xref: /freebsd/sys/kern/vfs_bio.c (revision f0157ce528a128e2abb181a5c766033a2ce49a5f)
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_drain_busy_pages(struct buf *bp);
112 static void vfs_clean_pages_dirty_buf(struct buf *bp);
113 static void vfs_setdirty_locked_object(struct buf *bp);
114 static void vfs_vmio_release(struct buf *bp);
115 static int vfs_bio_clcheck(struct vnode *vp, int size,
116 		daddr_t lblkno, daddr_t blkno);
117 static int buf_flush(struct vnode *vp, int);
118 static int flushbufqueues(struct vnode *, int, int);
119 static void buf_daemon(void);
120 static void bremfreel(struct buf *bp);
121 static __inline void bd_wakeup(void);
122 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
123     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
124 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
125 #endif
126 
127 int vmiodirenable = TRUE;
128 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
129     "Use the VM system for directory writes");
130 long runningbufspace;
131 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
132     "Amount of presently outstanding async buffer io");
133 static long bufspace;
134 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
135     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
136 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
137     &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
138 #else
139 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
140     "Virtual memory used for buffers");
141 #endif
142 static long unmapped_bufspace;
143 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD,
144     &unmapped_bufspace, 0,
145     "Amount of unmapped buffers, inclusive in the bufspace");
146 static long maxbufspace;
147 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
148     "Maximum allowed value of bufspace (including buf_daemon)");
149 static long bufmallocspace;
150 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
151     "Amount of malloced memory for buffers");
152 static long maxbufmallocspace;
153 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
154     "Maximum amount of malloced memory for buffers");
155 static long lobufspace;
156 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
157     "Minimum amount of buffers we want to have");
158 long hibufspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
160     "Maximum allowed value of bufspace (excluding buf_daemon)");
161 static int bufreusecnt;
162 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
163     "Number of times we have reused a buffer");
164 static int buffreekvacnt;
165 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
166     "Number of times we have freed the KVA space from some buffer");
167 static int bufdefragcnt;
168 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
169     "Number of times we have had to repeat buffer allocation to defragment");
170 static long lorunningspace;
171 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
172     "Minimum preferred space used for in-progress I/O");
173 static long hirunningspace;
174 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
175     "Maximum amount of space to use for in-progress I/O");
176 int dirtybufferflushes;
177 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
178     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
179 int bdwriteskip;
180 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
181     0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
182 int altbufferflushes;
183 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
184     0, "Number of fsync flushes to limit dirty buffers");
185 static int recursiveflushes;
186 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
187     0, "Number of flushes skipped due to being recursive");
188 static int numdirtybuffers;
189 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
190     "Number of buffers that are dirty (has unwritten changes) at the moment");
191 static int lodirtybuffers;
192 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
193     "How many buffers we want to have free before bufdaemon can sleep");
194 static int hidirtybuffers;
195 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
196     "When the number of dirty buffers is considered severe");
197 int dirtybufthresh;
198 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
199     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
200 static int numfreebuffers;
201 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
202     "Number of free buffers");
203 static int lofreebuffers;
204 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
205    "XXX Unused");
206 static int hifreebuffers;
207 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
208    "XXX Complicatedly unused");
209 static int getnewbufcalls;
210 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
211    "Number of calls to getnewbuf");
212 static int getnewbufrestarts;
213 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
214     "Number of times getnewbuf has had to restart a buffer aquisition");
215 static int mappingrestarts;
216 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
217     "Number of times getblk has had to restart a buffer mapping for "
218     "unmapped buffer");
219 static int flushbufqtarget = 100;
220 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
221     "Amount of work to do in flushbufqueues when helping bufdaemon");
222 static long notbufdflushes;
223 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, &notbufdflushes, 0,
224     "Number of dirty buffer flushes done by the bufdaemon helpers");
225 static long barrierwrites;
226 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
227     "Number of barrier writes");
228 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
229     &unmapped_buf_allowed, 0,
230     "Permit the use of the unmapped i/o");
231 
232 /*
233  * Lock for the non-dirty bufqueues
234  */
235 static struct mtx_padalign bqclean;
236 
237 /*
238  * Lock for the dirty queue.
239  */
240 static struct mtx_padalign bqdirty;
241 
242 /*
243  * This lock synchronizes access to bd_request.
244  */
245 static struct mtx_padalign bdlock;
246 
247 /*
248  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
249  * waitrunningbufspace().
250  */
251 static struct mtx_padalign rbreqlock;
252 
253 /*
254  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
255  */
256 static struct mtx_padalign nblock;
257 
258 /*
259  * Lock that protects bdirtywait.
260  */
261 static struct mtx_padalign bdirtylock;
262 
263 /*
264  * Wakeup point for bufdaemon, as well as indicator of whether it is already
265  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
266  * is idling.
267  */
268 static int bd_request;
269 
270 /*
271  * Request for the buf daemon to write more buffers than is indicated by
272  * lodirtybuf.  This may be necessary to push out excess dependencies or
273  * defragment the address space where a simple count of the number of dirty
274  * buffers is insufficient to characterize the demand for flushing them.
275  */
276 static int bd_speedupreq;
277 
278 /*
279  * bogus page -- for I/O to/from partially complete buffers
280  * this is a temporary solution to the problem, but it is not
281  * really that bad.  it would be better to split the buffer
282  * for input in the case of buffers partially already in memory,
283  * but the code is intricate enough already.
284  */
285 vm_page_t bogus_page;
286 
287 /*
288  * Synchronization (sleep/wakeup) variable for active buffer space requests.
289  * Set when wait starts, cleared prior to wakeup().
290  * Used in runningbufwakeup() and waitrunningbufspace().
291  */
292 static int runningbufreq;
293 
294 /*
295  * Synchronization (sleep/wakeup) variable for buffer requests.
296  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
297  * by and/or.
298  * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
299  * getnewbuf(), and getblk().
300  */
301 static int needsbuffer;
302 
303 /*
304  * Synchronization for bwillwrite() waiters.
305  */
306 static int bdirtywait;
307 
308 /*
309  * Definitions for the buffer free lists.
310  */
311 #define BUFFER_QUEUES	5	/* number of free buffer queues */
312 
313 #define QUEUE_NONE	0	/* on no queue */
314 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
315 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
316 #define QUEUE_EMPTYKVA	3	/* empty buffer headers w/KVA assignment */
317 #define QUEUE_EMPTY	4	/* empty buffer headers */
318 #define QUEUE_SENTINEL	1024	/* not an queue index, but mark for sentinel */
319 
320 /* Queues for free buffers with various properties */
321 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
322 #ifdef INVARIANTS
323 static int bq_len[BUFFER_QUEUES];
324 #endif
325 
326 /*
327  * Single global constant for BUF_WMESG, to avoid getting multiple references.
328  * buf_wmesg is referred from macros.
329  */
330 const char *buf_wmesg = BUF_WMESG;
331 
332 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
333 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
334 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
335 
336 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
337     defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
338 static int
339 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
340 {
341 	long lvalue;
342 	int ivalue;
343 
344 	if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
345 		return (sysctl_handle_long(oidp, arg1, arg2, req));
346 	lvalue = *(long *)arg1;
347 	if (lvalue > INT_MAX)
348 		/* On overflow, still write out a long to trigger ENOMEM. */
349 		return (sysctl_handle_long(oidp, &lvalue, 0, req));
350 	ivalue = lvalue;
351 	return (sysctl_handle_int(oidp, &ivalue, 0, req));
352 }
353 #endif
354 
355 #ifdef DIRECTIO
356 extern void ffs_rawread_setup(void);
357 #endif /* DIRECTIO */
358 
359 /*
360  *	bqlock:
361  *
362  *	Return the appropriate queue lock based on the index.
363  */
364 static inline struct mtx *
365 bqlock(int qindex)
366 {
367 
368 	if (qindex == QUEUE_DIRTY)
369 		return (struct mtx *)(&bqdirty);
370 	return (struct mtx *)(&bqclean);
371 }
372 
373 /*
374  *	bdirtywakeup:
375  *
376  *	Wakeup any bwillwrite() waiters.
377  */
378 static void
379 bdirtywakeup(void)
380 {
381 	mtx_lock(&bdirtylock);
382 	if (bdirtywait) {
383 		bdirtywait = 0;
384 		wakeup(&bdirtywait);
385 	}
386 	mtx_unlock(&bdirtylock);
387 }
388 
389 /*
390  *	bdirtysub:
391  *
392  *	Decrement the numdirtybuffers count by one and wakeup any
393  *	threads blocked in bwillwrite().
394  */
395 static void
396 bdirtysub(void)
397 {
398 
399 	if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
400 	    (lodirtybuffers + hidirtybuffers) / 2)
401 		bdirtywakeup();
402 }
403 
404 /*
405  *	bdirtyadd:
406  *
407  *	Increment the numdirtybuffers count by one and wakeup the buf
408  *	daemon if needed.
409  */
410 static void
411 bdirtyadd(void)
412 {
413 
414 	/*
415 	 * Only do the wakeup once as we cross the boundary.  The
416 	 * buf daemon will keep running until the condition clears.
417 	 */
418 	if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
419 	    (lodirtybuffers + hidirtybuffers) / 2)
420 		bd_wakeup();
421 }
422 
423 /*
424  *	bufspacewakeup:
425  *
426  *	Called when buffer space is potentially available for recovery.
427  *	getnewbuf() will block on this flag when it is unable to free
428  *	sufficient buffer space.  Buffer space becomes recoverable when
429  *	bp's get placed back in the queues.
430  */
431 
432 static __inline void
433 bufspacewakeup(void)
434 {
435 
436 	/*
437 	 * If someone is waiting for BUF space, wake them up.  Even
438 	 * though we haven't freed the kva space yet, the waiting
439 	 * process will be able to now.
440 	 */
441 	mtx_lock(&nblock);
442 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
443 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
444 		wakeup(&needsbuffer);
445 	}
446 	mtx_unlock(&nblock);
447 }
448 
449 /*
450  *	runningwakeup:
451  *
452  *	Wake up processes that are waiting on asynchronous writes to fall
453  *	below lorunningspace.
454  */
455 static void
456 runningwakeup(void)
457 {
458 
459 	mtx_lock(&rbreqlock);
460 	if (runningbufreq) {
461 		runningbufreq = 0;
462 		wakeup(&runningbufreq);
463 	}
464 	mtx_unlock(&rbreqlock);
465 }
466 
467 /*
468  *	runningbufwakeup:
469  *
470  *	Decrement the outstanding write count according.
471  */
472 void
473 runningbufwakeup(struct buf *bp)
474 {
475 	long space, bspace;
476 
477 	if (bp->b_runningbufspace == 0)
478 		return;
479 	space = atomic_fetchadd_long(&runningbufspace, -bp->b_runningbufspace);
480 	bspace = bp->b_runningbufspace;
481 	bp->b_runningbufspace = 0;
482 	/*
483 	 * Only acquire the lock and wakeup on the transition from exceeding
484 	 * the threshold to falling below it.
485 	 */
486 	if (space < lorunningspace)
487 		return;
488 	if (space - bspace > lorunningspace)
489 		return;
490 	runningwakeup();
491 }
492 
493 /*
494  *	bufcountadd:
495  *
496  *	Called when a buffer has been added to one of the free queues to
497  *	account for the buffer and to wakeup anyone waiting for free buffers.
498  *	This typically occurs when large amounts of metadata are being handled
499  *	by the buffer cache ( else buffer space runs out first, usually ).
500  */
501 static __inline void
502 bufcountadd(struct buf *bp)
503 {
504 	int old;
505 
506 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
507 	    ("buf %p already counted as free", bp));
508 	bp->b_flags |= B_INFREECNT;
509 	old = atomic_fetchadd_int(&numfreebuffers, 1);
510 	KASSERT(old >= 0 && old < nbuf,
511 	    ("numfreebuffers climbed to %d", old + 1));
512 	mtx_lock(&nblock);
513 	if (needsbuffer) {
514 		needsbuffer &= ~VFS_BIO_NEED_ANY;
515 		if (numfreebuffers >= hifreebuffers)
516 			needsbuffer &= ~VFS_BIO_NEED_FREE;
517 		wakeup(&needsbuffer);
518 	}
519 	mtx_unlock(&nblock);
520 }
521 
522 /*
523  *	bufcountsub:
524  *
525  *	Decrement the numfreebuffers count as needed.
526  */
527 static void
528 bufcountsub(struct buf *bp)
529 {
530 	int old;
531 
532 	/*
533 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
534 	 * delayed-write, the buffer was free and we must decrement
535 	 * numfreebuffers.
536 	 */
537 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
538 		KASSERT((bp->b_flags & B_INFREECNT) != 0,
539 		    ("buf %p not counted in numfreebuffers", bp));
540 		bp->b_flags &= ~B_INFREECNT;
541 		old = atomic_fetchadd_int(&numfreebuffers, -1);
542 		KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
543 	}
544 }
545 
546 /*
547  *	waitrunningbufspace()
548  *
549  *	runningbufspace is a measure of the amount of I/O currently
550  *	running.  This routine is used in async-write situations to
551  *	prevent creating huge backups of pending writes to a device.
552  *	Only asynchronous writes are governed by this function.
553  *
554  *	This does NOT turn an async write into a sync write.  It waits
555  *	for earlier writes to complete and generally returns before the
556  *	caller's write has reached the device.
557  */
558 void
559 waitrunningbufspace(void)
560 {
561 
562 	mtx_lock(&rbreqlock);
563 	while (runningbufspace > hirunningspace) {
564 		++runningbufreq;
565 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
566 	}
567 	mtx_unlock(&rbreqlock);
568 }
569 
570 
571 /*
572  *	vfs_buf_test_cache:
573  *
574  *	Called when a buffer is extended.  This function clears the B_CACHE
575  *	bit if the newly extended portion of the buffer does not contain
576  *	valid data.
577  */
578 static __inline
579 void
580 vfs_buf_test_cache(struct buf *bp,
581 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
582 		  vm_page_t m)
583 {
584 
585 	VM_OBJECT_ASSERT_WLOCKED(m->object);
586 	if (bp->b_flags & B_CACHE) {
587 		int base = (foff + off) & PAGE_MASK;
588 		if (vm_page_is_valid(m, base, size) == 0)
589 			bp->b_flags &= ~B_CACHE;
590 	}
591 }
592 
593 /* Wake up the buffer daemon if necessary */
594 static __inline void
595 bd_wakeup(void)
596 {
597 
598 	mtx_lock(&bdlock);
599 	if (bd_request == 0) {
600 		bd_request = 1;
601 		wakeup(&bd_request);
602 	}
603 	mtx_unlock(&bdlock);
604 }
605 
606 /*
607  * bd_speedup - speedup the buffer cache flushing code
608  */
609 void
610 bd_speedup(void)
611 {
612 	int needwake;
613 
614 	mtx_lock(&bdlock);
615 	needwake = 0;
616 	if (bd_speedupreq == 0 || bd_request == 0)
617 		needwake = 1;
618 	bd_speedupreq = 1;
619 	bd_request = 1;
620 	if (needwake)
621 		wakeup(&bd_request);
622 	mtx_unlock(&bdlock);
623 }
624 
625 #ifdef __i386__
626 #define	TRANSIENT_DENOM	5
627 #else
628 #define	TRANSIENT_DENOM 10
629 #endif
630 
631 /*
632  * Calculating buffer cache scaling values and reserve space for buffer
633  * headers.  This is called during low level kernel initialization and
634  * may be called more then once.  We CANNOT write to the memory area
635  * being reserved at this time.
636  */
637 caddr_t
638 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
639 {
640 	int tuned_nbuf;
641 	long maxbuf, maxbuf_sz, buf_sz,	biotmap_sz;
642 
643 	/*
644 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
645 	 * PAGE_SIZE is >= 1K)
646 	 */
647 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
648 
649 	/*
650 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
651 	 * For the first 64MB of ram nominally allocate sufficient buffers to
652 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
653 	 * buffers to cover 1/10 of our ram over 64MB.  When auto-sizing
654 	 * the buffer cache we limit the eventual kva reservation to
655 	 * maxbcache bytes.
656 	 *
657 	 * factor represents the 1/4 x ram conversion.
658 	 */
659 	if (nbuf == 0) {
660 		int factor = 4 * BKVASIZE / 1024;
661 
662 		nbuf = 50;
663 		if (physmem_est > 4096)
664 			nbuf += min((physmem_est - 4096) / factor,
665 			    65536 / factor);
666 		if (physmem_est > 65536)
667 			nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
668 			    32 * 1024 * 1024 / (factor * 5));
669 
670 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
671 			nbuf = maxbcache / BKVASIZE;
672 		tuned_nbuf = 1;
673 	} else
674 		tuned_nbuf = 0;
675 
676 	/* XXX Avoid unsigned long overflows later on with maxbufspace. */
677 	maxbuf = (LONG_MAX / 3) / BKVASIZE;
678 	if (nbuf > maxbuf) {
679 		if (!tuned_nbuf)
680 			printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
681 			    maxbuf);
682 		nbuf = maxbuf;
683 	}
684 
685 	/*
686 	 * Ideal allocation size for the transient bio submap if 10%
687 	 * of the maximal space buffer map.  This roughly corresponds
688 	 * to the amount of the buffer mapped for typical UFS load.
689 	 *
690 	 * Clip the buffer map to reserve space for the transient
691 	 * BIOs, if its extent is bigger than 90% (80% on i386) of the
692 	 * maximum buffer map extent on the platform.
693 	 *
694 	 * The fall-back to the maxbuf in case of maxbcache unset,
695 	 * allows to not trim the buffer KVA for the architectures
696 	 * with ample KVA space.
697 	 */
698 	if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
699 		maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
700 		buf_sz = (long)nbuf * BKVASIZE;
701 		if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
702 		    (TRANSIENT_DENOM - 1)) {
703 			/*
704 			 * There is more KVA than memory.  Do not
705 			 * adjust buffer map size, and assign the rest
706 			 * of maxbuf to transient map.
707 			 */
708 			biotmap_sz = maxbuf_sz - buf_sz;
709 		} else {
710 			/*
711 			 * Buffer map spans all KVA we could afford on
712 			 * this platform.  Give 10% (20% on i386) of
713 			 * the buffer map to the transient bio map.
714 			 */
715 			biotmap_sz = buf_sz / TRANSIENT_DENOM;
716 			buf_sz -= biotmap_sz;
717 		}
718 		if (biotmap_sz / INT_MAX > MAXPHYS)
719 			bio_transient_maxcnt = INT_MAX;
720 		else
721 			bio_transient_maxcnt = biotmap_sz / MAXPHYS;
722 		/*
723 		 * Artifically limit to 1024 simultaneous in-flight I/Os
724 		 * using the transient mapping.
725 		 */
726 		if (bio_transient_maxcnt > 1024)
727 			bio_transient_maxcnt = 1024;
728 		if (tuned_nbuf)
729 			nbuf = buf_sz / BKVASIZE;
730 	}
731 
732 	/*
733 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
734 	 * We have no less then 16 and no more then 256.
735 	 */
736 	nswbuf = max(min(nbuf/4, 256), 16);
737 #ifdef NSWBUF_MIN
738 	if (nswbuf < NSWBUF_MIN)
739 		nswbuf = NSWBUF_MIN;
740 #endif
741 #ifdef DIRECTIO
742 	ffs_rawread_setup();
743 #endif
744 
745 	/*
746 	 * Reserve space for the buffer cache buffers
747 	 */
748 	swbuf = (void *)v;
749 	v = (caddr_t)(swbuf + nswbuf);
750 	buf = (void *)v;
751 	v = (caddr_t)(buf + nbuf);
752 
753 	return(v);
754 }
755 
756 /* Initialize the buffer subsystem.  Called before use of any buffers. */
757 void
758 bufinit(void)
759 {
760 	struct buf *bp;
761 	int i;
762 
763 	mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
764 	mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
765 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
766 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
767 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
768 	mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
769 
770 	/* next, make a null set of free lists */
771 	for (i = 0; i < BUFFER_QUEUES; i++)
772 		TAILQ_INIT(&bufqueues[i]);
773 
774 	/* finally, initialize each buffer header and stick on empty q */
775 	for (i = 0; i < nbuf; i++) {
776 		bp = &buf[i];
777 		bzero(bp, sizeof *bp);
778 		bp->b_flags = B_INVAL | B_INFREECNT;
779 		bp->b_rcred = NOCRED;
780 		bp->b_wcred = NOCRED;
781 		bp->b_qindex = QUEUE_EMPTY;
782 		bp->b_xflags = 0;
783 		LIST_INIT(&bp->b_dep);
784 		BUF_LOCKINIT(bp);
785 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
786 #ifdef INVARIANTS
787 		bq_len[QUEUE_EMPTY]++;
788 #endif
789 	}
790 
791 	/*
792 	 * maxbufspace is the absolute maximum amount of buffer space we are
793 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
794 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
795 	 * used by most other processes.  The differential is required to
796 	 * ensure that buf_daemon is able to run when other processes might
797 	 * be blocked waiting for buffer space.
798 	 *
799 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
800 	 * this may result in KVM fragmentation which is not handled optimally
801 	 * by the system.
802 	 */
803 	maxbufspace = (long)nbuf * BKVASIZE;
804 	hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
805 	lobufspace = hibufspace - MAXBSIZE;
806 
807 	/*
808 	 * Note: The 16 MiB upper limit for hirunningspace was chosen
809 	 * arbitrarily and may need further tuning. It corresponds to
810 	 * 128 outstanding write IO requests (if IO size is 128 KiB),
811 	 * which fits with many RAID controllers' tagged queuing limits.
812 	 * The lower 1 MiB limit is the historical upper limit for
813 	 * hirunningspace.
814 	 */
815 	hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
816 	    16 * 1024 * 1024), 1024 * 1024);
817 	lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
818 
819 /*
820  * Limit the amount of malloc memory since it is wired permanently into
821  * the kernel space.  Even though this is accounted for in the buffer
822  * allocation, we don't want the malloced region to grow uncontrolled.
823  * The malloc scheme improves memory utilization significantly on average
824  * (small) directories.
825  */
826 	maxbufmallocspace = hibufspace / 20;
827 
828 /*
829  * Reduce the chance of a deadlock occuring by limiting the number
830  * of delayed-write dirty buffers we allow to stack up.
831  */
832 	hidirtybuffers = nbuf / 4 + 20;
833 	dirtybufthresh = hidirtybuffers * 9 / 10;
834 	numdirtybuffers = 0;
835 /*
836  * To support extreme low-memory systems, make sure hidirtybuffers cannot
837  * eat up all available buffer space.  This occurs when our minimum cannot
838  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
839  * BKVASIZE'd buffers.
840  */
841 	while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
842 		hidirtybuffers >>= 1;
843 	}
844 	lodirtybuffers = hidirtybuffers / 2;
845 
846 /*
847  * Try to keep the number of free buffers in the specified range,
848  * and give special processes (e.g. like buf_daemon) access to an
849  * emergency reserve.
850  */
851 	lofreebuffers = nbuf / 18 + 5;
852 	hifreebuffers = 2 * lofreebuffers;
853 	numfreebuffers = nbuf;
854 
855 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
856 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
857 	unmapped_buf = (caddr_t)kmem_alloc_nofault(kernel_map, MAXPHYS);
858 }
859 
860 #ifdef INVARIANTS
861 static inline void
862 vfs_buf_check_mapped(struct buf *bp)
863 {
864 
865 	KASSERT((bp->b_flags & B_UNMAPPED) == 0,
866 	    ("mapped buf %p %x", bp, bp->b_flags));
867 	KASSERT(bp->b_kvabase != unmapped_buf,
868 	    ("mapped buf: b_kvabase was not updated %p", bp));
869 	KASSERT(bp->b_data != unmapped_buf,
870 	    ("mapped buf: b_data was not updated %p", bp));
871 }
872 
873 static inline void
874 vfs_buf_check_unmapped(struct buf *bp)
875 {
876 
877 	KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
878 	    ("unmapped buf %p %x", bp, bp->b_flags));
879 	KASSERT(bp->b_kvabase == unmapped_buf,
880 	    ("unmapped buf: corrupted b_kvabase %p", bp));
881 	KASSERT(bp->b_data == unmapped_buf,
882 	    ("unmapped buf: corrupted b_data %p", bp));
883 }
884 
885 #define	BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
886 #define	BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
887 #else
888 #define	BUF_CHECK_MAPPED(bp) do {} while (0)
889 #define	BUF_CHECK_UNMAPPED(bp) do {} while (0)
890 #endif
891 
892 static void
893 bpmap_qenter(struct buf *bp)
894 {
895 
896 	BUF_CHECK_MAPPED(bp);
897 
898 	/*
899 	 * bp->b_data is relative to bp->b_offset, but
900 	 * bp->b_offset may be offset into the first page.
901 	 */
902 	bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
903 	pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
904 	bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
905 	    (vm_offset_t)(bp->b_offset & PAGE_MASK));
906 }
907 
908 /*
909  * bfreekva() - free the kva allocation for a buffer.
910  *
911  *	Since this call frees up buffer space, we call bufspacewakeup().
912  */
913 static void
914 bfreekva(struct buf *bp)
915 {
916 
917 	if (bp->b_kvasize == 0)
918 		return;
919 
920 	atomic_add_int(&buffreekvacnt, 1);
921 	atomic_subtract_long(&bufspace, bp->b_kvasize);
922 	if ((bp->b_flags & B_UNMAPPED) == 0) {
923 		BUF_CHECK_MAPPED(bp);
924 		vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
925 		    bp->b_kvasize);
926 	} else {
927 		BUF_CHECK_UNMAPPED(bp);
928 		if ((bp->b_flags & B_KVAALLOC) != 0) {
929 			vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
930 			    bp->b_kvasize);
931 		}
932 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
933 		bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
934 	}
935 	bp->b_kvasize = 0;
936 	bufspacewakeup();
937 }
938 
939 /*
940  *	binsfree:
941  *
942  *	Insert the buffer into the appropriate free list.
943  */
944 static void
945 binsfree(struct buf *bp, int qindex)
946 {
947 	struct mtx *olock, *nlock;
948 
949 	BUF_ASSERT_XLOCKED(bp);
950 
951 	olock = bqlock(bp->b_qindex);
952 	nlock = bqlock(qindex);
953 	mtx_lock(olock);
954 	/* Handle delayed bremfree() processing. */
955 	if (bp->b_flags & B_REMFREE)
956 		bremfreel(bp);
957 
958 	if (bp->b_qindex != QUEUE_NONE)
959 		panic("binsfree: free buffer onto another queue???");
960 
961 	bp->b_qindex = qindex;
962 	if (olock != nlock) {
963 		mtx_unlock(olock);
964 		mtx_lock(nlock);
965 	}
966 	if (bp->b_flags & B_AGE)
967 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
968 	else
969 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
970 #ifdef INVARIANTS
971 	bq_len[bp->b_qindex]++;
972 #endif
973 	mtx_unlock(nlock);
974 
975 	/*
976 	 * Something we can maybe free or reuse.
977 	 */
978 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
979 		bufspacewakeup();
980 
981 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
982 		bufcountadd(bp);
983 }
984 
985 /*
986  *	bremfree:
987  *
988  *	Mark the buffer for removal from the appropriate free list.
989  *
990  */
991 void
992 bremfree(struct buf *bp)
993 {
994 
995 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
996 	KASSERT((bp->b_flags & B_REMFREE) == 0,
997 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
998 	KASSERT(bp->b_qindex != QUEUE_NONE,
999 	    ("bremfree: buffer %p not on a queue.", bp));
1000 	BUF_ASSERT_XLOCKED(bp);
1001 
1002 	bp->b_flags |= B_REMFREE;
1003 	bufcountsub(bp);
1004 }
1005 
1006 /*
1007  *	bremfreef:
1008  *
1009  *	Force an immediate removal from a free list.  Used only in nfs when
1010  *	it abuses the b_freelist pointer.
1011  */
1012 void
1013 bremfreef(struct buf *bp)
1014 {
1015 	struct mtx *qlock;
1016 
1017 	qlock = bqlock(bp->b_qindex);
1018 	mtx_lock(qlock);
1019 	bremfreel(bp);
1020 	mtx_unlock(qlock);
1021 }
1022 
1023 /*
1024  *	bremfreel:
1025  *
1026  *	Removes a buffer from the free list, must be called with the
1027  *	correct qlock held.
1028  */
1029 static void
1030 bremfreel(struct buf *bp)
1031 {
1032 
1033 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1034 	    bp, bp->b_vp, bp->b_flags);
1035 	KASSERT(bp->b_qindex != QUEUE_NONE,
1036 	    ("bremfreel: buffer %p not on a queue.", bp));
1037 	BUF_ASSERT_XLOCKED(bp);
1038 	mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1039 
1040 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1041 #ifdef INVARIANTS
1042 	KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1043 	    bp->b_qindex));
1044 	bq_len[bp->b_qindex]--;
1045 #endif
1046 	bp->b_qindex = QUEUE_NONE;
1047 	/*
1048 	 * If this was a delayed bremfree() we only need to remove the buffer
1049 	 * from the queue and return the stats are already done.
1050 	 */
1051 	if (bp->b_flags & B_REMFREE) {
1052 		bp->b_flags &= ~B_REMFREE;
1053 		return;
1054 	}
1055 	bufcountsub(bp);
1056 }
1057 
1058 /*
1059  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
1060  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1061  * the buffer is valid and we do not have to do anything.
1062  */
1063 void
1064 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1065     int cnt, struct ucred * cred)
1066 {
1067 	struct buf *rabp;
1068 	int i;
1069 
1070 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1071 		if (inmem(vp, *rablkno))
1072 			continue;
1073 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1074 
1075 		if ((rabp->b_flags & B_CACHE) == 0) {
1076 			if (!TD_IS_IDLETHREAD(curthread))
1077 				curthread->td_ru.ru_inblock++;
1078 			rabp->b_flags |= B_ASYNC;
1079 			rabp->b_flags &= ~B_INVAL;
1080 			rabp->b_ioflags &= ~BIO_ERROR;
1081 			rabp->b_iocmd = BIO_READ;
1082 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
1083 				rabp->b_rcred = crhold(cred);
1084 			vfs_busy_pages(rabp, 0);
1085 			BUF_KERNPROC(rabp);
1086 			rabp->b_iooffset = dbtob(rabp->b_blkno);
1087 			bstrategy(rabp);
1088 		} else {
1089 			brelse(rabp);
1090 		}
1091 	}
1092 }
1093 
1094 /*
1095  * Entry point for bread() and breadn() via #defines in sys/buf.h.
1096  *
1097  * Get a buffer with the specified data.  Look in the cache first.  We
1098  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
1099  * is set, the buffer is valid and we do not have to do anything, see
1100  * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1101  */
1102 int
1103 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1104     int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1105 {
1106 	struct buf *bp;
1107 	int rv = 0, readwait = 0;
1108 
1109 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1110 	/*
1111 	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1112 	 */
1113 	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1114 	if (bp == NULL)
1115 		return (EBUSY);
1116 
1117 	/* if not found in cache, do some I/O */
1118 	if ((bp->b_flags & B_CACHE) == 0) {
1119 		if (!TD_IS_IDLETHREAD(curthread))
1120 			curthread->td_ru.ru_inblock++;
1121 		bp->b_iocmd = BIO_READ;
1122 		bp->b_flags &= ~B_INVAL;
1123 		bp->b_ioflags &= ~BIO_ERROR;
1124 		if (bp->b_rcred == NOCRED && cred != NOCRED)
1125 			bp->b_rcred = crhold(cred);
1126 		vfs_busy_pages(bp, 0);
1127 		bp->b_iooffset = dbtob(bp->b_blkno);
1128 		bstrategy(bp);
1129 		++readwait;
1130 	}
1131 
1132 	breada(vp, rablkno, rabsize, cnt, cred);
1133 
1134 	if (readwait) {
1135 		rv = bufwait(bp);
1136 	}
1137 	return (rv);
1138 }
1139 
1140 /*
1141  * Write, release buffer on completion.  (Done by iodone
1142  * if async).  Do not bother writing anything if the buffer
1143  * is invalid.
1144  *
1145  * Note that we set B_CACHE here, indicating that buffer is
1146  * fully valid and thus cacheable.  This is true even of NFS
1147  * now so we set it generally.  This could be set either here
1148  * or in biodone() since the I/O is synchronous.  We put it
1149  * here.
1150  */
1151 int
1152 bufwrite(struct buf *bp)
1153 {
1154 	int oldflags;
1155 	struct vnode *vp;
1156 	long space;
1157 	int vp_md;
1158 
1159 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1160 	if (bp->b_flags & B_INVAL) {
1161 		brelse(bp);
1162 		return (0);
1163 	}
1164 
1165 	if (bp->b_flags & B_BARRIER)
1166 		barrierwrites++;
1167 
1168 	oldflags = bp->b_flags;
1169 
1170 	BUF_ASSERT_HELD(bp);
1171 
1172 	if (bp->b_pin_count > 0)
1173 		bunpin_wait(bp);
1174 
1175 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1176 	    ("FFS background buffer should not get here %p", bp));
1177 
1178 	vp = bp->b_vp;
1179 	if (vp)
1180 		vp_md = vp->v_vflag & VV_MD;
1181 	else
1182 		vp_md = 0;
1183 
1184 	/*
1185 	 * Mark the buffer clean.  Increment the bufobj write count
1186 	 * before bundirty() call, to prevent other thread from seeing
1187 	 * empty dirty list and zero counter for writes in progress,
1188 	 * falsely indicating that the bufobj is clean.
1189 	 */
1190 	bufobj_wref(bp->b_bufobj);
1191 	bundirty(bp);
1192 
1193 	bp->b_flags &= ~B_DONE;
1194 	bp->b_ioflags &= ~BIO_ERROR;
1195 	bp->b_flags |= B_CACHE;
1196 	bp->b_iocmd = BIO_WRITE;
1197 
1198 	vfs_busy_pages(bp, 1);
1199 
1200 	/*
1201 	 * Normal bwrites pipeline writes
1202 	 */
1203 	bp->b_runningbufspace = bp->b_bufsize;
1204 	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1205 
1206 	if (!TD_IS_IDLETHREAD(curthread))
1207 		curthread->td_ru.ru_oublock++;
1208 	if (oldflags & B_ASYNC)
1209 		BUF_KERNPROC(bp);
1210 	bp->b_iooffset = dbtob(bp->b_blkno);
1211 	bstrategy(bp);
1212 
1213 	if ((oldflags & B_ASYNC) == 0) {
1214 		int rtval = bufwait(bp);
1215 		brelse(bp);
1216 		return (rtval);
1217 	} else if (space > hirunningspace) {
1218 		/*
1219 		 * don't allow the async write to saturate the I/O
1220 		 * system.  We will not deadlock here because
1221 		 * we are blocking waiting for I/O that is already in-progress
1222 		 * to complete. We do not block here if it is the update
1223 		 * or syncer daemon trying to clean up as that can lead
1224 		 * to deadlock.
1225 		 */
1226 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1227 			waitrunningbufspace();
1228 	}
1229 
1230 	return (0);
1231 }
1232 
1233 void
1234 bufbdflush(struct bufobj *bo, struct buf *bp)
1235 {
1236 	struct buf *nbp;
1237 
1238 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1239 		(void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1240 		altbufferflushes++;
1241 	} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1242 		BO_LOCK(bo);
1243 		/*
1244 		 * Try to find a buffer to flush.
1245 		 */
1246 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1247 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1248 			    BUF_LOCK(nbp,
1249 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
1250 				continue;
1251 			if (bp == nbp)
1252 				panic("bdwrite: found ourselves");
1253 			BO_UNLOCK(bo);
1254 			/* Don't countdeps with the bo lock held. */
1255 			if (buf_countdeps(nbp, 0)) {
1256 				BO_LOCK(bo);
1257 				BUF_UNLOCK(nbp);
1258 				continue;
1259 			}
1260 			if (nbp->b_flags & B_CLUSTEROK) {
1261 				vfs_bio_awrite(nbp);
1262 			} else {
1263 				bremfree(nbp);
1264 				bawrite(nbp);
1265 			}
1266 			dirtybufferflushes++;
1267 			break;
1268 		}
1269 		if (nbp == NULL)
1270 			BO_UNLOCK(bo);
1271 	}
1272 }
1273 
1274 /*
1275  * Delayed write. (Buffer is marked dirty).  Do not bother writing
1276  * anything if the buffer is marked invalid.
1277  *
1278  * Note that since the buffer must be completely valid, we can safely
1279  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
1280  * biodone() in order to prevent getblk from writing the buffer
1281  * out synchronously.
1282  */
1283 void
1284 bdwrite(struct buf *bp)
1285 {
1286 	struct thread *td = curthread;
1287 	struct vnode *vp;
1288 	struct bufobj *bo;
1289 
1290 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1291 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1292 	KASSERT((bp->b_flags & B_BARRIER) == 0,
1293 	    ("Barrier request in delayed write %p", bp));
1294 	BUF_ASSERT_HELD(bp);
1295 
1296 	if (bp->b_flags & B_INVAL) {
1297 		brelse(bp);
1298 		return;
1299 	}
1300 
1301 	/*
1302 	 * If we have too many dirty buffers, don't create any more.
1303 	 * If we are wildly over our limit, then force a complete
1304 	 * cleanup. Otherwise, just keep the situation from getting
1305 	 * out of control. Note that we have to avoid a recursive
1306 	 * disaster and not try to clean up after our own cleanup!
1307 	 */
1308 	vp = bp->b_vp;
1309 	bo = bp->b_bufobj;
1310 	if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1311 		td->td_pflags |= TDP_INBDFLUSH;
1312 		BO_BDFLUSH(bo, bp);
1313 		td->td_pflags &= ~TDP_INBDFLUSH;
1314 	} else
1315 		recursiveflushes++;
1316 
1317 	bdirty(bp);
1318 	/*
1319 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
1320 	 * true even of NFS now.
1321 	 */
1322 	bp->b_flags |= B_CACHE;
1323 
1324 	/*
1325 	 * This bmap keeps the system from needing to do the bmap later,
1326 	 * perhaps when the system is attempting to do a sync.  Since it
1327 	 * is likely that the indirect block -- or whatever other datastructure
1328 	 * that the filesystem needs is still in memory now, it is a good
1329 	 * thing to do this.  Note also, that if the pageout daemon is
1330 	 * requesting a sync -- there might not be enough memory to do
1331 	 * the bmap then...  So, this is important to do.
1332 	 */
1333 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1334 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1335 	}
1336 
1337 	/*
1338 	 * Set the *dirty* buffer range based upon the VM system dirty
1339 	 * pages.
1340 	 *
1341 	 * Mark the buffer pages as clean.  We need to do this here to
1342 	 * satisfy the vnode_pager and the pageout daemon, so that it
1343 	 * thinks that the pages have been "cleaned".  Note that since
1344 	 * the pages are in a delayed write buffer -- the VFS layer
1345 	 * "will" see that the pages get written out on the next sync,
1346 	 * or perhaps the cluster will be completed.
1347 	 */
1348 	vfs_clean_pages_dirty_buf(bp);
1349 	bqrelse(bp);
1350 
1351 	/*
1352 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1353 	 * due to the softdep code.
1354 	 */
1355 }
1356 
1357 /*
1358  *	bdirty:
1359  *
1360  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1361  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1362  *	itself to properly update it in the dirty/clean lists.  We mark it
1363  *	B_DONE to ensure that any asynchronization of the buffer properly
1364  *	clears B_DONE ( else a panic will occur later ).
1365  *
1366  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1367  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1368  *	should only be called if the buffer is known-good.
1369  *
1370  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1371  *	count.
1372  *
1373  *	The buffer must be on QUEUE_NONE.
1374  */
1375 void
1376 bdirty(struct buf *bp)
1377 {
1378 
1379 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1380 	    bp, bp->b_vp, bp->b_flags);
1381 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1382 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1383 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1384 	BUF_ASSERT_HELD(bp);
1385 	bp->b_flags &= ~(B_RELBUF);
1386 	bp->b_iocmd = BIO_WRITE;
1387 
1388 	if ((bp->b_flags & B_DELWRI) == 0) {
1389 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1390 		reassignbuf(bp);
1391 		bdirtyadd();
1392 	}
1393 }
1394 
1395 /*
1396  *	bundirty:
1397  *
1398  *	Clear B_DELWRI for buffer.
1399  *
1400  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1401  *	count.
1402  *
1403  *	The buffer must be on QUEUE_NONE.
1404  */
1405 
1406 void
1407 bundirty(struct buf *bp)
1408 {
1409 
1410 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1411 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1412 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1413 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1414 	BUF_ASSERT_HELD(bp);
1415 
1416 	if (bp->b_flags & B_DELWRI) {
1417 		bp->b_flags &= ~B_DELWRI;
1418 		reassignbuf(bp);
1419 		bdirtysub();
1420 	}
1421 	/*
1422 	 * Since it is now being written, we can clear its deferred write flag.
1423 	 */
1424 	bp->b_flags &= ~B_DEFERRED;
1425 }
1426 
1427 /*
1428  *	bawrite:
1429  *
1430  *	Asynchronous write.  Start output on a buffer, but do not wait for
1431  *	it to complete.  The buffer is released when the output completes.
1432  *
1433  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1434  *	B_INVAL buffers.  Not us.
1435  */
1436 void
1437 bawrite(struct buf *bp)
1438 {
1439 
1440 	bp->b_flags |= B_ASYNC;
1441 	(void) bwrite(bp);
1442 }
1443 
1444 /*
1445  *	babarrierwrite:
1446  *
1447  *	Asynchronous barrier write.  Start output on a buffer, but do not
1448  *	wait for it to complete.  Place a write barrier after this write so
1449  *	that this buffer and all buffers written before it are committed to
1450  *	the disk before any buffers written after this write are committed
1451  *	to the disk.  The buffer is released when the output completes.
1452  */
1453 void
1454 babarrierwrite(struct buf *bp)
1455 {
1456 
1457 	bp->b_flags |= B_ASYNC | B_BARRIER;
1458 	(void) bwrite(bp);
1459 }
1460 
1461 /*
1462  *	bbarrierwrite:
1463  *
1464  *	Synchronous barrier write.  Start output on a buffer and wait for
1465  *	it to complete.  Place a write barrier after this write so that
1466  *	this buffer and all buffers written before it are committed to
1467  *	the disk before any buffers written after this write are committed
1468  *	to the disk.  The buffer is released when the output completes.
1469  */
1470 int
1471 bbarrierwrite(struct buf *bp)
1472 {
1473 
1474 	bp->b_flags |= B_BARRIER;
1475 	return (bwrite(bp));
1476 }
1477 
1478 /*
1479  *	bwillwrite:
1480  *
1481  *	Called prior to the locking of any vnodes when we are expecting to
1482  *	write.  We do not want to starve the buffer cache with too many
1483  *	dirty buffers so we block here.  By blocking prior to the locking
1484  *	of any vnodes we attempt to avoid the situation where a locked vnode
1485  *	prevents the various system daemons from flushing related buffers.
1486  */
1487 void
1488 bwillwrite(void)
1489 {
1490 
1491 	if (numdirtybuffers >= hidirtybuffers) {
1492 		mtx_lock(&bdirtylock);
1493 		while (numdirtybuffers >= hidirtybuffers) {
1494 			bdirtywait = 1;
1495 			msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1496 			    "flswai", 0);
1497 		}
1498 		mtx_unlock(&bdirtylock);
1499 	}
1500 }
1501 
1502 /*
1503  * Return true if we have too many dirty buffers.
1504  */
1505 int
1506 buf_dirty_count_severe(void)
1507 {
1508 
1509 	return(numdirtybuffers >= hidirtybuffers);
1510 }
1511 
1512 static __noinline int
1513 buf_vm_page_count_severe(void)
1514 {
1515 
1516 	KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1517 
1518 	return vm_page_count_severe();
1519 }
1520 
1521 /*
1522  *	brelse:
1523  *
1524  *	Release a busy buffer and, if requested, free its resources.  The
1525  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1526  *	to be accessed later as a cache entity or reused for other purposes.
1527  */
1528 void
1529 brelse(struct buf *bp)
1530 {
1531 	int qindex;
1532 
1533 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1534 	    bp, bp->b_vp, bp->b_flags);
1535 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1536 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1537 
1538 	if (BUF_LOCKRECURSED(bp)) {
1539 		/*
1540 		 * Do not process, in particular, do not handle the
1541 		 * B_INVAL/B_RELBUF and do not release to free list.
1542 		 */
1543 		BUF_UNLOCK(bp);
1544 		return;
1545 	}
1546 
1547 	if (bp->b_flags & B_MANAGED) {
1548 		bqrelse(bp);
1549 		return;
1550 	}
1551 
1552 	if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1553 	    bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1554 		/*
1555 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1556 		 * pages from being scrapped.  If the error is anything
1557 		 * other than an I/O error (EIO), assume that retrying
1558 		 * is futile.
1559 		 */
1560 		bp->b_ioflags &= ~BIO_ERROR;
1561 		bdirty(bp);
1562 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1563 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1564 		/*
1565 		 * Either a failed I/O or we were asked to free or not
1566 		 * cache the buffer.
1567 		 */
1568 		bp->b_flags |= B_INVAL;
1569 		if (!LIST_EMPTY(&bp->b_dep))
1570 			buf_deallocate(bp);
1571 		if (bp->b_flags & B_DELWRI)
1572 			bdirtysub();
1573 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1574 		if ((bp->b_flags & B_VMIO) == 0) {
1575 			if (bp->b_bufsize)
1576 				allocbuf(bp, 0);
1577 			if (bp->b_vp)
1578 				brelvp(bp);
1579 		}
1580 	}
1581 
1582 	/*
1583 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1584 	 * is called with B_DELWRI set, the underlying pages may wind up
1585 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1586 	 * because pages associated with a B_DELWRI bp are marked clean.
1587 	 *
1588 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1589 	 * if B_DELWRI is set.
1590 	 *
1591 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1592 	 * on pages to return pages to the VM page queues.
1593 	 */
1594 	if (bp->b_flags & B_DELWRI)
1595 		bp->b_flags &= ~B_RELBUF;
1596 	else if (buf_vm_page_count_severe()) {
1597 		/*
1598 		 * BKGRDINPROG can only be set with the buf and bufobj
1599 		 * locks both held.  We tolerate a race to clear it here.
1600 		 */
1601 		if (!(bp->b_vflags & BV_BKGRDINPROG))
1602 			bp->b_flags |= B_RELBUF;
1603 	}
1604 
1605 	/*
1606 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1607 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1608 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1609 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1610 	 *
1611 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1612 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1613 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1614 	 *
1615 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1616 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1617 	 * the commit state and we cannot afford to lose the buffer. If the
1618 	 * buffer has a background write in progress, we need to keep it
1619 	 * around to prevent it from being reconstituted and starting a second
1620 	 * background write.
1621 	 */
1622 	if ((bp->b_flags & B_VMIO)
1623 	    && !(bp->b_vp->v_mount != NULL &&
1624 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1625 		 !vn_isdisk(bp->b_vp, NULL) &&
1626 		 (bp->b_flags & B_DELWRI))
1627 	    ) {
1628 
1629 		int i, j, resid;
1630 		vm_page_t m;
1631 		off_t foff;
1632 		vm_pindex_t poff;
1633 		vm_object_t obj;
1634 
1635 		obj = bp->b_bufobj->bo_object;
1636 
1637 		/*
1638 		 * Get the base offset and length of the buffer.  Note that
1639 		 * in the VMIO case if the buffer block size is not
1640 		 * page-aligned then b_data pointer may not be page-aligned.
1641 		 * But our b_pages[] array *IS* page aligned.
1642 		 *
1643 		 * block sizes less then DEV_BSIZE (usually 512) are not
1644 		 * supported due to the page granularity bits (m->valid,
1645 		 * m->dirty, etc...).
1646 		 *
1647 		 * See man buf(9) for more information
1648 		 */
1649 		resid = bp->b_bufsize;
1650 		foff = bp->b_offset;
1651 		for (i = 0; i < bp->b_npages; i++) {
1652 			int had_bogus = 0;
1653 
1654 			m = bp->b_pages[i];
1655 
1656 			/*
1657 			 * If we hit a bogus page, fixup *all* the bogus pages
1658 			 * now.
1659 			 */
1660 			if (m == bogus_page) {
1661 				poff = OFF_TO_IDX(bp->b_offset);
1662 				had_bogus = 1;
1663 
1664 				VM_OBJECT_RLOCK(obj);
1665 				for (j = i; j < bp->b_npages; j++) {
1666 					vm_page_t mtmp;
1667 					mtmp = bp->b_pages[j];
1668 					if (mtmp == bogus_page) {
1669 						mtmp = vm_page_lookup(obj, poff + j);
1670 						if (!mtmp) {
1671 							panic("brelse: page missing\n");
1672 						}
1673 						bp->b_pages[j] = mtmp;
1674 					}
1675 				}
1676 				VM_OBJECT_RUNLOCK(obj);
1677 
1678 				if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1679 					BUF_CHECK_MAPPED(bp);
1680 					pmap_qenter(
1681 					    trunc_page((vm_offset_t)bp->b_data),
1682 					    bp->b_pages, bp->b_npages);
1683 				}
1684 				m = bp->b_pages[i];
1685 			}
1686 			if ((bp->b_flags & B_NOCACHE) ||
1687 			    (bp->b_ioflags & BIO_ERROR &&
1688 			     bp->b_iocmd == BIO_READ)) {
1689 				int poffset = foff & PAGE_MASK;
1690 				int presid = resid > (PAGE_SIZE - poffset) ?
1691 					(PAGE_SIZE - poffset) : resid;
1692 
1693 				KASSERT(presid >= 0, ("brelse: extra page"));
1694 				VM_OBJECT_WLOCK(obj);
1695 				vm_page_set_invalid(m, poffset, presid);
1696 				VM_OBJECT_WUNLOCK(obj);
1697 				if (had_bogus)
1698 					printf("avoided corruption bug in bogus_page/brelse code\n");
1699 			}
1700 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1701 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1702 		}
1703 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1704 			vfs_vmio_release(bp);
1705 
1706 	} else if (bp->b_flags & B_VMIO) {
1707 
1708 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1709 			vfs_vmio_release(bp);
1710 		}
1711 
1712 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1713 		if (bp->b_bufsize != 0)
1714 			allocbuf(bp, 0);
1715 		if (bp->b_vp != NULL)
1716 			brelvp(bp);
1717 	}
1718 
1719 	/*
1720 	 * If the buffer has junk contents signal it and eventually
1721 	 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1722 	 * doesn't find it.
1723 	 */
1724 	if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1725 	    (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1726 		bp->b_flags |= B_INVAL;
1727 	if (bp->b_flags & B_INVAL) {
1728 		if (bp->b_flags & B_DELWRI)
1729 			bundirty(bp);
1730 		if (bp->b_vp)
1731 			brelvp(bp);
1732 	}
1733 
1734 	/* buffers with no memory */
1735 	if (bp->b_bufsize == 0) {
1736 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1737 		if (bp->b_vflags & BV_BKGRDINPROG)
1738 			panic("losing buffer 1");
1739 		if (bp->b_kvasize)
1740 			qindex = QUEUE_EMPTYKVA;
1741 		else
1742 			qindex = QUEUE_EMPTY;
1743 		bp->b_flags |= B_AGE;
1744 	/* buffers with junk contents */
1745 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1746 	    (bp->b_ioflags & BIO_ERROR)) {
1747 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1748 		if (bp->b_vflags & BV_BKGRDINPROG)
1749 			panic("losing buffer 2");
1750 		qindex = QUEUE_CLEAN;
1751 		bp->b_flags |= B_AGE;
1752 	/* remaining buffers */
1753 	} else if (bp->b_flags & B_DELWRI)
1754 		qindex = QUEUE_DIRTY;
1755 	else
1756 		qindex = QUEUE_CLEAN;
1757 
1758 	binsfree(bp, qindex);
1759 
1760 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1761 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1762 		panic("brelse: not dirty");
1763 	/* unlock */
1764 	BUF_UNLOCK(bp);
1765 }
1766 
1767 /*
1768  * Release a buffer back to the appropriate queue but do not try to free
1769  * it.  The buffer is expected to be used again soon.
1770  *
1771  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1772  * biodone() to requeue an async I/O on completion.  It is also used when
1773  * known good buffers need to be requeued but we think we may need the data
1774  * again soon.
1775  *
1776  * XXX we should be able to leave the B_RELBUF hint set on completion.
1777  */
1778 void
1779 bqrelse(struct buf *bp)
1780 {
1781 	int qindex;
1782 
1783 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1784 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1785 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1786 
1787 	if (BUF_LOCKRECURSED(bp)) {
1788 		/* do not release to free list */
1789 		BUF_UNLOCK(bp);
1790 		return;
1791 	}
1792 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1793 
1794 	if (bp->b_flags & B_MANAGED) {
1795 		if (bp->b_flags & B_REMFREE)
1796 			bremfreef(bp);
1797 		goto out;
1798 	}
1799 
1800 	/* buffers with stale but valid contents */
1801 	if (bp->b_flags & B_DELWRI) {
1802 		qindex = QUEUE_DIRTY;
1803 	} else {
1804 		if ((bp->b_flags & B_DELWRI) == 0 &&
1805 		    (bp->b_xflags & BX_VNDIRTY))
1806 			panic("bqrelse: not dirty");
1807 		/*
1808 		 * BKGRDINPROG can only be set with the buf and bufobj
1809 		 * locks both held.  We tolerate a race to clear it here.
1810 		 */
1811 		if (buf_vm_page_count_severe() &&
1812 		    (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1813 			/*
1814 			 * We are too low on memory, we have to try to free
1815 			 * the buffer (most importantly: the wired pages
1816 			 * making up its backing store) *now*.
1817 			 */
1818 			brelse(bp);
1819 			return;
1820 		}
1821 		qindex = QUEUE_CLEAN;
1822 	}
1823 	binsfree(bp, qindex);
1824 
1825 out:
1826 	/* unlock */
1827 	BUF_UNLOCK(bp);
1828 }
1829 
1830 /* Give pages used by the bp back to the VM system (where possible) */
1831 static void
1832 vfs_vmio_release(struct buf *bp)
1833 {
1834 	int i;
1835 	vm_page_t m;
1836 
1837 	if ((bp->b_flags & B_UNMAPPED) == 0) {
1838 		BUF_CHECK_MAPPED(bp);
1839 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1840 	} else
1841 		BUF_CHECK_UNMAPPED(bp);
1842 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
1843 	for (i = 0; i < bp->b_npages; i++) {
1844 		m = bp->b_pages[i];
1845 		bp->b_pages[i] = NULL;
1846 		/*
1847 		 * In order to keep page LRU ordering consistent, put
1848 		 * everything on the inactive queue.
1849 		 */
1850 		vm_page_lock(m);
1851 		vm_page_unwire(m, 0);
1852 		/*
1853 		 * We don't mess with busy pages, it is
1854 		 * the responsibility of the process that
1855 		 * busied the pages to deal with them.
1856 		 */
1857 		if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 &&
1858 		    m->wire_count == 0) {
1859 			/*
1860 			 * Might as well free the page if we can and it has
1861 			 * no valid data.  We also free the page if the
1862 			 * buffer was used for direct I/O
1863 			 */
1864 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1865 				vm_page_free(m);
1866 			} else if (bp->b_flags & B_DIRECT) {
1867 				vm_page_try_to_free(m);
1868 			} else if (buf_vm_page_count_severe()) {
1869 				vm_page_try_to_cache(m);
1870 			}
1871 		}
1872 		vm_page_unlock(m);
1873 	}
1874 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
1875 
1876 	if (bp->b_bufsize) {
1877 		bufspacewakeup();
1878 		bp->b_bufsize = 0;
1879 	}
1880 	bp->b_npages = 0;
1881 	bp->b_flags &= ~B_VMIO;
1882 	if (bp->b_vp)
1883 		brelvp(bp);
1884 }
1885 
1886 /*
1887  * Check to see if a block at a particular lbn is available for a clustered
1888  * write.
1889  */
1890 static int
1891 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1892 {
1893 	struct buf *bpa;
1894 	int match;
1895 
1896 	match = 0;
1897 
1898 	/* If the buf isn't in core skip it */
1899 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1900 		return (0);
1901 
1902 	/* If the buf is busy we don't want to wait for it */
1903 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1904 		return (0);
1905 
1906 	/* Only cluster with valid clusterable delayed write buffers */
1907 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1908 	    (B_DELWRI | B_CLUSTEROK))
1909 		goto done;
1910 
1911 	if (bpa->b_bufsize != size)
1912 		goto done;
1913 
1914 	/*
1915 	 * Check to see if it is in the expected place on disk and that the
1916 	 * block has been mapped.
1917 	 */
1918 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1919 		match = 1;
1920 done:
1921 	BUF_UNLOCK(bpa);
1922 	return (match);
1923 }
1924 
1925 /*
1926  *	vfs_bio_awrite:
1927  *
1928  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1929  *	This is much better then the old way of writing only one buffer at
1930  *	a time.  Note that we may not be presented with the buffers in the
1931  *	correct order, so we search for the cluster in both directions.
1932  */
1933 int
1934 vfs_bio_awrite(struct buf *bp)
1935 {
1936 	struct bufobj *bo;
1937 	int i;
1938 	int j;
1939 	daddr_t lblkno = bp->b_lblkno;
1940 	struct vnode *vp = bp->b_vp;
1941 	int ncl;
1942 	int nwritten;
1943 	int size;
1944 	int maxcl;
1945 	int gbflags;
1946 
1947 	bo = &vp->v_bufobj;
1948 	gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1949 	/*
1950 	 * right now we support clustered writing only to regular files.  If
1951 	 * we find a clusterable block we could be in the middle of a cluster
1952 	 * rather then at the beginning.
1953 	 */
1954 	if ((vp->v_type == VREG) &&
1955 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1956 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1957 
1958 		size = vp->v_mount->mnt_stat.f_iosize;
1959 		maxcl = MAXPHYS / size;
1960 
1961 		BO_RLOCK(bo);
1962 		for (i = 1; i < maxcl; i++)
1963 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1964 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1965 				break;
1966 
1967 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1968 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1969 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1970 				break;
1971 		BO_RUNLOCK(bo);
1972 		--j;
1973 		ncl = i + j;
1974 		/*
1975 		 * this is a possible cluster write
1976 		 */
1977 		if (ncl != 1) {
1978 			BUF_UNLOCK(bp);
1979 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
1980 			    gbflags);
1981 			return (nwritten);
1982 		}
1983 	}
1984 	bremfree(bp);
1985 	bp->b_flags |= B_ASYNC;
1986 	/*
1987 	 * default (old) behavior, writing out only one block
1988 	 *
1989 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1990 	 */
1991 	nwritten = bp->b_bufsize;
1992 	(void) bwrite(bp);
1993 
1994 	return (nwritten);
1995 }
1996 
1997 static void
1998 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
1999 {
2000 
2001 	KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2002 	    bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2003 	if ((gbflags & GB_UNMAPPED) == 0) {
2004 		bp->b_kvabase = (caddr_t)addr;
2005 	} else if ((gbflags & GB_KVAALLOC) != 0) {
2006 		KASSERT((gbflags & GB_UNMAPPED) != 0,
2007 		    ("GB_KVAALLOC without GB_UNMAPPED"));
2008 		bp->b_kvaalloc = (caddr_t)addr;
2009 		bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2010 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2011 	}
2012 	bp->b_kvasize = maxsize;
2013 }
2014 
2015 /*
2016  * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2017  * needed.
2018  */
2019 static int
2020 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2021 {
2022 	vm_offset_t addr;
2023 
2024 	bfreekva(bp);
2025 	addr = 0;
2026 
2027 	if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2028 		/*
2029 		 * Buffer map is too fragmented.  Request the caller
2030 		 * to defragment the map.
2031 		 */
2032 		atomic_add_int(&bufdefragcnt, 1);
2033 		return (1);
2034 	}
2035 	setbufkva(bp, addr, maxsize, gbflags);
2036 	atomic_add_long(&bufspace, bp->b_kvasize);
2037 	return (0);
2038 }
2039 
2040 /*
2041  * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2042  * locked vnode is supplied.
2043  */
2044 static void
2045 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2046     int defrag)
2047 {
2048 	struct thread *td;
2049 	char *waitmsg;
2050 	int fl, flags, norunbuf;
2051 
2052 	mtx_assert(&bqclean, MA_OWNED);
2053 
2054 	if (defrag) {
2055 		flags = VFS_BIO_NEED_BUFSPACE;
2056 		waitmsg = "nbufkv";
2057 	} else if (bufspace >= hibufspace) {
2058 		waitmsg = "nbufbs";
2059 		flags = VFS_BIO_NEED_BUFSPACE;
2060 	} else {
2061 		waitmsg = "newbuf";
2062 		flags = VFS_BIO_NEED_ANY;
2063 	}
2064 	mtx_lock(&nblock);
2065 	needsbuffer |= flags;
2066 	mtx_unlock(&nblock);
2067 	mtx_unlock(&bqclean);
2068 
2069 	bd_speedup();	/* heeeelp */
2070 	if ((gbflags & GB_NOWAIT_BD) != 0)
2071 		return;
2072 
2073 	td = curthread;
2074 	mtx_lock(&nblock);
2075 	while (needsbuffer & flags) {
2076 		if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2077 			mtx_unlock(&nblock);
2078 			/*
2079 			 * getblk() is called with a vnode locked, and
2080 			 * some majority of the dirty buffers may as
2081 			 * well belong to the vnode.  Flushing the
2082 			 * buffers there would make a progress that
2083 			 * cannot be achieved by the buf_daemon, that
2084 			 * cannot lock the vnode.
2085 			 */
2086 			norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2087 			    (td->td_pflags & TDP_NORUNNINGBUF);
2088 			/* play bufdaemon */
2089 			td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2090 			fl = buf_flush(vp, flushbufqtarget);
2091 			td->td_pflags &= norunbuf;
2092 			mtx_lock(&nblock);
2093 			if (fl != 0)
2094 				continue;
2095 			if ((needsbuffer & flags) == 0)
2096 				break;
2097 		}
2098 		if (msleep(&needsbuffer, &nblock, (PRIBIO + 4) | slpflag,
2099 		    waitmsg, slptimeo))
2100 			break;
2101 	}
2102 	mtx_unlock(&nblock);
2103 }
2104 
2105 static void
2106 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2107 {
2108 
2109 	CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2110 	    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2111 	     bp->b_kvasize, bp->b_bufsize, qindex);
2112 	mtx_assert(&bqclean, MA_NOTOWNED);
2113 
2114 	/*
2115 	 * Note: we no longer distinguish between VMIO and non-VMIO
2116 	 * buffers.
2117 	 */
2118 	KASSERT((bp->b_flags & B_DELWRI) == 0,
2119 	    ("delwri buffer %p found in queue %d", bp, qindex));
2120 
2121 	if (qindex == QUEUE_CLEAN) {
2122 		if (bp->b_flags & B_VMIO) {
2123 			bp->b_flags &= ~B_ASYNC;
2124 			vfs_vmio_release(bp);
2125 		}
2126 		if (bp->b_vp != NULL)
2127 			brelvp(bp);
2128 	}
2129 
2130 	/*
2131 	 * Get the rest of the buffer freed up.  b_kva* is still valid
2132 	 * after this operation.
2133 	 */
2134 
2135 	if (bp->b_rcred != NOCRED) {
2136 		crfree(bp->b_rcred);
2137 		bp->b_rcred = NOCRED;
2138 	}
2139 	if (bp->b_wcred != NOCRED) {
2140 		crfree(bp->b_wcred);
2141 		bp->b_wcred = NOCRED;
2142 	}
2143 	if (!LIST_EMPTY(&bp->b_dep))
2144 		buf_deallocate(bp);
2145 	if (bp->b_vflags & BV_BKGRDINPROG)
2146 		panic("losing buffer 3");
2147 	KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.  qindex: %d",
2148 	    bp, bp->b_vp, qindex));
2149 	KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2150 	    ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2151 
2152 	if (bp->b_bufsize)
2153 		allocbuf(bp, 0);
2154 
2155 	bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2156 	bp->b_ioflags = 0;
2157 	bp->b_xflags = 0;
2158 	KASSERT((bp->b_flags & B_INFREECNT) == 0,
2159 	    ("buf %p still counted as free?", bp));
2160 	bp->b_vflags = 0;
2161 	bp->b_vp = NULL;
2162 	bp->b_blkno = bp->b_lblkno = 0;
2163 	bp->b_offset = NOOFFSET;
2164 	bp->b_iodone = 0;
2165 	bp->b_error = 0;
2166 	bp->b_resid = 0;
2167 	bp->b_bcount = 0;
2168 	bp->b_npages = 0;
2169 	bp->b_dirtyoff = bp->b_dirtyend = 0;
2170 	bp->b_bufobj = NULL;
2171 	bp->b_pin_count = 0;
2172 	bp->b_fsprivate1 = NULL;
2173 	bp->b_fsprivate2 = NULL;
2174 	bp->b_fsprivate3 = NULL;
2175 
2176 	LIST_INIT(&bp->b_dep);
2177 }
2178 
2179 static int flushingbufs;
2180 
2181 static struct buf *
2182 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2183 {
2184 	struct buf *bp, *nbp;
2185 	int nqindex, qindex, pass;
2186 
2187 	KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2188 
2189 	pass = 1;
2190 restart:
2191 	atomic_add_int(&getnewbufrestarts, 1);
2192 
2193 	/*
2194 	 * Setup for scan.  If we do not have enough free buffers,
2195 	 * we setup a degenerate case that immediately fails.  Note
2196 	 * that if we are specially marked process, we are allowed to
2197 	 * dip into our reserves.
2198 	 *
2199 	 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2200 	 * for the allocation of the mapped buffer.  For unmapped, the
2201 	 * easiest is to start with EMPTY outright.
2202 	 *
2203 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
2204 	 * However, there are a number of cases (defragging, reusing, ...)
2205 	 * where we cannot backup.
2206 	 */
2207 	nbp = NULL;
2208 	mtx_lock(&bqclean);
2209 	if (!defrag && unmapped) {
2210 		nqindex = QUEUE_EMPTY;
2211 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2212 	}
2213 	if (nbp == NULL) {
2214 		nqindex = QUEUE_EMPTYKVA;
2215 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2216 	}
2217 
2218 	/*
2219 	 * If no EMPTYKVA buffers and we are either defragging or
2220 	 * reusing, locate a CLEAN buffer to free or reuse.  If
2221 	 * bufspace useage is low skip this step so we can allocate a
2222 	 * new buffer.
2223 	 */
2224 	if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2225 		nqindex = QUEUE_CLEAN;
2226 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2227 	}
2228 
2229 	/*
2230 	 * If we could not find or were not allowed to reuse a CLEAN
2231 	 * buffer, check to see if it is ok to use an EMPTY buffer.
2232 	 * We can only use an EMPTY buffer if allocating its KVA would
2233 	 * not otherwise run us out of buffer space.  No KVA is needed
2234 	 * for the unmapped allocation.
2235 	 */
2236 	if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2237 	    metadata)) {
2238 		nqindex = QUEUE_EMPTY;
2239 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2240 	}
2241 
2242 	/*
2243 	 * All available buffers might be clean, retry ignoring the
2244 	 * lobufspace as the last resort.
2245 	 */
2246 	if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2247 		nqindex = QUEUE_CLEAN;
2248 		nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2249 	}
2250 
2251 	/*
2252 	 * Run scan, possibly freeing data and/or kva mappings on the fly
2253 	 * depending.
2254 	 */
2255 	while ((bp = nbp) != NULL) {
2256 		qindex = nqindex;
2257 
2258 		/*
2259 		 * Calculate next bp (we can only use it if we do not
2260 		 * block or do other fancy things).
2261 		 */
2262 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2263 			switch (qindex) {
2264 			case QUEUE_EMPTY:
2265 				nqindex = QUEUE_EMPTYKVA;
2266 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2267 				if (nbp != NULL)
2268 					break;
2269 				/* FALLTHROUGH */
2270 			case QUEUE_EMPTYKVA:
2271 				nqindex = QUEUE_CLEAN;
2272 				nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2273 				if (nbp != NULL)
2274 					break;
2275 				/* FALLTHROUGH */
2276 			case QUEUE_CLEAN:
2277 				if (metadata && pass == 1) {
2278 					pass = 2;
2279 					nqindex = QUEUE_EMPTY;
2280 					nbp = TAILQ_FIRST(
2281 					    &bufqueues[QUEUE_EMPTY]);
2282 				}
2283 				/*
2284 				 * nbp is NULL.
2285 				 */
2286 				break;
2287 			}
2288 		}
2289 		/*
2290 		 * If we are defragging then we need a buffer with
2291 		 * b_kvasize != 0.  XXX this situation should no longer
2292 		 * occur, if defrag is non-zero the buffer's b_kvasize
2293 		 * should also be non-zero at this point.  XXX
2294 		 */
2295 		if (defrag && bp->b_kvasize == 0) {
2296 			printf("Warning: defrag empty buffer %p\n", bp);
2297 			continue;
2298 		}
2299 
2300 		/*
2301 		 * Start freeing the bp.  This is somewhat involved.  nbp
2302 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2303 		 */
2304 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2305 			continue;
2306 		/*
2307 		 * BKGRDINPROG can only be set with the buf and bufobj
2308 		 * locks both held.  We tolerate a race to clear it here.
2309 		 */
2310 		if (bp->b_vflags & BV_BKGRDINPROG) {
2311 			BUF_UNLOCK(bp);
2312 			continue;
2313 		}
2314 
2315 		KASSERT(bp->b_qindex == qindex,
2316 		    ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2317 
2318 		bremfreel(bp);
2319 		mtx_unlock(&bqclean);
2320 		/*
2321 		 * NOTE:  nbp is now entirely invalid.  We can only restart
2322 		 * the scan from this point on.
2323 		 */
2324 
2325 		getnewbuf_reuse_bp(bp, qindex);
2326 		mtx_assert(&bqclean, MA_NOTOWNED);
2327 
2328 		/*
2329 		 * If we are defragging then free the buffer.
2330 		 */
2331 		if (defrag) {
2332 			bp->b_flags |= B_INVAL;
2333 			bfreekva(bp);
2334 			brelse(bp);
2335 			defrag = 0;
2336 			goto restart;
2337 		}
2338 
2339 		/*
2340 		 * Notify any waiters for the buffer lock about
2341 		 * identity change by freeing the buffer.
2342 		 */
2343 		if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2344 			bp->b_flags |= B_INVAL;
2345 			bfreekva(bp);
2346 			brelse(bp);
2347 			goto restart;
2348 		}
2349 
2350 		if (metadata)
2351 			break;
2352 
2353 		/*
2354 		 * If we are overcomitted then recover the buffer and its
2355 		 * KVM space.  This occurs in rare situations when multiple
2356 		 * processes are blocked in getnewbuf() or allocbuf().
2357 		 */
2358 		if (bufspace >= hibufspace)
2359 			flushingbufs = 1;
2360 		if (flushingbufs && bp->b_kvasize != 0) {
2361 			bp->b_flags |= B_INVAL;
2362 			bfreekva(bp);
2363 			brelse(bp);
2364 			goto restart;
2365 		}
2366 		if (bufspace < lobufspace)
2367 			flushingbufs = 0;
2368 		break;
2369 	}
2370 	return (bp);
2371 }
2372 
2373 /*
2374  *	getnewbuf:
2375  *
2376  *	Find and initialize a new buffer header, freeing up existing buffers
2377  *	in the bufqueues as necessary.  The new buffer is returned locked.
2378  *
2379  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
2380  *	buffer away, the caller must set B_INVAL prior to calling brelse().
2381  *
2382  *	We block if:
2383  *		We have insufficient buffer headers
2384  *		We have insufficient buffer space
2385  *		buffer_arena is too fragmented ( space reservation fails )
2386  *		If we have to flush dirty buffers ( but we try to avoid this )
2387  */
2388 static struct buf *
2389 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2390     int gbflags)
2391 {
2392 	struct buf *bp;
2393 	int defrag, metadata;
2394 
2395 	KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2396 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2397 	if (!unmapped_buf_allowed)
2398 		gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2399 
2400 	defrag = 0;
2401 	if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2402 	    vp->v_type == VCHR)
2403 		metadata = 1;
2404 	else
2405 		metadata = 0;
2406 	/*
2407 	 * We can't afford to block since we might be holding a vnode lock,
2408 	 * which may prevent system daemons from running.  We deal with
2409 	 * low-memory situations by proactively returning memory and running
2410 	 * async I/O rather then sync I/O.
2411 	 */
2412 	atomic_add_int(&getnewbufcalls, 1);
2413 	atomic_subtract_int(&getnewbufrestarts, 1);
2414 restart:
2415 	bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2416 	    GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2417 	if (bp != NULL)
2418 		defrag = 0;
2419 
2420 	/*
2421 	 * If we exhausted our list, sleep as appropriate.  We may have to
2422 	 * wakeup various daemons and write out some dirty buffers.
2423 	 *
2424 	 * Generally we are sleeping due to insufficient buffer space.
2425 	 */
2426 	if (bp == NULL) {
2427 		mtx_assert(&bqclean, MA_OWNED);
2428 		getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2429 		mtx_assert(&bqclean, MA_NOTOWNED);
2430 	} else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2431 		mtx_assert(&bqclean, MA_NOTOWNED);
2432 
2433 		bfreekva(bp);
2434 		bp->b_flags |= B_UNMAPPED;
2435 		bp->b_kvabase = bp->b_data = unmapped_buf;
2436 		bp->b_kvasize = maxsize;
2437 		atomic_add_long(&bufspace, bp->b_kvasize);
2438 		atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2439 		atomic_add_int(&bufreusecnt, 1);
2440 	} else {
2441 		mtx_assert(&bqclean, MA_NOTOWNED);
2442 
2443 		/*
2444 		 * We finally have a valid bp.  We aren't quite out of the
2445 		 * woods, we still have to reserve kva space.  In order
2446 		 * to keep fragmentation sane we only allocate kva in
2447 		 * BKVASIZE chunks.
2448 		 */
2449 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2450 
2451 		if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2452 		    B_KVAALLOC)) == B_UNMAPPED) {
2453 			if (allocbufkva(bp, maxsize, gbflags)) {
2454 				defrag = 1;
2455 				bp->b_flags |= B_INVAL;
2456 				brelse(bp);
2457 				goto restart;
2458 			}
2459 			atomic_add_int(&bufreusecnt, 1);
2460 		} else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2461 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2462 			/*
2463 			 * If the reused buffer has KVA allocated,
2464 			 * reassign b_kvaalloc to b_kvabase.
2465 			 */
2466 			bp->b_kvabase = bp->b_kvaalloc;
2467 			bp->b_flags &= ~B_KVAALLOC;
2468 			atomic_subtract_long(&unmapped_bufspace,
2469 			    bp->b_kvasize);
2470 			atomic_add_int(&bufreusecnt, 1);
2471 		} else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2472 		    (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2473 		    GB_KVAALLOC)) {
2474 			/*
2475 			 * The case of reused buffer already have KVA
2476 			 * mapped, but the request is for unmapped
2477 			 * buffer with KVA allocated.
2478 			 */
2479 			bp->b_kvaalloc = bp->b_kvabase;
2480 			bp->b_data = bp->b_kvabase = unmapped_buf;
2481 			bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2482 			atomic_add_long(&unmapped_bufspace,
2483 			    bp->b_kvasize);
2484 			atomic_add_int(&bufreusecnt, 1);
2485 		}
2486 		if ((gbflags & GB_UNMAPPED) == 0) {
2487 			bp->b_saveaddr = bp->b_kvabase;
2488 			bp->b_data = bp->b_saveaddr;
2489 			bp->b_flags &= ~B_UNMAPPED;
2490 			BUF_CHECK_MAPPED(bp);
2491 		}
2492 	}
2493 	return (bp);
2494 }
2495 
2496 /*
2497  *	buf_daemon:
2498  *
2499  *	buffer flushing daemon.  Buffers are normally flushed by the
2500  *	update daemon but if it cannot keep up this process starts to
2501  *	take the load in an attempt to prevent getnewbuf() from blocking.
2502  */
2503 
2504 static struct kproc_desc buf_kp = {
2505 	"bufdaemon",
2506 	buf_daemon,
2507 	&bufdaemonproc
2508 };
2509 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2510 
2511 static int
2512 buf_flush(struct vnode *vp, int target)
2513 {
2514 	int flushed;
2515 
2516 	flushed = flushbufqueues(vp, target, 0);
2517 	if (flushed == 0) {
2518 		/*
2519 		 * Could not find any buffers without rollback
2520 		 * dependencies, so just write the first one
2521 		 * in the hopes of eventually making progress.
2522 		 */
2523 		if (vp != NULL && target > 2)
2524 			target /= 2;
2525 		flushbufqueues(vp, target, 1);
2526 	}
2527 	return (flushed);
2528 }
2529 
2530 static void
2531 buf_daemon()
2532 {
2533 	int lodirty;
2534 
2535 	/*
2536 	 * This process needs to be suspended prior to shutdown sync.
2537 	 */
2538 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2539 	    SHUTDOWN_PRI_LAST);
2540 
2541 	/*
2542 	 * This process is allowed to take the buffer cache to the limit
2543 	 */
2544 	curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2545 	mtx_lock(&bdlock);
2546 	for (;;) {
2547 		bd_request = 0;
2548 		mtx_unlock(&bdlock);
2549 
2550 		kproc_suspend_check(bufdaemonproc);
2551 		lodirty = lodirtybuffers;
2552 		if (bd_speedupreq) {
2553 			lodirty = numdirtybuffers / 2;
2554 			bd_speedupreq = 0;
2555 		}
2556 		/*
2557 		 * Do the flush.  Limit the amount of in-transit I/O we
2558 		 * allow to build up, otherwise we would completely saturate
2559 		 * the I/O system.
2560 		 */
2561 		while (numdirtybuffers > lodirty) {
2562 			if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
2563 				break;
2564 			kern_yield(PRI_USER);
2565 		}
2566 
2567 		/*
2568 		 * Only clear bd_request if we have reached our low water
2569 		 * mark.  The buf_daemon normally waits 1 second and
2570 		 * then incrementally flushes any dirty buffers that have
2571 		 * built up, within reason.
2572 		 *
2573 		 * If we were unable to hit our low water mark and couldn't
2574 		 * find any flushable buffers, we sleep for a short period
2575 		 * to avoid endless loops on unlockable buffers.
2576 		 */
2577 		mtx_lock(&bdlock);
2578 		if (numdirtybuffers <= lodirtybuffers) {
2579 			/*
2580 			 * We reached our low water mark, reset the
2581 			 * request and sleep until we are needed again.
2582 			 * The sleep is just so the suspend code works.
2583 			 */
2584 			bd_request = 0;
2585 			/*
2586 			 * Do an extra wakeup in case dirty threshold
2587 			 * changed via sysctl and the explicit transition
2588 			 * out of shortfall was missed.
2589 			 */
2590 			bdirtywakeup();
2591 			if (runningbufspace <= lorunningspace)
2592 				runningwakeup();
2593 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2594 		} else {
2595 			/*
2596 			 * We couldn't find any flushable dirty buffers but
2597 			 * still have too many dirty buffers, we
2598 			 * have to sleep and try again.  (rare)
2599 			 */
2600 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2601 		}
2602 	}
2603 }
2604 
2605 /*
2606  *	flushbufqueues:
2607  *
2608  *	Try to flush a buffer in the dirty queue.  We must be careful to
2609  *	free up B_INVAL buffers instead of write them, which NFS is
2610  *	particularly sensitive to.
2611  */
2612 static int flushwithdeps = 0;
2613 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2614     0, "Number of buffers flushed with dependecies that require rollbacks");
2615 
2616 static int
2617 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
2618 {
2619 	struct buf *sentinel;
2620 	struct vnode *vp;
2621 	struct mount *mp;
2622 	struct buf *bp;
2623 	int hasdeps;
2624 	int flushed;
2625 	int queue;
2626 
2627 	flushed = 0;
2628 	queue = QUEUE_DIRTY;
2629 	bp = NULL;
2630 	sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2631 	sentinel->b_qindex = QUEUE_SENTINEL;
2632 	mtx_lock(&bqdirty);
2633 	TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2634 	while (flushed != target) {
2635 		bp = TAILQ_NEXT(sentinel, b_freelist);
2636 		if (bp != NULL) {
2637 			TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2638 			TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2639 			    b_freelist);
2640 		} else
2641 			break;
2642 		/*
2643 		 * Skip sentinels inserted by other invocations of the
2644 		 * flushbufqueues(), taking care to not reorder them.
2645 		 */
2646 		if (bp->b_qindex == QUEUE_SENTINEL)
2647 			continue;
2648 		/*
2649 		 * Only flush the buffers that belong to the
2650 		 * vnode locked by the curthread.
2651 		 */
2652 		if (lvp != NULL && bp->b_vp != lvp)
2653 			continue;
2654 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2655 			continue;
2656 		if (bp->b_pin_count > 0) {
2657 			BUF_UNLOCK(bp);
2658 			continue;
2659 		}
2660 		/*
2661 		 * BKGRDINPROG can only be set with the buf and bufobj
2662 		 * locks both held.  We tolerate a race to clear it here.
2663 		 */
2664 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2665 		    (bp->b_flags & B_DELWRI) == 0) {
2666 			BUF_UNLOCK(bp);
2667 			continue;
2668 		}
2669 		if (bp->b_flags & B_INVAL) {
2670 			bremfreel(bp);
2671 			mtx_unlock(&bqdirty);
2672 			brelse(bp);
2673 			flushed++;
2674 			mtx_lock(&bqdirty);
2675 			continue;
2676 		}
2677 
2678 		if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2679 			if (flushdeps == 0) {
2680 				BUF_UNLOCK(bp);
2681 				continue;
2682 			}
2683 			hasdeps = 1;
2684 		} else
2685 			hasdeps = 0;
2686 		/*
2687 		 * We must hold the lock on a vnode before writing
2688 		 * one of its buffers. Otherwise we may confuse, or
2689 		 * in the case of a snapshot vnode, deadlock the
2690 		 * system.
2691 		 *
2692 		 * The lock order here is the reverse of the normal
2693 		 * of vnode followed by buf lock.  This is ok because
2694 		 * the NOWAIT will prevent deadlock.
2695 		 */
2696 		vp = bp->b_vp;
2697 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2698 			BUF_UNLOCK(bp);
2699 			continue;
2700 		}
2701 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2702 			mtx_unlock(&bqdirty);
2703 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2704 			    bp, bp->b_vp, bp->b_flags);
2705 			if (curproc == bufdaemonproc)
2706 				vfs_bio_awrite(bp);
2707 			else {
2708 				bremfree(bp);
2709 				bwrite(bp);
2710 				notbufdflushes++;
2711 			}
2712 			vn_finished_write(mp);
2713 			VOP_UNLOCK(vp, 0);
2714 			flushwithdeps += hasdeps;
2715 			flushed++;
2716 
2717 			/*
2718 			 * Sleeping on runningbufspace while holding
2719 			 * vnode lock leads to deadlock.
2720 			 */
2721 			if (curproc == bufdaemonproc &&
2722 			    runningbufspace > hirunningspace)
2723 				waitrunningbufspace();
2724 			mtx_lock(&bqdirty);
2725 			continue;
2726 		}
2727 		vn_finished_write(mp);
2728 		BUF_UNLOCK(bp);
2729 	}
2730 	TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2731 	mtx_unlock(&bqdirty);
2732 	free(sentinel, M_TEMP);
2733 	return (flushed);
2734 }
2735 
2736 /*
2737  * Check to see if a block is currently memory resident.
2738  */
2739 struct buf *
2740 incore(struct bufobj *bo, daddr_t blkno)
2741 {
2742 	struct buf *bp;
2743 
2744 	BO_RLOCK(bo);
2745 	bp = gbincore(bo, blkno);
2746 	BO_RUNLOCK(bo);
2747 	return (bp);
2748 }
2749 
2750 /*
2751  * Returns true if no I/O is needed to access the
2752  * associated VM object.  This is like incore except
2753  * it also hunts around in the VM system for the data.
2754  */
2755 
2756 static int
2757 inmem(struct vnode * vp, daddr_t blkno)
2758 {
2759 	vm_object_t obj;
2760 	vm_offset_t toff, tinc, size;
2761 	vm_page_t m;
2762 	vm_ooffset_t off;
2763 
2764 	ASSERT_VOP_LOCKED(vp, "inmem");
2765 
2766 	if (incore(&vp->v_bufobj, blkno))
2767 		return 1;
2768 	if (vp->v_mount == NULL)
2769 		return 0;
2770 	obj = vp->v_object;
2771 	if (obj == NULL)
2772 		return (0);
2773 
2774 	size = PAGE_SIZE;
2775 	if (size > vp->v_mount->mnt_stat.f_iosize)
2776 		size = vp->v_mount->mnt_stat.f_iosize;
2777 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2778 
2779 	VM_OBJECT_RLOCK(obj);
2780 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2781 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2782 		if (!m)
2783 			goto notinmem;
2784 		tinc = size;
2785 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2786 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2787 		if (vm_page_is_valid(m,
2788 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2789 			goto notinmem;
2790 	}
2791 	VM_OBJECT_RUNLOCK(obj);
2792 	return 1;
2793 
2794 notinmem:
2795 	VM_OBJECT_RUNLOCK(obj);
2796 	return (0);
2797 }
2798 
2799 /*
2800  * Set the dirty range for a buffer based on the status of the dirty
2801  * bits in the pages comprising the buffer.  The range is limited
2802  * to the size of the buffer.
2803  *
2804  * Tell the VM system that the pages associated with this buffer
2805  * are clean.  This is used for delayed writes where the data is
2806  * going to go to disk eventually without additional VM intevention.
2807  *
2808  * Note that while we only really need to clean through to b_bcount, we
2809  * just go ahead and clean through to b_bufsize.
2810  */
2811 static void
2812 vfs_clean_pages_dirty_buf(struct buf *bp)
2813 {
2814 	vm_ooffset_t foff, noff, eoff;
2815 	vm_page_t m;
2816 	int i;
2817 
2818 	if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2819 		return;
2820 
2821 	foff = bp->b_offset;
2822 	KASSERT(bp->b_offset != NOOFFSET,
2823 	    ("vfs_clean_pages_dirty_buf: no buffer offset"));
2824 
2825 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2826 	vfs_drain_busy_pages(bp);
2827 	vfs_setdirty_locked_object(bp);
2828 	for (i = 0; i < bp->b_npages; i++) {
2829 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2830 		eoff = noff;
2831 		if (eoff > bp->b_offset + bp->b_bufsize)
2832 			eoff = bp->b_offset + bp->b_bufsize;
2833 		m = bp->b_pages[i];
2834 		vfs_page_set_validclean(bp, foff, m);
2835 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2836 		foff = noff;
2837 	}
2838 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2839 }
2840 
2841 static void
2842 vfs_setdirty_locked_object(struct buf *bp)
2843 {
2844 	vm_object_t object;
2845 	int i;
2846 
2847 	object = bp->b_bufobj->bo_object;
2848 	VM_OBJECT_ASSERT_WLOCKED(object);
2849 
2850 	/*
2851 	 * We qualify the scan for modified pages on whether the
2852 	 * object has been flushed yet.
2853 	 */
2854 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2855 		vm_offset_t boffset;
2856 		vm_offset_t eoffset;
2857 
2858 		/*
2859 		 * test the pages to see if they have been modified directly
2860 		 * by users through the VM system.
2861 		 */
2862 		for (i = 0; i < bp->b_npages; i++)
2863 			vm_page_test_dirty(bp->b_pages[i]);
2864 
2865 		/*
2866 		 * Calculate the encompassing dirty range, boffset and eoffset,
2867 		 * (eoffset - boffset) bytes.
2868 		 */
2869 
2870 		for (i = 0; i < bp->b_npages; i++) {
2871 			if (bp->b_pages[i]->dirty)
2872 				break;
2873 		}
2874 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2875 
2876 		for (i = bp->b_npages - 1; i >= 0; --i) {
2877 			if (bp->b_pages[i]->dirty) {
2878 				break;
2879 			}
2880 		}
2881 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2882 
2883 		/*
2884 		 * Fit it to the buffer.
2885 		 */
2886 
2887 		if (eoffset > bp->b_bcount)
2888 			eoffset = bp->b_bcount;
2889 
2890 		/*
2891 		 * If we have a good dirty range, merge with the existing
2892 		 * dirty range.
2893 		 */
2894 
2895 		if (boffset < eoffset) {
2896 			if (bp->b_dirtyoff > boffset)
2897 				bp->b_dirtyoff = boffset;
2898 			if (bp->b_dirtyend < eoffset)
2899 				bp->b_dirtyend = eoffset;
2900 		}
2901 	}
2902 }
2903 
2904 /*
2905  * Allocate the KVA mapping for an existing buffer. It handles the
2906  * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2907  * KVA which is not mapped (B_KVAALLOC).
2908  */
2909 static void
2910 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2911 {
2912 	struct buf *scratch_bp;
2913 	int bsize, maxsize, need_mapping, need_kva;
2914 	off_t offset;
2915 
2916 	need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2917 	    (gbflags & GB_UNMAPPED) == 0;
2918 	need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2919 	    (gbflags & GB_KVAALLOC) != 0;
2920 	if (!need_mapping && !need_kva)
2921 		return;
2922 
2923 	BUF_CHECK_UNMAPPED(bp);
2924 
2925 	if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2926 		/*
2927 		 * Buffer is not mapped, but the KVA was already
2928 		 * reserved at the time of the instantiation.  Use the
2929 		 * allocated space.
2930 		 */
2931 		bp->b_flags &= ~B_KVAALLOC;
2932 		KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2933 		bp->b_kvabase = bp->b_kvaalloc;
2934 		atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
2935 		goto has_addr;
2936 	}
2937 
2938 	/*
2939 	 * Calculate the amount of the address space we would reserve
2940 	 * if the buffer was mapped.
2941 	 */
2942 	bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
2943 	offset = blkno * bsize;
2944 	maxsize = size + (offset & PAGE_MASK);
2945 	maxsize = imax(maxsize, bsize);
2946 
2947 mapping_loop:
2948 	if (allocbufkva(bp, maxsize, gbflags)) {
2949 		/*
2950 		 * Request defragmentation. getnewbuf() returns us the
2951 		 * allocated space by the scratch buffer KVA.
2952 		 */
2953 		scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
2954 		    (GB_UNMAPPED | GB_KVAALLOC));
2955 		if (scratch_bp == NULL) {
2956 			if ((gbflags & GB_NOWAIT_BD) != 0) {
2957 				/*
2958 				 * XXXKIB: defragmentation cannot
2959 				 * succeed, not sure what else to do.
2960 				 */
2961 				panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
2962 			}
2963 			atomic_add_int(&mappingrestarts, 1);
2964 			goto mapping_loop;
2965 		}
2966 		KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
2967 		    ("scratch bp !B_KVAALLOC %p", scratch_bp));
2968 		setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
2969 		    scratch_bp->b_kvasize, gbflags);
2970 
2971 		/* Get rid of the scratch buffer. */
2972 		scratch_bp->b_kvasize = 0;
2973 		scratch_bp->b_flags |= B_INVAL;
2974 		scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
2975 		brelse(scratch_bp);
2976 	}
2977 	if (!need_mapping)
2978 		return;
2979 
2980 has_addr:
2981 	bp->b_saveaddr = bp->b_kvabase;
2982 	bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
2983 	bp->b_flags &= ~B_UNMAPPED;
2984 	BUF_CHECK_MAPPED(bp);
2985 	bpmap_qenter(bp);
2986 }
2987 
2988 /*
2989  *	getblk:
2990  *
2991  *	Get a block given a specified block and offset into a file/device.
2992  *	The buffers B_DONE bit will be cleared on return, making it almost
2993  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2994  *	return.  The caller should clear B_INVAL prior to initiating a
2995  *	READ.
2996  *
2997  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2998  *	an existing buffer.
2999  *
3000  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
3001  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3002  *	and then cleared based on the backing VM.  If the previous buffer is
3003  *	non-0-sized but invalid, B_CACHE will be cleared.
3004  *
3005  *	If getblk() must create a new buffer, the new buffer is returned with
3006  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3007  *	case it is returned with B_INVAL clear and B_CACHE set based on the
3008  *	backing VM.
3009  *
3010  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
3011  *	B_CACHE bit is clear.
3012  *
3013  *	What this means, basically, is that the caller should use B_CACHE to
3014  *	determine whether the buffer is fully valid or not and should clear
3015  *	B_INVAL prior to issuing a read.  If the caller intends to validate
3016  *	the buffer by loading its data area with something, the caller needs
3017  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
3018  *	the caller should set B_CACHE ( as an optimization ), else the caller
3019  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
3020  *	a write attempt or if it was a successfull read.  If the caller
3021  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3022  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
3023  */
3024 struct buf *
3025 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3026     int flags)
3027 {
3028 	struct buf *bp;
3029 	struct bufobj *bo;
3030 	int bsize, error, maxsize, vmio;
3031 	off_t offset;
3032 
3033 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3034 	KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3035 	    ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3036 	ASSERT_VOP_LOCKED(vp, "getblk");
3037 	if (size > MAXBSIZE)
3038 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
3039 	if (!unmapped_buf_allowed)
3040 		flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3041 
3042 	bo = &vp->v_bufobj;
3043 loop:
3044 	BO_RLOCK(bo);
3045 	bp = gbincore(bo, blkno);
3046 	if (bp != NULL) {
3047 		int lockflags;
3048 		/*
3049 		 * Buffer is in-core.  If the buffer is not busy nor managed,
3050 		 * it must be on a queue.
3051 		 */
3052 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3053 
3054 		if (flags & GB_LOCK_NOWAIT)
3055 			lockflags |= LK_NOWAIT;
3056 
3057 		error = BUF_TIMELOCK(bp, lockflags,
3058 		    BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3059 
3060 		/*
3061 		 * If we slept and got the lock we have to restart in case
3062 		 * the buffer changed identities.
3063 		 */
3064 		if (error == ENOLCK)
3065 			goto loop;
3066 		/* We timed out or were interrupted. */
3067 		else if (error)
3068 			return (NULL);
3069 		/* If recursed, assume caller knows the rules. */
3070 		else if (BUF_LOCKRECURSED(bp))
3071 			goto end;
3072 
3073 		/*
3074 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
3075 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
3076 		 * and for a VMIO buffer B_CACHE is adjusted according to the
3077 		 * backing VM cache.
3078 		 */
3079 		if (bp->b_flags & B_INVAL)
3080 			bp->b_flags &= ~B_CACHE;
3081 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3082 			bp->b_flags |= B_CACHE;
3083 		if (bp->b_flags & B_MANAGED)
3084 			MPASS(bp->b_qindex == QUEUE_NONE);
3085 		else
3086 			bremfree(bp);
3087 
3088 		/*
3089 		 * check for size inconsistencies for non-VMIO case.
3090 		 */
3091 		if (bp->b_bcount != size) {
3092 			if ((bp->b_flags & B_VMIO) == 0 ||
3093 			    (size > bp->b_kvasize)) {
3094 				if (bp->b_flags & B_DELWRI) {
3095 					/*
3096 					 * If buffer is pinned and caller does
3097 					 * not want sleep  waiting for it to be
3098 					 * unpinned, bail out
3099 					 * */
3100 					if (bp->b_pin_count > 0) {
3101 						if (flags & GB_LOCK_NOWAIT) {
3102 							bqrelse(bp);
3103 							return (NULL);
3104 						} else {
3105 							bunpin_wait(bp);
3106 						}
3107 					}
3108 					bp->b_flags |= B_NOCACHE;
3109 					bwrite(bp);
3110 				} else {
3111 					if (LIST_EMPTY(&bp->b_dep)) {
3112 						bp->b_flags |= B_RELBUF;
3113 						brelse(bp);
3114 					} else {
3115 						bp->b_flags |= B_NOCACHE;
3116 						bwrite(bp);
3117 					}
3118 				}
3119 				goto loop;
3120 			}
3121 		}
3122 
3123 		/*
3124 		 * Handle the case of unmapped buffer which should
3125 		 * become mapped, or the buffer for which KVA
3126 		 * reservation is requested.
3127 		 */
3128 		bp_unmapped_get_kva(bp, blkno, size, flags);
3129 
3130 		/*
3131 		 * If the size is inconsistant in the VMIO case, we can resize
3132 		 * the buffer.  This might lead to B_CACHE getting set or
3133 		 * cleared.  If the size has not changed, B_CACHE remains
3134 		 * unchanged from its previous state.
3135 		 */
3136 		if (bp->b_bcount != size)
3137 			allocbuf(bp, size);
3138 
3139 		KASSERT(bp->b_offset != NOOFFSET,
3140 		    ("getblk: no buffer offset"));
3141 
3142 		/*
3143 		 * A buffer with B_DELWRI set and B_CACHE clear must
3144 		 * be committed before we can return the buffer in
3145 		 * order to prevent the caller from issuing a read
3146 		 * ( due to B_CACHE not being set ) and overwriting
3147 		 * it.
3148 		 *
3149 		 * Most callers, including NFS and FFS, need this to
3150 		 * operate properly either because they assume they
3151 		 * can issue a read if B_CACHE is not set, or because
3152 		 * ( for example ) an uncached B_DELWRI might loop due
3153 		 * to softupdates re-dirtying the buffer.  In the latter
3154 		 * case, B_CACHE is set after the first write completes,
3155 		 * preventing further loops.
3156 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
3157 		 * above while extending the buffer, we cannot allow the
3158 		 * buffer to remain with B_CACHE set after the write
3159 		 * completes or it will represent a corrupt state.  To
3160 		 * deal with this we set B_NOCACHE to scrap the buffer
3161 		 * after the write.
3162 		 *
3163 		 * We might be able to do something fancy, like setting
3164 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
3165 		 * so the below call doesn't set B_CACHE, but that gets real
3166 		 * confusing.  This is much easier.
3167 		 */
3168 
3169 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3170 			bp->b_flags |= B_NOCACHE;
3171 			bwrite(bp);
3172 			goto loop;
3173 		}
3174 		bp->b_flags &= ~B_DONE;
3175 	} else {
3176 		/*
3177 		 * Buffer is not in-core, create new buffer.  The buffer
3178 		 * returned by getnewbuf() is locked.  Note that the returned
3179 		 * buffer is also considered valid (not marked B_INVAL).
3180 		 */
3181 		BO_RUNLOCK(bo);
3182 		/*
3183 		 * If the user does not want us to create the buffer, bail out
3184 		 * here.
3185 		 */
3186 		if (flags & GB_NOCREAT)
3187 			return NULL;
3188 		if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3189 			return NULL;
3190 
3191 		bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3192 		offset = blkno * bsize;
3193 		vmio = vp->v_object != NULL;
3194 		if (vmio) {
3195 			maxsize = size + (offset & PAGE_MASK);
3196 		} else {
3197 			maxsize = size;
3198 			/* Do not allow non-VMIO notmapped buffers. */
3199 			flags &= ~GB_UNMAPPED;
3200 		}
3201 		maxsize = imax(maxsize, bsize);
3202 
3203 		bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3204 		if (bp == NULL) {
3205 			if (slpflag || slptimeo)
3206 				return NULL;
3207 			goto loop;
3208 		}
3209 
3210 		/*
3211 		 * This code is used to make sure that a buffer is not
3212 		 * created while the getnewbuf routine is blocked.
3213 		 * This can be a problem whether the vnode is locked or not.
3214 		 * If the buffer is created out from under us, we have to
3215 		 * throw away the one we just created.
3216 		 *
3217 		 * Note: this must occur before we associate the buffer
3218 		 * with the vp especially considering limitations in
3219 		 * the splay tree implementation when dealing with duplicate
3220 		 * lblkno's.
3221 		 */
3222 		BO_LOCK(bo);
3223 		if (gbincore(bo, blkno)) {
3224 			BO_UNLOCK(bo);
3225 			bp->b_flags |= B_INVAL;
3226 			brelse(bp);
3227 			goto loop;
3228 		}
3229 
3230 		/*
3231 		 * Insert the buffer into the hash, so that it can
3232 		 * be found by incore.
3233 		 */
3234 		bp->b_blkno = bp->b_lblkno = blkno;
3235 		bp->b_offset = offset;
3236 		bgetvp(vp, bp);
3237 		BO_UNLOCK(bo);
3238 
3239 		/*
3240 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
3241 		 * buffer size starts out as 0, B_CACHE will be set by
3242 		 * allocbuf() for the VMIO case prior to it testing the
3243 		 * backing store for validity.
3244 		 */
3245 
3246 		if (vmio) {
3247 			bp->b_flags |= B_VMIO;
3248 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3249 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3250 			    bp, vp->v_object, bp->b_bufobj->bo_object));
3251 		} else {
3252 			bp->b_flags &= ~B_VMIO;
3253 			KASSERT(bp->b_bufobj->bo_object == NULL,
3254 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
3255 			    bp, bp->b_bufobj->bo_object));
3256 			BUF_CHECK_MAPPED(bp);
3257 		}
3258 
3259 		allocbuf(bp, size);
3260 		bp->b_flags &= ~B_DONE;
3261 	}
3262 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3263 	BUF_ASSERT_HELD(bp);
3264 end:
3265 	KASSERT(bp->b_bufobj == bo,
3266 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3267 	return (bp);
3268 }
3269 
3270 /*
3271  * Get an empty, disassociated buffer of given size.  The buffer is initially
3272  * set to B_INVAL.
3273  */
3274 struct buf *
3275 geteblk(int size, int flags)
3276 {
3277 	struct buf *bp;
3278 	int maxsize;
3279 
3280 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
3281 	while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3282 		if ((flags & GB_NOWAIT_BD) &&
3283 		    (curthread->td_pflags & TDP_BUFNEED) != 0)
3284 			return (NULL);
3285 	}
3286 	allocbuf(bp, size);
3287 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
3288 	BUF_ASSERT_HELD(bp);
3289 	return (bp);
3290 }
3291 
3292 
3293 /*
3294  * This code constitutes the buffer memory from either anonymous system
3295  * memory (in the case of non-VMIO operations) or from an associated
3296  * VM object (in the case of VMIO operations).  This code is able to
3297  * resize a buffer up or down.
3298  *
3299  * Note that this code is tricky, and has many complications to resolve
3300  * deadlock or inconsistant data situations.  Tread lightly!!!
3301  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3302  * the caller.  Calling this code willy nilly can result in the loss of data.
3303  *
3304  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
3305  * B_CACHE for the non-VMIO case.
3306  */
3307 
3308 int
3309 allocbuf(struct buf *bp, int size)
3310 {
3311 	int newbsize, mbsize;
3312 	int i;
3313 
3314 	BUF_ASSERT_HELD(bp);
3315 
3316 	if (bp->b_kvasize < size)
3317 		panic("allocbuf: buffer too small");
3318 
3319 	if ((bp->b_flags & B_VMIO) == 0) {
3320 		caddr_t origbuf;
3321 		int origbufsize;
3322 		/*
3323 		 * Just get anonymous memory from the kernel.  Don't
3324 		 * mess with B_CACHE.
3325 		 */
3326 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3327 		if (bp->b_flags & B_MALLOC)
3328 			newbsize = mbsize;
3329 		else
3330 			newbsize = round_page(size);
3331 
3332 		if (newbsize < bp->b_bufsize) {
3333 			/*
3334 			 * malloced buffers are not shrunk
3335 			 */
3336 			if (bp->b_flags & B_MALLOC) {
3337 				if (newbsize) {
3338 					bp->b_bcount = size;
3339 				} else {
3340 					free(bp->b_data, M_BIOBUF);
3341 					if (bp->b_bufsize) {
3342 						atomic_subtract_long(
3343 						    &bufmallocspace,
3344 						    bp->b_bufsize);
3345 						bufspacewakeup();
3346 						bp->b_bufsize = 0;
3347 					}
3348 					bp->b_saveaddr = bp->b_kvabase;
3349 					bp->b_data = bp->b_saveaddr;
3350 					bp->b_bcount = 0;
3351 					bp->b_flags &= ~B_MALLOC;
3352 				}
3353 				return 1;
3354 			}
3355 			vm_hold_free_pages(bp, newbsize);
3356 		} else if (newbsize > bp->b_bufsize) {
3357 			/*
3358 			 * We only use malloced memory on the first allocation.
3359 			 * and revert to page-allocated memory when the buffer
3360 			 * grows.
3361 			 */
3362 			/*
3363 			 * There is a potential smp race here that could lead
3364 			 * to bufmallocspace slightly passing the max.  It
3365 			 * is probably extremely rare and not worth worrying
3366 			 * over.
3367 			 */
3368 			if ( (bufmallocspace < maxbufmallocspace) &&
3369 				(bp->b_bufsize == 0) &&
3370 				(mbsize <= PAGE_SIZE/2)) {
3371 
3372 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3373 				bp->b_bufsize = mbsize;
3374 				bp->b_bcount = size;
3375 				bp->b_flags |= B_MALLOC;
3376 				atomic_add_long(&bufmallocspace, mbsize);
3377 				return 1;
3378 			}
3379 			origbuf = NULL;
3380 			origbufsize = 0;
3381 			/*
3382 			 * If the buffer is growing on its other-than-first allocation,
3383 			 * then we revert to the page-allocation scheme.
3384 			 */
3385 			if (bp->b_flags & B_MALLOC) {
3386 				origbuf = bp->b_data;
3387 				origbufsize = bp->b_bufsize;
3388 				bp->b_data = bp->b_kvabase;
3389 				if (bp->b_bufsize) {
3390 					atomic_subtract_long(&bufmallocspace,
3391 					    bp->b_bufsize);
3392 					bufspacewakeup();
3393 					bp->b_bufsize = 0;
3394 				}
3395 				bp->b_flags &= ~B_MALLOC;
3396 				newbsize = round_page(newbsize);
3397 			}
3398 			vm_hold_load_pages(
3399 			    bp,
3400 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
3401 			    (vm_offset_t) bp->b_data + newbsize);
3402 			if (origbuf) {
3403 				bcopy(origbuf, bp->b_data, origbufsize);
3404 				free(origbuf, M_BIOBUF);
3405 			}
3406 		}
3407 	} else {
3408 		int desiredpages;
3409 
3410 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3411 		desiredpages = (size == 0) ? 0 :
3412 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3413 
3414 		if (bp->b_flags & B_MALLOC)
3415 			panic("allocbuf: VMIO buffer can't be malloced");
3416 		/*
3417 		 * Set B_CACHE initially if buffer is 0 length or will become
3418 		 * 0-length.
3419 		 */
3420 		if (size == 0 || bp->b_bufsize == 0)
3421 			bp->b_flags |= B_CACHE;
3422 
3423 		if (newbsize < bp->b_bufsize) {
3424 			/*
3425 			 * DEV_BSIZE aligned new buffer size is less then the
3426 			 * DEV_BSIZE aligned existing buffer size.  Figure out
3427 			 * if we have to remove any pages.
3428 			 */
3429 			if (desiredpages < bp->b_npages) {
3430 				vm_page_t m;
3431 
3432 				if ((bp->b_flags & B_UNMAPPED) == 0) {
3433 					BUF_CHECK_MAPPED(bp);
3434 					pmap_qremove((vm_offset_t)trunc_page(
3435 					    (vm_offset_t)bp->b_data) +
3436 					    (desiredpages << PAGE_SHIFT),
3437 					    (bp->b_npages - desiredpages));
3438 				} else
3439 					BUF_CHECK_UNMAPPED(bp);
3440 				VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3441 				for (i = desiredpages; i < bp->b_npages; i++) {
3442 					/*
3443 					 * the page is not freed here -- it
3444 					 * is the responsibility of
3445 					 * vnode_pager_setsize
3446 					 */
3447 					m = bp->b_pages[i];
3448 					KASSERT(m != bogus_page,
3449 					    ("allocbuf: bogus page found"));
3450 					while (vm_page_sleep_if_busy(m, TRUE,
3451 					    "biodep"))
3452 						continue;
3453 
3454 					bp->b_pages[i] = NULL;
3455 					vm_page_lock(m);
3456 					vm_page_unwire(m, 0);
3457 					vm_page_unlock(m);
3458 				}
3459 				VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3460 				bp->b_npages = desiredpages;
3461 			}
3462 		} else if (size > bp->b_bcount) {
3463 			/*
3464 			 * We are growing the buffer, possibly in a
3465 			 * byte-granular fashion.
3466 			 */
3467 			vm_object_t obj;
3468 			vm_offset_t toff;
3469 			vm_offset_t tinc;
3470 
3471 			/*
3472 			 * Step 1, bring in the VM pages from the object,
3473 			 * allocating them if necessary.  We must clear
3474 			 * B_CACHE if these pages are not valid for the
3475 			 * range covered by the buffer.
3476 			 */
3477 
3478 			obj = bp->b_bufobj->bo_object;
3479 
3480 			VM_OBJECT_WLOCK(obj);
3481 			while (bp->b_npages < desiredpages) {
3482 				vm_page_t m;
3483 
3484 				/*
3485 				 * We must allocate system pages since blocking
3486 				 * here could interfere with paging I/O, no
3487 				 * matter which process we are.
3488 				 *
3489 				 * We can only test VPO_BUSY here.  Blocking on
3490 				 * m->busy might lead to a deadlock:
3491 				 *  vm_fault->getpages->cluster_read->allocbuf
3492 				 * Thus, we specify VM_ALLOC_IGN_SBUSY.
3493 				 */
3494 				m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3495 				    bp->b_npages, VM_ALLOC_NOBUSY |
3496 				    VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3497 				    VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY |
3498 				    VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3499 				if (m->valid == 0)
3500 					bp->b_flags &= ~B_CACHE;
3501 				bp->b_pages[bp->b_npages] = m;
3502 				++bp->b_npages;
3503 			}
3504 
3505 			/*
3506 			 * Step 2.  We've loaded the pages into the buffer,
3507 			 * we have to figure out if we can still have B_CACHE
3508 			 * set.  Note that B_CACHE is set according to the
3509 			 * byte-granular range ( bcount and size ), new the
3510 			 * aligned range ( newbsize ).
3511 			 *
3512 			 * The VM test is against m->valid, which is DEV_BSIZE
3513 			 * aligned.  Needless to say, the validity of the data
3514 			 * needs to also be DEV_BSIZE aligned.  Note that this
3515 			 * fails with NFS if the server or some other client
3516 			 * extends the file's EOF.  If our buffer is resized,
3517 			 * B_CACHE may remain set! XXX
3518 			 */
3519 
3520 			toff = bp->b_bcount;
3521 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3522 
3523 			while ((bp->b_flags & B_CACHE) && toff < size) {
3524 				vm_pindex_t pi;
3525 
3526 				if (tinc > (size - toff))
3527 					tinc = size - toff;
3528 
3529 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3530 				    PAGE_SHIFT;
3531 
3532 				vfs_buf_test_cache(
3533 				    bp,
3534 				    bp->b_offset,
3535 				    toff,
3536 				    tinc,
3537 				    bp->b_pages[pi]
3538 				);
3539 				toff += tinc;
3540 				tinc = PAGE_SIZE;
3541 			}
3542 			VM_OBJECT_WUNLOCK(obj);
3543 
3544 			/*
3545 			 * Step 3, fixup the KVM pmap.
3546 			 */
3547 			if ((bp->b_flags & B_UNMAPPED) == 0)
3548 				bpmap_qenter(bp);
3549 			else
3550 				BUF_CHECK_UNMAPPED(bp);
3551 		}
3552 	}
3553 	if (newbsize < bp->b_bufsize)
3554 		bufspacewakeup();
3555 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
3556 	bp->b_bcount = size;		/* requested buffer size	*/
3557 	return 1;
3558 }
3559 
3560 extern int inflight_transient_maps;
3561 
3562 void
3563 biodone(struct bio *bp)
3564 {
3565 	struct mtx *mtxp;
3566 	void (*done)(struct bio *);
3567 	vm_offset_t start, end;
3568 	int transient;
3569 
3570 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3571 	mtx_lock(mtxp);
3572 	bp->bio_flags |= BIO_DONE;
3573 	if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3574 		start = trunc_page((vm_offset_t)bp->bio_data);
3575 		end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3576 		transient = 1;
3577 	} else {
3578 		transient = 0;
3579 		start = end = 0;
3580 	}
3581 	done = bp->bio_done;
3582 	if (done == NULL)
3583 		wakeup(bp);
3584 	mtx_unlock(mtxp);
3585 	if (done != NULL)
3586 		done(bp);
3587 	if (transient) {
3588 		pmap_qremove(start, OFF_TO_IDX(end - start));
3589 		vmem_free(transient_arena, start, end - start);
3590 		atomic_add_int(&inflight_transient_maps, -1);
3591 	}
3592 }
3593 
3594 /*
3595  * Wait for a BIO to finish.
3596  *
3597  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
3598  * case is not yet clear.
3599  */
3600 int
3601 biowait(struct bio *bp, const char *wchan)
3602 {
3603 	struct mtx *mtxp;
3604 
3605 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
3606 	mtx_lock(mtxp);
3607 	while ((bp->bio_flags & BIO_DONE) == 0)
3608 		msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3609 	mtx_unlock(mtxp);
3610 	if (bp->bio_error != 0)
3611 		return (bp->bio_error);
3612 	if (!(bp->bio_flags & BIO_ERROR))
3613 		return (0);
3614 	return (EIO);
3615 }
3616 
3617 void
3618 biofinish(struct bio *bp, struct devstat *stat, int error)
3619 {
3620 
3621 	if (error) {
3622 		bp->bio_error = error;
3623 		bp->bio_flags |= BIO_ERROR;
3624 	}
3625 	if (stat != NULL)
3626 		devstat_end_transaction_bio(stat, bp);
3627 	biodone(bp);
3628 }
3629 
3630 /*
3631  *	bufwait:
3632  *
3633  *	Wait for buffer I/O completion, returning error status.  The buffer
3634  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3635  *	error and cleared.
3636  */
3637 int
3638 bufwait(struct buf *bp)
3639 {
3640 	if (bp->b_iocmd == BIO_READ)
3641 		bwait(bp, PRIBIO, "biord");
3642 	else
3643 		bwait(bp, PRIBIO, "biowr");
3644 	if (bp->b_flags & B_EINTR) {
3645 		bp->b_flags &= ~B_EINTR;
3646 		return (EINTR);
3647 	}
3648 	if (bp->b_ioflags & BIO_ERROR) {
3649 		return (bp->b_error ? bp->b_error : EIO);
3650 	} else {
3651 		return (0);
3652 	}
3653 }
3654 
3655  /*
3656   * Call back function from struct bio back up to struct buf.
3657   */
3658 static void
3659 bufdonebio(struct bio *bip)
3660 {
3661 	struct buf *bp;
3662 
3663 	bp = bip->bio_caller2;
3664 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3665 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3666 	bp->b_ioflags = bip->bio_flags;
3667 	bp->b_error = bip->bio_error;
3668 	if (bp->b_error)
3669 		bp->b_ioflags |= BIO_ERROR;
3670 	bufdone(bp);
3671 	g_destroy_bio(bip);
3672 }
3673 
3674 void
3675 dev_strategy(struct cdev *dev, struct buf *bp)
3676 {
3677 	struct cdevsw *csw;
3678 	int ref;
3679 
3680 	KASSERT(dev->si_refcount > 0,
3681 	    ("dev_strategy on un-referenced struct cdev *(%s) %p",
3682 	    devtoname(dev), dev));
3683 
3684 	csw = dev_refthread(dev, &ref);
3685 	dev_strategy_csw(dev, csw, bp);
3686 	dev_relthread(dev, ref);
3687 }
3688 
3689 void
3690 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3691 {
3692 	struct bio *bip;
3693 
3694 	KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3695 	    ("b_iocmd botch"));
3696 	KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3697 	    dev->si_threadcount > 0,
3698 	    ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3699 	    dev));
3700 	if (csw == NULL) {
3701 		bp->b_error = ENXIO;
3702 		bp->b_ioflags = BIO_ERROR;
3703 		bufdone(bp);
3704 		return;
3705 	}
3706 	for (;;) {
3707 		bip = g_new_bio();
3708 		if (bip != NULL)
3709 			break;
3710 		/* Try again later */
3711 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3712 	}
3713 	bip->bio_cmd = bp->b_iocmd;
3714 	bip->bio_offset = bp->b_iooffset;
3715 	bip->bio_length = bp->b_bcount;
3716 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3717 	bdata2bio(bp, bip);
3718 	bip->bio_done = bufdonebio;
3719 	bip->bio_caller2 = bp;
3720 	bip->bio_dev = dev;
3721 	(*csw->d_strategy)(bip);
3722 }
3723 
3724 /*
3725  *	bufdone:
3726  *
3727  *	Finish I/O on a buffer, optionally calling a completion function.
3728  *	This is usually called from an interrupt so process blocking is
3729  *	not allowed.
3730  *
3731  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3732  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3733  *	assuming B_INVAL is clear.
3734  *
3735  *	For the VMIO case, we set B_CACHE if the op was a read and no
3736  *	read error occured, or if the op was a write.  B_CACHE is never
3737  *	set if the buffer is invalid or otherwise uncacheable.
3738  *
3739  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3740  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3741  *	in the biodone routine.
3742  */
3743 void
3744 bufdone(struct buf *bp)
3745 {
3746 	struct bufobj *dropobj;
3747 	void    (*biodone)(struct buf *);
3748 
3749 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3750 	dropobj = NULL;
3751 
3752 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3753 	BUF_ASSERT_HELD(bp);
3754 
3755 	runningbufwakeup(bp);
3756 	if (bp->b_iocmd == BIO_WRITE)
3757 		dropobj = bp->b_bufobj;
3758 	/* call optional completion function if requested */
3759 	if (bp->b_iodone != NULL) {
3760 		biodone = bp->b_iodone;
3761 		bp->b_iodone = NULL;
3762 		(*biodone) (bp);
3763 		if (dropobj)
3764 			bufobj_wdrop(dropobj);
3765 		return;
3766 	}
3767 
3768 	bufdone_finish(bp);
3769 
3770 	if (dropobj)
3771 		bufobj_wdrop(dropobj);
3772 }
3773 
3774 void
3775 bufdone_finish(struct buf *bp)
3776 {
3777 	BUF_ASSERT_HELD(bp);
3778 
3779 	if (!LIST_EMPTY(&bp->b_dep))
3780 		buf_complete(bp);
3781 
3782 	if (bp->b_flags & B_VMIO) {
3783 		vm_ooffset_t foff;
3784 		vm_page_t m;
3785 		vm_object_t obj;
3786 		struct vnode *vp;
3787 		int bogus, i, iosize;
3788 
3789 		obj = bp->b_bufobj->bo_object;
3790 		KASSERT(obj->paging_in_progress >= bp->b_npages,
3791 		    ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3792 		    obj->paging_in_progress, bp->b_npages));
3793 
3794 		vp = bp->b_vp;
3795 		KASSERT(vp->v_holdcnt > 0,
3796 		    ("biodone_finish: vnode %p has zero hold count", vp));
3797 		KASSERT(vp->v_object != NULL,
3798 		    ("biodone_finish: vnode %p has no vm_object", vp));
3799 
3800 		foff = bp->b_offset;
3801 		KASSERT(bp->b_offset != NOOFFSET,
3802 		    ("biodone_finish: bp %p has no buffer offset", bp));
3803 
3804 		/*
3805 		 * Set B_CACHE if the op was a normal read and no error
3806 		 * occured.  B_CACHE is set for writes in the b*write()
3807 		 * routines.
3808 		 */
3809 		iosize = bp->b_bcount - bp->b_resid;
3810 		if (bp->b_iocmd == BIO_READ &&
3811 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3812 		    !(bp->b_ioflags & BIO_ERROR)) {
3813 			bp->b_flags |= B_CACHE;
3814 		}
3815 		bogus = 0;
3816 		VM_OBJECT_WLOCK(obj);
3817 		for (i = 0; i < bp->b_npages; i++) {
3818 			int bogusflag = 0;
3819 			int resid;
3820 
3821 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3822 			if (resid > iosize)
3823 				resid = iosize;
3824 
3825 			/*
3826 			 * cleanup bogus pages, restoring the originals
3827 			 */
3828 			m = bp->b_pages[i];
3829 			if (m == bogus_page) {
3830 				bogus = bogusflag = 1;
3831 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3832 				if (m == NULL)
3833 					panic("biodone: page disappeared!");
3834 				bp->b_pages[i] = m;
3835 			}
3836 			KASSERT(OFF_TO_IDX(foff) == m->pindex,
3837 			    ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3838 			    (intmax_t)foff, (uintmax_t)m->pindex));
3839 
3840 			/*
3841 			 * In the write case, the valid and clean bits are
3842 			 * already changed correctly ( see bdwrite() ), so we
3843 			 * only need to do this here in the read case.
3844 			 */
3845 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3846 				KASSERT((m->dirty & vm_page_bits(foff &
3847 				    PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3848 				    " page %p has unexpected dirty bits", m));
3849 				vfs_page_set_valid(bp, foff, m);
3850 			}
3851 
3852 			vm_page_io_finish(m);
3853 			vm_object_pip_subtract(obj, 1);
3854 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3855 			iosize -= resid;
3856 		}
3857 		vm_object_pip_wakeupn(obj, 0);
3858 		VM_OBJECT_WUNLOCK(obj);
3859 		if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3860 			BUF_CHECK_MAPPED(bp);
3861 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3862 			    bp->b_pages, bp->b_npages);
3863 		}
3864 	}
3865 
3866 	/*
3867 	 * For asynchronous completions, release the buffer now. The brelse
3868 	 * will do a wakeup there if necessary - so no need to do a wakeup
3869 	 * here in the async case. The sync case always needs to do a wakeup.
3870 	 */
3871 
3872 	if (bp->b_flags & B_ASYNC) {
3873 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3874 			brelse(bp);
3875 		else
3876 			bqrelse(bp);
3877 	} else
3878 		bdone(bp);
3879 }
3880 
3881 /*
3882  * This routine is called in lieu of iodone in the case of
3883  * incomplete I/O.  This keeps the busy status for pages
3884  * consistant.
3885  */
3886 void
3887 vfs_unbusy_pages(struct buf *bp)
3888 {
3889 	int i;
3890 	vm_object_t obj;
3891 	vm_page_t m;
3892 
3893 	runningbufwakeup(bp);
3894 	if (!(bp->b_flags & B_VMIO))
3895 		return;
3896 
3897 	obj = bp->b_bufobj->bo_object;
3898 	VM_OBJECT_WLOCK(obj);
3899 	for (i = 0; i < bp->b_npages; i++) {
3900 		m = bp->b_pages[i];
3901 		if (m == bogus_page) {
3902 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3903 			if (!m)
3904 				panic("vfs_unbusy_pages: page missing\n");
3905 			bp->b_pages[i] = m;
3906 			if ((bp->b_flags & B_UNMAPPED) == 0) {
3907 				BUF_CHECK_MAPPED(bp);
3908 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3909 				    bp->b_pages, bp->b_npages);
3910 			} else
3911 				BUF_CHECK_UNMAPPED(bp);
3912 		}
3913 		vm_object_pip_subtract(obj, 1);
3914 		vm_page_io_finish(m);
3915 	}
3916 	vm_object_pip_wakeupn(obj, 0);
3917 	VM_OBJECT_WUNLOCK(obj);
3918 }
3919 
3920 /*
3921  * vfs_page_set_valid:
3922  *
3923  *	Set the valid bits in a page based on the supplied offset.   The
3924  *	range is restricted to the buffer's size.
3925  *
3926  *	This routine is typically called after a read completes.
3927  */
3928 static void
3929 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3930 {
3931 	vm_ooffset_t eoff;
3932 
3933 	/*
3934 	 * Compute the end offset, eoff, such that [off, eoff) does not span a
3935 	 * page boundary and eoff is not greater than the end of the buffer.
3936 	 * The end of the buffer, in this case, is our file EOF, not the
3937 	 * allocation size of the buffer.
3938 	 */
3939 	eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3940 	if (eoff > bp->b_offset + bp->b_bcount)
3941 		eoff = bp->b_offset + bp->b_bcount;
3942 
3943 	/*
3944 	 * Set valid range.  This is typically the entire buffer and thus the
3945 	 * entire page.
3946 	 */
3947 	if (eoff > off)
3948 		vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3949 }
3950 
3951 /*
3952  * vfs_page_set_validclean:
3953  *
3954  *	Set the valid bits and clear the dirty bits in a page based on the
3955  *	supplied offset.   The range is restricted to the buffer's size.
3956  */
3957 static void
3958 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3959 {
3960 	vm_ooffset_t soff, eoff;
3961 
3962 	/*
3963 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3964 	 * page boundry or cross the end of the buffer.  The end of the
3965 	 * buffer, in this case, is our file EOF, not the allocation size
3966 	 * of the buffer.
3967 	 */
3968 	soff = off;
3969 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3970 	if (eoff > bp->b_offset + bp->b_bcount)
3971 		eoff = bp->b_offset + bp->b_bcount;
3972 
3973 	/*
3974 	 * Set valid range.  This is typically the entire buffer and thus the
3975 	 * entire page.
3976 	 */
3977 	if (eoff > soff) {
3978 		vm_page_set_validclean(
3979 		    m,
3980 		   (vm_offset_t) (soff & PAGE_MASK),
3981 		   (vm_offset_t) (eoff - soff)
3982 		);
3983 	}
3984 }
3985 
3986 /*
3987  * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3988  * any page is busy, drain the flag.
3989  */
3990 static void
3991 vfs_drain_busy_pages(struct buf *bp)
3992 {
3993 	vm_page_t m;
3994 	int i, last_busied;
3995 
3996 	VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
3997 	last_busied = 0;
3998 	for (i = 0; i < bp->b_npages; i++) {
3999 		m = bp->b_pages[i];
4000 		if ((m->oflags & VPO_BUSY) != 0) {
4001 			for (; last_busied < i; last_busied++)
4002 				vm_page_busy(bp->b_pages[last_busied]);
4003 			while ((m->oflags & VPO_BUSY) != 0)
4004 				vm_page_sleep(m, "vbpage");
4005 		}
4006 	}
4007 	for (i = 0; i < last_busied; i++)
4008 		vm_page_wakeup(bp->b_pages[i]);
4009 }
4010 
4011 /*
4012  * This routine is called before a device strategy routine.
4013  * It is used to tell the VM system that paging I/O is in
4014  * progress, and treat the pages associated with the buffer
4015  * almost as being VPO_BUSY.  Also the object paging_in_progress
4016  * flag is handled to make sure that the object doesn't become
4017  * inconsistant.
4018  *
4019  * Since I/O has not been initiated yet, certain buffer flags
4020  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4021  * and should be ignored.
4022  */
4023 void
4024 vfs_busy_pages(struct buf *bp, int clear_modify)
4025 {
4026 	int i, bogus;
4027 	vm_object_t obj;
4028 	vm_ooffset_t foff;
4029 	vm_page_t m;
4030 
4031 	if (!(bp->b_flags & B_VMIO))
4032 		return;
4033 
4034 	obj = bp->b_bufobj->bo_object;
4035 	foff = bp->b_offset;
4036 	KASSERT(bp->b_offset != NOOFFSET,
4037 	    ("vfs_busy_pages: no buffer offset"));
4038 	VM_OBJECT_WLOCK(obj);
4039 	vfs_drain_busy_pages(bp);
4040 	if (bp->b_bufsize != 0)
4041 		vfs_setdirty_locked_object(bp);
4042 	bogus = 0;
4043 	for (i = 0; i < bp->b_npages; i++) {
4044 		m = bp->b_pages[i];
4045 
4046 		if ((bp->b_flags & B_CLUSTER) == 0) {
4047 			vm_object_pip_add(obj, 1);
4048 			vm_page_io_start(m);
4049 		}
4050 		/*
4051 		 * When readying a buffer for a read ( i.e
4052 		 * clear_modify == 0 ), it is important to do
4053 		 * bogus_page replacement for valid pages in
4054 		 * partially instantiated buffers.  Partially
4055 		 * instantiated buffers can, in turn, occur when
4056 		 * reconstituting a buffer from its VM backing store
4057 		 * base.  We only have to do this if B_CACHE is
4058 		 * clear ( which causes the I/O to occur in the
4059 		 * first place ).  The replacement prevents the read
4060 		 * I/O from overwriting potentially dirty VM-backed
4061 		 * pages.  XXX bogus page replacement is, uh, bogus.
4062 		 * It may not work properly with small-block devices.
4063 		 * We need to find a better way.
4064 		 */
4065 		if (clear_modify) {
4066 			pmap_remove_write(m);
4067 			vfs_page_set_validclean(bp, foff, m);
4068 		} else if (m->valid == VM_PAGE_BITS_ALL &&
4069 		    (bp->b_flags & B_CACHE) == 0) {
4070 			bp->b_pages[i] = bogus_page;
4071 			bogus++;
4072 		}
4073 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4074 	}
4075 	VM_OBJECT_WUNLOCK(obj);
4076 	if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4077 		BUF_CHECK_MAPPED(bp);
4078 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4079 		    bp->b_pages, bp->b_npages);
4080 	}
4081 }
4082 
4083 /*
4084  *	vfs_bio_set_valid:
4085  *
4086  *	Set the range within the buffer to valid.  The range is
4087  *	relative to the beginning of the buffer, b_offset.  Note that
4088  *	b_offset itself may be offset from the beginning of the first
4089  *	page.
4090  */
4091 void
4092 vfs_bio_set_valid(struct buf *bp, int base, int size)
4093 {
4094 	int i, n;
4095 	vm_page_t m;
4096 
4097 	if (!(bp->b_flags & B_VMIO))
4098 		return;
4099 
4100 	/*
4101 	 * Fixup base to be relative to beginning of first page.
4102 	 * Set initial n to be the maximum number of bytes in the
4103 	 * first page that can be validated.
4104 	 */
4105 	base += (bp->b_offset & PAGE_MASK);
4106 	n = PAGE_SIZE - (base & PAGE_MASK);
4107 
4108 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4109 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4110 		m = bp->b_pages[i];
4111 		if (n > size)
4112 			n = size;
4113 		vm_page_set_valid_range(m, base & PAGE_MASK, n);
4114 		base += n;
4115 		size -= n;
4116 		n = PAGE_SIZE;
4117 	}
4118 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4119 }
4120 
4121 /*
4122  *	vfs_bio_clrbuf:
4123  *
4124  *	If the specified buffer is a non-VMIO buffer, clear the entire
4125  *	buffer.  If the specified buffer is a VMIO buffer, clear and
4126  *	validate only the previously invalid portions of the buffer.
4127  *	This routine essentially fakes an I/O, so we need to clear
4128  *	BIO_ERROR and B_INVAL.
4129  *
4130  *	Note that while we only theoretically need to clear through b_bcount,
4131  *	we go ahead and clear through b_bufsize.
4132  */
4133 void
4134 vfs_bio_clrbuf(struct buf *bp)
4135 {
4136 	int i, j, mask, sa, ea, slide;
4137 
4138 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4139 		clrbuf(bp);
4140 		return;
4141 	}
4142 	bp->b_flags &= ~B_INVAL;
4143 	bp->b_ioflags &= ~BIO_ERROR;
4144 	VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4145 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4146 	    (bp->b_offset & PAGE_MASK) == 0) {
4147 		if (bp->b_pages[0] == bogus_page)
4148 			goto unlock;
4149 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4150 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4151 		if ((bp->b_pages[0]->valid & mask) == mask)
4152 			goto unlock;
4153 		if ((bp->b_pages[0]->valid & mask) == 0) {
4154 			pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4155 			bp->b_pages[0]->valid |= mask;
4156 			goto unlock;
4157 		}
4158 	}
4159 	sa = bp->b_offset & PAGE_MASK;
4160 	slide = 0;
4161 	for (i = 0; i < bp->b_npages; i++, sa = 0) {
4162 		slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4163 		ea = slide & PAGE_MASK;
4164 		if (ea == 0)
4165 			ea = PAGE_SIZE;
4166 		if (bp->b_pages[i] == bogus_page)
4167 			continue;
4168 		j = sa / DEV_BSIZE;
4169 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4170 		VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4171 		if ((bp->b_pages[i]->valid & mask) == mask)
4172 			continue;
4173 		if ((bp->b_pages[i]->valid & mask) == 0)
4174 			pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4175 		else {
4176 			for (; sa < ea; sa += DEV_BSIZE, j++) {
4177 				if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4178 					pmap_zero_page_area(bp->b_pages[i],
4179 					    sa, DEV_BSIZE);
4180 				}
4181 			}
4182 		}
4183 		bp->b_pages[i]->valid |= mask;
4184 	}
4185 unlock:
4186 	VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4187 	bp->b_resid = 0;
4188 }
4189 
4190 void
4191 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4192 {
4193 	vm_page_t m;
4194 	int i, n;
4195 
4196 	if ((bp->b_flags & B_UNMAPPED) == 0) {
4197 		BUF_CHECK_MAPPED(bp);
4198 		bzero(bp->b_data + base, size);
4199 	} else {
4200 		BUF_CHECK_UNMAPPED(bp);
4201 		n = PAGE_SIZE - (base & PAGE_MASK);
4202 		for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4203 			m = bp->b_pages[i];
4204 			if (n > size)
4205 				n = size;
4206 			pmap_zero_page_area(m, base & PAGE_MASK, n);
4207 			base += n;
4208 			size -= n;
4209 			n = PAGE_SIZE;
4210 		}
4211 	}
4212 }
4213 
4214 /*
4215  * vm_hold_load_pages and vm_hold_free_pages get pages into
4216  * a buffers address space.  The pages are anonymous and are
4217  * not associated with a file object.
4218  */
4219 static void
4220 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4221 {
4222 	vm_offset_t pg;
4223 	vm_page_t p;
4224 	int index;
4225 
4226 	BUF_CHECK_MAPPED(bp);
4227 
4228 	to = round_page(to);
4229 	from = round_page(from);
4230 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4231 
4232 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4233 tryagain:
4234 		/*
4235 		 * note: must allocate system pages since blocking here
4236 		 * could interfere with paging I/O, no matter which
4237 		 * process we are.
4238 		 */
4239 		p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4240 		    VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4241 		if (p == NULL) {
4242 			VM_WAIT;
4243 			goto tryagain;
4244 		}
4245 		pmap_qenter(pg, &p, 1);
4246 		bp->b_pages[index] = p;
4247 	}
4248 	bp->b_npages = index;
4249 }
4250 
4251 /* Return pages associated with this buf to the vm system */
4252 static void
4253 vm_hold_free_pages(struct buf *bp, int newbsize)
4254 {
4255 	vm_offset_t from;
4256 	vm_page_t p;
4257 	int index, newnpages;
4258 
4259 	BUF_CHECK_MAPPED(bp);
4260 
4261 	from = round_page((vm_offset_t)bp->b_data + newbsize);
4262 	newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4263 	if (bp->b_npages > newnpages)
4264 		pmap_qremove(from, bp->b_npages - newnpages);
4265 	for (index = newnpages; index < bp->b_npages; index++) {
4266 		p = bp->b_pages[index];
4267 		bp->b_pages[index] = NULL;
4268 		if (p->busy != 0)
4269 			printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4270 			    (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4271 		p->wire_count--;
4272 		vm_page_free(p);
4273 		atomic_subtract_int(&cnt.v_wire_count, 1);
4274 	}
4275 	bp->b_npages = newnpages;
4276 }
4277 
4278 /*
4279  * Map an IO request into kernel virtual address space.
4280  *
4281  * All requests are (re)mapped into kernel VA space.
4282  * Notice that we use b_bufsize for the size of the buffer
4283  * to be mapped.  b_bcount might be modified by the driver.
4284  *
4285  * Note that even if the caller determines that the address space should
4286  * be valid, a race or a smaller-file mapped into a larger space may
4287  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4288  * check the return value.
4289  */
4290 int
4291 vmapbuf(struct buf *bp, int mapbuf)
4292 {
4293 	caddr_t kva;
4294 	vm_prot_t prot;
4295 	int pidx;
4296 
4297 	if (bp->b_bufsize < 0)
4298 		return (-1);
4299 	prot = VM_PROT_READ;
4300 	if (bp->b_iocmd == BIO_READ)
4301 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
4302 	if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4303 	    (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4304 	    btoc(MAXPHYS))) < 0)
4305 		return (-1);
4306 	bp->b_npages = pidx;
4307 	if (mapbuf || !unmapped_buf_allowed) {
4308 		pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4309 		kva = bp->b_saveaddr;
4310 		bp->b_saveaddr = bp->b_data;
4311 		bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4312 		bp->b_flags &= ~B_UNMAPPED;
4313 	} else {
4314 		bp->b_flags |= B_UNMAPPED;
4315 		bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4316 		bp->b_saveaddr = bp->b_data;
4317 		bp->b_data = unmapped_buf;
4318 	}
4319 	return(0);
4320 }
4321 
4322 /*
4323  * Free the io map PTEs associated with this IO operation.
4324  * We also invalidate the TLB entries and restore the original b_addr.
4325  */
4326 void
4327 vunmapbuf(struct buf *bp)
4328 {
4329 	int npages;
4330 
4331 	npages = bp->b_npages;
4332 	if (bp->b_flags & B_UNMAPPED)
4333 		bp->b_flags &= ~B_UNMAPPED;
4334 	else
4335 		pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4336 	vm_page_unhold_pages(bp->b_pages, npages);
4337 
4338 	bp->b_data = bp->b_saveaddr;
4339 }
4340 
4341 void
4342 bdone(struct buf *bp)
4343 {
4344 	struct mtx *mtxp;
4345 
4346 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4347 	mtx_lock(mtxp);
4348 	bp->b_flags |= B_DONE;
4349 	wakeup(bp);
4350 	mtx_unlock(mtxp);
4351 }
4352 
4353 void
4354 bwait(struct buf *bp, u_char pri, const char *wchan)
4355 {
4356 	struct mtx *mtxp;
4357 
4358 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4359 	mtx_lock(mtxp);
4360 	while ((bp->b_flags & B_DONE) == 0)
4361 		msleep(bp, mtxp, pri, wchan, 0);
4362 	mtx_unlock(mtxp);
4363 }
4364 
4365 int
4366 bufsync(struct bufobj *bo, int waitfor)
4367 {
4368 
4369 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4370 }
4371 
4372 void
4373 bufstrategy(struct bufobj *bo, struct buf *bp)
4374 {
4375 	int i = 0;
4376 	struct vnode *vp;
4377 
4378 	vp = bp->b_vp;
4379 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4380 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4381 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4382 	i = VOP_STRATEGY(vp, bp);
4383 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4384 }
4385 
4386 void
4387 bufobj_wrefl(struct bufobj *bo)
4388 {
4389 
4390 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4391 	ASSERT_BO_WLOCKED(bo);
4392 	bo->bo_numoutput++;
4393 }
4394 
4395 void
4396 bufobj_wref(struct bufobj *bo)
4397 {
4398 
4399 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4400 	BO_LOCK(bo);
4401 	bo->bo_numoutput++;
4402 	BO_UNLOCK(bo);
4403 }
4404 
4405 void
4406 bufobj_wdrop(struct bufobj *bo)
4407 {
4408 
4409 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4410 	BO_LOCK(bo);
4411 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4412 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4413 		bo->bo_flag &= ~BO_WWAIT;
4414 		wakeup(&bo->bo_numoutput);
4415 	}
4416 	BO_UNLOCK(bo);
4417 }
4418 
4419 int
4420 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4421 {
4422 	int error;
4423 
4424 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4425 	ASSERT_BO_WLOCKED(bo);
4426 	error = 0;
4427 	while (bo->bo_numoutput) {
4428 		bo->bo_flag |= BO_WWAIT;
4429 		error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4430 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4431 		if (error)
4432 			break;
4433 	}
4434 	return (error);
4435 }
4436 
4437 void
4438 bpin(struct buf *bp)
4439 {
4440 	struct mtx *mtxp;
4441 
4442 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4443 	mtx_lock(mtxp);
4444 	bp->b_pin_count++;
4445 	mtx_unlock(mtxp);
4446 }
4447 
4448 void
4449 bunpin(struct buf *bp)
4450 {
4451 	struct mtx *mtxp;
4452 
4453 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4454 	mtx_lock(mtxp);
4455 	if (--bp->b_pin_count == 0)
4456 		wakeup(bp);
4457 	mtx_unlock(mtxp);
4458 }
4459 
4460 void
4461 bunpin_wait(struct buf *bp)
4462 {
4463 	struct mtx *mtxp;
4464 
4465 	mtxp = mtx_pool_find(mtxpool_sleep, bp);
4466 	mtx_lock(mtxp);
4467 	while (bp->b_pin_count > 0)
4468 		msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4469 	mtx_unlock(mtxp);
4470 }
4471 
4472 /*
4473  * Set bio_data or bio_ma for struct bio from the struct buf.
4474  */
4475 void
4476 bdata2bio(struct buf *bp, struct bio *bip)
4477 {
4478 
4479 	if ((bp->b_flags & B_UNMAPPED) != 0) {
4480 		KASSERT(unmapped_buf_allowed, ("unmapped"));
4481 		bip->bio_ma = bp->b_pages;
4482 		bip->bio_ma_n = bp->b_npages;
4483 		bip->bio_data = unmapped_buf;
4484 		bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4485 		bip->bio_flags |= BIO_UNMAPPED;
4486 		KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4487 		    PAGE_SIZE == bp->b_npages,
4488 		    ("Buffer %p too short: %d %d %d", bp, bip->bio_ma_offset,
4489 		    bip->bio_length, bip->bio_ma_n));
4490 	} else {
4491 		bip->bio_data = bp->b_data;
4492 		bip->bio_ma = NULL;
4493 	}
4494 }
4495 
4496 #include "opt_ddb.h"
4497 #ifdef DDB
4498 #include <ddb/ddb.h>
4499 
4500 /* DDB command to show buffer data */
4501 DB_SHOW_COMMAND(buffer, db_show_buffer)
4502 {
4503 	/* get args */
4504 	struct buf *bp = (struct buf *)addr;
4505 
4506 	if (!have_addr) {
4507 		db_printf("usage: show buffer <addr>\n");
4508 		return;
4509 	}
4510 
4511 	db_printf("buf at %p\n", bp);
4512 	db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4513 	    (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4514 	    PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4515 	db_printf(
4516 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4517 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4518 	    "b_dep = %p\n",
4519 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4520 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4521 	    (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4522 	if (bp->b_npages) {
4523 		int i;
4524 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4525 		for (i = 0; i < bp->b_npages; i++) {
4526 			vm_page_t m;
4527 			m = bp->b_pages[i];
4528 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4529 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4530 			if ((i + 1) < bp->b_npages)
4531 				db_printf(",");
4532 		}
4533 		db_printf("\n");
4534 	}
4535 	db_printf(" ");
4536 	BUF_LOCKPRINTINFO(bp);
4537 }
4538 
4539 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4540 {
4541 	struct buf *bp;
4542 	int i;
4543 
4544 	for (i = 0; i < nbuf; i++) {
4545 		bp = &buf[i];
4546 		if (BUF_ISLOCKED(bp)) {
4547 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4548 			db_printf("\n");
4549 		}
4550 	}
4551 }
4552 
4553 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4554 {
4555 	struct vnode *vp;
4556 	struct buf *bp;
4557 
4558 	if (!have_addr) {
4559 		db_printf("usage: show vnodebufs <addr>\n");
4560 		return;
4561 	}
4562 	vp = (struct vnode *)addr;
4563 	db_printf("Clean buffers:\n");
4564 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4565 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4566 		db_printf("\n");
4567 	}
4568 	db_printf("Dirty buffers:\n");
4569 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4570 		db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4571 		db_printf("\n");
4572 	}
4573 }
4574 
4575 DB_COMMAND(countfreebufs, db_coundfreebufs)
4576 {
4577 	struct buf *bp;
4578 	int i, used = 0, nfree = 0;
4579 
4580 	if (have_addr) {
4581 		db_printf("usage: countfreebufs\n");
4582 		return;
4583 	}
4584 
4585 	for (i = 0; i < nbuf; i++) {
4586 		bp = &buf[i];
4587 		if ((bp->b_flags & B_INFREECNT) != 0)
4588 			nfree++;
4589 		else
4590 			used++;
4591 	}
4592 
4593 	db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4594 	    nfree + used);
4595 	db_printf("numfreebuffers is %d\n", numfreebuffers);
4596 }
4597 #endif /* DDB */
4598