xref: /freebsd/sys/kern/vfs_bio.c (revision 3fc9e2c36555140de248a0b4def91bbfa44d7c2c)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * Portions of this software were developed by Konstantin Belousov
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * this file contains a new buffer I/O scheme implementing a coherent
34  * VM object and buffer cache scheme.  Pains have been taken to make
35  * sure that the performance degradation associated with schemes such
36  * as this is not realized.
37  *
38  * Author:  John S. Dyson
39  * Significant help during the development and debugging phases
40  * had been provided by David Greenman, also of the FreeBSD core team.
41  *
42  * see man buf(9) for more info.
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/conf.h>
52 #include <sys/buf.h>
53 #include <sys/devicestat.h>
54 #include <sys/eventhandler.h>
55 #include <sys/fail.h>
56 #include <sys/limits.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/kernel.h>
62 #include <sys/kthread.h>
63 #include <sys/proc.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
66 #include <sys/sysctl.h>
67 #include <sys/vmem.h>
68 #include <sys/vmmeter.h>
69 #include <sys/vnode.h>
70 #include <geom/geom.h>
71 #include <vm/vm.h>
72 #include <vm/vm_param.h>
73 #include <vm/vm_kern.h>
74 #include <vm/vm_pageout.h>
75 #include <vm/vm_page.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/vm_map.h>
79 #include "opt_compat.h"
80 #include "opt_directio.h"
81 #include "opt_swap.h"
82 
83 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
84 
85 struct	bio_ops bioops;		/* I/O operation notification */
86 
87 struct	buf_ops buf_ops_bio = {
88 	.bop_name	=	"buf_ops_bio",
89 	.bop_write	=	bufwrite,
90 	.bop_strategy	=	bufstrategy,
91 	.bop_sync	=	bufsync,
92 	.bop_bdflush	=	bufbdflush,
93 };
94 
95 /*
96  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
97  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
98  */
99 struct buf *buf;		/* buffer header pool */
100 caddr_t unmapped_buf;
101 
102 static struct proc *bufdaemonproc;
103 
104 static int inmem(struct vnode *vp, daddr_t blkno);
105 static void vm_hold_free_pages(struct buf *bp, int newbsize);
106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
107 		vm_offset_t to);
108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
110 		vm_page_t m);
111 static void vfs_clean_pages_dirty_buf(struct buf *bp);
112 static void vfs_setdirty_locked_object(struct buf *bp);
113 static void vfs_vmio_release(struct buf *bp);
114 static int vfs_bio_clcheck(struct vnode *vp, int size,
115 		daddr_t lblkno, daddr_t blkno);
116 static int buf_flush(struct vnode *vp, int);
117 static int flushbufqueues(struct vnode *, int, int);
118 static void buf_daemon(void);
119 static void bremfreel(struct buf *bp);
120 static __inline void bd_wakeup(void);
121 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
122     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
123 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
124 #endif
125 
126 int vmiodirenable = TRUE;
127 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
128     "Use the VM system for directory writes");
129 long runningbufspace;
130 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
131     "Amount of presently outstanding async buffer io");
132 static long bufspace;
133 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
134     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
135 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
136     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
137 #else
138 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
139     "Virtual memory used for buffers");
140 #endif
141 static long unmapped_bufspace;
142 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD,
143     &unmapped_bufspace, 0,
144     "Amount of unmapped buffers, inclusive in the bufspace");
145 static long maxbufspace;
146 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
147     "Maximum allowed value of bufspace (including buf_daemon)");
148 static long bufmallocspace;
149 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
150     "Amount of malloced memory for buffers");
151 static long maxbufmallocspace;
152 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
153     "Maximum amount of malloced memory for buffers");
154 static long lobufspace;
155 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
156     "Minimum amount of buffers we want to have");
157 long hibufspace;
158 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
159     "Maximum allowed value of bufspace (excluding buf_daemon)");
160 static int bufreusecnt;
161 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
162     "Number of times we have reused a buffer");
163 static int buffreekvacnt;
164 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
165     "Number of times we have freed the KVA space from some buffer");
166 static int bufdefragcnt;
167 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
168     "Number of times we have had to repeat buffer allocation to defragment");
169 static long lorunningspace;
170 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
171     "Minimum preferred space used for in-progress I/O");
172 static long hirunningspace;
173 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
174     "Maximum amount of space to use for in-progress I/O");
175 int dirtybufferflushes;
176 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
177     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
178 int bdwriteskip;
179 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
180     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
181 int altbufferflushes;
182 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
183     0, "Number of fsync flushes to limit dirty buffers");
184 static int recursiveflushes;
185 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
186     0, "Number of flushes skipped due to being recursive");
187 static int numdirtybuffers;
188 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
189     "Number of buffers that are dirty (has unwritten changes) at the moment");
190 static int lodirtybuffers;
191 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
192     "How many buffers we want to have free before bufdaemon can sleep");
193 static int hidirtybuffers;
194 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
195     "When the number of dirty buffers is considered severe");
196 int dirtybufthresh;
197 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
198     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
199 static int numfreebuffers;
200 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
201     "Number of free buffers");
202 static int lofreebuffers;
203 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
204    "XXX Unused");
205 static int hifreebuffers;
206 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
207    "XXX Complicatedly unused");
208 static int getnewbufcalls;
209 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
210    "Number of calls to getnewbuf");
211 static int getnewbufrestarts;
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
213     "Number of times getnewbuf has had to restart a buffer aquisition");
214 static int mappingrestarts;
215 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
216     "Number of times getblk has had to restart a buffer mapping for "
217     "unmapped buffer");
218 static int flushbufqtarget = 100;
219 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
220     "Amount of work to do in flushbufqueues when helping bufdaemon");
221 static long notbufdflushes;
222 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes, 0,
223     "Number of dirty buffer flushes done by the bufdaemon helpers");
224 static long barrierwrites;
225 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
226     "Number of barrier writes");
227 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
228     &unmapped_buf_allowed, 0,
229     "Permit the use of the unmapped i/o");
230 
231 /*
232  * Lock for the non-dirty bufqueues
233  */
234 static struct mtx_padalign bqclean;
235 
236 /*
237  * Lock for the dirty queue.
238  */
239 static struct mtx_padalign bqdirty;
240 
241 /*
242  * This lock synchronizes access to bd_request.
243  */
244 static struct mtx_padalign bdlock;
245 
246 /*
247  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
248  * waitrunningbufspace().
249  */
250 static struct mtx_padalign rbreqlock;
251 
252 /*
253  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
254  */
255 static struct mtx_padalign nblock;
256 
257 /*
258  * Lock that protects bdirtywait.
259  */
260 static struct mtx_padalign bdirtylock;
261 
262 /*
263  * Wakeup point for bufdaemon, as well as indicator of whether it is already
264  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
265  * is idling.
266  */
267 static int bd_request;
268 
269 /*
270  * Request for the buf daemon to write more buffers than is indicated by
271  * lodirtybuf.  This may be necessary to push out excess dependencies or
272  * defragment the address space where a simple count of the number of dirty
273  * buffers is insufficient to characterize the demand for flushing them.
274  */
275 static int bd_speedupreq;
276 
277 /*
278  * bogus page -- for I/O to/from partially complete buffers
279  * this is a temporary solution to the problem, but it is not
280  * really that bad.  it would be better to split the buffer
281  * for input in the case of buffers partially already in memory,
282  * but the code is intricate enough already.
283  */
284 vm_page_t bogus_page;
285 
286 /*
287  * Synchronization (sleep/wakeup) variable for active buffer space requests.
288  * Set when wait starts, cleared prior to wakeup().
289  * Used in runningbufwakeup() and waitrunningbufspace().
290  */
291 static int runningbufreq;
292 
293 /*
294  * Synchronization (sleep/wakeup) variable for buffer requests.
295  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
296  * by and/or.
297  * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
298  * getnewbuf(), and getblk().
299  */
300 static int needsbuffer;
301 
302 /*
303  * Synchronization for bwillwrite() waiters.
304  */
305 static int bdirtywait;
306 
307 /*
308  * Definitions for the buffer free lists.
309  */
310 #define BUFFER_QUEUES	5	/* number of free buffer queues */
311 
312 #define QUEUE_NONE	0	/* on no queue */
313 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
314 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
315 #define QUEUE_EMPTYKVA	3	/* empty buffer headers w/KVA assignment */
316 #define QUEUE_EMPTY	4	/* empty buffer headers */
317 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
318 
319 /* Queues for free buffers with various properties */
320 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
321 #ifdef INVARIANTS
322 static int bq_len[BUFFER_QUEUES];
323 #endif
324 
325 /*
326  * Single global constant for BUF_WMESG, to avoid getting multiple references.
327  * buf_wmesg is referred from macros.
328  */
329 const char *buf_wmesg = BUF_WMESG;
330 
331 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
332 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
333 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
334 
335 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
336     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
337 static int
338 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
339 {
340 	long lvalue;
341 	int ivalue;
342 
343 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
344 		return (sysctl_handle_long(oidp, arg1, arg2, req));
345 	lvalue = *(long *)arg1;
346 	if (lvalue > INT_MAX)
347 		/* On overflow, still write out a long to trigger ENOMEM. */
348 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
349 	ivalue = lvalue;
350 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
351 }
352 #endif
353 
354 #ifdef DIRECTIO
355 extern void ffs_rawread_setup(void);
356 #endif /* DIRECTIO */
357 
358 /*
359  *	bqlock:
360  *
361  *	Return the appropriate queue lock based on the index.
362  */
363 static inline struct mtx *
364 bqlock(int qindex)
365 {
366 
367 	if (qindex == QUEUE_DIRTY)
368 		return (struct mtx *)(&bqdirty);
369 	return (struct mtx *)(&bqclean);
370 }
371 
372 /*
373  *	bdirtywakeup:
374  *
375  *	Wakeup any bwillwrite() waiters.
376  */
377 static void
378 bdirtywakeup(void)
379 {
380 	mtx_lock(&bdirtylock);
381 	if (bdirtywait) {
382 		bdirtywait = 0;
383 		wakeup(&bdirtywait);
384 	}
385 	mtx_unlock(&bdirtylock);
386 }
387 
388 /*
389  *	bdirtysub:
390  *
391  *	Decrement the numdirtybuffers count by one and wakeup any
392  *	threads blocked in bwillwrite().
393  */
394 static void
395 bdirtysub(void)
396 {
397 
398 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
399 	    (lodirtybuffers + hidirtybuffers) / 2)
400 		bdirtywakeup();
401 }
402 
403 /*
404  *	bdirtyadd:
405  *
406  *	Increment the numdirtybuffers count by one and wakeup the buf
407  *	daemon if needed.
408  */
409 static void
410 bdirtyadd(void)
411 {
412 
413 	/*
414 	 * Only do the wakeup once as we cross the boundary.  The
415 	 * buf daemon will keep running until the condition clears.
416 	 */
417 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
418 	    (lodirtybuffers + hidirtybuffers) / 2)
419 		bd_wakeup();
420 }
421 
422 /*
423  *	bufspacewakeup:
424  *
425  *	Called when buffer space is potentially available for recovery.
426  *	getnewbuf() will block on this flag when it is unable to free
427  *	sufficient buffer space.  Buffer space becomes recoverable when
428  *	bp's get placed back in the queues.
429  */
430 
431 static __inline void
432 bufspacewakeup(void)
433 {
434 
435 	/*
436 	 * If someone is waiting for BUF space, wake them up.  Even
437 	 * though we haven't freed the kva space yet, the waiting
438 	 * process will be able to now.
439 	 */
440 	mtx_lock(&nblock);
441 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
442 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
443 		wakeup(&needsbuffer);
444 	}
445 	mtx_unlock(&nblock);
446 }
447 
448 /*
449  *	runningwakeup:
450  *
451  *	Wake up processes that are waiting on asynchronous writes to fall
452  *	below lorunningspace.
453  */
454 static void
455 runningwakeup(void)
456 {
457 
458 	mtx_lock(&rbreqlock);
459 	if (runningbufreq) {
460 		runningbufreq = 0;
461 		wakeup(&runningbufreq);
462 	}
463 	mtx_unlock(&rbreqlock);
464 }
465 
466 /*
467  *	runningbufwakeup:
468  *
469  *	Decrement the outstanding write count according.
470  */
471 void
472 runningbufwakeup(struct buf *bp)
473 {
474 	long space, bspace;
475 
476 	bspace = bp->b_runningbufspace;
477 	if (bspace == 0)
478 		return;
479 	space = atomic_fetchadd_long(&runningbufspace, -bspace);
480 	KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
481 	    space, bspace));
482 	bp->b_runningbufspace = 0;
483 	/*
484 	 * Only acquire the lock and wakeup on the transition from exceeding
485 	 * the threshold to falling below it.
486 	 */
487 	if (space < lorunningspace)
488 		return;
489 	if (space - bspace > lorunningspace)
490 		return;
491 	runningwakeup();
492 }
493 
494 /*
495  *	bufcountadd:
496  *
497  *	Called when a buffer has been added to one of the free queues to
498  *	account for the buffer and to wakeup anyone waiting for free buffers.
499  *	This typically occurs when large amounts of metadata are being handled
500  *	by the buffer cache ( else buffer space runs out first, usually ).
501  */
502 static __inline void
503 bufcountadd(struct buf *bp)
504 {
505 	int old;
506 
507 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
508 	    ("buf %p already counted as free", bp));
509 	bp->b_flags |= B_INFREECNT;
510 	old = atomic_fetchadd_int(&numfreebuffers, 1);
511 	KASSERT(old >= 0 && old < nbuf,
512 	    ("numfreebuffers climbed to %d", old + 1));
513 	mtx_lock(&nblock);
514 	if (needsbuffer) {
515 		needsbuffer &= ~VFS_BIO_NEED_ANY;
516 		if (numfreebuffers >= hifreebuffers)
517 			needsbuffer &= ~VFS_BIO_NEED_FREE;
518 		wakeup(&needsbuffer);
519 	}
520 	mtx_unlock(&nblock);
521 }
522 
523 /*
524  *	bufcountsub:
525  *
526  *	Decrement the numfreebuffers count as needed.
527  */
528 static void
529 bufcountsub(struct buf *bp)
530 {
531 	int old;
532 
533 	/*
534 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
535 	 * delayed-write, the buffer was free and we must decrement
536 	 * numfreebuffers.
537 	 */
538 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
539 		KASSERT((bp->b_flags & B_INFREECNT) != 0,
540 		    ("buf %p not counted in numfreebuffers", bp));
541 		bp->b_flags &= ~B_INFREECNT;
542 		old = atomic_fetchadd_int(&numfreebuffers, -1);
543 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
544 	}
545 }
546 
547 /*
548  *	waitrunningbufspace()
549  *
550  *	runningbufspace is a measure of the amount of I/O currently
551  *	running.  This routine is used in async-write situations to
552  *	prevent creating huge backups of pending writes to a device.
553  *	Only asynchronous writes are governed by this function.
554  *
555  *	This does NOT turn an async write into a sync write.  It waits
556  *	for earlier writes to complete and generally returns before the
557  *	caller's write has reached the device.
558  */
559 void
560 waitrunningbufspace(void)
561 {
562 
563 	mtx_lock(&rbreqlock);
564 	while (runningbufspace > hirunningspace) {
565 		runningbufreq = 1;
566 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
567 	}
568 	mtx_unlock(&rbreqlock);
569 }
570 
571 
572 /*
573  *	vfs_buf_test_cache:
574  *
575  *	Called when a buffer is extended.  This function clears the B_CACHE
576  *	bit if the newly extended portion of the buffer does not contain
577  *	valid data.
578  */
579 static __inline
580 void
581 vfs_buf_test_cache(struct buf *bp,
582 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
583 		  vm_page_t m)
584 {
585 
586 	VM_OBJECT_ASSERT_LOCKED(m->object);
587 	if (bp->b_flags & B_CACHE) {
588 		int base = (foff + off) & PAGE_MASK;
589 		if (vm_page_is_valid(m, base, size) == 0)
590 			bp->b_flags &= ~B_CACHE;
591 	}
592 }
593 
594 /* Wake up the buffer daemon if necessary */
595 static __inline void
596 bd_wakeup(void)
597 {
598 
599 	mtx_lock(&bdlock);
600 	if (bd_request == 0) {
601 		bd_request = 1;
602 		wakeup(&bd_request);
603 	}
604 	mtx_unlock(&bdlock);
605 }
606 
607 /*
608  * bd_speedup - speedup the buffer cache flushing code
609  */
610 void
611 bd_speedup(void)
612 {
613 	int needwake;
614 
615 	mtx_lock(&bdlock);
616 	needwake = 0;
617 	if (bd_speedupreq == 0 || bd_request == 0)
618 		needwake = 1;
619 	bd_speedupreq = 1;
620 	bd_request = 1;
621 	if (needwake)
622 		wakeup(&bd_request);
623 	mtx_unlock(&bdlock);
624 }
625 
626 #ifdef __i386__
627 #define	TRANSIENT_DENOM	5
628 #else
629 #define	TRANSIENT_DENOM 10
630 #endif
631 
632 /*
633  * Calculating buffer cache scaling values and reserve space for buffer
634  * headers.  This is called during low level kernel initialization and
635  * may be called more then once.  We CANNOT write to the memory area
636  * being reserved at this time.
637  */
638 caddr_t
639 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
640 {
641 	int tuned_nbuf;
642 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
643 
644 	/*
645 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
646 	 * PAGE_SIZE is >= 1K)
647 	 */
648 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
649 
650 	/*
651 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
652 	 * For the first 64MB of ram nominally allocate sufficient buffers to
653 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
654 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
655 	 * the buffer cache we limit the eventual kva reservation to
656 	 * maxbcache bytes.
657 	 *
658 	 * factor represents the 1/4 x ram conversion.
659 	 */
660 	if (nbuf == 0) {
661 		int factor = 4 * BKVASIZE / 1024;
662 
663 		nbuf = 50;
664 		if (physmem_est > 4096)
665 			nbuf += min((physmem_est - 4096) / factor,
666 			    65536 / factor);
667 		if (physmem_est > 65536)
668 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
669 			    32 * 1024 * 1024 / (factor * 5));
670 
671 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
672 			nbuf = maxbcache / BKVASIZE;
673 		tuned_nbuf = 1;
674 	} else
675 		tuned_nbuf = 0;
676 
677 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
678 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
679 	if (nbuf > maxbuf) {
680 		if (!tuned_nbuf)
681 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
682 			    maxbuf);
683 		nbuf = maxbuf;
684 	}
685 
686 	/*
687 	 * Ideal allocation size for the transient bio submap if 10%
688 	 * of the maximal space buffer map.  This roughly corresponds
689 	 * to the amount of the buffer mapped for typical UFS load.
690 	 *
691 	 * Clip the buffer map to reserve space for the transient
692 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
693 	 * maximum buffer map extent on the platform.
694 	 *
695 	 * The fall-back to the maxbuf in case of maxbcache unset,
696 	 * allows to not trim the buffer KVA for the architectures
697 	 * with ample KVA space.
698 	 */
699 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
700 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
701 		buf_sz = (long)nbuf * BKVASIZE;
702 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
703 		    (TRANSIENT_DENOM - 1)) {
704 			/*
705 			 * There is more KVA than memory.  Do not
706 			 * adjust buffer map size, and assign the rest
707 			 * of maxbuf to transient map.
708 			 */
709 			biotmap_sz = maxbuf_sz - buf_sz;
710 		} else {
711 			/*
712 			 * Buffer map spans all KVA we could afford on
713 			 * this platform.  Give 10% (20% on i386) of
714 			 * the buffer map to the transient bio map.
715 			 */
716 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
717 			buf_sz -= biotmap_sz;
718 		}
719 		if (biotmap_sz / INT_MAX > MAXPHYS)
720 			bio_transient_maxcnt = INT_MAX;
721 		else
722 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
723 		/*
724 		 * Artifically limit to 1024 simultaneous in-flight I/Os
725 		 * using the transient mapping.
726 		 */
727 		if (bio_transient_maxcnt > 1024)
728 			bio_transient_maxcnt = 1024;
729 		if (tuned_nbuf)
730 			nbuf = buf_sz / BKVASIZE;
731 	}
732 
733 	/*
734 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
735 	 * We have no less then 16 and no more then 256.
736 	 */
737 	nswbuf = max(min(nbuf/4, 256), 16);
738 #ifdef NSWBUF_MIN
739 	if (nswbuf < NSWBUF_MIN)
740 		nswbuf = NSWBUF_MIN;
741 #endif
742 #ifdef DIRECTIO
743 	ffs_rawread_setup();
744 #endif
745 
746 	/*
747 	 * Reserve space for the buffer cache buffers
748 	 */
749 	swbuf = (void *)v;
750 	v = (caddr_t)(swbuf + nswbuf);
751 	buf = (void *)v;
752 	v = (caddr_t)(buf + nbuf);
753 
754 	return(v);
755 }
756 
757 /* Initialize the buffer subsystem.  Called before use of any buffers. */
758 void
759 bufinit(void)
760 {
761 	struct buf *bp;
762 	int i;
763 
764 	mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
765 	mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
766 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
767 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
768 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
769 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
770 
771 	/* next, make a null set of free lists */
772 	for (i = 0; i < BUFFER_QUEUES; i++)
773 		TAILQ_INIT(&bufqueues[i]);
774 
775 	/* finally, initialize each buffer header and stick on empty q */
776 	for (i = 0; i < nbuf; i++) {
777 		bp = &buf[i];
778 		bzero(bp, sizeof *bp);
779 		bp->b_flags = B_INVAL | B_INFREECNT;
780 		bp->b_rcred = NOCRED;
781 		bp->b_wcred = NOCRED;
782 		bp->b_qindex = QUEUE_EMPTY;
783 		bp->b_xflags = 0;
784 		LIST_INIT(&bp->b_dep);
785 		BUF_LOCKINIT(bp);
786 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
787 #ifdef INVARIANTS
788 		bq_len[QUEUE_EMPTY]++;
789 #endif
790 	}
791 
792 	/*
793 	 * maxbufspace is the absolute maximum amount of buffer space we are
794 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
795 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
796 	 * used by most other processes.  The differential is required to
797 	 * ensure that buf_daemon is able to run when other processes might
798 	 * be blocked waiting for buffer space.
799 	 *
800 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
801 	 * this may result in KVM fragmentation which is not handled optimally
802 	 * by the system.
803 	 */
804 	maxbufspace = (long)nbuf * BKVASIZE;
805 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
806 	lobufspace = hibufspace - MAXBSIZE;
807 
808 	/*
809 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
810 	 * arbitrarily and may need further tuning. It corresponds to
811 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
812 	 * which fits with many RAID controllers' tagged queuing limits.
813 	 * The lower 1 MiB limit is the historical upper limit for
814 	 * hirunningspace.
815 	 */
816 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
817 	    16 * 1024 * 1024), 1024 * 1024);
818 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
819 
820 /*
821  * Limit the amount of malloc memory since it is wired permanently into
822  * the kernel space.  Even though this is accounted for in the buffer
823  * allocation, we don't want the malloced region to grow uncontrolled.
824  * The malloc scheme improves memory utilization significantly on average
825  * (small) directories.
826  */
827 	maxbufmallocspace = hibufspace / 20;
828 
829 /*
830  * Reduce the chance of a deadlock occuring by limiting the number
831  * of delayed-write dirty buffers we allow to stack up.
832  */
833 	hidirtybuffers = nbuf / 4 + 20;
834 	dirtybufthresh = hidirtybuffers * 9 / 10;
835 	numdirtybuffers = 0;
836 /*
837  * To support extreme low-memory systems, make sure hidirtybuffers cannot
838  * eat up all available buffer space.  This occurs when our minimum cannot
839  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
840  * BKVASIZE'd buffers.
841  */
842 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
843 		hidirtybuffers >>= 1;
844 	}
845 	lodirtybuffers = hidirtybuffers / 2;
846 
847 /*
848  * Try to keep the number of free buffers in the specified range,
849  * and give special processes (e.g. like buf_daemon) access to an
850  * emergency reserve.
851  */
852 	lofreebuffers = nbuf / 18 + 5;
853 	hifreebuffers = 2 * lofreebuffers;
854 	numfreebuffers = nbuf;
855 
856 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
857 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
858 	unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
859 }
860 
861 #ifdef INVARIANTS
862 static inline void
863 vfs_buf_check_mapped(struct buf *bp)
864 {
865 
866 	KASSERT((bp->b_flags & B_UNMAPPED) == 0,
867 	    ("mapped buf %p %x", bp, bp->b_flags));
868 	KASSERT(bp->b_kvabase != unmapped_buf,
869 	    ("mapped buf: b_kvabase was not updated %p", bp));
870 	KASSERT(bp->b_data != unmapped_buf,
871 	    ("mapped buf: b_data was not updated %p", bp));
872 }
873 
874 static inline void
875 vfs_buf_check_unmapped(struct buf *bp)
876 {
877 
878 	KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
879 	    ("unmapped buf %p %x", bp, bp->b_flags));
880 	KASSERT(bp->b_kvabase == unmapped_buf,
881 	    ("unmapped buf: corrupted b_kvabase %p", bp));
882 	KASSERT(bp->b_data == unmapped_buf,
883 	    ("unmapped buf: corrupted b_data %p", bp));
884 }
885 
886 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
887 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
888 #else
889 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
890 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
891 #endif
892 
893 static void
894 bpmap_qenter(struct buf *bp)
895 {
896 
897 	BUF_CHECK_MAPPED(bp);
898 
899 	/*
900 	 * bp->b_data is relative to bp->b_offset, but
901 	 * bp->b_offset may be offset into the first page.
902 	 */
903 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
904 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
905 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
906 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
907 }
908 
909 /*
910  * bfreekva() - free the kva allocation for a buffer.
911  *
912  *	Since this call frees up buffer space, we call bufspacewakeup().
913  */
914 static void
915 bfreekva(struct buf *bp)
916 {
917 
918 	if (bp->b_kvasize == 0)
919 		return;
920 
921 	atomic_add_int(&buffreekvacnt, 1);
922 	atomic_subtract_long(&bufspace, bp->b_kvasize);
923 	if ((bp->b_flags & B_UNMAPPED) == 0) {
924 		BUF_CHECK_MAPPED(bp);
925 		vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
926 		    bp->b_kvasize);
927 	} else {
928 		BUF_CHECK_UNMAPPED(bp);
929 		if ((bp->b_flags & B_KVAALLOC) != 0) {
930 			vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
931 			    bp->b_kvasize);
932 		}
933 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
934 		bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
935 	}
936 	bp->b_kvasize = 0;
937 	bufspacewakeup();
938 }
939 
940 /*
941  *	binsfree:
942  *
943  *	Insert the buffer into the appropriate free list.
944  */
945 static void
946 binsfree(struct buf *bp, int qindex)
947 {
948 	struct mtx *olock, *nlock;
949 
950 	BUF_ASSERT_XLOCKED(bp);
951 
952 	olock = bqlock(bp->b_qindex);
953 	nlock = bqlock(qindex);
954 	mtx_lock(olock);
955 	/* Handle delayed bremfree() processing. */
956 	if (bp->b_flags & B_REMFREE)
957 		bremfreel(bp);
958 
959 	if (bp->b_qindex != QUEUE_NONE)
960 		panic("binsfree: free buffer onto another queue???");
961 
962 	bp->b_qindex = qindex;
963 	if (olock != nlock) {
964 		mtx_unlock(olock);
965 		mtx_lock(nlock);
966 	}
967 	if (bp->b_flags & B_AGE)
968 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
969 	else
970 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
971 #ifdef INVARIANTS
972 	bq_len[bp->b_qindex]++;
973 #endif
974 	mtx_unlock(nlock);
975 
976 	/*
977 	 * Something we can maybe free or reuse.
978 	 */
979 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
980 		bufspacewakeup();
981 
982 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
983 		bufcountadd(bp);
984 }
985 
986 /*
987  *	bremfree:
988  *
989  *	Mark the buffer for removal from the appropriate free list.
990  *
991  */
992 void
993 bremfree(struct buf *bp)
994 {
995 
996 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
997 	KASSERT((bp->b_flags & B_REMFREE) == 0,
998 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
999 	KASSERT(bp->b_qindex != QUEUE_NONE,
1000 	    ("bremfree: buffer %p not on a queue.", bp));
1001 	BUF_ASSERT_XLOCKED(bp);
1002 
1003 	bp->b_flags |= B_REMFREE;
1004 	bufcountsub(bp);
1005 }
1006 
1007 /*
1008  *	bremfreef:
1009  *
1010  *	Force an immediate removal from a free list.  Used only in nfs when
1011  *	it abuses the b_freelist pointer.
1012  */
1013 void
1014 bremfreef(struct buf *bp)
1015 {
1016 	struct mtx *qlock;
1017 
1018 	qlock = bqlock(bp->b_qindex);
1019 	mtx_lock(qlock);
1020 	bremfreel(bp);
1021 	mtx_unlock(qlock);
1022 }
1023 
1024 /*
1025  *	bremfreel:
1026  *
1027  *	Removes a buffer from the free list, must be called with the
1028  *	correct qlock held.
1029  */
1030 static void
1031 bremfreel(struct buf *bp)
1032 {
1033 
1034 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1035 	    bp, bp->b_vp, bp->b_flags);
1036 	KASSERT(bp->b_qindex != QUEUE_NONE,
1037 	    ("bremfreel: buffer %p not on a queue.", bp));
1038 	BUF_ASSERT_XLOCKED(bp);
1039 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1040 
1041 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1042 #ifdef INVARIANTS
1043 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1044 	    bp->b_qindex));
1045 	bq_len[bp->b_qindex]--;
1046 #endif
1047 	bp->b_qindex = QUEUE_NONE;
1048 	/*
1049 	 * If this was a delayed bremfree() we only need to remove the buffer
1050 	 * from the queue and return the stats are already done.
1051 	 */
1052 	if (bp->b_flags & B_REMFREE) {
1053 		bp->b_flags &= ~B_REMFREE;
1054 		return;
1055 	}
1056 	bufcountsub(bp);
1057 }
1058 
1059 /*
1060  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1061  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1062  * the buffer is valid and we do not have to do anything.
1063  */
1064 void
1065 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1066     int cnt, struct ucred * cred)
1067 {
1068 	struct buf *rabp;
1069 	int i;
1070 
1071 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1072 		if (inmem(vp, *rablkno))
1073 			continue;
1074 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1075 
1076 		if ((rabp->b_flags & B_CACHE) == 0) {
1077 			if (!TD_IS_IDLETHREAD(curthread))
1078 				curthread->td_ru.ru_inblock++;
1079 			rabp->b_flags |= B_ASYNC;
1080 			rabp->b_flags &= ~B_INVAL;
1081 			rabp->b_ioflags &= ~BIO_ERROR;
1082 			rabp->b_iocmd = BIO_READ;
1083 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1084 				rabp->b_rcred = crhold(cred);
1085 			vfs_busy_pages(rabp, 0);
1086 			BUF_KERNPROC(rabp);
1087 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1088 			bstrategy(rabp);
1089 		} else {
1090 			brelse(rabp);
1091 		}
1092 	}
1093 }
1094 
1095 /*
1096  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1097  *
1098  * Get a buffer with the specified data.  Look in the cache first.  We
1099  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1100  * is set, the buffer is valid and we do not have to do anything, see
1101  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1102  */
1103 int
1104 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1105     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1106 {
1107 	struct buf *bp;
1108 	int rv = 0, readwait = 0;
1109 
1110 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1111 	/*
1112 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1113 	 */
1114 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1115 	if (bp == NULL)
1116 		return (EBUSY);
1117 
1118 	/* if not found in cache, do some I/O */
1119 	if ((bp->b_flags & B_CACHE) == 0) {
1120 		if (!TD_IS_IDLETHREAD(curthread))
1121 			curthread->td_ru.ru_inblock++;
1122 		bp->b_iocmd = BIO_READ;
1123 		bp->b_flags &= ~B_INVAL;
1124 		bp->b_ioflags &= ~BIO_ERROR;
1125 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1126 			bp->b_rcred = crhold(cred);
1127 		vfs_busy_pages(bp, 0);
1128 		bp->b_iooffset = dbtob(bp->b_blkno);
1129 		bstrategy(bp);
1130 		++readwait;
1131 	}
1132 
1133 	breada(vp, rablkno, rabsize, cnt, cred);
1134 
1135 	if (readwait) {
1136 		rv = bufwait(bp);
1137 	}
1138 	return (rv);
1139 }
1140 
1141 /*
1142  * Write, release buffer on completion.  (Done by iodone
1143  * if async).  Do not bother writing anything if the buffer
1144  * is invalid.
1145  *
1146  * Note that we set B_CACHE here, indicating that buffer is
1147  * fully valid and thus cacheable.  This is true even of NFS
1148  * now so we set it generally.  This could be set either here
1149  * or in biodone() since the I/O is synchronous.  We put it
1150  * here.
1151  */
1152 int
1153 bufwrite(struct buf *bp)
1154 {
1155 	int oldflags;
1156 	struct vnode *vp;
1157 	long space;
1158 	int vp_md;
1159 
1160 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1161 	if (bp->b_flags & B_INVAL) {
1162 		brelse(bp);
1163 		return (0);
1164 	}
1165 
1166 	if (bp->b_flags & B_BARRIER)
1167 		barrierwrites++;
1168 
1169 	oldflags = bp->b_flags;
1170 
1171 	BUF_ASSERT_HELD(bp);
1172 
1173 	if (bp->b_pin_count > 0)
1174 		bunpin_wait(bp);
1175 
1176 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1177 	    ("FFS background buffer should not get here %p", bp));
1178 
1179 	vp = bp->b_vp;
1180 	if (vp)
1181 		vp_md = vp->v_vflag & VV_MD;
1182 	else
1183 		vp_md = 0;
1184 
1185 	/*
1186 	 * Mark the buffer clean.  Increment the bufobj write count
1187 	 * before bundirty() call, to prevent other thread from seeing
1188 	 * empty dirty list and zero counter for writes in progress,
1189 	 * falsely indicating that the bufobj is clean.
1190 	 */
1191 	bufobj_wref(bp->b_bufobj);
1192 	bundirty(bp);
1193 
1194 	bp->b_flags &= ~B_DONE;
1195 	bp->b_ioflags &= ~BIO_ERROR;
1196 	bp->b_flags |= B_CACHE;
1197 	bp->b_iocmd = BIO_WRITE;
1198 
1199 	vfs_busy_pages(bp, 1);
1200 
1201 	/*
1202 	 * Normal bwrites pipeline writes
1203 	 */
1204 	bp->b_runningbufspace = bp->b_bufsize;
1205 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1206 
1207 	if (!TD_IS_IDLETHREAD(curthread))
1208 		curthread->td_ru.ru_oublock++;
1209 	if (oldflags & B_ASYNC)
1210 		BUF_KERNPROC(bp);
1211 	bp->b_iooffset = dbtob(bp->b_blkno);
1212 	bstrategy(bp);
1213 
1214 	if ((oldflags & B_ASYNC) == 0) {
1215 		int rtval = bufwait(bp);
1216 		brelse(bp);
1217 		return (rtval);
1218 	} else if (space > hirunningspace) {
1219 		/*
1220 		 * don't allow the async write to saturate the I/O
1221 		 * system.  We will not deadlock here because
1222 		 * we are blocking waiting for I/O that is already in-progress
1223 		 * to complete. We do not block here if it is the update
1224 		 * or syncer daemon trying to clean up as that can lead
1225 		 * to deadlock.
1226 		 */
1227 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1228 			waitrunningbufspace();
1229 	}
1230 
1231 	return (0);
1232 }
1233 
1234 void
1235 bufbdflush(struct bufobj *bo, struct buf *bp)
1236 {
1237 	struct buf *nbp;
1238 
1239 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1240 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1241 		altbufferflushes++;
1242 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1243 		BO_LOCK(bo);
1244 		/*
1245 		 * Try to find a buffer to flush.
1246 		 */
1247 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1248 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1249 			    BUF_LOCK(nbp,
1250 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1251 				continue;
1252 			if (bp == nbp)
1253 				panic("bdwrite: found ourselves");
1254 			BO_UNLOCK(bo);
1255 			/* Don't countdeps with the bo lock held. */
1256 			if (buf_countdeps(nbp, 0)) {
1257 				BO_LOCK(bo);
1258 				BUF_UNLOCK(nbp);
1259 				continue;
1260 			}
1261 			if (nbp->b_flags & B_CLUSTEROK) {
1262 				vfs_bio_awrite(nbp);
1263 			} else {
1264 				bremfree(nbp);
1265 				bawrite(nbp);
1266 			}
1267 			dirtybufferflushes++;
1268 			break;
1269 		}
1270 		if (nbp == NULL)
1271 			BO_UNLOCK(bo);
1272 	}
1273 }
1274 
1275 /*
1276  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1277  * anything if the buffer is marked invalid.
1278  *
1279  * Note that since the buffer must be completely valid, we can safely
1280  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1281  * biodone() in order to prevent getblk from writing the buffer
1282  * out synchronously.
1283  */
1284 void
1285 bdwrite(struct buf *bp)
1286 {
1287 	struct thread *td = curthread;
1288 	struct vnode *vp;
1289 	struct bufobj *bo;
1290 
1291 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1292 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1293 	KASSERT((bp->b_flags & B_BARRIER) == 0,
1294 	    ("Barrier request in delayed write %p", bp));
1295 	BUF_ASSERT_HELD(bp);
1296 
1297 	if (bp->b_flags & B_INVAL) {
1298 		brelse(bp);
1299 		return;
1300 	}
1301 
1302 	/*
1303 	 * If we have too many dirty buffers, don't create any more.
1304 	 * If we are wildly over our limit, then force a complete
1305 	 * cleanup. Otherwise, just keep the situation from getting
1306 	 * out of control. Note that we have to avoid a recursive
1307 	 * disaster and not try to clean up after our own cleanup!
1308 	 */
1309 	vp = bp->b_vp;
1310 	bo = bp->b_bufobj;
1311 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1312 		td->td_pflags |= TDP_INBDFLUSH;
1313 		BO_BDFLUSH(bo, bp);
1314 		td->td_pflags &= ~TDP_INBDFLUSH;
1315 	} else
1316 		recursiveflushes++;
1317 
1318 	bdirty(bp);
1319 	/*
1320 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1321 	 * true even of NFS now.
1322 	 */
1323 	bp->b_flags |= B_CACHE;
1324 
1325 	/*
1326 	 * This bmap keeps the system from needing to do the bmap later,
1327 	 * perhaps when the system is attempting to do a sync.  Since it
1328 	 * is likely that the indirect block -- or whatever other datastructure
1329 	 * that the filesystem needs is still in memory now, it is a good
1330 	 * thing to do this.  Note also, that if the pageout daemon is
1331 	 * requesting a sync -- there might not be enough memory to do
1332 	 * the bmap then...  So, this is important to do.
1333 	 */
1334 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1335 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1336 	}
1337 
1338 	/*
1339 	 * Set the *dirty* buffer range based upon the VM system dirty
1340 	 * pages.
1341 	 *
1342 	 * Mark the buffer pages as clean.  We need to do this here to
1343 	 * satisfy the vnode_pager and the pageout daemon, so that it
1344 	 * thinks that the pages have been "cleaned".  Note that since
1345 	 * the pages are in a delayed write buffer -- the VFS layer
1346 	 * "will" see that the pages get written out on the next sync,
1347 	 * or perhaps the cluster will be completed.
1348 	 */
1349 	vfs_clean_pages_dirty_buf(bp);
1350 	bqrelse(bp);
1351 
1352 	/*
1353 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1354 	 * due to the softdep code.
1355 	 */
1356 }
1357 
1358 /*
1359  *	bdirty:
1360  *
1361  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1362  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1363  *	itself to properly update it in the dirty/clean lists.  We mark it
1364  *	B_DONE to ensure that any asynchronization of the buffer properly
1365  *	clears B_DONE ( else a panic will occur later ).
1366  *
1367  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1368  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1369  *	should only be called if the buffer is known-good.
1370  *
1371  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1372  *	count.
1373  *
1374  *	The buffer must be on QUEUE_NONE.
1375  */
1376 void
1377 bdirty(struct buf *bp)
1378 {
1379 
1380 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1381 	    bp, bp->b_vp, bp->b_flags);
1382 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1383 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1384 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1385 	BUF_ASSERT_HELD(bp);
1386 	bp->b_flags &= ~(B_RELBUF);
1387 	bp->b_iocmd = BIO_WRITE;
1388 
1389 	if ((bp->b_flags & B_DELWRI) == 0) {
1390 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1391 		reassignbuf(bp);
1392 		bdirtyadd();
1393 	}
1394 }
1395 
1396 /*
1397  *	bundirty:
1398  *
1399  *	Clear B_DELWRI for buffer.
1400  *
1401  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1402  *	count.
1403  *
1404  *	The buffer must be on QUEUE_NONE.
1405  */
1406 
1407 void
1408 bundirty(struct buf *bp)
1409 {
1410 
1411 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1412 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1413 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1414 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1415 	BUF_ASSERT_HELD(bp);
1416 
1417 	if (bp->b_flags & B_DELWRI) {
1418 		bp->b_flags &= ~B_DELWRI;
1419 		reassignbuf(bp);
1420 		bdirtysub();
1421 	}
1422 	/*
1423 	 * Since it is now being written, we can clear its deferred write flag.
1424 	 */
1425 	bp->b_flags &= ~B_DEFERRED;
1426 }
1427 
1428 /*
1429  *	bawrite:
1430  *
1431  *	Asynchronous write.  Start output on a buffer, but do not wait for
1432  *	it to complete.  The buffer is released when the output completes.
1433  *
1434  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1435  *	B_INVAL buffers.  Not us.
1436  */
1437 void
1438 bawrite(struct buf *bp)
1439 {
1440 
1441 	bp->b_flags |= B_ASYNC;
1442 	(void) bwrite(bp);
1443 }
1444 
1445 /*
1446  *	babarrierwrite:
1447  *
1448  *	Asynchronous barrier write.  Start output on a buffer, but do not
1449  *	wait for it to complete.  Place a write barrier after this write so
1450  *	that this buffer and all buffers written before it are committed to
1451  *	the disk before any buffers written after this write are committed
1452  *	to the disk.  The buffer is released when the output completes.
1453  */
1454 void
1455 babarrierwrite(struct buf *bp)
1456 {
1457 
1458 	bp->b_flags |= B_ASYNC | B_BARRIER;
1459 	(void) bwrite(bp);
1460 }
1461 
1462 /*
1463  *	bbarrierwrite:
1464  *
1465  *	Synchronous barrier write.  Start output on a buffer and wait for
1466  *	it to complete.  Place a write barrier after this write so that
1467  *	this buffer and all buffers written before it are committed to
1468  *	the disk before any buffers written after this write are committed
1469  *	to the disk.  The buffer is released when the output completes.
1470  */
1471 int
1472 bbarrierwrite(struct buf *bp)
1473 {
1474 
1475 	bp->b_flags |= B_BARRIER;
1476 	return (bwrite(bp));
1477 }
1478 
1479 /*
1480  *	bwillwrite:
1481  *
1482  *	Called prior to the locking of any vnodes when we are expecting to
1483  *	write.  We do not want to starve the buffer cache with too many
1484  *	dirty buffers so we block here.  By blocking prior to the locking
1485  *	of any vnodes we attempt to avoid the situation where a locked vnode
1486  *	prevents the various system daemons from flushing related buffers.
1487  */
1488 void
1489 bwillwrite(void)
1490 {
1491 
1492 	if (numdirtybuffers >= hidirtybuffers) {
1493 		mtx_lock(&bdirtylock);
1494 		while (numdirtybuffers >= hidirtybuffers) {
1495 			bdirtywait = 1;
1496 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1497 			    "flswai", 0);
1498 		}
1499 		mtx_unlock(&bdirtylock);
1500 	}
1501 }
1502 
1503 /*
1504  * Return true if we have too many dirty buffers.
1505  */
1506 int
1507 buf_dirty_count_severe(void)
1508 {
1509 
1510 	return(numdirtybuffers >= hidirtybuffers);
1511 }
1512 
1513 static __noinline int
1514 buf_vm_page_count_severe(void)
1515 {
1516 
1517 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1518 
1519 	return vm_page_count_severe();
1520 }
1521 
1522 /*
1523  *	brelse:
1524  *
1525  *	Release a busy buffer and, if requested, free its resources.  The
1526  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1527  *	to be accessed later as a cache entity or reused for other purposes.
1528  */
1529 void
1530 brelse(struct buf *bp)
1531 {
1532 	int qindex;
1533 
1534 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1535 	    bp, bp->b_vp, bp->b_flags);
1536 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1537 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1538 
1539 	if (BUF_LOCKRECURSED(bp)) {
1540 		/*
1541 		 * Do not process, in particular, do not handle the
1542 		 * B_INVAL/B_RELBUF and do not release to free list.
1543 		 */
1544 		BUF_UNLOCK(bp);
1545 		return;
1546 	}
1547 
1548 	if (bp->b_flags & B_MANAGED) {
1549 		bqrelse(bp);
1550 		return;
1551 	}
1552 
1553 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1554 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1555 		/*
1556 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1557 		 * pages from being scrapped.  If the error is anything
1558 		 * other than an I/O error (EIO), assume that retrying
1559 		 * is futile.
1560 		 */
1561 		bp->b_ioflags &= ~BIO_ERROR;
1562 		bdirty(bp);
1563 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1564 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1565 		/*
1566 		 * Either a failed I/O or we were asked to free or not
1567 		 * cache the buffer.
1568 		 */
1569 		bp->b_flags |= B_INVAL;
1570 		if (!LIST_EMPTY(&bp->b_dep))
1571 			buf_deallocate(bp);
1572 		if (bp->b_flags & B_DELWRI)
1573 			bdirtysub();
1574 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1575 		if ((bp->b_flags & B_VMIO) == 0) {
1576 			if (bp->b_bufsize)
1577 				allocbuf(bp, 0);
1578 			if (bp->b_vp)
1579 				brelvp(bp);
1580 		}
1581 	}
1582 
1583 	/*
1584 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1585 	 * is called with B_DELWRI set, the underlying pages may wind up
1586 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1587 	 * because pages associated with a B_DELWRI bp are marked clean.
1588 	 *
1589 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1590 	 * if B_DELWRI is set.
1591 	 *
1592 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1593 	 * on pages to return pages to the VM page queues.
1594 	 */
1595 	if (bp->b_flags & B_DELWRI)
1596 		bp->b_flags &= ~B_RELBUF;
1597 	else if (buf_vm_page_count_severe()) {
1598 		/*
1599 		 * BKGRDINPROG can only be set with the buf and bufobj
1600 		 * locks both held.  We tolerate a race to clear it here.
1601 		 */
1602 		if (!(bp->b_vflags & BV_BKGRDINPROG))
1603 			bp->b_flags |= B_RELBUF;
1604 	}
1605 
1606 	/*
1607 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1608 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1609 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1610 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1611 	 *
1612 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1613 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1614 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1615 	 *
1616 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1617 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1618 	 * the commit state and we cannot afford to lose the buffer. If the
1619 	 * buffer has a background write in progress, we need to keep it
1620 	 * around to prevent it from being reconstituted and starting a second
1621 	 * background write.
1622 	 */
1623 	if ((bp->b_flags & B_VMIO)
1624 	    && !(bp->b_vp->v_mount != NULL &&
1625 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1626 		 !vn_isdisk(bp->b_vp, NULL) &&
1627 		 (bp->b_flags & B_DELWRI))
1628 	    ) {
1629 
1630 		int i, j, resid;
1631 		vm_page_t m;
1632 		off_t foff;
1633 		vm_pindex_t poff;
1634 		vm_object_t obj;
1635 
1636 		obj = bp->b_bufobj->bo_object;
1637 
1638 		/*
1639 		 * Get the base offset and length of the buffer.  Note that
1640 		 * in the VMIO case if the buffer block size is not
1641 		 * page-aligned then b_data pointer may not be page-aligned.
1642 		 * But our b_pages[] array *IS* page aligned.
1643 		 *
1644 		 * block sizes less then DEV_BSIZE (usually 512) are not
1645 		 * supported due to the page granularity bits (m->valid,
1646 		 * m->dirty, etc...).
1647 		 *
1648 		 * See man buf(9) for more information
1649 		 */
1650 		resid = bp->b_bufsize;
1651 		foff = bp->b_offset;
1652 		for (i = 0; i < bp->b_npages; i++) {
1653 			int had_bogus = 0;
1654 
1655 			m = bp->b_pages[i];
1656 
1657 			/*
1658 			 * If we hit a bogus page, fixup *all* the bogus pages
1659 			 * now.
1660 			 */
1661 			if (m == bogus_page) {
1662 				poff = OFF_TO_IDX(bp->b_offset);
1663 				had_bogus = 1;
1664 
1665 				VM_OBJECT_RLOCK(obj);
1666 				for (j = i; j < bp->b_npages; j++) {
1667 					vm_page_t mtmp;
1668 					mtmp = bp->b_pages[j];
1669 					if (mtmp == bogus_page) {
1670 						mtmp = vm_page_lookup(obj, poff + j);
1671 						if (!mtmp) {
1672 							panic("brelse: page missing\n");
1673 						}
1674 						bp->b_pages[j] = mtmp;
1675 					}
1676 				}
1677 				VM_OBJECT_RUNLOCK(obj);
1678 
1679 				if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1680 					BUF_CHECK_MAPPED(bp);
1681 					pmap_qenter(
1682 					    trunc_page((vm_offset_t)bp->b_data),
1683 					    bp->b_pages, bp->b_npages);
1684 				}
1685 				m = bp->b_pages[i];
1686 			}
1687 			if ((bp->b_flags & B_NOCACHE) ||
1688 			    (bp->b_ioflags & BIO_ERROR &&
1689 			     bp->b_iocmd == BIO_READ)) {
1690 				int poffset = foff & PAGE_MASK;
1691 				int presid = resid > (PAGE_SIZE - poffset) ?
1692 					(PAGE_SIZE - poffset) : resid;
1693 
1694 				KASSERT(presid >= 0, ("brelse: extra page"));
1695 				VM_OBJECT_WLOCK(obj);
1696 				while (vm_page_xbusied(m)) {
1697 					vm_page_lock(m);
1698 					VM_OBJECT_WUNLOCK(obj);
1699 					vm_page_busy_sleep(m, "mbncsh");
1700 					VM_OBJECT_WLOCK(obj);
1701 				}
1702 				if (pmap_page_wired_mappings(m) == 0)
1703 					vm_page_set_invalid(m, poffset, presid);
1704 				VM_OBJECT_WUNLOCK(obj);
1705 				if (had_bogus)
1706 					printf("avoided corruption bug in bogus_page/brelse code\n");
1707 			}
1708 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1709 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1710 		}
1711 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1712 			vfs_vmio_release(bp);
1713 
1714 	} else if (bp->b_flags & B_VMIO) {
1715 
1716 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1717 			vfs_vmio_release(bp);
1718 		}
1719 
1720 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1721 		if (bp->b_bufsize != 0)
1722 			allocbuf(bp, 0);
1723 		if (bp->b_vp != NULL)
1724 			brelvp(bp);
1725 	}
1726 
1727 	/*
1728 	 * If the buffer has junk contents signal it and eventually
1729 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1730 	 * doesn't find it.
1731 	 */
1732 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1733 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1734 		bp->b_flags |= B_INVAL;
1735 	if (bp->b_flags & B_INVAL) {
1736 		if (bp->b_flags & B_DELWRI)
1737 			bundirty(bp);
1738 		if (bp->b_vp)
1739 			brelvp(bp);
1740 	}
1741 
1742 	/* buffers with no memory */
1743 	if (bp->b_bufsize == 0) {
1744 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1745 		if (bp->b_vflags & BV_BKGRDINPROG)
1746 			panic("losing buffer 1");
1747 		if (bp->b_kvasize)
1748 			qindex = QUEUE_EMPTYKVA;
1749 		else
1750 			qindex = QUEUE_EMPTY;
1751 		bp->b_flags |= B_AGE;
1752 	/* buffers with junk contents */
1753 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1754 	    (bp->b_ioflags & BIO_ERROR)) {
1755 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1756 		if (bp->b_vflags & BV_BKGRDINPROG)
1757 			panic("losing buffer 2");
1758 		qindex = QUEUE_CLEAN;
1759 		bp->b_flags |= B_AGE;
1760 	/* remaining buffers */
1761 	} else if (bp->b_flags & B_DELWRI)
1762 		qindex = QUEUE_DIRTY;
1763 	else
1764 		qindex = QUEUE_CLEAN;
1765 
1766 	binsfree(bp, qindex);
1767 
1768 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1769 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1770 		panic("brelse: not dirty");
1771 	/* unlock */
1772 	BUF_UNLOCK(bp);
1773 }
1774 
1775 /*
1776  * Release a buffer back to the appropriate queue but do not try to free
1777  * it.  The buffer is expected to be used again soon.
1778  *
1779  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1780  * biodone() to requeue an async I/O on completion.  It is also used when
1781  * known good buffers need to be requeued but we think we may need the data
1782  * again soon.
1783  *
1784  * XXX we should be able to leave the B_RELBUF hint set on completion.
1785  */
1786 void
1787 bqrelse(struct buf *bp)
1788 {
1789 	int qindex;
1790 
1791 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1792 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1793 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1794 
1795 	if (BUF_LOCKRECURSED(bp)) {
1796 		/* do not release to free list */
1797 		BUF_UNLOCK(bp);
1798 		return;
1799 	}
1800 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1801 
1802 	if (bp->b_flags & B_MANAGED) {
1803 		if (bp->b_flags & B_REMFREE)
1804 			bremfreef(bp);
1805 		goto out;
1806 	}
1807 
1808 	/* buffers with stale but valid contents */
1809 	if (bp->b_flags & B_DELWRI) {
1810 		qindex = QUEUE_DIRTY;
1811 	} else {
1812 		if ((bp->b_flags & B_DELWRI) == 0 &&
1813 		    (bp->b_xflags & BX_VNDIRTY))
1814 			panic("bqrelse: not dirty");
1815 		/*
1816 		 * BKGRDINPROG can only be set with the buf and bufobj
1817 		 * locks both held.  We tolerate a race to clear it here.
1818 		 */
1819 		if (buf_vm_page_count_severe() &&
1820 		    (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1821 			/*
1822 			 * We are too low on memory, we have to try to free
1823 			 * the buffer (most importantly: the wired pages
1824 			 * making up its backing store) *now*.
1825 			 */
1826 			brelse(bp);
1827 			return;
1828 		}
1829 		qindex = QUEUE_CLEAN;
1830 	}
1831 	binsfree(bp, qindex);
1832 
1833 out:
1834 	/* unlock */
1835 	BUF_UNLOCK(bp);
1836 }
1837 
1838 /* Give pages used by the bp back to the VM system (where possible) */
1839 static void
1840 vfs_vmio_release(struct buf *bp)
1841 {
1842 	int i;
1843 	vm_page_t m;
1844 
1845 	if ((bp->b_flags & B_UNMAPPED) == 0) {
1846 		BUF_CHECK_MAPPED(bp);
1847 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1848 	} else
1849 		BUF_CHECK_UNMAPPED(bp);
1850 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
1851 	for (i = 0; i < bp->b_npages; i++) {
1852 		m = bp->b_pages[i];
1853 		bp->b_pages[i] = NULL;
1854 		/*
1855 		 * In order to keep page LRU ordering consistent, put
1856 		 * everything on the inactive queue.
1857 		 */
1858 		vm_page_lock(m);
1859 		vm_page_unwire(m, 0);
1860 
1861 		/*
1862 		 * Might as well free the page if we can and it has
1863 		 * no valid data.  We also free the page if the
1864 		 * buffer was used for direct I/O
1865 		 */
1866 		if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1867 			if (m->wire_count == 0 && !vm_page_busied(m))
1868 				vm_page_free(m);
1869 		} else if (bp->b_flags & B_DIRECT)
1870 			vm_page_try_to_free(m);
1871 		else if (buf_vm_page_count_severe())
1872 			vm_page_try_to_cache(m);
1873 		vm_page_unlock(m);
1874 	}
1875 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
1876 
1877 	if (bp->b_bufsize) {
1878 		bufspacewakeup();
1879 		bp->b_bufsize = 0;
1880 	}
1881 	bp->b_npages = 0;
1882 	bp->b_flags &= ~B_VMIO;
1883 	if (bp->b_vp)
1884 		brelvp(bp);
1885 }
1886 
1887 /*
1888  * Check to see if a block at a particular lbn is available for a clustered
1889  * write.
1890  */
1891 static int
1892 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1893 {
1894 	struct buf *bpa;
1895 	int match;
1896 
1897 	match = 0;
1898 
1899 	/* If the buf isn't in core skip it */
1900 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1901 		return (0);
1902 
1903 	/* If the buf is busy we don't want to wait for it */
1904 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1905 		return (0);
1906 
1907 	/* Only cluster with valid clusterable delayed write buffers */
1908 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1909 	    (B_DELWRI | B_CLUSTEROK))
1910 		goto done;
1911 
1912 	if (bpa->b_bufsize != size)
1913 		goto done;
1914 
1915 	/*
1916 	 * Check to see if it is in the expected place on disk and that the
1917 	 * block has been mapped.
1918 	 */
1919 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1920 		match = 1;
1921 done:
1922 	BUF_UNLOCK(bpa);
1923 	return (match);
1924 }
1925 
1926 /*
1927  *	vfs_bio_awrite:
1928  *
1929  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1930  *	This is much better then the old way of writing only one buffer at
1931  *	a time.  Note that we may not be presented with the buffers in the
1932  *	correct order, so we search for the cluster in both directions.
1933  */
1934 int
1935 vfs_bio_awrite(struct buf *bp)
1936 {
1937 	struct bufobj *bo;
1938 	int i;
1939 	int j;
1940 	daddr_t lblkno = bp->b_lblkno;
1941 	struct vnode *vp = bp->b_vp;
1942 	int ncl;
1943 	int nwritten;
1944 	int size;
1945 	int maxcl;
1946 	int gbflags;
1947 
1948 	bo = &vp->v_bufobj;
1949 	gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1950 	/*
1951 	 * right now we support clustered writing only to regular files.  If
1952 	 * we find a clusterable block we could be in the middle of a cluster
1953 	 * rather then at the beginning.
1954 	 */
1955 	if ((vp->v_type == VREG) &&
1956 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1957 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1958 
1959 		size = vp->v_mount->mnt_stat.f_iosize;
1960 		maxcl = MAXPHYS / size;
1961 
1962 		BO_RLOCK(bo);
1963 		for (i = 1; i < maxcl; i++)
1964 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1965 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1966 				break;
1967 
1968 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1969 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1970 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1971 				break;
1972 		BO_RUNLOCK(bo);
1973 		--j;
1974 		ncl = i + j;
1975 		/*
1976 		 * this is a possible cluster write
1977 		 */
1978 		if (ncl != 1) {
1979 			BUF_UNLOCK(bp);
1980 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
1981 			    gbflags);
1982 			return (nwritten);
1983 		}
1984 	}
1985 	bremfree(bp);
1986 	bp->b_flags |= B_ASYNC;
1987 	/*
1988 	 * default (old) behavior, writing out only one block
1989 	 *
1990 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1991 	 */
1992 	nwritten = bp->b_bufsize;
1993 	(void) bwrite(bp);
1994 
1995 	return (nwritten);
1996 }
1997 
1998 static void
1999 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2000 {
2001 
2002 	KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2003 	    bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2004 	if ((gbflags & GB_UNMAPPED) == 0) {
2005 		bp->b_kvabase = (caddr_t)addr;
2006 	} else if ((gbflags & GB_KVAALLOC) != 0) {
2007 		KASSERT((gbflags & GB_UNMAPPED) != 0,
2008 		    ("GB_KVAALLOC without GB_UNMAPPED"));
2009 		bp->b_kvaalloc = (caddr_t)addr;
2010 		bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2011 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2012 	}
2013 	bp->b_kvasize = maxsize;
2014 }
2015 
2016 /*
2017  * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2018  * needed.
2019  */
2020 static int
2021 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2022 {
2023 	vm_offset_t addr;
2024 
2025 	bfreekva(bp);
2026 	addr = 0;
2027 
2028 	if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2029 		/*
2030 		 * Buffer map is too fragmented.  Request the caller
2031 		 * to defragment the map.
2032 		 */
2033 		atomic_add_int(&bufdefragcnt, 1);
2034 		return (1);
2035 	}
2036 	setbufkva(bp, addr, maxsize, gbflags);
2037 	atomic_add_long(&bufspace, bp->b_kvasize);
2038 	return (0);
2039 }
2040 
2041 /*
2042  * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2043  * locked vnode is supplied.
2044  */
2045 static void
2046 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2047     int defrag)
2048 {
2049 	struct thread *td;
2050 	char *waitmsg;
2051 	int fl, flags, norunbuf;
2052 
2053 	mtx_assert(&bqclean, MA_OWNED);
2054 
2055 	if (defrag) {
2056 		flags = VFS_BIO_NEED_BUFSPACE;
2057 		waitmsg = "nbufkv";
2058 	} else if (bufspace >= hibufspace) {
2059 		waitmsg = "nbufbs";
2060 		flags = VFS_BIO_NEED_BUFSPACE;
2061 	} else {
2062 		waitmsg = "newbuf";
2063 		flags = VFS_BIO_NEED_ANY;
2064 	}
2065 	mtx_lock(&nblock);
2066 	needsbuffer |= flags;
2067 	mtx_unlock(&nblock);
2068 	mtx_unlock(&bqclean);
2069 
2070 	bd_speedup();	/* heeeelp */
2071 	if ((gbflags & GB_NOWAIT_BD) != 0)
2072 		return;
2073 
2074 	td = curthread;
2075 	mtx_lock(&nblock);
2076 	while (needsbuffer & flags) {
2077 		if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2078 			mtx_unlock(&nblock);
2079 			/*
2080 			 * getblk() is called with a vnode locked, and
2081 			 * some majority of the dirty buffers may as
2082 			 * well belong to the vnode.  Flushing the
2083 			 * buffers there would make a progress that
2084 			 * cannot be achieved by the buf_daemon, that
2085 			 * cannot lock the vnode.
2086 			 */
2087 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2088 			    (td->td_pflags & TDP_NORUNNINGBUF);
2089 			/* play bufdaemon */
2090 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2091 			fl = buf_flush(vp, flushbufqtarget);
2092 			td->td_pflags &= norunbuf;
2093 			mtx_lock(&nblock);
2094 			if (fl != 0)
2095 				continue;
2096 			if ((needsbuffer & flags) == 0)
2097 				break;
2098 		}
2099 		if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag,
2100 		    waitmsg, slptimeo))
2101 			break;
2102 	}
2103 	mtx_unlock(&nblock);
2104 }
2105 
2106 static void
2107 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2108 {
2109 
2110 	CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2111 	    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2112 	     bp->b_kvasize, bp->b_bufsize, qindex);
2113 	mtx_assert(&bqclean, MA_NOTOWNED);
2114 
2115 	/*
2116 	 * Note: we no longer distinguish between VMIO and non-VMIO
2117 	 * buffers.
2118 	 */
2119 	KASSERT((bp->b_flags & B_DELWRI) == 0,
2120 	    ("delwri buffer %p found in queue %d", bp, qindex));
2121 
2122 	if (qindex == QUEUE_CLEAN) {
2123 		if (bp->b_flags & B_VMIO) {
2124 			bp->b_flags &= ~B_ASYNC;
2125 			vfs_vmio_release(bp);
2126 		}
2127 		if (bp->b_vp != NULL)
2128 			brelvp(bp);
2129 	}
2130 
2131 	/*
2132 	 * Get the rest of the buffer freed up.  b_kva* is still valid
2133 	 * after this operation.
2134 	 */
2135 
2136 	if (bp->b_rcred != NOCRED) {
2137 		crfree(bp->b_rcred);
2138 		bp->b_rcred = NOCRED;
2139 	}
2140 	if (bp->b_wcred != NOCRED) {
2141 		crfree(bp->b_wcred);
2142 		bp->b_wcred = NOCRED;
2143 	}
2144 	if (!LIST_EMPTY(&bp->b_dep))
2145 		buf_deallocate(bp);
2146 	if (bp->b_vflags & BV_BKGRDINPROG)
2147 		panic("losing buffer 3");
2148 	KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.  qindex: %d",
2149 	    bp, bp->b_vp, qindex));
2150 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2151 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2152 
2153 	if (bp->b_bufsize)
2154 		allocbuf(bp, 0);
2155 
2156 	bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2157 	bp->b_ioflags = 0;
2158 	bp->b_xflags = 0;
2159 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
2160 	    ("buf %p still counted as free?", bp));
2161 	bp->b_vflags = 0;
2162 	bp->b_vp = NULL;
2163 	bp->b_blkno = bp->b_lblkno = 0;
2164 	bp->b_offset = NOOFFSET;
2165 	bp->b_iodone = 0;
2166 	bp->b_error = 0;
2167 	bp->b_resid = 0;
2168 	bp->b_bcount = 0;
2169 	bp->b_npages = 0;
2170 	bp->b_dirtyoff = bp->b_dirtyend = 0;
2171 	bp->b_bufobj = NULL;
2172 	bp->b_pin_count = 0;
2173 	bp->b_fsprivate1 = NULL;
2174 	bp->b_fsprivate2 = NULL;
2175 	bp->b_fsprivate3 = NULL;
2176 
2177 	LIST_INIT(&bp->b_dep);
2178 }
2179 
2180 static int flushingbufs;
2181 
2182 static struct buf *
2183 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2184 {
2185 	struct buf *bp, *nbp;
2186 	int nqindex, qindex, pass;
2187 
2188 	KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2189 
2190 	pass = 1;
2191 restart:
2192 	atomic_add_int(&getnewbufrestarts, 1);
2193 
2194 	/*
2195 	 * Setup for scan.  If we do not have enough free buffers,
2196 	 * we setup a degenerate case that immediately fails.  Note
2197 	 * that if we are specially marked process, we are allowed to
2198 	 * dip into our reserves.
2199 	 *
2200 	 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2201 	 * for the allocation of the mapped buffer.  For unmapped, the
2202 	 * easiest is to start with EMPTY outright.
2203 	 *
2204 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
2205 	 * However, there are a number of cases (defragging, reusing, ...)
2206 	 * where we cannot backup.
2207 	 */
2208 	nbp = NULL;
2209 	mtx_lock(&bqclean);
2210 	if (!defrag && unmapped) {
2211 		nqindex = QUEUE_EMPTY;
2212 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2213 	}
2214 	if (nbp == NULL) {
2215 		nqindex = QUEUE_EMPTYKVA;
2216 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2217 	}
2218 
2219 	/*
2220 	 * If no EMPTYKVA buffers and we are either defragging or
2221 	 * reusing, locate a CLEAN buffer to free or reuse.  If
2222 	 * bufspace useage is low skip this step so we can allocate a
2223 	 * new buffer.
2224 	 */
2225 	if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2226 		nqindex = QUEUE_CLEAN;
2227 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2228 	}
2229 
2230 	/*
2231 	 * If we could not find or were not allowed to reuse a CLEAN
2232 	 * buffer, check to see if it is ok to use an EMPTY buffer.
2233 	 * We can only use an EMPTY buffer if allocating its KVA would
2234 	 * not otherwise run us out of buffer space.  No KVA is needed
2235 	 * for the unmapped allocation.
2236 	 */
2237 	if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2238 	    metadata)) {
2239 		nqindex = QUEUE_EMPTY;
2240 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2241 	}
2242 
2243 	/*
2244 	 * All available buffers might be clean, retry ignoring the
2245 	 * lobufspace as the last resort.
2246 	 */
2247 	if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2248 		nqindex = QUEUE_CLEAN;
2249 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2250 	}
2251 
2252 	/*
2253 	 * Run scan, possibly freeing data and/or kva mappings on the fly
2254 	 * depending.
2255 	 */
2256 	while ((bp = nbp) != NULL) {
2257 		qindex = nqindex;
2258 
2259 		/*
2260 		 * Calculate next bp (we can only use it if we do not
2261 		 * block or do other fancy things).
2262 		 */
2263 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2264 			switch (qindex) {
2265 			case QUEUE_EMPTY:
2266 				nqindex = QUEUE_EMPTYKVA;
2267 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2268 				if (nbp != NULL)
2269 					break;
2270 				/* FALLTHROUGH */
2271 			case QUEUE_EMPTYKVA:
2272 				nqindex = QUEUE_CLEAN;
2273 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2274 				if (nbp != NULL)
2275 					break;
2276 				/* FALLTHROUGH */
2277 			case QUEUE_CLEAN:
2278 				if (metadata && pass == 1) {
2279 					pass = 2;
2280 					nqindex = QUEUE_EMPTY;
2281 					nbp = TAILQ_FIRST(
2282 					    &bufqueues[QUEUE_EMPTY]);
2283 				}
2284 				/*
2285 				 * nbp is NULL.
2286 				 */
2287 				break;
2288 			}
2289 		}
2290 		/*
2291 		 * If we are defragging then we need a buffer with
2292 		 * b_kvasize != 0.  XXX this situation should no longer
2293 		 * occur, if defrag is non-zero the buffer's b_kvasize
2294 		 * should also be non-zero at this point.  XXX
2295 		 */
2296 		if (defrag && bp->b_kvasize == 0) {
2297 			printf("Warning: defrag empty buffer %p\n", bp);
2298 			continue;
2299 		}
2300 
2301 		/*
2302 		 * Start freeing the bp.  This is somewhat involved.  nbp
2303 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2304 		 */
2305 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2306 			continue;
2307 		/*
2308 		 * BKGRDINPROG can only be set with the buf and bufobj
2309 		 * locks both held.  We tolerate a race to clear it here.
2310 		 */
2311 		if (bp->b_vflags & BV_BKGRDINPROG) {
2312 			BUF_UNLOCK(bp);
2313 			continue;
2314 		}
2315 
2316 		KASSERT(bp->b_qindex == qindex,
2317 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2318 
2319 		bremfreel(bp);
2320 		mtx_unlock(&bqclean);
2321 		/*
2322 		 * NOTE:  nbp is now entirely invalid.  We can only restart
2323 		 * the scan from this point on.
2324 		 */
2325 
2326 		getnewbuf_reuse_bp(bp, qindex);
2327 		mtx_assert(&bqclean, MA_NOTOWNED);
2328 
2329 		/*
2330 		 * If we are defragging then free the buffer.
2331 		 */
2332 		if (defrag) {
2333 			bp->b_flags |= B_INVAL;
2334 			bfreekva(bp);
2335 			brelse(bp);
2336 			defrag = 0;
2337 			goto restart;
2338 		}
2339 
2340 		/*
2341 		 * Notify any waiters for the buffer lock about
2342 		 * identity change by freeing the buffer.
2343 		 */
2344 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2345 			bp->b_flags |= B_INVAL;
2346 			bfreekva(bp);
2347 			brelse(bp);
2348 			goto restart;
2349 		}
2350 
2351 		if (metadata)
2352 			break;
2353 
2354 		/*
2355 		 * If we are overcomitted then recover the buffer and its
2356 		 * KVM space.  This occurs in rare situations when multiple
2357 		 * processes are blocked in getnewbuf() or allocbuf().
2358 		 */
2359 		if (bufspace >= hibufspace)
2360 			flushingbufs = 1;
2361 		if (flushingbufs && bp->b_kvasize != 0) {
2362 			bp->b_flags |= B_INVAL;
2363 			bfreekva(bp);
2364 			brelse(bp);
2365 			goto restart;
2366 		}
2367 		if (bufspace < lobufspace)
2368 			flushingbufs = 0;
2369 		break;
2370 	}
2371 	return (bp);
2372 }
2373 
2374 /*
2375  *	getnewbuf:
2376  *
2377  *	Find and initialize a new buffer header, freeing up existing buffers
2378  *	in the bufqueues as necessary.  The new buffer is returned locked.
2379  *
2380  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
2381  *	buffer away, the caller must set B_INVAL prior to calling brelse().
2382  *
2383  *	We block if:
2384  *		We have insufficient buffer headers
2385  *		We have insufficient buffer space
2386  *		buffer_arena is too fragmented ( space reservation fails )
2387  *		If we have to flush dirty buffers ( but we try to avoid this )
2388  */
2389 static struct buf *
2390 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2391     int gbflags)
2392 {
2393 	struct buf *bp;
2394 	int defrag, metadata;
2395 
2396 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2397 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2398 	if (!unmapped_buf_allowed)
2399 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2400 
2401 	defrag = 0;
2402 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2403 	    vp->v_type == VCHR)
2404 		metadata = 1;
2405 	else
2406 		metadata = 0;
2407 	/*
2408 	 * We can't afford to block since we might be holding a vnode lock,
2409 	 * which may prevent system daemons from running.  We deal with
2410 	 * low-memory situations by proactively returning memory and running
2411 	 * async I/O rather then sync I/O.
2412 	 */
2413 	atomic_add_int(&getnewbufcalls, 1);
2414 	atomic_subtract_int(&getnewbufrestarts, 1);
2415 restart:
2416 	bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2417 	    GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2418 	if (bp != NULL)
2419 		defrag = 0;
2420 
2421 	/*
2422 	 * If we exhausted our list, sleep as appropriate.  We may have to
2423 	 * wakeup various daemons and write out some dirty buffers.
2424 	 *
2425 	 * Generally we are sleeping due to insufficient buffer space.
2426 	 */
2427 	if (bp == NULL) {
2428 		mtx_assert(&bqclean, MA_OWNED);
2429 		getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2430 		mtx_assert(&bqclean, MA_NOTOWNED);
2431 	} else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2432 		mtx_assert(&bqclean, MA_NOTOWNED);
2433 
2434 		bfreekva(bp);
2435 		bp->b_flags |= B_UNMAPPED;
2436 		bp->b_kvabase = bp->b_data = unmapped_buf;
2437 		bp->b_kvasize = maxsize;
2438 		atomic_add_long(&bufspace, bp->b_kvasize);
2439 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2440 		atomic_add_int(&bufreusecnt, 1);
2441 	} else {
2442 		mtx_assert(&bqclean, MA_NOTOWNED);
2443 
2444 		/*
2445 		 * We finally have a valid bp.  We aren't quite out of the
2446 		 * woods, we still have to reserve kva space.  In order
2447 		 * to keep fragmentation sane we only allocate kva in
2448 		 * BKVASIZE chunks.
2449 		 */
2450 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2451 
2452 		if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2453 		    B_KVAALLOC)) == B_UNMAPPED) {
2454 			if (allocbufkva(bp, maxsize, gbflags)) {
2455 				defrag = 1;
2456 				bp->b_flags |= B_INVAL;
2457 				brelse(bp);
2458 				goto restart;
2459 			}
2460 			atomic_add_int(&bufreusecnt, 1);
2461 		} else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2462 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2463 			/*
2464 			 * If the reused buffer has KVA allocated,
2465 			 * reassign b_kvaalloc to b_kvabase.
2466 			 */
2467 			bp->b_kvabase = bp->b_kvaalloc;
2468 			bp->b_flags &= ~B_KVAALLOC;
2469 			atomic_subtract_long(&unmapped_bufspace,
2470 			    bp->b_kvasize);
2471 			atomic_add_int(&bufreusecnt, 1);
2472 		} else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2473 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2474 		    GB_KVAALLOC)) {
2475 			/*
2476 			 * The case of reused buffer already have KVA
2477 			 * mapped, but the request is for unmapped
2478 			 * buffer with KVA allocated.
2479 			 */
2480 			bp->b_kvaalloc = bp->b_kvabase;
2481 			bp->b_data = bp->b_kvabase = unmapped_buf;
2482 			bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2483 			atomic_add_long(&unmapped_bufspace,
2484 			    bp->b_kvasize);
2485 			atomic_add_int(&bufreusecnt, 1);
2486 		}
2487 		if ((gbflags & GB_UNMAPPED) == 0) {
2488 			bp->b_saveaddr = bp->b_kvabase;
2489 			bp->b_data = bp->b_saveaddr;
2490 			bp->b_flags &= ~B_UNMAPPED;
2491 			BUF_CHECK_MAPPED(bp);
2492 		}
2493 	}
2494 	return (bp);
2495 }
2496 
2497 /*
2498  *	buf_daemon:
2499  *
2500  *	buffer flushing daemon.  Buffers are normally flushed by the
2501  *	update daemon but if it cannot keep up this process starts to
2502  *	take the load in an attempt to prevent getnewbuf() from blocking.
2503  */
2504 
2505 static struct kproc_desc buf_kp = {
2506 	"bufdaemon",
2507 	buf_daemon,
2508 	&bufdaemonproc
2509 };
2510 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2511 
2512 static int
2513 buf_flush(struct vnode *vp, int target)
2514 {
2515 	int flushed;
2516 
2517 	flushed = flushbufqueues(vp, target, 0);
2518 	if (flushed == 0) {
2519 		/*
2520 		 * Could not find any buffers without rollback
2521 		 * dependencies, so just write the first one
2522 		 * in the hopes of eventually making progress.
2523 		 */
2524 		if (vp != NULL && target > 2)
2525 			target /= 2;
2526 		flushbufqueues(vp, target, 1);
2527 	}
2528 	return (flushed);
2529 }
2530 
2531 static void
2532 buf_daemon()
2533 {
2534 	int lodirty;
2535 
2536 	/*
2537 	 * This process needs to be suspended prior to shutdown sync.
2538 	 */
2539 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2540 	    SHUTDOWN_PRI_LAST);
2541 
2542 	/*
2543 	 * This process is allowed to take the buffer cache to the limit
2544 	 */
2545 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2546 	mtx_lock(&bdlock);
2547 	for (;;) {
2548 		bd_request = 0;
2549 		mtx_unlock(&bdlock);
2550 
2551 		kproc_suspend_check(bufdaemonproc);
2552 		lodirty = lodirtybuffers;
2553 		if (bd_speedupreq) {
2554 			lodirty = numdirtybuffers / 2;
2555 			bd_speedupreq = 0;
2556 		}
2557 		/*
2558 		 * Do the flush.  Limit the amount of in-transit I/O we
2559 		 * allow to build up, otherwise we would completely saturate
2560 		 * the I/O system.
2561 		 */
2562 		while (numdirtybuffers > lodirty) {
2563 			if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
2564 				break;
2565 			kern_yield(PRI_USER);
2566 		}
2567 
2568 		/*
2569 		 * Only clear bd_request if we have reached our low water
2570 		 * mark.  The buf_daemon normally waits 1 second and
2571 		 * then incrementally flushes any dirty buffers that have
2572 		 * built up, within reason.
2573 		 *
2574 		 * If we were unable to hit our low water mark and couldn't
2575 		 * find any flushable buffers, we sleep for a short period
2576 		 * to avoid endless loops on unlockable buffers.
2577 		 */
2578 		mtx_lock(&bdlock);
2579 		if (numdirtybuffers <= lodirtybuffers) {
2580 			/*
2581 			 * We reached our low water mark, reset the
2582 			 * request and sleep until we are needed again.
2583 			 * The sleep is just so the suspend code works.
2584 			 */
2585 			bd_request = 0;
2586 			/*
2587 			 * Do an extra wakeup in case dirty threshold
2588 			 * changed via sysctl and the explicit transition
2589 			 * out of shortfall was missed.
2590 			 */
2591 			bdirtywakeup();
2592 			if (runningbufspace <= lorunningspace)
2593 				runningwakeup();
2594 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2595 		} else {
2596 			/*
2597 			 * We couldn't find any flushable dirty buffers but
2598 			 * still have too many dirty buffers, we
2599 			 * have to sleep and try again.  (rare)
2600 			 */
2601 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2602 		}
2603 	}
2604 }
2605 
2606 /*
2607  *	flushbufqueues:
2608  *
2609  *	Try to flush a buffer in the dirty queue.  We must be careful to
2610  *	free up B_INVAL buffers instead of write them, which NFS is
2611  *	particularly sensitive to.
2612  */
2613 static int flushwithdeps = 0;
2614 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2615     0, "Number of buffers flushed with dependecies that require rollbacks");
2616 
2617 static int
2618 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
2619 {
2620 	struct buf *sentinel;
2621 	struct vnode *vp;
2622 	struct mount *mp;
2623 	struct buf *bp;
2624 	int hasdeps;
2625 	int flushed;
2626 	int queue;
2627 
2628 	flushed = 0;
2629 	queue = QUEUE_DIRTY;
2630 	bp = NULL;
2631 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2632 	sentinel->b_qindex = QUEUE_SENTINEL;
2633 	mtx_lock(&bqdirty);
2634 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2635 	while (flushed != target) {
2636 		bp = TAILQ_NEXT(sentinel, b_freelist);
2637 		if (bp != NULL) {
2638 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2639 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2640 			    b_freelist);
2641 		} else
2642 			break;
2643 		/*
2644 		 * Skip sentinels inserted by other invocations of the
2645 		 * flushbufqueues(), taking care to not reorder them.
2646 		 */
2647 		if (bp->b_qindex == QUEUE_SENTINEL)
2648 			continue;
2649 		/*
2650 		 * Only flush the buffers that belong to the
2651 		 * vnode locked by the curthread.
2652 		 */
2653 		if (lvp != NULL && bp->b_vp != lvp)
2654 			continue;
2655 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2656 			continue;
2657 		if (bp->b_pin_count > 0) {
2658 			BUF_UNLOCK(bp);
2659 			continue;
2660 		}
2661 		/*
2662 		 * BKGRDINPROG can only be set with the buf and bufobj
2663 		 * locks both held.  We tolerate a race to clear it here.
2664 		 */
2665 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2666 		    (bp->b_flags & B_DELWRI) == 0) {
2667 			BUF_UNLOCK(bp);
2668 			continue;
2669 		}
2670 		if (bp->b_flags & B_INVAL) {
2671 			bremfreel(bp);
2672 			mtx_unlock(&bqdirty);
2673 			brelse(bp);
2674 			flushed++;
2675 			mtx_lock(&bqdirty);
2676 			continue;
2677 		}
2678 
2679 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2680 			if (flushdeps == 0) {
2681 				BUF_UNLOCK(bp);
2682 				continue;
2683 			}
2684 			hasdeps = 1;
2685 		} else
2686 			hasdeps = 0;
2687 		/*
2688 		 * We must hold the lock on a vnode before writing
2689 		 * one of its buffers. Otherwise we may confuse, or
2690 		 * in the case of a snapshot vnode, deadlock the
2691 		 * system.
2692 		 *
2693 		 * The lock order here is the reverse of the normal
2694 		 * of vnode followed by buf lock.  This is ok because
2695 		 * the NOWAIT will prevent deadlock.
2696 		 */
2697 		vp = bp->b_vp;
2698 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2699 			BUF_UNLOCK(bp);
2700 			continue;
2701 		}
2702 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2703 			mtx_unlock(&bqdirty);
2704 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2705 			    bp, bp->b_vp, bp->b_flags);
2706 			if (curproc == bufdaemonproc)
2707 				vfs_bio_awrite(bp);
2708 			else {
2709 				bremfree(bp);
2710 				bwrite(bp);
2711 				notbufdflushes++;
2712 			}
2713 			vn_finished_write(mp);
2714 			VOP_UNLOCK(vp, 0);
2715 			flushwithdeps += hasdeps;
2716 			flushed++;
2717 
2718 			/*
2719 			 * Sleeping on runningbufspace while holding
2720 			 * vnode lock leads to deadlock.
2721 			 */
2722 			if (curproc == bufdaemonproc &&
2723 			    runningbufspace > hirunningspace)
2724 				waitrunningbufspace();
2725 			mtx_lock(&bqdirty);
2726 			continue;
2727 		}
2728 		vn_finished_write(mp);
2729 		BUF_UNLOCK(bp);
2730 	}
2731 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2732 	mtx_unlock(&bqdirty);
2733 	free(sentinel, M_TEMP);
2734 	return (flushed);
2735 }
2736 
2737 /*
2738  * Check to see if a block is currently memory resident.
2739  */
2740 struct buf *
2741 incore(struct bufobj *bo, daddr_t blkno)
2742 {
2743 	struct buf *bp;
2744 
2745 	BO_RLOCK(bo);
2746 	bp = gbincore(bo, blkno);
2747 	BO_RUNLOCK(bo);
2748 	return (bp);
2749 }
2750 
2751 /*
2752  * Returns true if no I/O is needed to access the
2753  * associated VM object.  This is like incore except
2754  * it also hunts around in the VM system for the data.
2755  */
2756 
2757 static int
2758 inmem(struct vnode * vp, daddr_t blkno)
2759 {
2760 	vm_object_t obj;
2761 	vm_offset_t toff, tinc, size;
2762 	vm_page_t m;
2763 	vm_ooffset_t off;
2764 
2765 	ASSERT_VOP_LOCKED(vp, "inmem");
2766 
2767 	if (incore(&vp->v_bufobj, blkno))
2768 		return 1;
2769 	if (vp->v_mount == NULL)
2770 		return 0;
2771 	obj = vp->v_object;
2772 	if (obj == NULL)
2773 		return (0);
2774 
2775 	size = PAGE_SIZE;
2776 	if (size > vp->v_mount->mnt_stat.f_iosize)
2777 		size = vp->v_mount->mnt_stat.f_iosize;
2778 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2779 
2780 	VM_OBJECT_RLOCK(obj);
2781 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2782 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2783 		if (!m)
2784 			goto notinmem;
2785 		tinc = size;
2786 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2787 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2788 		if (vm_page_is_valid(m,
2789 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2790 			goto notinmem;
2791 	}
2792 	VM_OBJECT_RUNLOCK(obj);
2793 	return 1;
2794 
2795 notinmem:
2796 	VM_OBJECT_RUNLOCK(obj);
2797 	return (0);
2798 }
2799 
2800 /*
2801  * Set the dirty range for a buffer based on the status of the dirty
2802  * bits in the pages comprising the buffer.  The range is limited
2803  * to the size of the buffer.
2804  *
2805  * Tell the VM system that the pages associated with this buffer
2806  * are clean.  This is used for delayed writes where the data is
2807  * going to go to disk eventually without additional VM intevention.
2808  *
2809  * Note that while we only really need to clean through to b_bcount, we
2810  * just go ahead and clean through to b_bufsize.
2811  */
2812 static void
2813 vfs_clean_pages_dirty_buf(struct buf *bp)
2814 {
2815 	vm_ooffset_t foff, noff, eoff;
2816 	vm_page_t m;
2817 	int i;
2818 
2819 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2820 		return;
2821 
2822 	foff = bp->b_offset;
2823 	KASSERT(bp->b_offset != NOOFFSET,
2824 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2825 
2826 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2827 	vfs_drain_busy_pages(bp);
2828 	vfs_setdirty_locked_object(bp);
2829 	for (i = 0; i < bp->b_npages; i++) {
2830 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2831 		eoff = noff;
2832 		if (eoff > bp->b_offset + bp->b_bufsize)
2833 			eoff = bp->b_offset + bp->b_bufsize;
2834 		m = bp->b_pages[i];
2835 		vfs_page_set_validclean(bp, foff, m);
2836 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2837 		foff = noff;
2838 	}
2839 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2840 }
2841 
2842 static void
2843 vfs_setdirty_locked_object(struct buf *bp)
2844 {
2845 	vm_object_t object;
2846 	int i;
2847 
2848 	object = bp->b_bufobj->bo_object;
2849 	VM_OBJECT_ASSERT_WLOCKED(object);
2850 
2851 	/*
2852 	 * We qualify the scan for modified pages on whether the
2853 	 * object has been flushed yet.
2854 	 */
2855 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2856 		vm_offset_t boffset;
2857 		vm_offset_t eoffset;
2858 
2859 		/*
2860 		 * test the pages to see if they have been modified directly
2861 		 * by users through the VM system.
2862 		 */
2863 		for (i = 0; i < bp->b_npages; i++)
2864 			vm_page_test_dirty(bp->b_pages[i]);
2865 
2866 		/*
2867 		 * Calculate the encompassing dirty range, boffset and eoffset,
2868 		 * (eoffset - boffset) bytes.
2869 		 */
2870 
2871 		for (i = 0; i < bp->b_npages; i++) {
2872 			if (bp->b_pages[i]->dirty)
2873 				break;
2874 		}
2875 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2876 
2877 		for (i = bp->b_npages - 1; i >= 0; --i) {
2878 			if (bp->b_pages[i]->dirty) {
2879 				break;
2880 			}
2881 		}
2882 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2883 
2884 		/*
2885 		 * Fit it to the buffer.
2886 		 */
2887 
2888 		if (eoffset > bp->b_bcount)
2889 			eoffset = bp->b_bcount;
2890 
2891 		/*
2892 		 * If we have a good dirty range, merge with the existing
2893 		 * dirty range.
2894 		 */
2895 
2896 		if (boffset < eoffset) {
2897 			if (bp->b_dirtyoff > boffset)
2898 				bp->b_dirtyoff = boffset;
2899 			if (bp->b_dirtyend < eoffset)
2900 				bp->b_dirtyend = eoffset;
2901 		}
2902 	}
2903 }
2904 
2905 /*
2906  * Allocate the KVA mapping for an existing buffer. It handles the
2907  * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2908  * KVA which is not mapped (B_KVAALLOC).
2909  */
2910 static void
2911 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2912 {
2913 	struct buf *scratch_bp;
2914 	int bsize, maxsize, need_mapping, need_kva;
2915 	off_t offset;
2916 
2917 	need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2918 	    (gbflags & GB_UNMAPPED) == 0;
2919 	need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2920 	    (gbflags & GB_KVAALLOC) != 0;
2921 	if (!need_mapping && !need_kva)
2922 		return;
2923 
2924 	BUF_CHECK_UNMAPPED(bp);
2925 
2926 	if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2927 		/*
2928 		 * Buffer is not mapped, but the KVA was already
2929 		 * reserved at the time of the instantiation.  Use the
2930 		 * allocated space.
2931 		 */
2932 		bp->b_flags &= ~B_KVAALLOC;
2933 		KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2934 		bp->b_kvabase = bp->b_kvaalloc;
2935 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
2936 		goto has_addr;
2937 	}
2938 
2939 	/*
2940 	 * Calculate the amount of the address space we would reserve
2941 	 * if the buffer was mapped.
2942 	 */
2943 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
2944 	offset = blkno * bsize;
2945 	maxsize = size + (offset & PAGE_MASK);
2946 	maxsize = imax(maxsize, bsize);
2947 
2948 mapping_loop:
2949 	if (allocbufkva(bp, maxsize, gbflags)) {
2950 		/*
2951 		 * Request defragmentation. getnewbuf() returns us the
2952 		 * allocated space by the scratch buffer KVA.
2953 		 */
2954 		scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
2955 		    (GB_UNMAPPED | GB_KVAALLOC));
2956 		if (scratch_bp == NULL) {
2957 			if ((gbflags & GB_NOWAIT_BD) != 0) {
2958 				/*
2959 				 * XXXKIB: defragmentation cannot
2960 				 * succeed, not sure what else to do.
2961 				 */
2962 				panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
2963 			}
2964 			atomic_add_int(&mappingrestarts, 1);
2965 			goto mapping_loop;
2966 		}
2967 		KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
2968 		    ("scratch bp !B_KVAALLOC %p", scratch_bp));
2969 		setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
2970 		    scratch_bp->b_kvasize, gbflags);
2971 
2972 		/* Get rid of the scratch buffer. */
2973 		scratch_bp->b_kvasize = 0;
2974 		scratch_bp->b_flags |= B_INVAL;
2975 		scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
2976 		brelse(scratch_bp);
2977 	}
2978 	if (!need_mapping)
2979 		return;
2980 
2981 has_addr:
2982 	bp->b_saveaddr = bp->b_kvabase;
2983 	bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
2984 	bp->b_flags &= ~B_UNMAPPED;
2985 	BUF_CHECK_MAPPED(bp);
2986 	bpmap_qenter(bp);
2987 }
2988 
2989 /*
2990  *	getblk:
2991  *
2992  *	Get a block given a specified block and offset into a file/device.
2993  *	The buffers B_DONE bit will be cleared on return, making it almost
2994  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2995  *	return.  The caller should clear B_INVAL prior to initiating a
2996  *	READ.
2997  *
2998  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2999  *	an existing buffer.
3000  *
3001  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3002  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3003  *	and then cleared based on the backing VM.  If the previous buffer is
3004  *	non-0-sized but invalid, B_CACHE will be cleared.
3005  *
3006  *	If getblk() must create a new buffer, the new buffer is returned with
3007  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3008  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3009  *	backing VM.
3010  *
3011  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3012  *	B_CACHE bit is clear.
3013  *
3014  *	What this means, basically, is that the caller should use B_CACHE to
3015  *	determine whether the buffer is fully valid or not and should clear
3016  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3017  *	the buffer by loading its data area with something, the caller needs
3018  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3019  *	the caller should set B_CACHE ( as an optimization ), else the caller
3020  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3021  *	a write attempt or if it was a successfull read.  If the caller
3022  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3023  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3024  */
3025 struct buf *
3026 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3027     int flags)
3028 {
3029 	struct buf *bp;
3030 	struct bufobj *bo;
3031 	int bsize, error, maxsize, vmio;
3032 	off_t offset;
3033 
3034 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3035 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3036 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3037 	ASSERT_VOP_LOCKED(vp, "getblk");
3038 	if (size > MAXBSIZE)
3039 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
3040 	if (!unmapped_buf_allowed)
3041 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3042 
3043 	bo = &vp->v_bufobj;
3044 loop:
3045 	BO_RLOCK(bo);
3046 	bp = gbincore(bo, blkno);
3047 	if (bp != NULL) {
3048 		int lockflags;
3049 		/*
3050 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3051 		 * it must be on a queue.
3052 		 */
3053 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3054 
3055 		if (flags & GB_LOCK_NOWAIT)
3056 			lockflags |= LK_NOWAIT;
3057 
3058 		error = BUF_TIMELOCK(bp, lockflags,
3059 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3060 
3061 		/*
3062 		 * If we slept and got the lock we have to restart in case
3063 		 * the buffer changed identities.
3064 		 */
3065 		if (error == ENOLCK)
3066 			goto loop;
3067 		/* We timed out or were interrupted. */
3068 		else if (error)
3069 			return (NULL);
3070 		/* If recursed, assume caller knows the rules. */
3071 		else if (BUF_LOCKRECURSED(bp))
3072 			goto end;
3073 
3074 		/*
3075 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3076 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3077 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3078 		 * backing VM cache.
3079 		 */
3080 		if (bp->b_flags & B_INVAL)
3081 			bp->b_flags &= ~B_CACHE;
3082 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3083 			bp->b_flags |= B_CACHE;
3084 		if (bp->b_flags & B_MANAGED)
3085 			MPASS(bp->b_qindex == QUEUE_NONE);
3086 		else
3087 			bremfree(bp);
3088 
3089 		/*
3090 		 * check for size inconsistencies for non-VMIO case.
3091 		 */
3092 		if (bp->b_bcount != size) {
3093 			if ((bp->b_flags & B_VMIO) == 0 ||
3094 			    (size > bp->b_kvasize)) {
3095 				if (bp->b_flags & B_DELWRI) {
3096 					/*
3097 					 * If buffer is pinned and caller does
3098 					 * not want sleep  waiting for it to be
3099 					 * unpinned, bail out
3100 					 * */
3101 					if (bp->b_pin_count > 0) {
3102 						if (flags & GB_LOCK_NOWAIT) {
3103 							bqrelse(bp);
3104 							return (NULL);
3105 						} else {
3106 							bunpin_wait(bp);
3107 						}
3108 					}
3109 					bp->b_flags |= B_NOCACHE;
3110 					bwrite(bp);
3111 				} else {
3112 					if (LIST_EMPTY(&bp->b_dep)) {
3113 						bp->b_flags |= B_RELBUF;
3114 						brelse(bp);
3115 					} else {
3116 						bp->b_flags |= B_NOCACHE;
3117 						bwrite(bp);
3118 					}
3119 				}
3120 				goto loop;
3121 			}
3122 		}
3123 
3124 		/*
3125 		 * Handle the case of unmapped buffer which should
3126 		 * become mapped, or the buffer for which KVA
3127 		 * reservation is requested.
3128 		 */
3129 		bp_unmapped_get_kva(bp, blkno, size, flags);
3130 
3131 		/*
3132 		 * If the size is inconsistant in the VMIO case, we can resize
3133 		 * the buffer.  This might lead to B_CACHE getting set or
3134 		 * cleared.  If the size has not changed, B_CACHE remains
3135 		 * unchanged from its previous state.
3136 		 */
3137 		if (bp->b_bcount != size)
3138 			allocbuf(bp, size);
3139 
3140 		KASSERT(bp->b_offset != NOOFFSET,
3141 		    ("getblk: no buffer offset"));
3142 
3143 		/*
3144 		 * A buffer with B_DELWRI set and B_CACHE clear must
3145 		 * be committed before we can return the buffer in
3146 		 * order to prevent the caller from issuing a read
3147 		 * ( due to B_CACHE not being set ) and overwriting
3148 		 * it.
3149 		 *
3150 		 * Most callers, including NFS and FFS, need this to
3151 		 * operate properly either because they assume they
3152 		 * can issue a read if B_CACHE is not set, or because
3153 		 * ( for example ) an uncached B_DELWRI might loop due
3154 		 * to softupdates re-dirtying the buffer.  In the latter
3155 		 * case, B_CACHE is set after the first write completes,
3156 		 * preventing further loops.
3157 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3158 		 * above while extending the buffer, we cannot allow the
3159 		 * buffer to remain with B_CACHE set after the write
3160 		 * completes or it will represent a corrupt state.  To
3161 		 * deal with this we set B_NOCACHE to scrap the buffer
3162 		 * after the write.
3163 		 *
3164 		 * We might be able to do something fancy, like setting
3165 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3166 		 * so the below call doesn't set B_CACHE, but that gets real
3167 		 * confusing.  This is much easier.
3168 		 */
3169 
3170 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3171 			bp->b_flags |= B_NOCACHE;
3172 			bwrite(bp);
3173 			goto loop;
3174 		}
3175 		bp->b_flags &= ~B_DONE;
3176 	} else {
3177 		/*
3178 		 * Buffer is not in-core, create new buffer.  The buffer
3179 		 * returned by getnewbuf() is locked.  Note that the returned
3180 		 * buffer is also considered valid (not marked B_INVAL).
3181 		 */
3182 		BO_RUNLOCK(bo);
3183 		/*
3184 		 * If the user does not want us to create the buffer, bail out
3185 		 * here.
3186 		 */
3187 		if (flags & GB_NOCREAT)
3188 			return NULL;
3189 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3190 			return NULL;
3191 
3192 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3193 		offset = blkno * bsize;
3194 		vmio = vp->v_object != NULL;
3195 		if (vmio) {
3196 			maxsize = size + (offset & PAGE_MASK);
3197 		} else {
3198 			maxsize = size;
3199 			/* Do not allow non-VMIO notmapped buffers. */
3200 			flags &= ~GB_UNMAPPED;
3201 		}
3202 		maxsize = imax(maxsize, bsize);
3203 
3204 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3205 		if (bp == NULL) {
3206 			if (slpflag || slptimeo)
3207 				return NULL;
3208 			goto loop;
3209 		}
3210 
3211 		/*
3212 		 * This code is used to make sure that a buffer is not
3213 		 * created while the getnewbuf routine is blocked.
3214 		 * This can be a problem whether the vnode is locked or not.
3215 		 * If the buffer is created out from under us, we have to
3216 		 * throw away the one we just created.
3217 		 *
3218 		 * Note: this must occur before we associate the buffer
3219 		 * with the vp especially considering limitations in
3220 		 * the splay tree implementation when dealing with duplicate
3221 		 * lblkno's.
3222 		 */
3223 		BO_LOCK(bo);
3224 		if (gbincore(bo, blkno)) {
3225 			BO_UNLOCK(bo);
3226 			bp->b_flags |= B_INVAL;
3227 			brelse(bp);
3228 			goto loop;
3229 		}
3230 
3231 		/*
3232 		 * Insert the buffer into the hash, so that it can
3233 		 * be found by incore.
3234 		 */
3235 		bp->b_blkno = bp->b_lblkno = blkno;
3236 		bp->b_offset = offset;
3237 		bgetvp(vp, bp);
3238 		BO_UNLOCK(bo);
3239 
3240 		/*
3241 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3242 		 * buffer size starts out as 0, B_CACHE will be set by
3243 		 * allocbuf() for the VMIO case prior to it testing the
3244 		 * backing store for validity.
3245 		 */
3246 
3247 		if (vmio) {
3248 			bp->b_flags |= B_VMIO;
3249 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3250 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3251 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3252 		} else {
3253 			bp->b_flags &= ~B_VMIO;
3254 			KASSERT(bp->b_bufobj->bo_object == NULL,
3255 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3256 			    bp, bp->b_bufobj->bo_object));
3257 			BUF_CHECK_MAPPED(bp);
3258 		}
3259 
3260 		allocbuf(bp, size);
3261 		bp->b_flags &= ~B_DONE;
3262 	}
3263 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3264 	BUF_ASSERT_HELD(bp);
3265 end:
3266 	KASSERT(bp->b_bufobj == bo,
3267 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3268 	return (bp);
3269 }
3270 
3271 /*
3272  * Get an empty, disassociated buffer of given size.  The buffer is initially
3273  * set to B_INVAL.
3274  */
3275 struct buf *
3276 geteblk(int size, int flags)
3277 {
3278 	struct buf *bp;
3279 	int maxsize;
3280 
3281 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3282 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3283 		if ((flags & GB_NOWAIT_BD) &&
3284 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3285 			return (NULL);
3286 	}
3287 	allocbuf(bp, size);
3288 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3289 	BUF_ASSERT_HELD(bp);
3290 	return (bp);
3291 }
3292 
3293 
3294 /*
3295  * This code constitutes the buffer memory from either anonymous system
3296  * memory (in the case of non-VMIO operations) or from an associated
3297  * VM object (in the case of VMIO operations).  This code is able to
3298  * resize a buffer up or down.
3299  *
3300  * Note that this code is tricky, and has many complications to resolve
3301  * deadlock or inconsistant data situations.  Tread lightly!!!
3302  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3303  * the caller.  Calling this code willy nilly can result in the loss of data.
3304  *
3305  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3306  * B_CACHE for the non-VMIO case.
3307  */
3308 
3309 int
3310 allocbuf(struct buf *bp, int size)
3311 {
3312 	int newbsize, mbsize;
3313 	int i;
3314 
3315 	BUF_ASSERT_HELD(bp);
3316 
3317 	if (bp->b_kvasize < size)
3318 		panic("allocbuf: buffer too small");
3319 
3320 	if ((bp->b_flags & B_VMIO) == 0) {
3321 		caddr_t origbuf;
3322 		int origbufsize;
3323 		/*
3324 		 * Just get anonymous memory from the kernel.  Don't
3325 		 * mess with B_CACHE.
3326 		 */
3327 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3328 		if (bp->b_flags & B_MALLOC)
3329 			newbsize = mbsize;
3330 		else
3331 			newbsize = round_page(size);
3332 
3333 		if (newbsize < bp->b_bufsize) {
3334 			/*
3335 			 * malloced buffers are not shrunk
3336 			 */
3337 			if (bp->b_flags & B_MALLOC) {
3338 				if (newbsize) {
3339 					bp->b_bcount = size;
3340 				} else {
3341 					free(bp->b_data, M_BIOBUF);
3342 					if (bp->b_bufsize) {
3343 						atomic_subtract_long(
3344 						    &bufmallocspace,
3345 						    bp->b_bufsize);
3346 						bufspacewakeup();
3347 						bp->b_bufsize = 0;
3348 					}
3349 					bp->b_saveaddr = bp->b_kvabase;
3350 					bp->b_data = bp->b_saveaddr;
3351 					bp->b_bcount = 0;
3352 					bp->b_flags &= ~B_MALLOC;
3353 				}
3354 				return 1;
3355 			}
3356 			vm_hold_free_pages(bp, newbsize);
3357 		} else if (newbsize > bp->b_bufsize) {
3358 			/*
3359 			 * We only use malloced memory on the first allocation.
3360 			 * and revert to page-allocated memory when the buffer
3361 			 * grows.
3362 			 */
3363 			/*
3364 			 * There is a potential smp race here that could lead
3365 			 * to bufmallocspace slightly passing the max.  It
3366 			 * is probably extremely rare and not worth worrying
3367 			 * over.
3368 			 */
3369 			if ( (bufmallocspace < maxbufmallocspace) &&
3370 				(bp->b_bufsize == 0) &&
3371 				(mbsize <= PAGE_SIZE/2)) {
3372 
3373 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3374 				bp->b_bufsize = mbsize;
3375 				bp->b_bcount = size;
3376 				bp->b_flags |= B_MALLOC;
3377 				atomic_add_long(&bufmallocspace, mbsize);
3378 				return 1;
3379 			}
3380 			origbuf = NULL;
3381 			origbufsize = 0;
3382 			/*
3383 			 * If the buffer is growing on its other-than-first allocation,
3384 			 * then we revert to the page-allocation scheme.
3385 			 */
3386 			if (bp->b_flags & B_MALLOC) {
3387 				origbuf = bp->b_data;
3388 				origbufsize = bp->b_bufsize;
3389 				bp->b_data = bp->b_kvabase;
3390 				if (bp->b_bufsize) {
3391 					atomic_subtract_long(&bufmallocspace,
3392 					    bp->b_bufsize);
3393 					bufspacewakeup();
3394 					bp->b_bufsize = 0;
3395 				}
3396 				bp->b_flags &= ~B_MALLOC;
3397 				newbsize = round_page(newbsize);
3398 			}
3399 			vm_hold_load_pages(
3400 			    bp,
3401 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
3402 			    (vm_offset_t) bp->b_data + newbsize);
3403 			if (origbuf) {
3404 				bcopy(origbuf, bp->b_data, origbufsize);
3405 				free(origbuf, M_BIOBUF);
3406 			}
3407 		}
3408 	} else {
3409 		int desiredpages;
3410 
3411 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3412 		desiredpages = (size == 0) ? 0 :
3413 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3414 
3415 		if (bp->b_flags & B_MALLOC)
3416 			panic("allocbuf: VMIO buffer can't be malloced");
3417 		/*
3418 		 * Set B_CACHE initially if buffer is 0 length or will become
3419 		 * 0-length.
3420 		 */
3421 		if (size == 0 || bp->b_bufsize == 0)
3422 			bp->b_flags |= B_CACHE;
3423 
3424 		if (newbsize < bp->b_bufsize) {
3425 			/*
3426 			 * DEV_BSIZE aligned new buffer size is less then the
3427 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3428 			 * if we have to remove any pages.
3429 			 */
3430 			if (desiredpages < bp->b_npages) {
3431 				vm_page_t m;
3432 
3433 				if ((bp->b_flags & B_UNMAPPED) == 0) {
3434 					BUF_CHECK_MAPPED(bp);
3435 					pmap_qremove((vm_offset_t)trunc_page(
3436 					    (vm_offset_t)bp->b_data) +
3437 					    (desiredpages << PAGE_SHIFT),
3438 					    (bp->b_npages - desiredpages));
3439 				} else
3440 					BUF_CHECK_UNMAPPED(bp);
3441 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3442 				for (i = desiredpages; i < bp->b_npages; i++) {
3443 					/*
3444 					 * the page is not freed here -- it
3445 					 * is the responsibility of
3446 					 * vnode_pager_setsize
3447 					 */
3448 					m = bp->b_pages[i];
3449 					KASSERT(m != bogus_page,
3450 					    ("allocbuf: bogus page found"));
3451 					while (vm_page_sleep_if_busy(m,
3452 					    "biodep"))
3453 						continue;
3454 
3455 					bp->b_pages[i] = NULL;
3456 					vm_page_lock(m);
3457 					vm_page_unwire(m, 0);
3458 					vm_page_unlock(m);
3459 				}
3460 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3461 				bp->b_npages = desiredpages;
3462 			}
3463 		} else if (size > bp->b_bcount) {
3464 			/*
3465 			 * We are growing the buffer, possibly in a
3466 			 * byte-granular fashion.
3467 			 */
3468 			vm_object_t obj;
3469 			vm_offset_t toff;
3470 			vm_offset_t tinc;
3471 
3472 			/*
3473 			 * Step 1, bring in the VM pages from the object,
3474 			 * allocating them if necessary.  We must clear
3475 			 * B_CACHE if these pages are not valid for the
3476 			 * range covered by the buffer.
3477 			 */
3478 
3479 			obj = bp->b_bufobj->bo_object;
3480 
3481 			VM_OBJECT_WLOCK(obj);
3482 			while (bp->b_npages < desiredpages) {
3483 				vm_page_t m;
3484 
3485 				/*
3486 				 * We must allocate system pages since blocking
3487 				 * here could interfere with paging I/O, no
3488 				 * matter which process we are.
3489 				 *
3490 				 * Only exclusive busy can be tested here.
3491 				 * Blocking on shared busy might lead to
3492 				 * deadlocks once allocbuf() is called after
3493 				 * pages are vfs_busy_pages().
3494 				 */
3495 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3496 				    bp->b_npages, VM_ALLOC_NOBUSY |
3497 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3498 				    VM_ALLOC_IGN_SBUSY |
3499 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3500 				if (m->valid == 0)
3501 					bp->b_flags &= ~B_CACHE;
3502 				bp->b_pages[bp->b_npages] = m;
3503 				++bp->b_npages;
3504 			}
3505 
3506 			/*
3507 			 * Step 2.  We've loaded the pages into the buffer,
3508 			 * we have to figure out if we can still have B_CACHE
3509 			 * set.  Note that B_CACHE is set according to the
3510 			 * byte-granular range ( bcount and size ), new the
3511 			 * aligned range ( newbsize ).
3512 			 *
3513 			 * The VM test is against m->valid, which is DEV_BSIZE
3514 			 * aligned.  Needless to say, the validity of the data
3515 			 * needs to also be DEV_BSIZE aligned.  Note that this
3516 			 * fails with NFS if the server or some other client
3517 			 * extends the file's EOF.  If our buffer is resized,
3518 			 * B_CACHE may remain set! XXX
3519 			 */
3520 
3521 			toff = bp->b_bcount;
3522 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3523 
3524 			while ((bp->b_flags & B_CACHE) && toff < size) {
3525 				vm_pindex_t pi;
3526 
3527 				if (tinc > (size - toff))
3528 					tinc = size - toff;
3529 
3530 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3531 				    PAGE_SHIFT;
3532 
3533 				vfs_buf_test_cache(
3534 				    bp,
3535 				    bp->b_offset,
3536 				    toff,
3537 				    tinc,
3538 				    bp->b_pages[pi]
3539 				);
3540 				toff += tinc;
3541 				tinc = PAGE_SIZE;
3542 			}
3543 			VM_OBJECT_WUNLOCK(obj);
3544 
3545 			/*
3546 			 * Step 3, fixup the KVM pmap.
3547 			 */
3548 			if ((bp->b_flags & B_UNMAPPED) == 0)
3549 				bpmap_qenter(bp);
3550 			else
3551 				BUF_CHECK_UNMAPPED(bp);
3552 		}
3553 	}
3554 	if (newbsize < bp->b_bufsize)
3555 		bufspacewakeup();
3556 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3557 	bp->b_bcount = size;		/* requested buffer size	*/
3558 	return 1;
3559 }
3560 
3561 extern int inflight_transient_maps;
3562 
3563 void
3564 biodone(struct bio *bp)
3565 {
3566 	struct mtx *mtxp;
3567 	void (*done)(struct bio *);
3568 	vm_offset_t start, end;
3569 	int transient;
3570 
3571 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3572 	mtx_lock(mtxp);
3573 	bp->bio_flags |= BIO_DONE;
3574 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3575 		start = trunc_page((vm_offset_t)bp->bio_data);
3576 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3577 		transient = 1;
3578 	} else {
3579 		transient = 0;
3580 		start = end = 0;
3581 	}
3582 	done = bp->bio_done;
3583 	if (done == NULL)
3584 		wakeup(bp);
3585 	mtx_unlock(mtxp);
3586 	if (done != NULL)
3587 		done(bp);
3588 	if (transient) {
3589 		pmap_qremove(start, OFF_TO_IDX(end - start));
3590 		vmem_free(transient_arena, start, end - start);
3591 		atomic_add_int(&inflight_transient_maps, -1);
3592 	}
3593 }
3594 
3595 /*
3596  * Wait for a BIO to finish.
3597  *
3598  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3599  * case is not yet clear.
3600  */
3601 int
3602 biowait(struct bio *bp, const char *wchan)
3603 {
3604 	struct mtx *mtxp;
3605 
3606 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3607 	mtx_lock(mtxp);
3608 	while ((bp->bio_flags & BIO_DONE) == 0)
3609 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3610 	mtx_unlock(mtxp);
3611 	if (bp->bio_error != 0)
3612 		return (bp->bio_error);
3613 	if (!(bp->bio_flags & BIO_ERROR))
3614 		return (0);
3615 	return (EIO);
3616 }
3617 
3618 void
3619 biofinish(struct bio *bp, struct devstat *stat, int error)
3620 {
3621 
3622 	if (error) {
3623 		bp->bio_error = error;
3624 		bp->bio_flags |= BIO_ERROR;
3625 	}
3626 	if (stat != NULL)
3627 		devstat_end_transaction_bio(stat, bp);
3628 	biodone(bp);
3629 }
3630 
3631 /*
3632  *	bufwait:
3633  *
3634  *	Wait for buffer I/O completion, returning error status.  The buffer
3635  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3636  *	error and cleared.
3637  */
3638 int
3639 bufwait(struct buf *bp)
3640 {
3641 	if (bp->b_iocmd == BIO_READ)
3642 		bwait(bp, PRIBIO, "biord");
3643 	else
3644 		bwait(bp, PRIBIO, "biowr");
3645 	if (bp->b_flags & B_EINTR) {
3646 		bp->b_flags &= ~B_EINTR;
3647 		return (EINTR);
3648 	}
3649 	if (bp->b_ioflags & BIO_ERROR) {
3650 		return (bp->b_error ? bp->b_error : EIO);
3651 	} else {
3652 		return (0);
3653 	}
3654 }
3655 
3656  /*
3657   * Call back function from struct bio back up to struct buf.
3658   */
3659 static void
3660 bufdonebio(struct bio *bip)
3661 {
3662 	struct buf *bp;
3663 
3664 	bp = bip->bio_caller2;
3665 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3666 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3667 	bp->b_ioflags = bip->bio_flags;
3668 	bp->b_error = bip->bio_error;
3669 	if (bp->b_error)
3670 		bp->b_ioflags |= BIO_ERROR;
3671 	bufdone(bp);
3672 	g_destroy_bio(bip);
3673 }
3674 
3675 void
3676 dev_strategy(struct cdev *dev, struct buf *bp)
3677 {
3678 	struct cdevsw *csw;
3679 	int ref;
3680 
3681 	KASSERT(dev->si_refcount > 0,
3682 	    ("dev_strategy on un-referenced struct cdev *(%s) %p",
3683 	    devtoname(dev), dev));
3684 
3685 	csw = dev_refthread(dev, &ref);
3686 	dev_strategy_csw(dev, csw, bp);
3687 	dev_relthread(dev, ref);
3688 }
3689 
3690 void
3691 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3692 {
3693 	struct bio *bip;
3694 
3695 	KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3696 	    ("b_iocmd botch"));
3697 	KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3698 	    dev->si_threadcount > 0,
3699 	    ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3700 	    dev));
3701 	if (csw == NULL) {
3702 		bp->b_error = ENXIO;
3703 		bp->b_ioflags = BIO_ERROR;
3704 		bufdone(bp);
3705 		return;
3706 	}
3707 	for (;;) {
3708 		bip = g_new_bio();
3709 		if (bip != NULL)
3710 			break;
3711 		/* Try again later */
3712 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3713 	}
3714 	bip->bio_cmd = bp->b_iocmd;
3715 	bip->bio_offset = bp->b_iooffset;
3716 	bip->bio_length = bp->b_bcount;
3717 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3718 	bdata2bio(bp, bip);
3719 	bip->bio_done = bufdonebio;
3720 	bip->bio_caller2 = bp;
3721 	bip->bio_dev = dev;
3722 	(*csw->d_strategy)(bip);
3723 }
3724 
3725 /*
3726  *	bufdone:
3727  *
3728  *	Finish I/O on a buffer, optionally calling a completion function.
3729  *	This is usually called from an interrupt so process blocking is
3730  *	not allowed.
3731  *
3732  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3733  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3734  *	assuming B_INVAL is clear.
3735  *
3736  *	For the VMIO case, we set B_CACHE if the op was a read and no
3737  *	read error occured, or if the op was a write.  B_CACHE is never
3738  *	set if the buffer is invalid or otherwise uncacheable.
3739  *
3740  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3741  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3742  *	in the biodone routine.
3743  */
3744 void
3745 bufdone(struct buf *bp)
3746 {
3747 	struct bufobj *dropobj;
3748 	void    (*biodone)(struct buf *);
3749 
3750 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3751 	dropobj = NULL;
3752 
3753 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3754 	BUF_ASSERT_HELD(bp);
3755 
3756 	runningbufwakeup(bp);
3757 	if (bp->b_iocmd == BIO_WRITE)
3758 		dropobj = bp->b_bufobj;
3759 	/* call optional completion function if requested */
3760 	if (bp->b_iodone != NULL) {
3761 		biodone = bp->b_iodone;
3762 		bp->b_iodone = NULL;
3763 		(*biodone) (bp);
3764 		if (dropobj)
3765 			bufobj_wdrop(dropobj);
3766 		return;
3767 	}
3768 
3769 	bufdone_finish(bp);
3770 
3771 	if (dropobj)
3772 		bufobj_wdrop(dropobj);
3773 }
3774 
3775 void
3776 bufdone_finish(struct buf *bp)
3777 {
3778 	BUF_ASSERT_HELD(bp);
3779 
3780 	if (!LIST_EMPTY(&bp->b_dep))
3781 		buf_complete(bp);
3782 
3783 	if (bp->b_flags & B_VMIO) {
3784 		vm_ooffset_t foff;
3785 		vm_page_t m;
3786 		vm_object_t obj;
3787 		struct vnode *vp;
3788 		int bogus, i, iosize;
3789 
3790 		obj = bp->b_bufobj->bo_object;
3791 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3792 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3793 		    obj->paging_in_progress, bp->b_npages));
3794 
3795 		vp = bp->b_vp;
3796 		KASSERT(vp->v_holdcnt > 0,
3797 		    ("biodone_finish: vnode %p has zero hold count", vp));
3798 		KASSERT(vp->v_object != NULL,
3799 		    ("biodone_finish: vnode %p has no vm_object", vp));
3800 
3801 		foff = bp->b_offset;
3802 		KASSERT(bp->b_offset != NOOFFSET,
3803 		    ("biodone_finish: bp %p has no buffer offset", bp));
3804 
3805 		/*
3806 		 * Set B_CACHE if the op was a normal read and no error
3807 		 * occured.  B_CACHE is set for writes in the b*write()
3808 		 * routines.
3809 		 */
3810 		iosize = bp->b_bcount - bp->b_resid;
3811 		if (bp->b_iocmd == BIO_READ &&
3812 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3813 		    !(bp->b_ioflags & BIO_ERROR)) {
3814 			bp->b_flags |= B_CACHE;
3815 		}
3816 		bogus = 0;
3817 		VM_OBJECT_WLOCK(obj);
3818 		for (i = 0; i < bp->b_npages; i++) {
3819 			int bogusflag = 0;
3820 			int resid;
3821 
3822 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3823 			if (resid > iosize)
3824 				resid = iosize;
3825 
3826 			/*
3827 			 * cleanup bogus pages, restoring the originals
3828 			 */
3829 			m = bp->b_pages[i];
3830 			if (m == bogus_page) {
3831 				bogus = bogusflag = 1;
3832 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3833 				if (m == NULL)
3834 					panic("biodone: page disappeared!");
3835 				bp->b_pages[i] = m;
3836 			}
3837 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3838 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3839 			    (intmax_t)foff, (uintmax_t)m->pindex));
3840 
3841 			/*
3842 			 * In the write case, the valid and clean bits are
3843 			 * already changed correctly ( see bdwrite() ), so we
3844 			 * only need to do this here in the read case.
3845 			 */
3846 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3847 				KASSERT((m->dirty & vm_page_bits(foff &
3848 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3849 				    " page %p has unexpected dirty bits", m));
3850 				vfs_page_set_valid(bp, foff, m);
3851 			}
3852 
3853 			vm_page_sunbusy(m);
3854 			vm_object_pip_subtract(obj, 1);
3855 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3856 			iosize -= resid;
3857 		}
3858 		vm_object_pip_wakeupn(obj, 0);
3859 		VM_OBJECT_WUNLOCK(obj);
3860 		if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3861 			BUF_CHECK_MAPPED(bp);
3862 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3863 			    bp->b_pages, bp->b_npages);
3864 		}
3865 	}
3866 
3867 	/*
3868 	 * For asynchronous completions, release the buffer now. The brelse
3869 	 * will do a wakeup there if necessary - so no need to do a wakeup
3870 	 * here in the async case. The sync case always needs to do a wakeup.
3871 	 */
3872 
3873 	if (bp->b_flags & B_ASYNC) {
3874 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3875 			brelse(bp);
3876 		else
3877 			bqrelse(bp);
3878 	} else
3879 		bdone(bp);
3880 }
3881 
3882 /*
3883  * This routine is called in lieu of iodone in the case of
3884  * incomplete I/O.  This keeps the busy status for pages
3885  * consistant.
3886  */
3887 void
3888 vfs_unbusy_pages(struct buf *bp)
3889 {
3890 	int i;
3891 	vm_object_t obj;
3892 	vm_page_t m;
3893 
3894 	runningbufwakeup(bp);
3895 	if (!(bp->b_flags & B_VMIO))
3896 		return;
3897 
3898 	obj = bp->b_bufobj->bo_object;
3899 	VM_OBJECT_WLOCK(obj);
3900 	for (i = 0; i < bp->b_npages; i++) {
3901 		m = bp->b_pages[i];
3902 		if (m == bogus_page) {
3903 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3904 			if (!m)
3905 				panic("vfs_unbusy_pages: page missing\n");
3906 			bp->b_pages[i] = m;
3907 			if ((bp->b_flags & B_UNMAPPED) == 0) {
3908 				BUF_CHECK_MAPPED(bp);
3909 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3910 				    bp->b_pages, bp->b_npages);
3911 			} else
3912 				BUF_CHECK_UNMAPPED(bp);
3913 		}
3914 		vm_object_pip_subtract(obj, 1);
3915 		vm_page_sunbusy(m);
3916 	}
3917 	vm_object_pip_wakeupn(obj, 0);
3918 	VM_OBJECT_WUNLOCK(obj);
3919 }
3920 
3921 /*
3922  * vfs_page_set_valid:
3923  *
3924  *	Set the valid bits in a page based on the supplied offset.   The
3925  *	range is restricted to the buffer's size.
3926  *
3927  *	This routine is typically called after a read completes.
3928  */
3929 static void
3930 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3931 {
3932 	vm_ooffset_t eoff;
3933 
3934 	/*
3935 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3936 	 * page boundary and eoff is not greater than the end of the buffer.
3937 	 * The end of the buffer, in this case, is our file EOF, not the
3938 	 * allocation size of the buffer.
3939 	 */
3940 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3941 	if (eoff > bp->b_offset + bp->b_bcount)
3942 		eoff = bp->b_offset + bp->b_bcount;
3943 
3944 	/*
3945 	 * Set valid range.  This is typically the entire buffer and thus the
3946 	 * entire page.
3947 	 */
3948 	if (eoff > off)
3949 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3950 }
3951 
3952 /*
3953  * vfs_page_set_validclean:
3954  *
3955  *	Set the valid bits and clear the dirty bits in a page based on the
3956  *	supplied offset.   The range is restricted to the buffer's size.
3957  */
3958 static void
3959 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3960 {
3961 	vm_ooffset_t soff, eoff;
3962 
3963 	/*
3964 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3965 	 * page boundry or cross the end of the buffer.  The end of the
3966 	 * buffer, in this case, is our file EOF, not the allocation size
3967 	 * of the buffer.
3968 	 */
3969 	soff = off;
3970 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3971 	if (eoff > bp->b_offset + bp->b_bcount)
3972 		eoff = bp->b_offset + bp->b_bcount;
3973 
3974 	/*
3975 	 * Set valid range.  This is typically the entire buffer and thus the
3976 	 * entire page.
3977 	 */
3978 	if (eoff > soff) {
3979 		vm_page_set_validclean(
3980 		    m,
3981 		   (vm_offset_t) (soff & PAGE_MASK),
3982 		   (vm_offset_t) (eoff - soff)
3983 		);
3984 	}
3985 }
3986 
3987 /*
3988  * Ensure that all buffer pages are not exclusive busied.  If any page is
3989  * exclusive busy, drain it.
3990  */
3991 void
3992 vfs_drain_busy_pages(struct buf *bp)
3993 {
3994 	vm_page_t m;
3995 	int i, last_busied;
3996 
3997 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
3998 	last_busied = 0;
3999 	for (i = 0; i < bp->b_npages; i++) {
4000 		m = bp->b_pages[i];
4001 		if (vm_page_xbusied(m)) {
4002 			for (; last_busied < i; last_busied++)
4003 				vm_page_sbusy(bp->b_pages[last_busied]);
4004 			while (vm_page_xbusied(m)) {
4005 				vm_page_lock(m);
4006 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4007 				vm_page_busy_sleep(m, "vbpage");
4008 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4009 			}
4010 		}
4011 	}
4012 	for (i = 0; i < last_busied; i++)
4013 		vm_page_sunbusy(bp->b_pages[i]);
4014 }
4015 
4016 /*
4017  * This routine is called before a device strategy routine.
4018  * It is used to tell the VM system that paging I/O is in
4019  * progress, and treat the pages associated with the buffer
4020  * almost as being exclusive busy.  Also the object paging_in_progress
4021  * flag is handled to make sure that the object doesn't become
4022  * inconsistant.
4023  *
4024  * Since I/O has not been initiated yet, certain buffer flags
4025  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4026  * and should be ignored.
4027  */
4028 void
4029 vfs_busy_pages(struct buf *bp, int clear_modify)
4030 {
4031 	int i, bogus;
4032 	vm_object_t obj;
4033 	vm_ooffset_t foff;
4034 	vm_page_t m;
4035 
4036 	if (!(bp->b_flags & B_VMIO))
4037 		return;
4038 
4039 	obj = bp->b_bufobj->bo_object;
4040 	foff = bp->b_offset;
4041 	KASSERT(bp->b_offset != NOOFFSET,
4042 	    ("vfs_busy_pages: no buffer offset"));
4043 	VM_OBJECT_WLOCK(obj);
4044 	vfs_drain_busy_pages(bp);
4045 	if (bp->b_bufsize != 0)
4046 		vfs_setdirty_locked_object(bp);
4047 	bogus = 0;
4048 	for (i = 0; i < bp->b_npages; i++) {
4049 		m = bp->b_pages[i];
4050 
4051 		if ((bp->b_flags & B_CLUSTER) == 0) {
4052 			vm_object_pip_add(obj, 1);
4053 			vm_page_sbusy(m);
4054 		}
4055 		/*
4056 		 * When readying a buffer for a read ( i.e
4057 		 * clear_modify == 0 ), it is important to do
4058 		 * bogus_page replacement for valid pages in
4059 		 * partially instantiated buffers.  Partially
4060 		 * instantiated buffers can, in turn, occur when
4061 		 * reconstituting a buffer from its VM backing store
4062 		 * base.  We only have to do this if B_CACHE is
4063 		 * clear ( which causes the I/O to occur in the
4064 		 * first place ).  The replacement prevents the read
4065 		 * I/O from overwriting potentially dirty VM-backed
4066 		 * pages.  XXX bogus page replacement is, uh, bogus.
4067 		 * It may not work properly with small-block devices.
4068 		 * We need to find a better way.
4069 		 */
4070 		if (clear_modify) {
4071 			pmap_remove_write(m);
4072 			vfs_page_set_validclean(bp, foff, m);
4073 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4074 		    (bp->b_flags & B_CACHE) == 0) {
4075 			bp->b_pages[i] = bogus_page;
4076 			bogus++;
4077 		}
4078 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4079 	}
4080 	VM_OBJECT_WUNLOCK(obj);
4081 	if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4082 		BUF_CHECK_MAPPED(bp);
4083 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4084 		    bp->b_pages, bp->b_npages);
4085 	}
4086 }
4087 
4088 /*
4089  *	vfs_bio_set_valid:
4090  *
4091  *	Set the range within the buffer to valid.  The range is
4092  *	relative to the beginning of the buffer, b_offset.  Note that
4093  *	b_offset itself may be offset from the beginning of the first
4094  *	page.
4095  */
4096 void
4097 vfs_bio_set_valid(struct buf *bp, int base, int size)
4098 {
4099 	int i, n;
4100 	vm_page_t m;
4101 
4102 	if (!(bp->b_flags & B_VMIO))
4103 		return;
4104 
4105 	/*
4106 	 * Fixup base to be relative to beginning of first page.
4107 	 * Set initial n to be the maximum number of bytes in the
4108 	 * first page that can be validated.
4109 	 */
4110 	base += (bp->b_offset & PAGE_MASK);
4111 	n = PAGE_SIZE - (base & PAGE_MASK);
4112 
4113 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4114 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4115 		m = bp->b_pages[i];
4116 		if (n > size)
4117 			n = size;
4118 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4119 		base += n;
4120 		size -= n;
4121 		n = PAGE_SIZE;
4122 	}
4123 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4124 }
4125 
4126 /*
4127  *	vfs_bio_clrbuf:
4128  *
4129  *	If the specified buffer is a non-VMIO buffer, clear the entire
4130  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4131  *	validate only the previously invalid portions of the buffer.
4132  *	This routine essentially fakes an I/O, so we need to clear
4133  *	BIO_ERROR and B_INVAL.
4134  *
4135  *	Note that while we only theoretically need to clear through b_bcount,
4136  *	we go ahead and clear through b_bufsize.
4137  */
4138 void
4139 vfs_bio_clrbuf(struct buf *bp)
4140 {
4141 	int i, j, mask, sa, ea, slide;
4142 
4143 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4144 		clrbuf(bp);
4145 		return;
4146 	}
4147 	bp->b_flags &= ~B_INVAL;
4148 	bp->b_ioflags &= ~BIO_ERROR;
4149 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4150 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4151 	    (bp->b_offset & PAGE_MASK) == 0) {
4152 		if (bp->b_pages[0] == bogus_page)
4153 			goto unlock;
4154 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4155 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4156 		if ((bp->b_pages[0]->valid & mask) == mask)
4157 			goto unlock;
4158 		if ((bp->b_pages[0]->valid & mask) == 0) {
4159 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4160 			bp->b_pages[0]->valid |= mask;
4161 			goto unlock;
4162 		}
4163 	}
4164 	sa = bp->b_offset & PAGE_MASK;
4165 	slide = 0;
4166 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4167 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4168 		ea = slide & PAGE_MASK;
4169 		if (ea == 0)
4170 			ea = PAGE_SIZE;
4171 		if (bp->b_pages[i] == bogus_page)
4172 			continue;
4173 		j = sa / DEV_BSIZE;
4174 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4175 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4176 		if ((bp->b_pages[i]->valid & mask) == mask)
4177 			continue;
4178 		if ((bp->b_pages[i]->valid & mask) == 0)
4179 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4180 		else {
4181 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4182 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4183 					pmap_zero_page_area(bp->b_pages[i],
4184 					    sa, DEV_BSIZE);
4185 				}
4186 			}
4187 		}
4188 		bp->b_pages[i]->valid |= mask;
4189 	}
4190 unlock:
4191 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4192 	bp->b_resid = 0;
4193 }
4194 
4195 void
4196 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4197 {
4198 	vm_page_t m;
4199 	int i, n;
4200 
4201 	if ((bp->b_flags & B_UNMAPPED) == 0) {
4202 		BUF_CHECK_MAPPED(bp);
4203 		bzero(bp->b_data + base, size);
4204 	} else {
4205 		BUF_CHECK_UNMAPPED(bp);
4206 		n = PAGE_SIZE - (base & PAGE_MASK);
4207 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4208 			m = bp->b_pages[i];
4209 			if (n > size)
4210 				n = size;
4211 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4212 			base += n;
4213 			size -= n;
4214 			n = PAGE_SIZE;
4215 		}
4216 	}
4217 }
4218 
4219 /*
4220  * vm_hold_load_pages and vm_hold_free_pages get pages into
4221  * a buffers address space.  The pages are anonymous and are
4222  * not associated with a file object.
4223  */
4224 static void
4225 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4226 {
4227 	vm_offset_t pg;
4228 	vm_page_t p;
4229 	int index;
4230 
4231 	BUF_CHECK_MAPPED(bp);
4232 
4233 	to = round_page(to);
4234 	from = round_page(from);
4235 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4236 
4237 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4238 tryagain:
4239 		/*
4240 		 * note: must allocate system pages since blocking here
4241 		 * could interfere with paging I/O, no matter which
4242 		 * process we are.
4243 		 */
4244 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4245 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4246 		if (p == NULL) {
4247 			VM_WAIT;
4248 			goto tryagain;
4249 		}
4250 		pmap_qenter(pg, &p, 1);
4251 		bp->b_pages[index] = p;
4252 	}
4253 	bp->b_npages = index;
4254 }
4255 
4256 /* Return pages associated with this buf to the vm system */
4257 static void
4258 vm_hold_free_pages(struct buf *bp, int newbsize)
4259 {
4260 	vm_offset_t from;
4261 	vm_page_t p;
4262 	int index, newnpages;
4263 
4264 	BUF_CHECK_MAPPED(bp);
4265 
4266 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4267 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4268 	if (bp->b_npages > newnpages)
4269 		pmap_qremove(from, bp->b_npages - newnpages);
4270 	for (index = newnpages; index < bp->b_npages; index++) {
4271 		p = bp->b_pages[index];
4272 		bp->b_pages[index] = NULL;
4273 		if (vm_page_sbusied(p))
4274 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4275 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4276 		p->wire_count--;
4277 		vm_page_free(p);
4278 		atomic_subtract_int(&cnt.v_wire_count, 1);
4279 	}
4280 	bp->b_npages = newnpages;
4281 }
4282 
4283 /*
4284  * Map an IO request into kernel virtual address space.
4285  *
4286  * All requests are (re)mapped into kernel VA space.
4287  * Notice that we use b_bufsize for the size of the buffer
4288  * to be mapped.  b_bcount might be modified by the driver.
4289  *
4290  * Note that even if the caller determines that the address space should
4291  * be valid, a race or a smaller-file mapped into a larger space may
4292  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4293  * check the return value.
4294  */
4295 int
4296 vmapbuf(struct buf *bp, int mapbuf)
4297 {
4298 	caddr_t kva;
4299 	vm_prot_t prot;
4300 	int pidx;
4301 
4302 	if (bp->b_bufsize < 0)
4303 		return (-1);
4304 	prot = VM_PROT_READ;
4305 	if (bp->b_iocmd == BIO_READ)
4306 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4307 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4308 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4309 	    btoc(MAXPHYS))) < 0)
4310 		return (-1);
4311 	bp->b_npages = pidx;
4312 	if (mapbuf || !unmapped_buf_allowed) {
4313 		pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4314 		kva = bp->b_saveaddr;
4315 		bp->b_saveaddr = bp->b_data;
4316 		bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4317 		bp->b_flags &= ~B_UNMAPPED;
4318 	} else {
4319 		bp->b_flags |= B_UNMAPPED;
4320 		bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4321 		bp->b_saveaddr = bp->b_data;
4322 		bp->b_data = unmapped_buf;
4323 	}
4324 	return(0);
4325 }
4326 
4327 /*
4328  * Free the io map PTEs associated with this IO operation.
4329  * We also invalidate the TLB entries and restore the original b_addr.
4330  */
4331 void
4332 vunmapbuf(struct buf *bp)
4333 {
4334 	int npages;
4335 
4336 	npages = bp->b_npages;
4337 	if (bp->b_flags & B_UNMAPPED)
4338 		bp->b_flags &= ~B_UNMAPPED;
4339 	else
4340 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4341 	vm_page_unhold_pages(bp->b_pages, npages);
4342 
4343 	bp->b_data = bp->b_saveaddr;
4344 }
4345 
4346 void
4347 bdone(struct buf *bp)
4348 {
4349 	struct mtx *mtxp;
4350 
4351 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4352 	mtx_lock(mtxp);
4353 	bp->b_flags |= B_DONE;
4354 	wakeup(bp);
4355 	mtx_unlock(mtxp);
4356 }
4357 
4358 void
4359 bwait(struct buf *bp, u_char pri, const char *wchan)
4360 {
4361 	struct mtx *mtxp;
4362 
4363 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4364 	mtx_lock(mtxp);
4365 	while ((bp->b_flags & B_DONE) == 0)
4366 		msleep(bp, mtxp, pri, wchan, 0);
4367 	mtx_unlock(mtxp);
4368 }
4369 
4370 int
4371 bufsync(struct bufobj *bo, int waitfor)
4372 {
4373 
4374 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4375 }
4376 
4377 void
4378 bufstrategy(struct bufobj *bo, struct buf *bp)
4379 {
4380 	int i = 0;
4381 	struct vnode *vp;
4382 
4383 	vp = bp->b_vp;
4384 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4385 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4386 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4387 	i = VOP_STRATEGY(vp, bp);
4388 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4389 }
4390 
4391 void
4392 bufobj_wrefl(struct bufobj *bo)
4393 {
4394 
4395 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4396 	ASSERT_BO_WLOCKED(bo);
4397 	bo->bo_numoutput++;
4398 }
4399 
4400 void
4401 bufobj_wref(struct bufobj *bo)
4402 {
4403 
4404 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4405 	BO_LOCK(bo);
4406 	bo->bo_numoutput++;
4407 	BO_UNLOCK(bo);
4408 }
4409 
4410 void
4411 bufobj_wdrop(struct bufobj *bo)
4412 {
4413 
4414 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4415 	BO_LOCK(bo);
4416 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4417 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4418 		bo->bo_flag &= ~BO_WWAIT;
4419 		wakeup(&bo->bo_numoutput);
4420 	}
4421 	BO_UNLOCK(bo);
4422 }
4423 
4424 int
4425 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4426 {
4427 	int error;
4428 
4429 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4430 	ASSERT_BO_WLOCKED(bo);
4431 	error = 0;
4432 	while (bo->bo_numoutput) {
4433 		bo->bo_flag |= BO_WWAIT;
4434 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4435 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4436 		if (error)
4437 			break;
4438 	}
4439 	return (error);
4440 }
4441 
4442 void
4443 bpin(struct buf *bp)
4444 {
4445 	struct mtx *mtxp;
4446 
4447 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4448 	mtx_lock(mtxp);
4449 	bp->b_pin_count++;
4450 	mtx_unlock(mtxp);
4451 }
4452 
4453 void
4454 bunpin(struct buf *bp)
4455 {
4456 	struct mtx *mtxp;
4457 
4458 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4459 	mtx_lock(mtxp);
4460 	if (--bp->b_pin_count == 0)
4461 		wakeup(bp);
4462 	mtx_unlock(mtxp);
4463 }
4464 
4465 void
4466 bunpin_wait(struct buf *bp)
4467 {
4468 	struct mtx *mtxp;
4469 
4470 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4471 	mtx_lock(mtxp);
4472 	while (bp->b_pin_count > 0)
4473 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4474 	mtx_unlock(mtxp);
4475 }
4476 
4477 /*
4478  * Set bio_data or bio_ma for struct bio from the struct buf.
4479  */
4480 void
4481 bdata2bio(struct buf *bp, struct bio *bip)
4482 {
4483 
4484 	if ((bp->b_flags & B_UNMAPPED) != 0) {
4485 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4486 		bip->bio_ma = bp->b_pages;
4487 		bip->bio_ma_n = bp->b_npages;
4488 		bip->bio_data = unmapped_buf;
4489 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4490 		bip->bio_flags |= BIO_UNMAPPED;
4491 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4492 		    PAGE_SIZE == bp->b_npages,
4493 		    ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4494 		    (long long)bip->bio_length, bip->bio_ma_n));
4495 	} else {
4496 		bip->bio_data = bp->b_data;
4497 		bip->bio_ma = NULL;
4498 	}
4499 }
4500 
4501 #include "opt_ddb.h"
4502 #ifdef DDB
4503 #include <ddb/ddb.h>
4504 
4505 /* DDB command to show buffer data */
4506 DB_SHOW_COMMAND(buffer, db_show_buffer)
4507 {
4508 	/* get args */
4509 	struct buf *bp = (struct buf *)addr;
4510 
4511 	if (!have_addr) {
4512 		db_printf("usage: show buffer <addr>\n");
4513 		return;
4514 	}
4515 
4516 	db_printf("buf at %p\n", bp);
4517 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4518 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4519 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4520 	db_printf(
4521 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4522 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4523 	    "b_dep = %p\n",
4524 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4525 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4526 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4527 	if (bp->b_npages) {
4528 		int i;
4529 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4530 		for (i = 0; i < bp->b_npages; i++) {
4531 			vm_page_t m;
4532 			m = bp->b_pages[i];
4533 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4534 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4535 			if ((i + 1) < bp->b_npages)
4536 				db_printf(",");
4537 		}
4538 		db_printf("\n");
4539 	}
4540 	db_printf(" ");
4541 	BUF_LOCKPRINTINFO(bp);
4542 }
4543 
4544 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4545 {
4546 	struct buf *bp;
4547 	int i;
4548 
4549 	for (i = 0; i < nbuf; i++) {
4550 		bp = &buf[i];
4551 		if (BUF_ISLOCKED(bp)) {
4552 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4553 			db_printf("\n");
4554 		}
4555 	}
4556 }
4557 
4558 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4559 {
4560 	struct vnode *vp;
4561 	struct buf *bp;
4562 
4563 	if (!have_addr) {
4564 		db_printf("usage: show vnodebufs <addr>\n");
4565 		return;
4566 	}
4567 	vp = (struct vnode *)addr;
4568 	db_printf("Clean buffers:\n");
4569 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4570 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4571 		db_printf("\n");
4572 	}
4573 	db_printf("Dirty buffers:\n");
4574 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4575 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4576 		db_printf("\n");
4577 	}
4578 }
4579 
4580 DB_COMMAND(countfreebufs, db_coundfreebufs)
4581 {
4582 	struct buf *bp;
4583 	int i, used = 0, nfree = 0;
4584 
4585 	if (have_addr) {
4586 		db_printf("usage: countfreebufs\n");
4587 		return;
4588 	}
4589 
4590 	for (i = 0; i < nbuf; i++) {
4591 		bp = &buf[i];
4592 		if ((bp->b_flags & B_INFREECNT) != 0)
4593 			nfree++;
4594 		else
4595 			used++;
4596 	}
4597 
4598 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4599 	    nfree + used);
4600 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4601 }
4602 #endif /* DDB */
4603