4216  * progress, and treat the pages associated with the buffer
4217  * almost as being exclusive busy.  Also the object paging_in_progress
4218  * flag is handled to make sure that the object doesn't become
4219  * inconsistent.
4220  *
4221  * Since I/O has not been initiated yet, certain buffer flags
4222  * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4223  * and should be ignored.
4224  */
4225 void
4226 vfs_busy_pages(struct buf *bp, int clear_modify)
4227 {
4228 	int i, bogus;
4229 	vm_object_t obj;
4230 	vm_ooffset_t foff;
4231 	vm_page_t m;
4232 
4233 	if (!(bp->b_flags & B_VMIO))
4234 		return;
4235 
4236 	obj = bp->b_bufobj->bo_object;
4237 	foff = bp->b_offset;
4238 	KASSERT(bp->b_offset != NOOFFSET,
4239 	    ("vfs_busy_pages: no buffer offset"));
4240 	VM_OBJECT_WLOCK(obj);
4241 	vfs_drain_busy_pages(bp);
4242 	if (bp->b_bufsize != 0)
4243 		vfs_setdirty_locked_object(bp);
4244 	bogus = 0;
4245 	for (i = 0; i < bp->b_npages; i++) {
4246 		m = bp->b_pages[i];
4247 
4248 		if ((bp->b_flags & B_CLUSTER) == 0) {
4249 			vm_object_pip_add(obj, 1);
4250 			vm_page_sbusy(m);
4251 		}
4252 		/*
4253 		 * When readying a buffer for a read ( i.e
4254 		 * clear_modify == 0 ), it is important to do
4255 		 * bogus_page replacement for valid pages in
4256 		 * partially instantiated buffers.  Partially
4257 		 * instantiated buffers can, in turn, occur when
4258 		 * reconstituting a buffer from its VM backing store
4259 		 * base.  We only have to do this if B_CACHE is
4260 		 * clear ( which causes the I/O to occur in the
4261 		 * first place ).  The replacement prevents the read
4262 		 * I/O from overwriting potentially dirty VM-backed
4263 		 * pages.  XXX bogus page replacement is, uh, bogus.
4264 		 * It may not work properly with small-block devices.
4265 		 * We need to find a better way.
4266 		 */
4267 		if (clear_modify) {
4268 			pmap_remove_write(m);
4269 			vfs_page_set_validclean(bp, foff, m);
4270 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4271 		    (bp->b_flags & B_CACHE) == 0) {
4272 			bp->b_pages[i] = bogus_page;
4273 			bogus++;
4274 		}
4275 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4276 	}
4277 	VM_OBJECT_WUNLOCK(obj);
4278 	if (bogus && buf_mapped(bp)) {
4279 		BUF_CHECK_MAPPED(bp);
4280 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4281 		    bp->b_pages, bp->b_npages);
4282 	}
4283 }
4284 
4285 /*
4286  *	vfs_bio_set_valid:
4287  *
4288  *	Set the range within the buffer to valid.  The range is
4289  *	relative to the beginning of the buffer, b_offset.  Note that
4290  *	b_offset itself may be offset from the beginning of the first
4291  *	page.
4292  */
4293 void
4294 vfs_bio_set_valid(struct buf *bp, int base, int size)
4295 {
4296 	int i, n;
4297 	vm_page_t m;
4298 
4299 	if (!(bp->b_flags & B_VMIO))
4300 		return;
4301 
4302 	/*
4303 	 * Fixup base to be relative to beginning of first page.
4304 	 * Set initial n to be the maximum number of bytes in the
4305 	 * first page that can be validated.
4306 	 */
4307 	base += (bp->b_offset & PAGE_MASK);
4308 	n = PAGE_SIZE - (base & PAGE_MASK);
4309 
4310 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4311 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4312 		m = bp->b_pages[i];
4313 		if (n > size)
4314 			n = size;
4315 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4316 		base += n;
4317 		size -= n;
4318 		n = PAGE_SIZE;
4319 	}
4320 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4321 }
4322 
4323 /*
4324  *	vfs_bio_clrbuf:
4325  *
4326  *	If the specified buffer is a non-VMIO buffer, clear the entire
4327  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4328  *	validate only the previously invalid portions of the buffer.
4329  *	This routine essentially fakes an I/O, so we need to clear
4330  *	BIO_ERROR and B_INVAL.
4331  *
4332  *	Note that while we only theoretically need to clear through b_bcount,
4333  *	we go ahead and clear through b_bufsize.
4334  */
4335 void
4336 vfs_bio_clrbuf(struct buf *bp)
4337 {
4338 	int i, j, mask, sa, ea, slide;
4339 
4340 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4341 		clrbuf(bp);
4342 		return;
4343 	}
4344 	bp->b_flags &= ~B_INVAL;
4345 	bp->b_ioflags &= ~BIO_ERROR;
4346 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4347 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4348 	    (bp->b_offset & PAGE_MASK) == 0) {
4349 		if (bp->b_pages[0] == bogus_page)
4350 			goto unlock;
4351 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4352 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4353 		if ((bp->b_pages[0]->valid & mask) == mask)
4354 			goto unlock;
4355 		if ((bp->b_pages[0]->valid & mask) == 0) {
4356 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4357 			bp->b_pages[0]->valid |= mask;
4358 			goto unlock;
4359 		}
4360 	}
4361 	sa = bp->b_offset & PAGE_MASK;
4362 	slide = 0;
4363 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4364 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4365 		ea = slide & PAGE_MASK;
4366 		if (ea == 0)
4367 			ea = PAGE_SIZE;
4368 		if (bp->b_pages[i] == bogus_page)
4369 			continue;
4370 		j = sa / DEV_BSIZE;
4371 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4372 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4373 		if ((bp->b_pages[i]->valid & mask) == mask)
4374 			continue;
4375 		if ((bp->b_pages[i]->valid & mask) == 0)
4376 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4377 		else {
4378 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4379 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4380 					pmap_zero_page_area(bp->b_pages[i],
4381 					    sa, DEV_BSIZE);
4382 				}
4383 			}
4384 		}
4385 		bp->b_pages[i]->valid |= mask;
4386 	}
4387 unlock:
4388 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4389 	bp->b_resid = 0;
4390 }
4391 
4392 void
4393 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4394 {
4395 	vm_page_t m;
4396 	int i, n;
4397 
4398 	if (buf_mapped(bp)) {
4399 		BUF_CHECK_MAPPED(bp);
4400 		bzero(bp->b_data + base, size);
4401 	} else {
4402 		BUF_CHECK_UNMAPPED(bp);
4403 		n = PAGE_SIZE - (base & PAGE_MASK);
4404 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4405 			m = bp->b_pages[i];
4406 			if (n > size)
4407 				n = size;
4408 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4409 			base += n;
4410 			size -= n;
4411 			n = PAGE_SIZE;
4412 		}
4413 	}
4414 }
4415 
4416 /*
4417  * Update buffer flags based on I/O request parameters, optionally releasing the
4418  * buffer.  If it's VMIO or direct I/O, the buffer pages are released to the VM,
4419  * where they may be placed on a page queue (VMIO) or freed immediately (direct
4420  * I/O).  Otherwise the buffer is released to the cache.
4421  */
4422 static void
4423 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4424 {
4425 
4426 	KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4427 	    ("buf %p non-VMIO noreuse", bp));
4428 
4429 	if ((ioflag & IO_DIRECT) != 0)
4430 		bp->b_flags |= B_DIRECT;
4431 	if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4432 		bp->b_flags |= B_RELBUF;
4433 		if ((ioflag & IO_NOREUSE) != 0)
4434 			bp->b_flags |= B_NOREUSE;
4435 		if (release)
4436 			brelse(bp);
4437 	} else if (release)
4438 		bqrelse(bp);
4439 }
4440 
4441 void
4442 vfs_bio_brelse(struct buf *bp, int ioflag)
4443 {
4444 
4445 	b_io_dismiss(bp, ioflag, true);
4446 }
4447 
4448 void
4449 vfs_bio_set_flags(struct buf *bp, int ioflag)
4450 {
4451 
4452 	b_io_dismiss(bp, ioflag, false);
4453 }
4454 
4455 /*
4456  * vm_hold_load_pages and vm_hold_free_pages get pages into
4457  * a buffers address space.  The pages are anonymous and are
4458  * not associated with a file object.
4459  */
4460 static void
4461 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4462 {
4463 	vm_offset_t pg;
4464 	vm_page_t p;
4465 	int index;
4466 
4467 	BUF_CHECK_MAPPED(bp);
4468 
4469 	to = round_page(to);
4470 	from = round_page(from);
4471 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4472 
4473 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4474 tryagain:
4475 		/*
4476 		 * note: must allocate system pages since blocking here
4477 		 * could interfere with paging I/O, no matter which
4478 		 * process we are.
4479 		 */
4480 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4481 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4482 		if (p == NULL) {
4483 			VM_WAIT;
4484 			goto tryagain;
4485 		}
4486 		pmap_qenter(pg, &p, 1);
4487 		bp->b_pages[index] = p;
4488 	}
4489 	bp->b_npages = index;
4490 }
4491 
4492 /* Return pages associated with this buf to the vm system */
4493 static void
4494 vm_hold_free_pages(struct buf *bp, int newbsize)
4495 {
4496 	vm_offset_t from;
4497 	vm_page_t p;
4498 	int index, newnpages;
4499 
4500 	BUF_CHECK_MAPPED(bp);
4501 
4502 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4503 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4504 	if (bp->b_npages > newnpages)
4505 		pmap_qremove(from, bp->b_npages - newnpages);
4506 	for (index = newnpages; index < bp->b_npages; index++) {
4507 		p = bp->b_pages[index];
4508 		bp->b_pages[index] = NULL;
4509 		if (vm_page_sbusied(p))
4510 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4511 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4512 		p->wire_count--;
4513 		vm_page_free(p);
4514 		atomic_subtract_int(&vm_cnt.v_wire_count, 1);
4515 	}
4516 	bp->b_npages = newnpages;
4517 }
4518 
4519 /*
4520  * Map an IO request into kernel virtual address space.
4521  *
4522  * All requests are (re)mapped into kernel VA space.
4523  * Notice that we use b_bufsize for the size of the buffer
4524  * to be mapped.  b_bcount might be modified by the driver.
4525  *
4526  * Note that even if the caller determines that the address space should
4527  * be valid, a race or a smaller-file mapped into a larger space may
4528  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4529  * check the return value.
4530  *
4531  * This function only works with pager buffers.
4532  */
4533 int
4534 vmapbuf(struct buf *bp, int mapbuf)
4535 {
4536 	vm_prot_t prot;
4537 	int pidx;
4538 
4539 	if (bp->b_bufsize < 0)
4540 		return (-1);
4541 	prot = VM_PROT_READ;
4542 	if (bp->b_iocmd == BIO_READ)
4543 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4544 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4545 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4546 	    btoc(MAXPHYS))) < 0)
4547 		return (-1);
4548 	bp->b_npages = pidx;
4549 	bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4550 	if (mapbuf || !unmapped_buf_allowed) {
4551 		pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4552 		bp->b_data = bp->b_kvabase + bp->b_offset;
4553 	} else
4554 		bp->b_data = unmapped_buf;
4555 	return(0);
4556 }
4557 
4558 /*
4559  * Free the io map PTEs associated with this IO operation.
4560  * We also invalidate the TLB entries and restore the original b_addr.
4561  *
4562  * This function only works with pager buffers.
4563  */
4564 void
4565 vunmapbuf(struct buf *bp)
4566 {
4567 	int npages;
4568 
4569 	npages = bp->b_npages;
4570 	if (buf_mapped(bp))
4571 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4572 	vm_page_unhold_pages(bp->b_pages, npages);
4573 
4574 	bp->b_data = unmapped_buf;
4575 }
4576 
4577 void
4578 bdone(struct buf *bp)
4579 {
4580 	struct mtx *mtxp;
4581 
4582 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4583 	mtx_lock(mtxp);
4584 	bp->b_flags |= B_DONE;
4585 	wakeup(bp);
4586 	mtx_unlock(mtxp);
4587 }
4588 
4589 void
4590 bwait(struct buf *bp, u_char pri, const char *wchan)
4591 {
4592 	struct mtx *mtxp;
4593 
4594 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4595 	mtx_lock(mtxp);
4596 	while ((bp->b_flags & B_DONE) == 0)
4597 		msleep(bp, mtxp, pri, wchan, 0);
4598 	mtx_unlock(mtxp);
4599 }
4600 
4601 int
4602 bufsync(struct bufobj *bo, int waitfor)
4603 {
4604 
4605 	return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4606 }
4607 
4608 void
4609 bufstrategy(struct bufobj *bo, struct buf *bp)
4610 {
4611 	int i = 0;
4612 	struct vnode *vp;
4613 
4614 	vp = bp->b_vp;
4615 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4616 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4617 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4618 	i = VOP_STRATEGY(vp, bp);
4619 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4620 }
4621 
4622 void
4623 bufobj_wrefl(struct bufobj *bo)
4624 {
4625 
4626 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4627 	ASSERT_BO_WLOCKED(bo);
4628 	bo->bo_numoutput++;
4629 }
4630 
4631 void
4632 bufobj_wref(struct bufobj *bo)
4633 {
4634 
4635 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4636 	BO_LOCK(bo);
4637 	bo->bo_numoutput++;
4638 	BO_UNLOCK(bo);
4639 }
4640 
4641 void
4642 bufobj_wdrop(struct bufobj *bo)
4643 {
4644 
4645 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4646 	BO_LOCK(bo);
4647 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4648 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4649 		bo->bo_flag &= ~BO_WWAIT;
4650 		wakeup(&bo->bo_numoutput);
4651 	}
4652 	BO_UNLOCK(bo);
4653 }
4654 
4655 int
4656 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4657 {
4658 	int error;
4659 
4660 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4661 	ASSERT_BO_WLOCKED(bo);
4662 	error = 0;
4663 	while (bo->bo_numoutput) {
4664 		bo->bo_flag |= BO_WWAIT;
4665 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4666 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4667 		if (error)
4668 			break;
4669 	}
4670 	return (error);
4671 }
4672 
4673 /*
4674  * Set bio_data or bio_ma for struct bio from the struct buf.
4675  */
4676 void
4677 bdata2bio(struct buf *bp, struct bio *bip)
4678 {
4679 
4680 	if (!buf_mapped(bp)) {
4681 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4682 		bip->bio_ma = bp->b_pages;
4683 		bip->bio_ma_n = bp->b_npages;
4684 		bip->bio_data = unmapped_buf;
4685 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4686 		bip->bio_flags |= BIO_UNMAPPED;
4687 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4688 		    PAGE_SIZE == bp->b_npages,
4689 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4690 		    (long long)bip->bio_length, bip->bio_ma_n));
4691 	} else {
4692 		bip->bio_data = bp->b_data;
4693 		bip->bio_ma = NULL;
4694 	}
4695 }
4696 
4697 /*
4698  * The MIPS pmap code currently doesn't handle aliased pages.
4699  * The VIPT caches may not handle page aliasing themselves, leading
4700  * to data corruption.
4701  *
4702  * As such, this code makes a system extremely unhappy if said
4703  * system doesn't support unaliasing the above situation in hardware.
4704  * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
4705  * this feature at build time, so it has to be handled in software.
4706  *
4707  * Once the MIPS pmap/cache code grows to support this function on
4708  * earlier chips, it should be flipped back off.
4709  */
4710 #ifdef	__mips__
4711 static int buf_pager_relbuf = 1;
4712 #else
4713 static int buf_pager_relbuf = 0;
4714 #endif
4715 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
4716     &buf_pager_relbuf, 0,
4717     "Make buffer pager release buffers after reading");
4718 
4719 /*
4720  * The buffer pager.  It uses buffer reads to validate pages.
4721  *
4722  * In contrast to the generic local pager from vm/vnode_pager.c, this
4723  * pager correctly and easily handles volumes where the underlying
4724  * device block size is greater than the machine page size.  The
4725  * buffer cache transparently extends the requested page run to be
4726  * aligned at the block boundary, and does the necessary bogus page
4727  * replacements in the addends to avoid obliterating already valid
4728  * pages.
4729  *
4730  * The only non-trivial issue is that the exclusive busy state for
4731  * pages, which is assumed by the vm_pager_getpages() interface, is
4732  * incompatible with the VMIO buffer cache's desire to share-busy the
4733  * pages.  This function performs a trivial downgrade of the pages'
4734  * state before reading buffers, and a less trivial upgrade from the
4735  * shared-busy to excl-busy state after the read.
4736  */
4737 int
4738 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
4739     int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
4740     vbg_get_blksize_t get_blksize)
4741 {
4742 	vm_page_t m;
4743 	vm_object_t object;
4744 	struct buf *bp;
4745 	struct mount *mp;
4746 	daddr_t lbn, lbnp;
4747 	vm_ooffset_t la, lb, poff, poffe;
4748 	long bsize;
4749 	int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
4750 	bool redo, lpart;
4751 
4752 	object = vp->v_object;
4753 	mp = vp->v_mount;
4754 	la = IDX_TO_OFF(ma[count - 1]->pindex);
4755 	if (la >= object->un_pager.vnp.vnp_size)
4756 		return (VM_PAGER_BAD);
4757 	lpart = la + PAGE_SIZE > object->un_pager.vnp.vnp_size;
4758 	bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
4759 
4760 	/*
4761 	 * Calculate read-ahead, behind and total pages.
4762 	 */
4763 	pgsin = count;
4764 	lb = IDX_TO_OFF(ma[0]->pindex);
4765 	pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
4766 	pgsin += pgsin_b;
4767 	if (rbehind != NULL)
4768 		*rbehind = pgsin_b;
4769 	pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
4770 	if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
4771 		pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
4772 		    PAGE_SIZE) - la);
4773 	pgsin += pgsin_a;
4774 	if (rahead != NULL)
4775 		*rahead = pgsin_a;
4776 	PCPU_INC(cnt.v_vnodein);
4777 	PCPU_ADD(cnt.v_vnodepgsin, pgsin);
4778 
4779 	br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
4780 	    != 0) ? GB_UNMAPPED : 0;
4781 	VM_OBJECT_WLOCK(object);
4782 again:
4783 	for (i = 0; i < count; i++)
4784 		vm_page_busy_downgrade(ma[i]);
4785 	VM_OBJECT_WUNLOCK(object);
4786 
4787 	lbnp = -1;
4788 	for (i = 0; i < count; i++) {
4789 		m = ma[i];
4790 
4791 		/*
4792 		 * Pages are shared busy and the object lock is not
4793 		 * owned, which together allow for the pages'
4794 		 * invalidation.  The racy test for validity avoids
4795 		 * useless creation of the buffer for the most typical
4796 		 * case when invalidation is not used in redo or for
4797 		 * parallel read.  The shared->excl upgrade loop at
4798 		 * the end of the function catches the race in a
4799 		 * reliable way (protected by the object lock).
4800 		 */
4801 		if (m->valid == VM_PAGE_BITS_ALL)
4802 			continue;
4803 
4804 		poff = IDX_TO_OFF(m->pindex);
4805 		poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
4806 		for (; poff < poffe; poff += bsize) {
4807 			lbn = get_lblkno(vp, poff);
4808 			if (lbn == lbnp)
4809 				goto next_page;
4810 			lbnp = lbn;
4811 
4812 			bsize = get_blksize(vp, lbn);
4813 			error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
4814 			    br_flags, &bp);
4815 			if (error != 0)
4816 				goto end_pages;
4817 			if (LIST_EMPTY(&bp->b_dep)) {
4818 				/*
4819 				 * Invalidation clears m->valid, but
4820 				 * may leave B_CACHE flag if the
4821 				 * buffer existed at the invalidation
4822 				 * time.  In this case, recycle the
4823 				 * buffer to do real read on next
4824 				 * bread() after redo.
4825 				 *
4826 				 * Otherwise B_RELBUF is not strictly
4827 				 * necessary, enable to reduce buf
4828 				 * cache pressure.
4829 				 */
4830 				if (buf_pager_relbuf ||
4831 				    m->valid != VM_PAGE_BITS_ALL)
4832 					bp->b_flags |= B_RELBUF;
4833 
4834 				bp->b_flags &= ~B_NOCACHE;
4835 				brelse(bp);
4836 			} else {
4837 				bqrelse(bp);
4838 			}
4839 		}
4840 		KASSERT(1 /* racy, enable for debugging */ ||
4841 		    m->valid == VM_PAGE_BITS_ALL || i == count - 1,
4842 		    ("buf %d %p invalid", i, m));
4843 		if (i == count - 1 && lpart) {
4844 			VM_OBJECT_WLOCK(object);
4845 			if (m->valid != 0 &&
4846 			    m->valid != VM_PAGE_BITS_ALL)
4847 				vm_page_zero_invalid(m, TRUE);
4848 			VM_OBJECT_WUNLOCK(object);
4849 		}
4850 next_page:;
4851 	}
4852 end_pages:
4853 
4854 	VM_OBJECT_WLOCK(object);
4855 	redo = false;
4856 	for (i = 0; i < count; i++) {
4857 		vm_page_sunbusy(ma[i]);
4858 		ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
4859 
4860 		/*
4861 		 * Since the pages were only sbusy while neither the
4862 		 * buffer nor the object lock was held by us, or
4863 		 * reallocated while vm_page_grab() slept for busy
4864 		 * relinguish, they could have been invalidated.
4865 		 * Recheck the valid bits and re-read as needed.
4866 		 *
4867 		 * Note that the last page is made fully valid in the
4868 		 * read loop, and partial validity for the page at
4869 		 * index count - 1 could mean that the page was
4870 		 * invalidated or removed, so we must restart for
4871 		 * safety as well.
4872 		 */
4873 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
4874 			redo = true;
4875 	}
4876 	if (redo && error == 0)
4877 		goto again;
4878 	VM_OBJECT_WUNLOCK(object);
4879 	return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
4880 }
4881 
4882 #include "opt_ddb.h"
4883 #ifdef DDB
4884 #include <ddb/ddb.h>
4885 
4886 /* DDB command to show buffer data */
4887 DB_SHOW_COMMAND(buffer, db_show_buffer)
4888 {
4889 	/* get args */
4890 	struct buf *bp = (struct buf *)addr;
4891 #ifdef FULL_BUF_TRACKING
4892 	uint32_t i, j;
4893 #endif
4894 
4895 	if (!have_addr) {
4896 		db_printf("usage: show buffer <addr>\n");
4897 		return;
4898 	}
4899 
4900 	db_printf("buf at %p\n", bp);
4901 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4902 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4903 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4904 	db_printf(
4905 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4906 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4907 	    "b_dep = %p\n",
4908 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4909 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4910 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4911 	db_printf("b_kvabase = %p, b_kvasize = %d\n",
4912 	    bp->b_kvabase, bp->b_kvasize);
4913 	if (bp->b_npages) {
4914 		int i;
4915 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4916 		for (i = 0; i < bp->b_npages; i++) {
4917 			vm_page_t m;
4918 			m = bp->b_pages[i];
4919 			if (m != NULL)
4920 				db_printf("(%p, 0x%lx, 0x%lx)", m->object,
4921 				    (u_long)m->pindex,
4922 				    (u_long)VM_PAGE_TO_PHYS(m));
4923 			else
4924 				db_printf("( ??? )");
4925 			if ((i + 1) < bp->b_npages)
4926 				db_printf(",");
4927 		}
4928 		db_printf("\n");
4929 	}
4930 #if defined(FULL_BUF_TRACKING)
4931 	db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
4932 
4933 	i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
4934 	for (j = 1; j <= BUF_TRACKING_SIZE; j++)
4935 		db_printf(" %2u: %s\n", j,
4936 		    bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
4937 #elif defined(BUF_TRACKING)
4938 	db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
4939 #endif
4940 	db_printf(" ");
4941 	BUF_LOCKPRINTINFO(bp);
4942 }
4943 
4944 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4945 {
4946 	struct buf *bp;
4947 	int i;
4948 
4949 	for (i = 0; i < nbuf; i++) {
4950 		bp = &buf[i];
4951 		if (BUF_ISLOCKED(bp)) {
4952 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4953 			db_printf("\n");
4954 		}
4955 	}
4956 }
4957 
4958 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4959 {
4960 	struct vnode *vp;
4961 	struct buf *bp;
4962 
4963 	if (!have_addr) {
4964 		db_printf("usage: show vnodebufs <addr>\n");
4965 		return;
4966 	}
4967 	vp = (struct vnode *)addr;
4968 	db_printf("Clean buffers:\n");
4969 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4970 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4971 		db_printf("\n");
4972 	}
4973 	db_printf("Dirty buffers:\n");
4974 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4975 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4976 		db_printf("\n");
4977 	}
4978 }
4979 
4980 DB_COMMAND(countfreebufs, db_coundfreebufs)
4981 {
4982 	struct buf *bp;
4983 	int i, used = 0, nfree = 0;
4984 
4985 	if (have_addr) {
4986 		db_printf("usage: countfreebufs\n");
4987 		return;
4988 	}
4989 
4990 	for (i = 0; i < nbuf; i++) {
4991 		bp = &buf[i];
4992 		if (bp->b_qindex == QUEUE_EMPTY)
4993 			nfree++;
4994 		else
4995 			used++;
4996 	}
4997 
4998 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4999 	    nfree + used);
5000 	db_printf("numfreebuffers is %d\n", numfreebuffers);
5001 }
5002 #endif /* DDB */
5003