xref: /freebsd/sys/kern/vfs_bio.c (revision e4e9813eb92cd7c4d4b819a8fbed5cbd3d92f5d8)
1 /*-
2  * Copyright (c) 2004 Poul-Henning Kamp
3  * Copyright (c) 1994,1997 John S. Dyson
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * this file contains a new buffer I/O scheme implementing a coherent
30  * VM object and buffer cache scheme.  Pains have been taken to make
31  * sure that the performance degradation associated with schemes such
32  * as this is not realized.
33  *
34  * Author:  John S. Dyson
35  * Significant help during the development and debugging phases
36  * had been provided by David Greenman, also of the FreeBSD core team.
37  *
38  * see man buf(9) for more info.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bio.h>
47 #include <sys/conf.h>
48 #include <sys/buf.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
51 #include <sys/lock.h>
52 #include <sys/malloc.h>
53 #include <sys/mount.h>
54 #include <sys/mutex.h>
55 #include <sys/kernel.h>
56 #include <sys/kthread.h>
57 #include <sys/proc.h>
58 #include <sys/resourcevar.h>
59 #include <sys/sysctl.h>
60 #include <sys/vmmeter.h>
61 #include <sys/vnode.h>
62 #include <geom/geom.h>
63 #include <vm/vm.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_pageout.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_extern.h>
70 #include <vm/vm_map.h>
71 #include "opt_directio.h"
72 #include "opt_swap.h"
73 
74 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
75 
76 struct	bio_ops bioops;		/* I/O operation notification */
77 
78 struct	buf_ops buf_ops_bio = {
79 	.bop_name	=	"buf_ops_bio",
80 	.bop_write	=	bufwrite,
81 	.bop_strategy	=	bufstrategy,
82 	.bop_sync	=	bufsync,
83 };
84 
85 /*
86  * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
87  * carnal knowledge of buffers.  This knowledge should be moved to vfs_bio.c.
88  */
89 struct buf *buf;		/* buffer header pool */
90 
91 static struct proc *bufdaemonproc;
92 
93 static int inmem(struct vnode *vp, daddr_t blkno);
94 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from,
95 		vm_offset_t to);
96 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
97 		vm_offset_t to);
98 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
99 			       int pageno, vm_page_t m);
100 static void vfs_clean_pages(struct buf *bp);
101 static void vfs_setdirty(struct buf *bp);
102 static void vfs_vmio_release(struct buf *bp);
103 static int vfs_bio_clcheck(struct vnode *vp, int size,
104 		daddr_t lblkno, daddr_t blkno);
105 static int flushbufqueues(int, int);
106 static void buf_daemon(void);
107 static void bremfreel(struct buf *bp);
108 
109 int vmiodirenable = TRUE;
110 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
111     "Use the VM system for directory writes");
112 int runningbufspace;
113 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
114     "Amount of presently outstanding async buffer io");
115 static int bufspace;
116 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
117     "KVA memory used for bufs");
118 static int maxbufspace;
119 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
120     "Maximum allowed value of bufspace (including buf_daemon)");
121 static int bufmallocspace;
122 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
123     "Amount of malloced memory for buffers");
124 static int maxbufmallocspace;
125 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
126     "Maximum amount of malloced memory for buffers");
127 static int lobufspace;
128 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
129     "Minimum amount of buffers we want to have");
130 int hibufspace;
131 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
132     "Maximum allowed value of bufspace (excluding buf_daemon)");
133 static int bufreusecnt;
134 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
135     "Number of times we have reused a buffer");
136 static int buffreekvacnt;
137 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
138     "Number of times we have freed the KVA space from some buffer");
139 static int bufdefragcnt;
140 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
141     "Number of times we have had to repeat buffer allocation to defragment");
142 static int lorunningspace;
143 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
144     "Minimum preferred space used for in-progress I/O");
145 static int hirunningspace;
146 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
147     "Maximum amount of space to use for in-progress I/O");
148 static int dirtybufferflushes;
149 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
150     0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
151 static int altbufferflushes;
152 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
153     0, "Number of fsync flushes to limit dirty buffers");
154 static int recursiveflushes;
155 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
156     0, "Number of flushes skipped due to being recursive");
157 static int numdirtybuffers;
158 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
159     "Number of buffers that are dirty (has unwritten changes) at the moment");
160 static int lodirtybuffers;
161 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
162     "How many buffers we want to have free before bufdaemon can sleep");
163 static int hidirtybuffers;
164 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
165     "When the number of dirty buffers is considered severe");
166 static int dirtybufthresh;
167 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
168     0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
169 static int numfreebuffers;
170 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
171     "Number of free buffers");
172 static int lofreebuffers;
173 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
174    "XXX Unused");
175 static int hifreebuffers;
176 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
177    "XXX Complicatedly unused");
178 static int getnewbufcalls;
179 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
180    "Number of calls to getnewbuf");
181 static int getnewbufrestarts;
182 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
183     "Number of times getnewbuf has had to restart a buffer aquisition");
184 
185 /*
186  * Wakeup point for bufdaemon, as well as indicator of whether it is already
187  * active.  Set to 1 when the bufdaemon is already "on" the queue, 0 when it
188  * is idling.
189  */
190 static int bd_request;
191 
192 /*
193  * This lock synchronizes access to bd_request.
194  */
195 static struct mtx bdlock;
196 
197 /*
198  * bogus page -- for I/O to/from partially complete buffers
199  * this is a temporary solution to the problem, but it is not
200  * really that bad.  it would be better to split the buffer
201  * for input in the case of buffers partially already in memory,
202  * but the code is intricate enough already.
203  */
204 vm_page_t bogus_page;
205 
206 /*
207  * Synchronization (sleep/wakeup) variable for active buffer space requests.
208  * Set when wait starts, cleared prior to wakeup().
209  * Used in runningbufwakeup() and waitrunningbufspace().
210  */
211 static int runningbufreq;
212 
213 /*
214  * This lock protects the runningbufreq and synchronizes runningbufwakeup and
215  * waitrunningbufspace().
216  */
217 static struct mtx rbreqlock;
218 
219 /*
220  * Synchronization (sleep/wakeup) variable for buffer requests.
221  * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
222  * by and/or.
223  * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
224  * getnewbuf(), and getblk().
225  */
226 static int needsbuffer;
227 
228 /*
229  * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
230  */
231 static struct mtx nblock;
232 
233 /*
234  * Lock that protects against bwait()/bdone()/B_DONE races.
235  */
236 
237 static struct mtx bdonelock;
238 
239 /*
240  * Lock that protects against bwait()/bdone()/B_DONE races.
241  */
242 static struct mtx bpinlock;
243 
244 /*
245  * Definitions for the buffer free lists.
246  */
247 #define BUFFER_QUEUES	6	/* number of free buffer queues */
248 
249 #define QUEUE_NONE	0	/* on no queue */
250 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
251 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
252 #define QUEUE_DIRTY_GIANT 3	/* B_DELWRI buffers that need giant */
253 #define QUEUE_EMPTYKVA	4	/* empty buffer headers w/KVA assignment */
254 #define QUEUE_EMPTY	5	/* empty buffer headers */
255 
256 /* Queues for free buffers with various properties */
257 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
258 
259 /* Lock for the bufqueues */
260 static struct mtx bqlock;
261 
262 /*
263  * Single global constant for BUF_WMESG, to avoid getting multiple references.
264  * buf_wmesg is referred from macros.
265  */
266 const char *buf_wmesg = BUF_WMESG;
267 
268 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
269 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
270 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
271 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
272 
273 #ifdef DIRECTIO
274 extern void ffs_rawread_setup(void);
275 #endif /* DIRECTIO */
276 /*
277  *	numdirtywakeup:
278  *
279  *	If someone is blocked due to there being too many dirty buffers,
280  *	and numdirtybuffers is now reasonable, wake them up.
281  */
282 
283 static __inline void
284 numdirtywakeup(int level)
285 {
286 
287 	if (numdirtybuffers <= level) {
288 		mtx_lock(&nblock);
289 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
290 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
291 			wakeup(&needsbuffer);
292 		}
293 		mtx_unlock(&nblock);
294 	}
295 }
296 
297 /*
298  *	bufspacewakeup:
299  *
300  *	Called when buffer space is potentially available for recovery.
301  *	getnewbuf() will block on this flag when it is unable to free
302  *	sufficient buffer space.  Buffer space becomes recoverable when
303  *	bp's get placed back in the queues.
304  */
305 
306 static __inline void
307 bufspacewakeup(void)
308 {
309 
310 	/*
311 	 * If someone is waiting for BUF space, wake them up.  Even
312 	 * though we haven't freed the kva space yet, the waiting
313 	 * process will be able to now.
314 	 */
315 	mtx_lock(&nblock);
316 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
317 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
318 		wakeup(&needsbuffer);
319 	}
320 	mtx_unlock(&nblock);
321 }
322 
323 /*
324  * runningbufwakeup() - in-progress I/O accounting.
325  *
326  */
327 void
328 runningbufwakeup(struct buf *bp)
329 {
330 
331 	if (bp->b_runningbufspace) {
332 		atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
333 		bp->b_runningbufspace = 0;
334 		mtx_lock(&rbreqlock);
335 		if (runningbufreq && runningbufspace <= lorunningspace) {
336 			runningbufreq = 0;
337 			wakeup(&runningbufreq);
338 		}
339 		mtx_unlock(&rbreqlock);
340 	}
341 }
342 
343 /*
344  *	bufcountwakeup:
345  *
346  *	Called when a buffer has been added to one of the free queues to
347  *	account for the buffer and to wakeup anyone waiting for free buffers.
348  *	This typically occurs when large amounts of metadata are being handled
349  *	by the buffer cache ( else buffer space runs out first, usually ).
350  */
351 
352 static __inline void
353 bufcountwakeup(void)
354 {
355 
356 	atomic_add_int(&numfreebuffers, 1);
357 	mtx_lock(&nblock);
358 	if (needsbuffer) {
359 		needsbuffer &= ~VFS_BIO_NEED_ANY;
360 		if (numfreebuffers >= hifreebuffers)
361 			needsbuffer &= ~VFS_BIO_NEED_FREE;
362 		wakeup(&needsbuffer);
363 	}
364 	mtx_unlock(&nblock);
365 }
366 
367 /*
368  *	waitrunningbufspace()
369  *
370  *	runningbufspace is a measure of the amount of I/O currently
371  *	running.  This routine is used in async-write situations to
372  *	prevent creating huge backups of pending writes to a device.
373  *	Only asynchronous writes are governed by this function.
374  *
375  *	Reads will adjust runningbufspace, but will not block based on it.
376  *	The read load has a side effect of reducing the allowed write load.
377  *
378  *	This does NOT turn an async write into a sync write.  It waits
379  *	for earlier writes to complete and generally returns before the
380  *	caller's write has reached the device.
381  */
382 void
383 waitrunningbufspace(void)
384 {
385 
386 	mtx_lock(&rbreqlock);
387 	while (runningbufspace > hirunningspace) {
388 		++runningbufreq;
389 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
390 	}
391 	mtx_unlock(&rbreqlock);
392 }
393 
394 
395 /*
396  *	vfs_buf_test_cache:
397  *
398  *	Called when a buffer is extended.  This function clears the B_CACHE
399  *	bit if the newly extended portion of the buffer does not contain
400  *	valid data.
401  */
402 static __inline
403 void
404 vfs_buf_test_cache(struct buf *bp,
405 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
406 		  vm_page_t m)
407 {
408 
409 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
410 	if (bp->b_flags & B_CACHE) {
411 		int base = (foff + off) & PAGE_MASK;
412 		if (vm_page_is_valid(m, base, size) == 0)
413 			bp->b_flags &= ~B_CACHE;
414 	}
415 }
416 
417 /* Wake up the buffer deamon if necessary */
418 static __inline
419 void
420 bd_wakeup(int dirtybuflevel)
421 {
422 
423 	mtx_lock(&bdlock);
424 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
425 		bd_request = 1;
426 		wakeup(&bd_request);
427 	}
428 	mtx_unlock(&bdlock);
429 }
430 
431 /*
432  * bd_speedup - speedup the buffer cache flushing code
433  */
434 
435 static __inline
436 void
437 bd_speedup(void)
438 {
439 
440 	bd_wakeup(1);
441 }
442 
443 /*
444  * Calculating buffer cache scaling values and reserve space for buffer
445  * headers.  This is called during low level kernel initialization and
446  * may be called more then once.  We CANNOT write to the memory area
447  * being reserved at this time.
448  */
449 caddr_t
450 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
451 {
452 
453 	/*
454 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
455 	 * PAGE_SIZE is >= 1K)
456 	 */
457 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
458 
459 	/*
460 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
461 	 * For the first 64MB of ram nominally allocate sufficient buffers to
462 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
463 	 * buffers to cover 1/20 of our ram over 64MB.  When auto-sizing
464 	 * the buffer cache we limit the eventual kva reservation to
465 	 * maxbcache bytes.
466 	 *
467 	 * factor represents the 1/4 x ram conversion.
468 	 */
469 	if (nbuf == 0) {
470 		int factor = 4 * BKVASIZE / 1024;
471 
472 		nbuf = 50;
473 		if (physmem_est > 4096)
474 			nbuf += min((physmem_est - 4096) / factor,
475 			    65536 / factor);
476 		if (physmem_est > 65536)
477 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
478 
479 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
480 			nbuf = maxbcache / BKVASIZE;
481 	}
482 
483 #if 0
484 	/*
485 	 * Do not allow the buffer_map to be more then 1/2 the size of the
486 	 * kernel_map.
487 	 */
488 	if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
489 	    (BKVASIZE * 2)) {
490 		nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
491 		    (BKVASIZE * 2);
492 		printf("Warning: nbufs capped at %d\n", nbuf);
493 	}
494 #endif
495 
496 	/*
497 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
498 	 * We have no less then 16 and no more then 256.
499 	 */
500 	nswbuf = max(min(nbuf/4, 256), 16);
501 #ifdef NSWBUF_MIN
502 	if (nswbuf < NSWBUF_MIN)
503 		nswbuf = NSWBUF_MIN;
504 #endif
505 #ifdef DIRECTIO
506 	ffs_rawread_setup();
507 #endif
508 
509 	/*
510 	 * Reserve space for the buffer cache buffers
511 	 */
512 	swbuf = (void *)v;
513 	v = (caddr_t)(swbuf + nswbuf);
514 	buf = (void *)v;
515 	v = (caddr_t)(buf + nbuf);
516 
517 	return(v);
518 }
519 
520 /* Initialize the buffer subsystem.  Called before use of any buffers. */
521 void
522 bufinit(void)
523 {
524 	struct buf *bp;
525 	int i;
526 
527 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
528 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
529 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
530 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
531 	mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
532 	mtx_init(&bpinlock, "bpin lock", NULL, MTX_DEF);
533 
534 	/* next, make a null set of free lists */
535 	for (i = 0; i < BUFFER_QUEUES; i++)
536 		TAILQ_INIT(&bufqueues[i]);
537 
538 	/* finally, initialize each buffer header and stick on empty q */
539 	for (i = 0; i < nbuf; i++) {
540 		bp = &buf[i];
541 		bzero(bp, sizeof *bp);
542 		bp->b_flags = B_INVAL;	/* we're just an empty header */
543 		bp->b_rcred = NOCRED;
544 		bp->b_wcred = NOCRED;
545 		bp->b_qindex = QUEUE_EMPTY;
546 		bp->b_vflags = 0;
547 		bp->b_xflags = 0;
548 		LIST_INIT(&bp->b_dep);
549 		BUF_LOCKINIT(bp);
550 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
551 	}
552 
553 	/*
554 	 * maxbufspace is the absolute maximum amount of buffer space we are
555 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
556 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
557 	 * used by most other processes.  The differential is required to
558 	 * ensure that buf_daemon is able to run when other processes might
559 	 * be blocked waiting for buffer space.
560 	 *
561 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
562 	 * this may result in KVM fragmentation which is not handled optimally
563 	 * by the system.
564 	 */
565 	maxbufspace = nbuf * BKVASIZE;
566 	hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
567 	lobufspace = hibufspace - MAXBSIZE;
568 
569 	lorunningspace = 512 * 1024;
570 	hirunningspace = 1024 * 1024;
571 
572 /*
573  * Limit the amount of malloc memory since it is wired permanently into
574  * the kernel space.  Even though this is accounted for in the buffer
575  * allocation, we don't want the malloced region to grow uncontrolled.
576  * The malloc scheme improves memory utilization significantly on average
577  * (small) directories.
578  */
579 	maxbufmallocspace = hibufspace / 20;
580 
581 /*
582  * Reduce the chance of a deadlock occuring by limiting the number
583  * of delayed-write dirty buffers we allow to stack up.
584  */
585 	hidirtybuffers = nbuf / 4 + 20;
586 	dirtybufthresh = hidirtybuffers * 9 / 10;
587 	numdirtybuffers = 0;
588 /*
589  * To support extreme low-memory systems, make sure hidirtybuffers cannot
590  * eat up all available buffer space.  This occurs when our minimum cannot
591  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
592  * BKVASIZE'd (8K) buffers.
593  */
594 	while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
595 		hidirtybuffers >>= 1;
596 	}
597 	lodirtybuffers = hidirtybuffers / 2;
598 
599 /*
600  * Try to keep the number of free buffers in the specified range,
601  * and give special processes (e.g. like buf_daemon) access to an
602  * emergency reserve.
603  */
604 	lofreebuffers = nbuf / 18 + 5;
605 	hifreebuffers = 2 * lofreebuffers;
606 	numfreebuffers = nbuf;
607 
608 /*
609  * Maximum number of async ops initiated per buf_daemon loop.  This is
610  * somewhat of a hack at the moment, we really need to limit ourselves
611  * based on the number of bytes of I/O in-transit that were initiated
612  * from buf_daemon.
613  */
614 
615 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
616 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
617 }
618 
619 /*
620  * bfreekva() - free the kva allocation for a buffer.
621  *
622  *	Since this call frees up buffer space, we call bufspacewakeup().
623  */
624 static void
625 bfreekva(struct buf *bp)
626 {
627 
628 	if (bp->b_kvasize) {
629 		atomic_add_int(&buffreekvacnt, 1);
630 		atomic_subtract_int(&bufspace, bp->b_kvasize);
631 		vm_map_lock(buffer_map);
632 		vm_map_delete(buffer_map,
633 		    (vm_offset_t) bp->b_kvabase,
634 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize
635 		);
636 		vm_map_unlock(buffer_map);
637 		bp->b_kvasize = 0;
638 		bufspacewakeup();
639 	}
640 }
641 
642 /*
643  *	bremfree:
644  *
645  *	Mark the buffer for removal from the appropriate free list in brelse.
646  *
647  */
648 void
649 bremfree(struct buf *bp)
650 {
651 
652 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
653 	KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked."));
654 	KASSERT((bp->b_flags & B_REMFREE) == 0,
655 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
656 	KASSERT(bp->b_qindex != QUEUE_NONE,
657 	    ("bremfree: buffer %p not on a queue.", bp));
658 
659 	bp->b_flags |= B_REMFREE;
660 	/* Fixup numfreebuffers count.  */
661 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
662 		atomic_subtract_int(&numfreebuffers, 1);
663 }
664 
665 /*
666  *	bremfreef:
667  *
668  *	Force an immediate removal from a free list.  Used only in nfs when
669  *	it abuses the b_freelist pointer.
670  */
671 void
672 bremfreef(struct buf *bp)
673 {
674 	mtx_lock(&bqlock);
675 	bremfreel(bp);
676 	mtx_unlock(&bqlock);
677 }
678 
679 /*
680  *	bremfreel:
681  *
682  *	Removes a buffer from the free list, must be called with the
683  *	bqlock held.
684  */
685 static void
686 bremfreel(struct buf *bp)
687 {
688 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
689 	    bp, bp->b_vp, bp->b_flags);
690 	KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp));
691 	KASSERT(bp->b_qindex != QUEUE_NONE,
692 	    ("bremfreel: buffer %p not on a queue.", bp));
693 	mtx_assert(&bqlock, MA_OWNED);
694 
695 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
696 	bp->b_qindex = QUEUE_NONE;
697 	/*
698 	 * If this was a delayed bremfree() we only need to remove the buffer
699 	 * from the queue and return the stats are already done.
700 	 */
701 	if (bp->b_flags & B_REMFREE) {
702 		bp->b_flags &= ~B_REMFREE;
703 		return;
704 	}
705 	/*
706 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
707 	 * delayed-write, the buffer was free and we must decrement
708 	 * numfreebuffers.
709 	 */
710 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
711 		atomic_subtract_int(&numfreebuffers, 1);
712 }
713 
714 
715 /*
716  * Get a buffer with the specified data.  Look in the cache first.  We
717  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
718  * is set, the buffer is valid and we do not have to do anything ( see
719  * getblk() ).  This is really just a special case of breadn().
720  */
721 int
722 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
723     struct buf **bpp)
724 {
725 
726 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
727 }
728 
729 /*
730  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
731  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
732  * the buffer is valid and we do not have to do anything.
733  */
734 void
735 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
736     int cnt, struct ucred * cred)
737 {
738 	struct buf *rabp;
739 	int i;
740 
741 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
742 		if (inmem(vp, *rablkno))
743 			continue;
744 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
745 
746 		if ((rabp->b_flags & B_CACHE) == 0) {
747 			if (curthread != PCPU_GET(idlethread))
748 				curthread->td_proc->p_stats->p_ru.ru_inblock++;
749 			rabp->b_flags |= B_ASYNC;
750 			rabp->b_flags &= ~B_INVAL;
751 			rabp->b_ioflags &= ~BIO_ERROR;
752 			rabp->b_iocmd = BIO_READ;
753 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
754 				rabp->b_rcred = crhold(cred);
755 			vfs_busy_pages(rabp, 0);
756 			BUF_KERNPROC(rabp);
757 			rabp->b_iooffset = dbtob(rabp->b_blkno);
758 			bstrategy(rabp);
759 		} else {
760 			brelse(rabp);
761 		}
762 	}
763 }
764 
765 /*
766  * Operates like bread, but also starts asynchronous I/O on
767  * read-ahead blocks.
768  */
769 int
770 breadn(struct vnode * vp, daddr_t blkno, int size,
771     daddr_t * rablkno, int *rabsize,
772     int cnt, struct ucred * cred, struct buf **bpp)
773 {
774 	struct buf *bp;
775 	int rv = 0, readwait = 0;
776 
777 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
778 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
779 
780 	/* if not found in cache, do some I/O */
781 	if ((bp->b_flags & B_CACHE) == 0) {
782 		if (curthread != PCPU_GET(idlethread))
783 			curthread->td_proc->p_stats->p_ru.ru_inblock++;
784 		bp->b_iocmd = BIO_READ;
785 		bp->b_flags &= ~B_INVAL;
786 		bp->b_ioflags &= ~BIO_ERROR;
787 		if (bp->b_rcred == NOCRED && cred != NOCRED)
788 			bp->b_rcred = crhold(cred);
789 		vfs_busy_pages(bp, 0);
790 		bp->b_iooffset = dbtob(bp->b_blkno);
791 		bstrategy(bp);
792 		++readwait;
793 	}
794 
795 	breada(vp, rablkno, rabsize, cnt, cred);
796 
797 	if (readwait) {
798 		rv = bufwait(bp);
799 	}
800 	return (rv);
801 }
802 
803 /*
804  * Write, release buffer on completion.  (Done by iodone
805  * if async).  Do not bother writing anything if the buffer
806  * is invalid.
807  *
808  * Note that we set B_CACHE here, indicating that buffer is
809  * fully valid and thus cacheable.  This is true even of NFS
810  * now so we set it generally.  This could be set either here
811  * or in biodone() since the I/O is synchronous.  We put it
812  * here.
813  */
814 int
815 bufwrite(struct buf *bp)
816 {
817 	int oldflags;
818 
819 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
820 	if (bp->b_flags & B_INVAL) {
821 		brelse(bp);
822 		return (0);
823 	}
824 
825 	oldflags = bp->b_flags;
826 
827 	if (BUF_REFCNT(bp) == 0)
828 		panic("bufwrite: buffer is not busy???");
829 
830 	if (bp->b_pin_count > 0)
831 		bunpin_wait(bp);
832 
833 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
834 	    ("FFS background buffer should not get here %p", bp));
835 
836 	/* Mark the buffer clean */
837 	bundirty(bp);
838 
839 	bp->b_flags &= ~B_DONE;
840 	bp->b_ioflags &= ~BIO_ERROR;
841 	bp->b_flags |= B_CACHE;
842 	bp->b_iocmd = BIO_WRITE;
843 
844 	bufobj_wref(bp->b_bufobj);
845 	vfs_busy_pages(bp, 1);
846 
847 	/*
848 	 * Normal bwrites pipeline writes
849 	 */
850 	bp->b_runningbufspace = bp->b_bufsize;
851 	atomic_add_int(&runningbufspace, bp->b_runningbufspace);
852 
853 	if (curthread != PCPU_GET(idlethread))
854 		curthread->td_proc->p_stats->p_ru.ru_oublock++;
855 	if (oldflags & B_ASYNC)
856 		BUF_KERNPROC(bp);
857 	bp->b_iooffset = dbtob(bp->b_blkno);
858 	bstrategy(bp);
859 
860 	if ((oldflags & B_ASYNC) == 0) {
861 		int rtval = bufwait(bp);
862 		brelse(bp);
863 		return (rtval);
864 	} else {
865 		/*
866 		 * don't allow the async write to saturate the I/O
867 		 * system.  We will not deadlock here because
868 		 * we are blocking waiting for I/O that is already in-progress
869 		 * to complete. We do not block here if it is the update
870 		 * or syncer daemon trying to clean up as that can lead
871 		 * to deadlock.
872 		 */
873 		if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0)
874 			waitrunningbufspace();
875 	}
876 
877 	return (0);
878 }
879 
880 /*
881  * Delayed write. (Buffer is marked dirty).  Do not bother writing
882  * anything if the buffer is marked invalid.
883  *
884  * Note that since the buffer must be completely valid, we can safely
885  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
886  * biodone() in order to prevent getblk from writing the buffer
887  * out synchronously.
888  */
889 void
890 bdwrite(struct buf *bp)
891 {
892 	struct thread *td = curthread;
893 	struct vnode *vp;
894 	struct buf *nbp;
895 	struct bufobj *bo;
896 
897 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
898 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
899 	KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy"));
900 
901 	if (bp->b_flags & B_INVAL) {
902 		brelse(bp);
903 		return;
904 	}
905 
906 	/*
907 	 * If we have too many dirty buffers, don't create any more.
908 	 * If we are wildly over our limit, then force a complete
909 	 * cleanup. Otherwise, just keep the situation from getting
910 	 * out of control. Note that we have to avoid a recursive
911 	 * disaster and not try to clean up after our own cleanup!
912 	 */
913 	vp = bp->b_vp;
914 	bo = bp->b_bufobj;
915 	if ((td->td_pflags & TDP_COWINPROGRESS) == 0) {
916 		BO_LOCK(bo);
917 		if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
918 			BO_UNLOCK(bo);
919 			(void) VOP_FSYNC(vp, MNT_NOWAIT, td);
920 			altbufferflushes++;
921 		} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
922 			/*
923 			 * Try to find a buffer to flush.
924 			 */
925 			TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
926 				if ((nbp->b_vflags & BV_BKGRDINPROG) ||
927 				    BUF_LOCK(nbp,
928 				    LK_EXCLUSIVE | LK_NOWAIT, NULL))
929 					continue;
930 				if (bp == nbp)
931 					panic("bdwrite: found ourselves");
932 				BO_UNLOCK(bo);
933 				/* Don't countdeps with the bo lock held. */
934 				if (buf_countdeps(nbp, 0)) {
935 					BO_LOCK(bo);
936 					BUF_UNLOCK(nbp);
937 					continue;
938 				}
939 				if (nbp->b_flags & B_CLUSTEROK) {
940 					vfs_bio_awrite(nbp);
941 				} else {
942 					bremfree(nbp);
943 					bawrite(nbp);
944 				}
945 				dirtybufferflushes++;
946 				break;
947 			}
948 			if (nbp == NULL)
949 				BO_UNLOCK(bo);
950 		} else
951 			BO_UNLOCK(bo);
952 	} else
953 		recursiveflushes++;
954 
955 	bdirty(bp);
956 	/*
957 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
958 	 * true even of NFS now.
959 	 */
960 	bp->b_flags |= B_CACHE;
961 
962 	/*
963 	 * This bmap keeps the system from needing to do the bmap later,
964 	 * perhaps when the system is attempting to do a sync.  Since it
965 	 * is likely that the indirect block -- or whatever other datastructure
966 	 * that the filesystem needs is still in memory now, it is a good
967 	 * thing to do this.  Note also, that if the pageout daemon is
968 	 * requesting a sync -- there might not be enough memory to do
969 	 * the bmap then...  So, this is important to do.
970 	 */
971 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
972 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
973 	}
974 
975 	/*
976 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
977 	 */
978 	vfs_setdirty(bp);
979 
980 	/*
981 	 * We need to do this here to satisfy the vnode_pager and the
982 	 * pageout daemon, so that it thinks that the pages have been
983 	 * "cleaned".  Note that since the pages are in a delayed write
984 	 * buffer -- the VFS layer "will" see that the pages get written
985 	 * out on the next sync, or perhaps the cluster will be completed.
986 	 */
987 	vfs_clean_pages(bp);
988 	bqrelse(bp);
989 
990 	/*
991 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
992 	 * buffers (midpoint between our recovery point and our stall
993 	 * point).
994 	 */
995 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
996 
997 	/*
998 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
999 	 * due to the softdep code.
1000 	 */
1001 }
1002 
1003 /*
1004  *	bdirty:
1005  *
1006  *	Turn buffer into delayed write request.  We must clear BIO_READ and
1007  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
1008  *	itself to properly update it in the dirty/clean lists.  We mark it
1009  *	B_DONE to ensure that any asynchronization of the buffer properly
1010  *	clears B_DONE ( else a panic will occur later ).
1011  *
1012  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1013  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
1014  *	should only be called if the buffer is known-good.
1015  *
1016  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1017  *	count.
1018  *
1019  *	The buffer must be on QUEUE_NONE.
1020  */
1021 void
1022 bdirty(struct buf *bp)
1023 {
1024 
1025 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1026 	    bp, bp->b_vp, bp->b_flags);
1027 	KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp));
1028 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1029 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1030 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1031 	bp->b_flags &= ~(B_RELBUF);
1032 	bp->b_iocmd = BIO_WRITE;
1033 
1034 	if ((bp->b_flags & B_DELWRI) == 0) {
1035 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1036 		reassignbuf(bp);
1037 		atomic_add_int(&numdirtybuffers, 1);
1038 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1039 	}
1040 }
1041 
1042 /*
1043  *	bundirty:
1044  *
1045  *	Clear B_DELWRI for buffer.
1046  *
1047  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1048  *	count.
1049  *
1050  *	The buffer must be on QUEUE_NONE.
1051  */
1052 
1053 void
1054 bundirty(struct buf *bp)
1055 {
1056 
1057 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1058 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1059 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1060 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1061 	KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp));
1062 
1063 	if (bp->b_flags & B_DELWRI) {
1064 		bp->b_flags &= ~B_DELWRI;
1065 		reassignbuf(bp);
1066 		atomic_subtract_int(&numdirtybuffers, 1);
1067 		numdirtywakeup(lodirtybuffers);
1068 	}
1069 	/*
1070 	 * Since it is now being written, we can clear its deferred write flag.
1071 	 */
1072 	bp->b_flags &= ~B_DEFERRED;
1073 }
1074 
1075 /*
1076  *	bawrite:
1077  *
1078  *	Asynchronous write.  Start output on a buffer, but do not wait for
1079  *	it to complete.  The buffer is released when the output completes.
1080  *
1081  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1082  *	B_INVAL buffers.  Not us.
1083  */
1084 void
1085 bawrite(struct buf *bp)
1086 {
1087 
1088 	bp->b_flags |= B_ASYNC;
1089 	(void) bwrite(bp);
1090 }
1091 
1092 /*
1093  *	bwillwrite:
1094  *
1095  *	Called prior to the locking of any vnodes when we are expecting to
1096  *	write.  We do not want to starve the buffer cache with too many
1097  *	dirty buffers so we block here.  By blocking prior to the locking
1098  *	of any vnodes we attempt to avoid the situation where a locked vnode
1099  *	prevents the various system daemons from flushing related buffers.
1100  */
1101 
1102 void
1103 bwillwrite(void)
1104 {
1105 
1106 	if (numdirtybuffers >= hidirtybuffers) {
1107 		mtx_lock(&nblock);
1108 		while (numdirtybuffers >= hidirtybuffers) {
1109 			bd_wakeup(1);
1110 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1111 			msleep(&needsbuffer, &nblock,
1112 			    (PRIBIO + 4), "flswai", 0);
1113 		}
1114 		mtx_unlock(&nblock);
1115 	}
1116 }
1117 
1118 /*
1119  * Return true if we have too many dirty buffers.
1120  */
1121 int
1122 buf_dirty_count_severe(void)
1123 {
1124 
1125 	return(numdirtybuffers >= hidirtybuffers);
1126 }
1127 
1128 /*
1129  *	brelse:
1130  *
1131  *	Release a busy buffer and, if requested, free its resources.  The
1132  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1133  *	to be accessed later as a cache entity or reused for other purposes.
1134  */
1135 void
1136 brelse(struct buf *bp)
1137 {
1138 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1139 	    bp, bp->b_vp, bp->b_flags);
1140 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1141 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1142 
1143 	if (bp->b_flags & B_MANAGED) {
1144 		bqrelse(bp);
1145 		return;
1146 	}
1147 
1148 	if (bp->b_iocmd == BIO_WRITE &&
1149 	    (bp->b_ioflags & BIO_ERROR) &&
1150 	    !(bp->b_flags & B_INVAL)) {
1151 		/*
1152 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1153 		 * pages from being scrapped.  If B_INVAL is set then
1154 		 * this case is not run and the next case is run to
1155 		 * destroy the buffer.  B_INVAL can occur if the buffer
1156 		 * is outside the range supported by the underlying device.
1157 		 */
1158 		bp->b_ioflags &= ~BIO_ERROR;
1159 		bdirty(bp);
1160 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1161 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1162 		/*
1163 		 * Either a failed I/O or we were asked to free or not
1164 		 * cache the buffer.
1165 		 */
1166 		bp->b_flags |= B_INVAL;
1167 		if (LIST_FIRST(&bp->b_dep) != NULL)
1168 			buf_deallocate(bp);
1169 		if (bp->b_flags & B_DELWRI) {
1170 			atomic_subtract_int(&numdirtybuffers, 1);
1171 			numdirtywakeup(lodirtybuffers);
1172 		}
1173 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1174 		if ((bp->b_flags & B_VMIO) == 0) {
1175 			if (bp->b_bufsize)
1176 				allocbuf(bp, 0);
1177 			if (bp->b_vp)
1178 				brelvp(bp);
1179 		}
1180 	}
1181 
1182 	/*
1183 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1184 	 * is called with B_DELWRI set, the underlying pages may wind up
1185 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1186 	 * because pages associated with a B_DELWRI bp are marked clean.
1187 	 *
1188 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1189 	 * if B_DELWRI is set.
1190 	 *
1191 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1192 	 * on pages to return pages to the VM page queues.
1193 	 */
1194 	if (bp->b_flags & B_DELWRI)
1195 		bp->b_flags &= ~B_RELBUF;
1196 	else if (vm_page_count_severe()) {
1197 		/*
1198 		 * XXX This lock may not be necessary since BKGRDINPROG
1199 		 * cannot be set while we hold the buf lock, it can only be
1200 		 * cleared if it is already pending.
1201 		 */
1202 		if (bp->b_vp) {
1203 			BO_LOCK(bp->b_bufobj);
1204 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1205 				bp->b_flags |= B_RELBUF;
1206 			BO_UNLOCK(bp->b_bufobj);
1207 		} else
1208 			bp->b_flags |= B_RELBUF;
1209 	}
1210 
1211 	/*
1212 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1213 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1214 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1215 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1216 	 *
1217 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1218 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1219 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1220 	 *
1221 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1222 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1223 	 * the commit state and we cannot afford to lose the buffer. If the
1224 	 * buffer has a background write in progress, we need to keep it
1225 	 * around to prevent it from being reconstituted and starting a second
1226 	 * background write.
1227 	 */
1228 	if ((bp->b_flags & B_VMIO)
1229 	    && !(bp->b_vp->v_mount != NULL &&
1230 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1231 		 !vn_isdisk(bp->b_vp, NULL) &&
1232 		 (bp->b_flags & B_DELWRI))
1233 	    ) {
1234 
1235 		int i, j, resid;
1236 		vm_page_t m;
1237 		off_t foff;
1238 		vm_pindex_t poff;
1239 		vm_object_t obj;
1240 
1241 		obj = bp->b_bufobj->bo_object;
1242 
1243 		/*
1244 		 * Get the base offset and length of the buffer.  Note that
1245 		 * in the VMIO case if the buffer block size is not
1246 		 * page-aligned then b_data pointer may not be page-aligned.
1247 		 * But our b_pages[] array *IS* page aligned.
1248 		 *
1249 		 * block sizes less then DEV_BSIZE (usually 512) are not
1250 		 * supported due to the page granularity bits (m->valid,
1251 		 * m->dirty, etc...).
1252 		 *
1253 		 * See man buf(9) for more information
1254 		 */
1255 		resid = bp->b_bufsize;
1256 		foff = bp->b_offset;
1257 		VM_OBJECT_LOCK(obj);
1258 		for (i = 0; i < bp->b_npages; i++) {
1259 			int had_bogus = 0;
1260 
1261 			m = bp->b_pages[i];
1262 
1263 			/*
1264 			 * If we hit a bogus page, fixup *all* the bogus pages
1265 			 * now.
1266 			 */
1267 			if (m == bogus_page) {
1268 				poff = OFF_TO_IDX(bp->b_offset);
1269 				had_bogus = 1;
1270 
1271 				for (j = i; j < bp->b_npages; j++) {
1272 					vm_page_t mtmp;
1273 					mtmp = bp->b_pages[j];
1274 					if (mtmp == bogus_page) {
1275 						mtmp = vm_page_lookup(obj, poff + j);
1276 						if (!mtmp) {
1277 							panic("brelse: page missing\n");
1278 						}
1279 						bp->b_pages[j] = mtmp;
1280 					}
1281 				}
1282 
1283 				if ((bp->b_flags & B_INVAL) == 0) {
1284 					pmap_qenter(
1285 					    trunc_page((vm_offset_t)bp->b_data),
1286 					    bp->b_pages, bp->b_npages);
1287 				}
1288 				m = bp->b_pages[i];
1289 			}
1290 			if ((bp->b_flags & B_NOCACHE) ||
1291 			    (bp->b_ioflags & BIO_ERROR)) {
1292 				int poffset = foff & PAGE_MASK;
1293 				int presid = resid > (PAGE_SIZE - poffset) ?
1294 					(PAGE_SIZE - poffset) : resid;
1295 
1296 				KASSERT(presid >= 0, ("brelse: extra page"));
1297 				vm_page_lock_queues();
1298 				vm_page_set_invalid(m, poffset, presid);
1299 				vm_page_unlock_queues();
1300 				if (had_bogus)
1301 					printf("avoided corruption bug in bogus_page/brelse code\n");
1302 			}
1303 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1304 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1305 		}
1306 		VM_OBJECT_UNLOCK(obj);
1307 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1308 			vfs_vmio_release(bp);
1309 
1310 	} else if (bp->b_flags & B_VMIO) {
1311 
1312 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1313 			vfs_vmio_release(bp);
1314 		}
1315 
1316 	} else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1317 		if (bp->b_bufsize != 0)
1318 			allocbuf(bp, 0);
1319 		if (bp->b_vp != NULL)
1320 			brelvp(bp);
1321 	}
1322 
1323 	if (BUF_REFCNT(bp) > 1) {
1324 		/* do not release to free list */
1325 		BUF_UNLOCK(bp);
1326 		return;
1327 	}
1328 
1329 	/* enqueue */
1330 	mtx_lock(&bqlock);
1331 	/* Handle delayed bremfree() processing. */
1332 	if (bp->b_flags & B_REMFREE)
1333 		bremfreel(bp);
1334 	if (bp->b_qindex != QUEUE_NONE)
1335 		panic("brelse: free buffer onto another queue???");
1336 
1337 	/* buffers with no memory */
1338 	if (bp->b_bufsize == 0) {
1339 		bp->b_flags |= B_INVAL;
1340 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1341 		if (bp->b_vflags & BV_BKGRDINPROG)
1342 			panic("losing buffer 1");
1343 		if (bp->b_kvasize) {
1344 			bp->b_qindex = QUEUE_EMPTYKVA;
1345 		} else {
1346 			bp->b_qindex = QUEUE_EMPTY;
1347 		}
1348 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1349 	/* buffers with junk contents */
1350 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1351 	    (bp->b_ioflags & BIO_ERROR)) {
1352 		bp->b_flags |= B_INVAL;
1353 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1354 		if (bp->b_vflags & BV_BKGRDINPROG)
1355 			panic("losing buffer 2");
1356 		bp->b_qindex = QUEUE_CLEAN;
1357 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1358 	/* remaining buffers */
1359 	} else {
1360 		if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1361 		    (B_DELWRI|B_NEEDSGIANT))
1362 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1363 		if (bp->b_flags & B_DELWRI)
1364 			bp->b_qindex = QUEUE_DIRTY;
1365 		else
1366 			bp->b_qindex = QUEUE_CLEAN;
1367 		if (bp->b_flags & B_AGE)
1368 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1369 		else
1370 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1371 	}
1372 	mtx_unlock(&bqlock);
1373 
1374 	/*
1375 	 * If B_INVAL and B_DELWRI is set, clear B_DELWRI.  We have already
1376 	 * placed the buffer on the correct queue.  We must also disassociate
1377 	 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1378 	 * find it.
1379 	 */
1380 	if (bp->b_flags & B_INVAL) {
1381 		if (bp->b_flags & B_DELWRI)
1382 			bundirty(bp);
1383 		if (bp->b_vp)
1384 			brelvp(bp);
1385 	}
1386 
1387 	/*
1388 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1389 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1390 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1391 	 * if B_INVAL is set ).
1392 	 */
1393 
1394 	if (!(bp->b_flags & B_DELWRI))
1395 		bufcountwakeup();
1396 
1397 	/*
1398 	 * Something we can maybe free or reuse
1399 	 */
1400 	if (bp->b_bufsize || bp->b_kvasize)
1401 		bufspacewakeup();
1402 
1403 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1404 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1405 		panic("brelse: not dirty");
1406 	/* unlock */
1407 	BUF_UNLOCK(bp);
1408 }
1409 
1410 /*
1411  * Release a buffer back to the appropriate queue but do not try to free
1412  * it.  The buffer is expected to be used again soon.
1413  *
1414  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1415  * biodone() to requeue an async I/O on completion.  It is also used when
1416  * known good buffers need to be requeued but we think we may need the data
1417  * again soon.
1418  *
1419  * XXX we should be able to leave the B_RELBUF hint set on completion.
1420  */
1421 void
1422 bqrelse(struct buf *bp)
1423 {
1424 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1425 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1426 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1427 
1428 	if (BUF_REFCNT(bp) > 1) {
1429 		/* do not release to free list */
1430 		BUF_UNLOCK(bp);
1431 		return;
1432 	}
1433 
1434 	if (bp->b_flags & B_MANAGED) {
1435 		if (bp->b_flags & B_REMFREE) {
1436 			mtx_lock(&bqlock);
1437 			bremfreel(bp);
1438 			mtx_unlock(&bqlock);
1439 		}
1440 		bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1441 		BUF_UNLOCK(bp);
1442 		return;
1443 	}
1444 
1445 	mtx_lock(&bqlock);
1446 	/* Handle delayed bremfree() processing. */
1447 	if (bp->b_flags & B_REMFREE)
1448 		bremfreel(bp);
1449 	if (bp->b_qindex != QUEUE_NONE)
1450 		panic("bqrelse: free buffer onto another queue???");
1451 	/* buffers with stale but valid contents */
1452 	if (bp->b_flags & B_DELWRI) {
1453 		if (bp->b_flags & B_NEEDSGIANT)
1454 			bp->b_qindex = QUEUE_DIRTY_GIANT;
1455 		else
1456 			bp->b_qindex = QUEUE_DIRTY;
1457 		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1458 	} else {
1459 		/*
1460 		 * XXX This lock may not be necessary since BKGRDINPROG
1461 		 * cannot be set while we hold the buf lock, it can only be
1462 		 * cleared if it is already pending.
1463 		 */
1464 		BO_LOCK(bp->b_bufobj);
1465 		if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) {
1466 			BO_UNLOCK(bp->b_bufobj);
1467 			bp->b_qindex = QUEUE_CLEAN;
1468 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1469 			    b_freelist);
1470 		} else {
1471 			/*
1472 			 * We are too low on memory, we have to try to free
1473 			 * the buffer (most importantly: the wired pages
1474 			 * making up its backing store) *now*.
1475 			 */
1476 			BO_UNLOCK(bp->b_bufobj);
1477 			mtx_unlock(&bqlock);
1478 			brelse(bp);
1479 			return;
1480 		}
1481 	}
1482 	mtx_unlock(&bqlock);
1483 
1484 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1485 		bufcountwakeup();
1486 
1487 	/*
1488 	 * Something we can maybe free or reuse.
1489 	 */
1490 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1491 		bufspacewakeup();
1492 
1493 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1494 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1495 		panic("bqrelse: not dirty");
1496 	/* unlock */
1497 	BUF_UNLOCK(bp);
1498 }
1499 
1500 /* Give pages used by the bp back to the VM system (where possible) */
1501 static void
1502 vfs_vmio_release(struct buf *bp)
1503 {
1504 	int i;
1505 	vm_page_t m;
1506 
1507 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1508 	vm_page_lock_queues();
1509 	for (i = 0; i < bp->b_npages; i++) {
1510 		m = bp->b_pages[i];
1511 		bp->b_pages[i] = NULL;
1512 		/*
1513 		 * In order to keep page LRU ordering consistent, put
1514 		 * everything on the inactive queue.
1515 		 */
1516 		vm_page_unwire(m, 0);
1517 		/*
1518 		 * We don't mess with busy pages, it is
1519 		 * the responsibility of the process that
1520 		 * busied the pages to deal with them.
1521 		 */
1522 		if ((m->flags & PG_BUSY) || (m->busy != 0))
1523 			continue;
1524 
1525 		if (m->wire_count == 0) {
1526 			/*
1527 			 * Might as well free the page if we can and it has
1528 			 * no valid data.  We also free the page if the
1529 			 * buffer was used for direct I/O
1530 			 */
1531 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1532 			    m->hold_count == 0) {
1533 				vm_page_free(m);
1534 			} else if (bp->b_flags & B_DIRECT) {
1535 				vm_page_try_to_free(m);
1536 			} else if (vm_page_count_severe()) {
1537 				vm_page_try_to_cache(m);
1538 			}
1539 		}
1540 	}
1541 	vm_page_unlock_queues();
1542 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1543 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1544 
1545 	if (bp->b_bufsize) {
1546 		bufspacewakeup();
1547 		bp->b_bufsize = 0;
1548 	}
1549 	bp->b_npages = 0;
1550 	bp->b_flags &= ~B_VMIO;
1551 	if (bp->b_vp)
1552 		brelvp(bp);
1553 }
1554 
1555 /*
1556  * Check to see if a block at a particular lbn is available for a clustered
1557  * write.
1558  */
1559 static int
1560 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1561 {
1562 	struct buf *bpa;
1563 	int match;
1564 
1565 	match = 0;
1566 
1567 	/* If the buf isn't in core skip it */
1568 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1569 		return (0);
1570 
1571 	/* If the buf is busy we don't want to wait for it */
1572 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1573 		return (0);
1574 
1575 	/* Only cluster with valid clusterable delayed write buffers */
1576 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1577 	    (B_DELWRI | B_CLUSTEROK))
1578 		goto done;
1579 
1580 	if (bpa->b_bufsize != size)
1581 		goto done;
1582 
1583 	/*
1584 	 * Check to see if it is in the expected place on disk and that the
1585 	 * block has been mapped.
1586 	 */
1587 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1588 		match = 1;
1589 done:
1590 	BUF_UNLOCK(bpa);
1591 	return (match);
1592 }
1593 
1594 /*
1595  *	vfs_bio_awrite:
1596  *
1597  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1598  *	This is much better then the old way of writing only one buffer at
1599  *	a time.  Note that we may not be presented with the buffers in the
1600  *	correct order, so we search for the cluster in both directions.
1601  */
1602 int
1603 vfs_bio_awrite(struct buf *bp)
1604 {
1605 	int i;
1606 	int j;
1607 	daddr_t lblkno = bp->b_lblkno;
1608 	struct vnode *vp = bp->b_vp;
1609 	int ncl;
1610 	int nwritten;
1611 	int size;
1612 	int maxcl;
1613 
1614 	/*
1615 	 * right now we support clustered writing only to regular files.  If
1616 	 * we find a clusterable block we could be in the middle of a cluster
1617 	 * rather then at the beginning.
1618 	 */
1619 	if ((vp->v_type == VREG) &&
1620 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1621 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1622 
1623 		size = vp->v_mount->mnt_stat.f_iosize;
1624 		maxcl = MAXPHYS / size;
1625 
1626 		VI_LOCK(vp);
1627 		for (i = 1; i < maxcl; i++)
1628 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1629 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1630 				break;
1631 
1632 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1633 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1634 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1635 				break;
1636 
1637 		VI_UNLOCK(vp);
1638 		--j;
1639 		ncl = i + j;
1640 		/*
1641 		 * this is a possible cluster write
1642 		 */
1643 		if (ncl != 1) {
1644 			BUF_UNLOCK(bp);
1645 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1646 			return nwritten;
1647 		}
1648 	}
1649 	bremfree(bp);
1650 	bp->b_flags |= B_ASYNC;
1651 	/*
1652 	 * default (old) behavior, writing out only one block
1653 	 *
1654 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1655 	 */
1656 	nwritten = bp->b_bufsize;
1657 	(void) bwrite(bp);
1658 
1659 	return nwritten;
1660 }
1661 
1662 /*
1663  *	getnewbuf:
1664  *
1665  *	Find and initialize a new buffer header, freeing up existing buffers
1666  *	in the bufqueues as necessary.  The new buffer is returned locked.
1667  *
1668  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1669  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1670  *
1671  *	We block if:
1672  *		We have insufficient buffer headers
1673  *		We have insufficient buffer space
1674  *		buffer_map is too fragmented ( space reservation fails )
1675  *		If we have to flush dirty buffers ( but we try to avoid this )
1676  *
1677  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1678  *	Instead we ask the buf daemon to do it for us.  We attempt to
1679  *	avoid piecemeal wakeups of the pageout daemon.
1680  */
1681 
1682 static struct buf *
1683 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1684 {
1685 	struct buf *bp;
1686 	struct buf *nbp;
1687 	int defrag = 0;
1688 	int nqindex;
1689 	static int flushingbufs;
1690 
1691 	/*
1692 	 * We can't afford to block since we might be holding a vnode lock,
1693 	 * which may prevent system daemons from running.  We deal with
1694 	 * low-memory situations by proactively returning memory and running
1695 	 * async I/O rather then sync I/O.
1696 	 */
1697 
1698 	atomic_add_int(&getnewbufcalls, 1);
1699 	atomic_subtract_int(&getnewbufrestarts, 1);
1700 restart:
1701 	atomic_add_int(&getnewbufrestarts, 1);
1702 
1703 	/*
1704 	 * Setup for scan.  If we do not have enough free buffers,
1705 	 * we setup a degenerate case that immediately fails.  Note
1706 	 * that if we are specially marked process, we are allowed to
1707 	 * dip into our reserves.
1708 	 *
1709 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1710 	 *
1711 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1712 	 * However, there are a number of cases (defragging, reusing, ...)
1713 	 * where we cannot backup.
1714 	 */
1715 	mtx_lock(&bqlock);
1716 	nqindex = QUEUE_EMPTYKVA;
1717 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1718 
1719 	if (nbp == NULL) {
1720 		/*
1721 		 * If no EMPTYKVA buffers and we are either
1722 		 * defragging or reusing, locate a CLEAN buffer
1723 		 * to free or reuse.  If bufspace useage is low
1724 		 * skip this step so we can allocate a new buffer.
1725 		 */
1726 		if (defrag || bufspace >= lobufspace) {
1727 			nqindex = QUEUE_CLEAN;
1728 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1729 		}
1730 
1731 		/*
1732 		 * If we could not find or were not allowed to reuse a
1733 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1734 		 * buffer.  We can only use an EMPTY buffer if allocating
1735 		 * its KVA would not otherwise run us out of buffer space.
1736 		 */
1737 		if (nbp == NULL && defrag == 0 &&
1738 		    bufspace + maxsize < hibufspace) {
1739 			nqindex = QUEUE_EMPTY;
1740 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1741 		}
1742 	}
1743 
1744 	/*
1745 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1746 	 * depending.
1747 	 */
1748 
1749 	while ((bp = nbp) != NULL) {
1750 		int qindex = nqindex;
1751 
1752 		/*
1753 		 * Calculate next bp ( we can only use it if we do not block
1754 		 * or do other fancy things ).
1755 		 */
1756 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1757 			switch(qindex) {
1758 			case QUEUE_EMPTY:
1759 				nqindex = QUEUE_EMPTYKVA;
1760 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1761 					break;
1762 				/* FALLTHROUGH */
1763 			case QUEUE_EMPTYKVA:
1764 				nqindex = QUEUE_CLEAN;
1765 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1766 					break;
1767 				/* FALLTHROUGH */
1768 			case QUEUE_CLEAN:
1769 				/*
1770 				 * nbp is NULL.
1771 				 */
1772 				break;
1773 			}
1774 		}
1775 		/*
1776 		 * If we are defragging then we need a buffer with
1777 		 * b_kvasize != 0.  XXX this situation should no longer
1778 		 * occur, if defrag is non-zero the buffer's b_kvasize
1779 		 * should also be non-zero at this point.  XXX
1780 		 */
1781 		if (defrag && bp->b_kvasize == 0) {
1782 			printf("Warning: defrag empty buffer %p\n", bp);
1783 			continue;
1784 		}
1785 
1786 		/*
1787 		 * Start freeing the bp.  This is somewhat involved.  nbp
1788 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1789 		 */
1790 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1791 			continue;
1792 		if (bp->b_vp) {
1793 			BO_LOCK(bp->b_bufobj);
1794 			if (bp->b_vflags & BV_BKGRDINPROG) {
1795 				BO_UNLOCK(bp->b_bufobj);
1796 				BUF_UNLOCK(bp);
1797 				continue;
1798 			}
1799 			BO_UNLOCK(bp->b_bufobj);
1800 		}
1801 		CTR6(KTR_BUF,
1802 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1803 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1804 		    bp->b_kvasize, bp->b_bufsize, qindex);
1805 
1806 		/*
1807 		 * Sanity Checks
1808 		 */
1809 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1810 
1811 		/*
1812 		 * Note: we no longer distinguish between VMIO and non-VMIO
1813 		 * buffers.
1814 		 */
1815 
1816 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1817 
1818 		bremfreel(bp);
1819 		mtx_unlock(&bqlock);
1820 
1821 		if (qindex == QUEUE_CLEAN) {
1822 			if (bp->b_flags & B_VMIO) {
1823 				bp->b_flags &= ~B_ASYNC;
1824 				vfs_vmio_release(bp);
1825 			}
1826 			if (bp->b_vp)
1827 				brelvp(bp);
1828 		}
1829 
1830 		/*
1831 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1832 		 * the scan from this point on.
1833 		 *
1834 		 * Get the rest of the buffer freed up.  b_kva* is still
1835 		 * valid after this operation.
1836 		 */
1837 
1838 		if (bp->b_rcred != NOCRED) {
1839 			crfree(bp->b_rcred);
1840 			bp->b_rcred = NOCRED;
1841 		}
1842 		if (bp->b_wcred != NOCRED) {
1843 			crfree(bp->b_wcred);
1844 			bp->b_wcred = NOCRED;
1845 		}
1846 		if (LIST_FIRST(&bp->b_dep) != NULL)
1847 			buf_deallocate(bp);
1848 		if (bp->b_vflags & BV_BKGRDINPROG)
1849 			panic("losing buffer 3");
1850 		KASSERT(bp->b_vp == NULL,
1851 		    ("bp: %p still has vnode %p.  qindex: %d",
1852 		    bp, bp->b_vp, qindex));
1853 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1854 		   ("bp: %p still on a buffer list. xflags %X",
1855 		    bp, bp->b_xflags));
1856 
1857 		if (bp->b_bufsize)
1858 			allocbuf(bp, 0);
1859 
1860 		bp->b_flags = 0;
1861 		bp->b_ioflags = 0;
1862 		bp->b_xflags = 0;
1863 		bp->b_vflags = 0;
1864 		bp->b_vp = NULL;
1865 		bp->b_blkno = bp->b_lblkno = 0;
1866 		bp->b_offset = NOOFFSET;
1867 		bp->b_iodone = 0;
1868 		bp->b_error = 0;
1869 		bp->b_resid = 0;
1870 		bp->b_bcount = 0;
1871 		bp->b_npages = 0;
1872 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1873 		bp->b_bufobj = NULL;
1874 		bp->b_pin_count = 0;
1875 		bp->b_fsprivate1 = NULL;
1876 		bp->b_fsprivate2 = NULL;
1877 		bp->b_fsprivate3 = NULL;
1878 
1879 		LIST_INIT(&bp->b_dep);
1880 
1881 		/*
1882 		 * If we are defragging then free the buffer.
1883 		 */
1884 		if (defrag) {
1885 			bp->b_flags |= B_INVAL;
1886 			bfreekva(bp);
1887 			brelse(bp);
1888 			defrag = 0;
1889 			goto restart;
1890 		}
1891 
1892 		/*
1893 		 * If we are overcomitted then recover the buffer and its
1894 		 * KVM space.  This occurs in rare situations when multiple
1895 		 * processes are blocked in getnewbuf() or allocbuf().
1896 		 */
1897 		if (bufspace >= hibufspace)
1898 			flushingbufs = 1;
1899 		if (flushingbufs && bp->b_kvasize != 0) {
1900 			bp->b_flags |= B_INVAL;
1901 			bfreekva(bp);
1902 			brelse(bp);
1903 			goto restart;
1904 		}
1905 		if (bufspace < lobufspace)
1906 			flushingbufs = 0;
1907 		break;
1908 	}
1909 
1910 	/*
1911 	 * If we exhausted our list, sleep as appropriate.  We may have to
1912 	 * wakeup various daemons and write out some dirty buffers.
1913 	 *
1914 	 * Generally we are sleeping due to insufficient buffer space.
1915 	 */
1916 
1917 	if (bp == NULL) {
1918 		int flags;
1919 		char *waitmsg;
1920 
1921 		if (defrag) {
1922 			flags = VFS_BIO_NEED_BUFSPACE;
1923 			waitmsg = "nbufkv";
1924 		} else if (bufspace >= hibufspace) {
1925 			waitmsg = "nbufbs";
1926 			flags = VFS_BIO_NEED_BUFSPACE;
1927 		} else {
1928 			waitmsg = "newbuf";
1929 			flags = VFS_BIO_NEED_ANY;
1930 		}
1931 		mtx_lock(&nblock);
1932 		needsbuffer |= flags;
1933 		mtx_unlock(&nblock);
1934 		mtx_unlock(&bqlock);
1935 
1936 		bd_speedup();	/* heeeelp */
1937 
1938 		mtx_lock(&nblock);
1939 		while (needsbuffer & flags) {
1940 			if (msleep(&needsbuffer, &nblock,
1941 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1942 				mtx_unlock(&nblock);
1943 				return (NULL);
1944 			}
1945 		}
1946 		mtx_unlock(&nblock);
1947 	} else {
1948 		/*
1949 		 * We finally have a valid bp.  We aren't quite out of the
1950 		 * woods, we still have to reserve kva space.  In order
1951 		 * to keep fragmentation sane we only allocate kva in
1952 		 * BKVASIZE chunks.
1953 		 */
1954 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1955 
1956 		if (maxsize != bp->b_kvasize) {
1957 			vm_offset_t addr = 0;
1958 
1959 			bfreekva(bp);
1960 
1961 			vm_map_lock(buffer_map);
1962 			if (vm_map_findspace(buffer_map,
1963 				vm_map_min(buffer_map), maxsize, &addr)) {
1964 				/*
1965 				 * Uh oh.  Buffer map is to fragmented.  We
1966 				 * must defragment the map.
1967 				 */
1968 				atomic_add_int(&bufdefragcnt, 1);
1969 				vm_map_unlock(buffer_map);
1970 				defrag = 1;
1971 				bp->b_flags |= B_INVAL;
1972 				brelse(bp);
1973 				goto restart;
1974 			}
1975 			if (addr) {
1976 				vm_map_insert(buffer_map, NULL, 0,
1977 					addr, addr + maxsize,
1978 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1979 
1980 				bp->b_kvabase = (caddr_t) addr;
1981 				bp->b_kvasize = maxsize;
1982 				atomic_add_int(&bufspace, bp->b_kvasize);
1983 				atomic_add_int(&bufreusecnt, 1);
1984 			}
1985 			vm_map_unlock(buffer_map);
1986 		}
1987 		bp->b_saveaddr = bp->b_kvabase;
1988 		bp->b_data = bp->b_saveaddr;
1989 	}
1990 	return(bp);
1991 }
1992 
1993 /*
1994  *	buf_daemon:
1995  *
1996  *	buffer flushing daemon.  Buffers are normally flushed by the
1997  *	update daemon but if it cannot keep up this process starts to
1998  *	take the load in an attempt to prevent getnewbuf() from blocking.
1999  */
2000 
2001 static struct kproc_desc buf_kp = {
2002 	"bufdaemon",
2003 	buf_daemon,
2004 	&bufdaemonproc
2005 };
2006 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
2007 
2008 static void
2009 buf_daemon()
2010 {
2011 
2012 	/*
2013 	 * This process needs to be suspended prior to shutdown sync.
2014 	 */
2015 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2016 	    SHUTDOWN_PRI_LAST);
2017 
2018 	/*
2019 	 * This process is allowed to take the buffer cache to the limit
2020 	 */
2021 	curthread->td_pflags |= TDP_NORUNNINGBUF;
2022 	mtx_lock(&bdlock);
2023 	for (;;) {
2024 		bd_request = 0;
2025 		mtx_unlock(&bdlock);
2026 
2027 		kthread_suspend_check(bufdaemonproc);
2028 
2029 		/*
2030 		 * Do the flush.  Limit the amount of in-transit I/O we
2031 		 * allow to build up, otherwise we would completely saturate
2032 		 * the I/O system.  Wakeup any waiting processes before we
2033 		 * normally would so they can run in parallel with our drain.
2034 		 */
2035 		while (numdirtybuffers > lodirtybuffers) {
2036 			int flushed;
2037 
2038 			flushed = flushbufqueues(QUEUE_DIRTY, 0);
2039 			/* The list empty check here is slightly racy */
2040 			if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2041 				mtx_lock(&Giant);
2042 				flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0);
2043 				mtx_unlock(&Giant);
2044 			}
2045 			if (flushed == 0) {
2046 				/*
2047 				 * Could not find any buffers without rollback
2048 				 * dependencies, so just write the first one
2049 				 * in the hopes of eventually making progress.
2050 				 */
2051 				flushbufqueues(QUEUE_DIRTY, 1);
2052 				if (!TAILQ_EMPTY(
2053 				    &bufqueues[QUEUE_DIRTY_GIANT])) {
2054 					mtx_lock(&Giant);
2055 					flushbufqueues(QUEUE_DIRTY_GIANT, 1);
2056 					mtx_unlock(&Giant);
2057 				}
2058 				break;
2059 			}
2060 			uio_yield();
2061 		}
2062 
2063 		/*
2064 		 * Only clear bd_request if we have reached our low water
2065 		 * mark.  The buf_daemon normally waits 1 second and
2066 		 * then incrementally flushes any dirty buffers that have
2067 		 * built up, within reason.
2068 		 *
2069 		 * If we were unable to hit our low water mark and couldn't
2070 		 * find any flushable buffers, we sleep half a second.
2071 		 * Otherwise we loop immediately.
2072 		 */
2073 		mtx_lock(&bdlock);
2074 		if (numdirtybuffers <= lodirtybuffers) {
2075 			/*
2076 			 * We reached our low water mark, reset the
2077 			 * request and sleep until we are needed again.
2078 			 * The sleep is just so the suspend code works.
2079 			 */
2080 			bd_request = 0;
2081 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2082 		} else {
2083 			/*
2084 			 * We couldn't find any flushable dirty buffers but
2085 			 * still have too many dirty buffers, we
2086 			 * have to sleep and try again.  (rare)
2087 			 */
2088 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2089 		}
2090 	}
2091 }
2092 
2093 /*
2094  *	flushbufqueues:
2095  *
2096  *	Try to flush a buffer in the dirty queue.  We must be careful to
2097  *	free up B_INVAL buffers instead of write them, which NFS is
2098  *	particularly sensitive to.
2099  */
2100 static int flushwithdeps = 0;
2101 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2102     0, "Number of buffers flushed with dependecies that require rollbacks");
2103 
2104 static int
2105 flushbufqueues(int queue, int flushdeps)
2106 {
2107 	struct thread *td = curthread;
2108 	struct buf sentinel;
2109 	struct vnode *vp;
2110 	struct mount *mp;
2111 	struct buf *bp;
2112 	int hasdeps;
2113 	int flushed;
2114 	int target;
2115 
2116 	target = numdirtybuffers - lodirtybuffers;
2117 	if (flushdeps && target > 2)
2118 		target /= 2;
2119 	flushed = 0;
2120 	bp = NULL;
2121 	mtx_lock(&bqlock);
2122 	TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist);
2123 	while (flushed != target) {
2124 		bp = TAILQ_FIRST(&bufqueues[queue]);
2125 		if (bp == &sentinel)
2126 			break;
2127 		TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist);
2128 		TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist);
2129 
2130 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2131 			continue;
2132 		if (bp->b_pin_count > 0) {
2133 			BUF_UNLOCK(bp);
2134 			continue;
2135 		}
2136 		BO_LOCK(bp->b_bufobj);
2137 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2138 		    (bp->b_flags & B_DELWRI) == 0) {
2139 			BO_UNLOCK(bp->b_bufobj);
2140 			BUF_UNLOCK(bp);
2141 			continue;
2142 		}
2143 		BO_UNLOCK(bp->b_bufobj);
2144 		if (bp->b_flags & B_INVAL) {
2145 			bremfreel(bp);
2146 			mtx_unlock(&bqlock);
2147 			brelse(bp);
2148 			flushed++;
2149 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2150 			mtx_lock(&bqlock);
2151 			continue;
2152 		}
2153 
2154 		if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) {
2155 			if (flushdeps == 0) {
2156 				BUF_UNLOCK(bp);
2157 				continue;
2158 			}
2159 			hasdeps = 1;
2160 		} else
2161 			hasdeps = 0;
2162 		/*
2163 		 * We must hold the lock on a vnode before writing
2164 		 * one of its buffers. Otherwise we may confuse, or
2165 		 * in the case of a snapshot vnode, deadlock the
2166 		 * system.
2167 		 *
2168 		 * The lock order here is the reverse of the normal
2169 		 * of vnode followed by buf lock.  This is ok because
2170 		 * the NOWAIT will prevent deadlock.
2171 		 */
2172 		vp = bp->b_vp;
2173 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2174 			BUF_UNLOCK(bp);
2175 			continue;
2176 		}
2177 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2178 			mtx_unlock(&bqlock);
2179 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2180 			    bp, bp->b_vp, bp->b_flags);
2181 			vfs_bio_awrite(bp);
2182 			vn_finished_write(mp);
2183 			VOP_UNLOCK(vp, 0, td);
2184 			flushwithdeps += hasdeps;
2185 			flushed++;
2186 			waitrunningbufspace();
2187 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2188 			mtx_lock(&bqlock);
2189 			continue;
2190 		}
2191 		vn_finished_write(mp);
2192 		BUF_UNLOCK(bp);
2193 	}
2194 	TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist);
2195 	mtx_unlock(&bqlock);
2196 	return (flushed);
2197 }
2198 
2199 /*
2200  * Check to see if a block is currently memory resident.
2201  */
2202 struct buf *
2203 incore(struct bufobj *bo, daddr_t blkno)
2204 {
2205 	struct buf *bp;
2206 
2207 	BO_LOCK(bo);
2208 	bp = gbincore(bo, blkno);
2209 	BO_UNLOCK(bo);
2210 	return (bp);
2211 }
2212 
2213 /*
2214  * Returns true if no I/O is needed to access the
2215  * associated VM object.  This is like incore except
2216  * it also hunts around in the VM system for the data.
2217  */
2218 
2219 static int
2220 inmem(struct vnode * vp, daddr_t blkno)
2221 {
2222 	vm_object_t obj;
2223 	vm_offset_t toff, tinc, size;
2224 	vm_page_t m;
2225 	vm_ooffset_t off;
2226 
2227 	ASSERT_VOP_LOCKED(vp, "inmem");
2228 
2229 	if (incore(&vp->v_bufobj, blkno))
2230 		return 1;
2231 	if (vp->v_mount == NULL)
2232 		return 0;
2233 	obj = vp->v_object;
2234 	if (obj == NULL)
2235 		return (0);
2236 
2237 	size = PAGE_SIZE;
2238 	if (size > vp->v_mount->mnt_stat.f_iosize)
2239 		size = vp->v_mount->mnt_stat.f_iosize;
2240 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2241 
2242 	VM_OBJECT_LOCK(obj);
2243 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2244 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2245 		if (!m)
2246 			goto notinmem;
2247 		tinc = size;
2248 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2249 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2250 		if (vm_page_is_valid(m,
2251 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2252 			goto notinmem;
2253 	}
2254 	VM_OBJECT_UNLOCK(obj);
2255 	return 1;
2256 
2257 notinmem:
2258 	VM_OBJECT_UNLOCK(obj);
2259 	return (0);
2260 }
2261 
2262 /*
2263  *	vfs_setdirty:
2264  *
2265  *	Sets the dirty range for a buffer based on the status of the dirty
2266  *	bits in the pages comprising the buffer.
2267  *
2268  *	The range is limited to the size of the buffer.
2269  *
2270  *	This routine is primarily used by NFS, but is generalized for the
2271  *	B_VMIO case.
2272  */
2273 static void
2274 vfs_setdirty(struct buf *bp)
2275 {
2276 	int i;
2277 	vm_object_t object;
2278 
2279 	/*
2280 	 * Degenerate case - empty buffer
2281 	 */
2282 
2283 	if (bp->b_bufsize == 0)
2284 		return;
2285 
2286 	/*
2287 	 * We qualify the scan for modified pages on whether the
2288 	 * object has been flushed yet.
2289 	 */
2290 
2291 	if ((bp->b_flags & B_VMIO) == 0)
2292 		return;
2293 
2294 	object = bp->b_pages[0]->object;
2295 	VM_OBJECT_LOCK(object);
2296 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2297 		vm_offset_t boffset;
2298 		vm_offset_t eoffset;
2299 
2300 		vm_page_lock_queues();
2301 		/*
2302 		 * test the pages to see if they have been modified directly
2303 		 * by users through the VM system.
2304 		 */
2305 		for (i = 0; i < bp->b_npages; i++)
2306 			vm_page_test_dirty(bp->b_pages[i]);
2307 
2308 		/*
2309 		 * Calculate the encompassing dirty range, boffset and eoffset,
2310 		 * (eoffset - boffset) bytes.
2311 		 */
2312 
2313 		for (i = 0; i < bp->b_npages; i++) {
2314 			if (bp->b_pages[i]->dirty)
2315 				break;
2316 		}
2317 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2318 
2319 		for (i = bp->b_npages - 1; i >= 0; --i) {
2320 			if (bp->b_pages[i]->dirty) {
2321 				break;
2322 			}
2323 		}
2324 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2325 
2326 		vm_page_unlock_queues();
2327 		/*
2328 		 * Fit it to the buffer.
2329 		 */
2330 
2331 		if (eoffset > bp->b_bcount)
2332 			eoffset = bp->b_bcount;
2333 
2334 		/*
2335 		 * If we have a good dirty range, merge with the existing
2336 		 * dirty range.
2337 		 */
2338 
2339 		if (boffset < eoffset) {
2340 			if (bp->b_dirtyoff > boffset)
2341 				bp->b_dirtyoff = boffset;
2342 			if (bp->b_dirtyend < eoffset)
2343 				bp->b_dirtyend = eoffset;
2344 		}
2345 	}
2346 	VM_OBJECT_UNLOCK(object);
2347 }
2348 
2349 /*
2350  *	getblk:
2351  *
2352  *	Get a block given a specified block and offset into a file/device.
2353  *	The buffers B_DONE bit will be cleared on return, making it almost
2354  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2355  *	return.  The caller should clear B_INVAL prior to initiating a
2356  *	READ.
2357  *
2358  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2359  *	an existing buffer.
2360  *
2361  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2362  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2363  *	and then cleared based on the backing VM.  If the previous buffer is
2364  *	non-0-sized but invalid, B_CACHE will be cleared.
2365  *
2366  *	If getblk() must create a new buffer, the new buffer is returned with
2367  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2368  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2369  *	backing VM.
2370  *
2371  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2372  *	B_CACHE bit is clear.
2373  *
2374  *	What this means, basically, is that the caller should use B_CACHE to
2375  *	determine whether the buffer is fully valid or not and should clear
2376  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2377  *	the buffer by loading its data area with something, the caller needs
2378  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2379  *	the caller should set B_CACHE ( as an optimization ), else the caller
2380  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2381  *	a write attempt or if it was a successfull read.  If the caller
2382  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2383  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2384  */
2385 struct buf *
2386 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2387     int flags)
2388 {
2389 	struct buf *bp;
2390 	struct bufobj *bo;
2391 	int error;
2392 
2393 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2394 	ASSERT_VOP_LOCKED(vp, "getblk");
2395 	if (size > MAXBSIZE)
2396 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2397 
2398 	bo = &vp->v_bufobj;
2399 loop:
2400 	/*
2401 	 * Block if we are low on buffers.   Certain processes are allowed
2402 	 * to completely exhaust the buffer cache.
2403          *
2404          * If this check ever becomes a bottleneck it may be better to
2405          * move it into the else, when gbincore() fails.  At the moment
2406          * it isn't a problem.
2407 	 *
2408 	 * XXX remove if 0 sections (clean this up after its proven)
2409          */
2410 	if (numfreebuffers == 0) {
2411 		if (curthread == PCPU_GET(idlethread))
2412 			return NULL;
2413 		mtx_lock(&nblock);
2414 		needsbuffer |= VFS_BIO_NEED_ANY;
2415 		mtx_unlock(&nblock);
2416 	}
2417 
2418 	VI_LOCK(vp);
2419 	bp = gbincore(bo, blkno);
2420 	if (bp != NULL) {
2421 		int lockflags;
2422 		/*
2423 		 * Buffer is in-core.  If the buffer is not busy, it must
2424 		 * be on a queue.
2425 		 */
2426 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2427 
2428 		if (flags & GB_LOCK_NOWAIT)
2429 			lockflags |= LK_NOWAIT;
2430 
2431 		error = BUF_TIMELOCK(bp, lockflags,
2432 		    VI_MTX(vp), "getblk", slpflag, slptimeo);
2433 
2434 		/*
2435 		 * If we slept and got the lock we have to restart in case
2436 		 * the buffer changed identities.
2437 		 */
2438 		if (error == ENOLCK)
2439 			goto loop;
2440 		/* We timed out or were interrupted. */
2441 		else if (error)
2442 			return (NULL);
2443 
2444 		/*
2445 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2446 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2447 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2448 		 * backing VM cache.
2449 		 */
2450 		if (bp->b_flags & B_INVAL)
2451 			bp->b_flags &= ~B_CACHE;
2452 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2453 			bp->b_flags |= B_CACHE;
2454 		bremfree(bp);
2455 
2456 		/*
2457 		 * check for size inconsistancies for non-VMIO case.
2458 		 */
2459 
2460 		if (bp->b_bcount != size) {
2461 			if ((bp->b_flags & B_VMIO) == 0 ||
2462 			    (size > bp->b_kvasize)) {
2463 				if (bp->b_flags & B_DELWRI) {
2464 					/*
2465 					 * If buffer is pinned and caller does
2466 					 * not want sleep  waiting for it to be
2467 					 * unpinned, bail out
2468 					 * */
2469 					if (bp->b_pin_count > 0) {
2470 						if (flags & GB_LOCK_NOWAIT) {
2471 							bqrelse(bp);
2472 							return (NULL);
2473 						} else {
2474 							bunpin_wait(bp);
2475 						}
2476 					}
2477 					bp->b_flags |= B_NOCACHE;
2478 					bwrite(bp);
2479 				} else {
2480 					if (LIST_FIRST(&bp->b_dep) == NULL) {
2481 						bp->b_flags |= B_RELBUF;
2482 						brelse(bp);
2483 					} else {
2484 						bp->b_flags |= B_NOCACHE;
2485 						bwrite(bp);
2486 					}
2487 				}
2488 				goto loop;
2489 			}
2490 		}
2491 
2492 		/*
2493 		 * If the size is inconsistant in the VMIO case, we can resize
2494 		 * the buffer.  This might lead to B_CACHE getting set or
2495 		 * cleared.  If the size has not changed, B_CACHE remains
2496 		 * unchanged from its previous state.
2497 		 */
2498 
2499 		if (bp->b_bcount != size)
2500 			allocbuf(bp, size);
2501 
2502 		KASSERT(bp->b_offset != NOOFFSET,
2503 		    ("getblk: no buffer offset"));
2504 
2505 		/*
2506 		 * A buffer with B_DELWRI set and B_CACHE clear must
2507 		 * be committed before we can return the buffer in
2508 		 * order to prevent the caller from issuing a read
2509 		 * ( due to B_CACHE not being set ) and overwriting
2510 		 * it.
2511 		 *
2512 		 * Most callers, including NFS and FFS, need this to
2513 		 * operate properly either because they assume they
2514 		 * can issue a read if B_CACHE is not set, or because
2515 		 * ( for example ) an uncached B_DELWRI might loop due
2516 		 * to softupdates re-dirtying the buffer.  In the latter
2517 		 * case, B_CACHE is set after the first write completes,
2518 		 * preventing further loops.
2519 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2520 		 * above while extending the buffer, we cannot allow the
2521 		 * buffer to remain with B_CACHE set after the write
2522 		 * completes or it will represent a corrupt state.  To
2523 		 * deal with this we set B_NOCACHE to scrap the buffer
2524 		 * after the write.
2525 		 *
2526 		 * We might be able to do something fancy, like setting
2527 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2528 		 * so the below call doesn't set B_CACHE, but that gets real
2529 		 * confusing.  This is much easier.
2530 		 */
2531 
2532 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2533 			bp->b_flags |= B_NOCACHE;
2534 			bwrite(bp);
2535 			goto loop;
2536 		}
2537 		bp->b_flags &= ~B_DONE;
2538 	} else {
2539 		int bsize, maxsize, vmio;
2540 		off_t offset;
2541 
2542 		/*
2543 		 * Buffer is not in-core, create new buffer.  The buffer
2544 		 * returned by getnewbuf() is locked.  Note that the returned
2545 		 * buffer is also considered valid (not marked B_INVAL).
2546 		 */
2547 		VI_UNLOCK(vp);
2548 		/*
2549 		 * If the user does not want us to create the buffer, bail out
2550 		 * here.
2551 		 */
2552 		if (flags & GB_NOCREAT)
2553 			return NULL;
2554 		bsize = bo->bo_bsize;
2555 		offset = blkno * bsize;
2556 		vmio = vp->v_object != NULL;
2557 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2558 		maxsize = imax(maxsize, bsize);
2559 
2560 		bp = getnewbuf(slpflag, slptimeo, size, maxsize);
2561 		if (bp == NULL) {
2562 			if (slpflag || slptimeo)
2563 				return NULL;
2564 			goto loop;
2565 		}
2566 
2567 		/*
2568 		 * This code is used to make sure that a buffer is not
2569 		 * created while the getnewbuf routine is blocked.
2570 		 * This can be a problem whether the vnode is locked or not.
2571 		 * If the buffer is created out from under us, we have to
2572 		 * throw away the one we just created.
2573 		 *
2574 		 * Note: this must occur before we associate the buffer
2575 		 * with the vp especially considering limitations in
2576 		 * the splay tree implementation when dealing with duplicate
2577 		 * lblkno's.
2578 		 */
2579 		BO_LOCK(bo);
2580 		if (gbincore(bo, blkno)) {
2581 			BO_UNLOCK(bo);
2582 			bp->b_flags |= B_INVAL;
2583 			brelse(bp);
2584 			goto loop;
2585 		}
2586 
2587 		/*
2588 		 * Insert the buffer into the hash, so that it can
2589 		 * be found by incore.
2590 		 */
2591 		bp->b_blkno = bp->b_lblkno = blkno;
2592 		bp->b_offset = offset;
2593 		bgetvp(vp, bp);
2594 		BO_UNLOCK(bo);
2595 
2596 		/*
2597 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2598 		 * buffer size starts out as 0, B_CACHE will be set by
2599 		 * allocbuf() for the VMIO case prior to it testing the
2600 		 * backing store for validity.
2601 		 */
2602 
2603 		if (vmio) {
2604 			bp->b_flags |= B_VMIO;
2605 #if defined(VFS_BIO_DEBUG)
2606 			if (vn_canvmio(vp) != TRUE)
2607 				printf("getblk: VMIO on vnode type %d\n",
2608 					vp->v_type);
2609 #endif
2610 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2611 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2612 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2613 		} else {
2614 			bp->b_flags &= ~B_VMIO;
2615 			KASSERT(bp->b_bufobj->bo_object == NULL,
2616 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2617 			    bp, bp->b_bufobj->bo_object));
2618 		}
2619 
2620 		allocbuf(bp, size);
2621 		bp->b_flags &= ~B_DONE;
2622 	}
2623 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2624 	KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2625 	KASSERT(bp->b_bufobj == bo,
2626 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2627 	return (bp);
2628 }
2629 
2630 /*
2631  * Get an empty, disassociated buffer of given size.  The buffer is initially
2632  * set to B_INVAL.
2633  */
2634 struct buf *
2635 geteblk(int size)
2636 {
2637 	struct buf *bp;
2638 	int maxsize;
2639 
2640 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2641 	while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2642 		continue;
2643 	allocbuf(bp, size);
2644 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2645 	KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2646 	return (bp);
2647 }
2648 
2649 
2650 /*
2651  * This code constitutes the buffer memory from either anonymous system
2652  * memory (in the case of non-VMIO operations) or from an associated
2653  * VM object (in the case of VMIO operations).  This code is able to
2654  * resize a buffer up or down.
2655  *
2656  * Note that this code is tricky, and has many complications to resolve
2657  * deadlock or inconsistant data situations.  Tread lightly!!!
2658  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2659  * the caller.  Calling this code willy nilly can result in the loss of data.
2660  *
2661  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2662  * B_CACHE for the non-VMIO case.
2663  */
2664 
2665 int
2666 allocbuf(struct buf *bp, int size)
2667 {
2668 	int newbsize, mbsize;
2669 	int i;
2670 
2671 	if (BUF_REFCNT(bp) == 0)
2672 		panic("allocbuf: buffer not busy");
2673 
2674 	if (bp->b_kvasize < size)
2675 		panic("allocbuf: buffer too small");
2676 
2677 	if ((bp->b_flags & B_VMIO) == 0) {
2678 		caddr_t origbuf;
2679 		int origbufsize;
2680 		/*
2681 		 * Just get anonymous memory from the kernel.  Don't
2682 		 * mess with B_CACHE.
2683 		 */
2684 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2685 		if (bp->b_flags & B_MALLOC)
2686 			newbsize = mbsize;
2687 		else
2688 			newbsize = round_page(size);
2689 
2690 		if (newbsize < bp->b_bufsize) {
2691 			/*
2692 			 * malloced buffers are not shrunk
2693 			 */
2694 			if (bp->b_flags & B_MALLOC) {
2695 				if (newbsize) {
2696 					bp->b_bcount = size;
2697 				} else {
2698 					free(bp->b_data, M_BIOBUF);
2699 					if (bp->b_bufsize) {
2700 						atomic_subtract_int(
2701 						    &bufmallocspace,
2702 						    bp->b_bufsize);
2703 						bufspacewakeup();
2704 						bp->b_bufsize = 0;
2705 					}
2706 					bp->b_saveaddr = bp->b_kvabase;
2707 					bp->b_data = bp->b_saveaddr;
2708 					bp->b_bcount = 0;
2709 					bp->b_flags &= ~B_MALLOC;
2710 				}
2711 				return 1;
2712 			}
2713 			vm_hold_free_pages(
2714 			    bp,
2715 			    (vm_offset_t) bp->b_data + newbsize,
2716 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2717 		} else if (newbsize > bp->b_bufsize) {
2718 			/*
2719 			 * We only use malloced memory on the first allocation.
2720 			 * and revert to page-allocated memory when the buffer
2721 			 * grows.
2722 			 */
2723 			/*
2724 			 * There is a potential smp race here that could lead
2725 			 * to bufmallocspace slightly passing the max.  It
2726 			 * is probably extremely rare and not worth worrying
2727 			 * over.
2728 			 */
2729 			if ( (bufmallocspace < maxbufmallocspace) &&
2730 				(bp->b_bufsize == 0) &&
2731 				(mbsize <= PAGE_SIZE/2)) {
2732 
2733 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2734 				bp->b_bufsize = mbsize;
2735 				bp->b_bcount = size;
2736 				bp->b_flags |= B_MALLOC;
2737 				atomic_add_int(&bufmallocspace, mbsize);
2738 				return 1;
2739 			}
2740 			origbuf = NULL;
2741 			origbufsize = 0;
2742 			/*
2743 			 * If the buffer is growing on its other-than-first allocation,
2744 			 * then we revert to the page-allocation scheme.
2745 			 */
2746 			if (bp->b_flags & B_MALLOC) {
2747 				origbuf = bp->b_data;
2748 				origbufsize = bp->b_bufsize;
2749 				bp->b_data = bp->b_kvabase;
2750 				if (bp->b_bufsize) {
2751 					atomic_subtract_int(&bufmallocspace,
2752 					    bp->b_bufsize);
2753 					bufspacewakeup();
2754 					bp->b_bufsize = 0;
2755 				}
2756 				bp->b_flags &= ~B_MALLOC;
2757 				newbsize = round_page(newbsize);
2758 			}
2759 			vm_hold_load_pages(
2760 			    bp,
2761 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2762 			    (vm_offset_t) bp->b_data + newbsize);
2763 			if (origbuf) {
2764 				bcopy(origbuf, bp->b_data, origbufsize);
2765 				free(origbuf, M_BIOBUF);
2766 			}
2767 		}
2768 	} else {
2769 		int desiredpages;
2770 
2771 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2772 		desiredpages = (size == 0) ? 0 :
2773 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2774 
2775 		if (bp->b_flags & B_MALLOC)
2776 			panic("allocbuf: VMIO buffer can't be malloced");
2777 		/*
2778 		 * Set B_CACHE initially if buffer is 0 length or will become
2779 		 * 0-length.
2780 		 */
2781 		if (size == 0 || bp->b_bufsize == 0)
2782 			bp->b_flags |= B_CACHE;
2783 
2784 		if (newbsize < bp->b_bufsize) {
2785 			/*
2786 			 * DEV_BSIZE aligned new buffer size is less then the
2787 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2788 			 * if we have to remove any pages.
2789 			 */
2790 			if (desiredpages < bp->b_npages) {
2791 				vm_page_t m;
2792 
2793 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2794 				vm_page_lock_queues();
2795 				for (i = desiredpages; i < bp->b_npages; i++) {
2796 					/*
2797 					 * the page is not freed here -- it
2798 					 * is the responsibility of
2799 					 * vnode_pager_setsize
2800 					 */
2801 					m = bp->b_pages[i];
2802 					KASSERT(m != bogus_page,
2803 					    ("allocbuf: bogus page found"));
2804 					while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2805 						vm_page_lock_queues();
2806 
2807 					bp->b_pages[i] = NULL;
2808 					vm_page_unwire(m, 0);
2809 				}
2810 				vm_page_unlock_queues();
2811 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2812 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2813 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2814 				bp->b_npages = desiredpages;
2815 			}
2816 		} else if (size > bp->b_bcount) {
2817 			/*
2818 			 * We are growing the buffer, possibly in a
2819 			 * byte-granular fashion.
2820 			 */
2821 			struct vnode *vp;
2822 			vm_object_t obj;
2823 			vm_offset_t toff;
2824 			vm_offset_t tinc;
2825 
2826 			/*
2827 			 * Step 1, bring in the VM pages from the object,
2828 			 * allocating them if necessary.  We must clear
2829 			 * B_CACHE if these pages are not valid for the
2830 			 * range covered by the buffer.
2831 			 */
2832 
2833 			vp = bp->b_vp;
2834 			obj = bp->b_bufobj->bo_object;
2835 
2836 			VM_OBJECT_LOCK(obj);
2837 			while (bp->b_npages < desiredpages) {
2838 				vm_page_t m;
2839 				vm_pindex_t pi;
2840 
2841 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2842 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2843 					/*
2844 					 * note: must allocate system pages
2845 					 * since blocking here could intefere
2846 					 * with paging I/O, no matter which
2847 					 * process we are.
2848 					 */
2849 					m = vm_page_alloc(obj, pi,
2850 					    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2851 					    VM_ALLOC_WIRED);
2852 					if (m == NULL) {
2853 						atomic_add_int(&vm_pageout_deficit,
2854 						    desiredpages - bp->b_npages);
2855 						VM_OBJECT_UNLOCK(obj);
2856 						VM_WAIT;
2857 						VM_OBJECT_LOCK(obj);
2858 					} else {
2859 						bp->b_flags &= ~B_CACHE;
2860 						bp->b_pages[bp->b_npages] = m;
2861 						++bp->b_npages;
2862 					}
2863 					continue;
2864 				}
2865 
2866 				/*
2867 				 * We found a page.  If we have to sleep on it,
2868 				 * retry because it might have gotten freed out
2869 				 * from under us.
2870 				 *
2871 				 * We can only test PG_BUSY here.  Blocking on
2872 				 * m->busy might lead to a deadlock:
2873 				 *
2874 				 *  vm_fault->getpages->cluster_read->allocbuf
2875 				 *
2876 				 */
2877 				vm_page_lock_queues();
2878 				if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2879 					continue;
2880 
2881 				/*
2882 				 * We have a good page.  Should we wakeup the
2883 				 * page daemon?
2884 				 */
2885 				if ((curproc != pageproc) &&
2886 				    (VM_PAGE_INQUEUE1(m, PQ_CACHE)) &&
2887 				    ((cnt.v_free_count + cnt.v_cache_count) <
2888 			 		(cnt.v_free_min + cnt.v_cache_min))) {
2889 					pagedaemon_wakeup();
2890 				}
2891 				vm_page_wire(m);
2892 				vm_page_unlock_queues();
2893 				bp->b_pages[bp->b_npages] = m;
2894 				++bp->b_npages;
2895 			}
2896 
2897 			/*
2898 			 * Step 2.  We've loaded the pages into the buffer,
2899 			 * we have to figure out if we can still have B_CACHE
2900 			 * set.  Note that B_CACHE is set according to the
2901 			 * byte-granular range ( bcount and size ), new the
2902 			 * aligned range ( newbsize ).
2903 			 *
2904 			 * The VM test is against m->valid, which is DEV_BSIZE
2905 			 * aligned.  Needless to say, the validity of the data
2906 			 * needs to also be DEV_BSIZE aligned.  Note that this
2907 			 * fails with NFS if the server or some other client
2908 			 * extends the file's EOF.  If our buffer is resized,
2909 			 * B_CACHE may remain set! XXX
2910 			 */
2911 
2912 			toff = bp->b_bcount;
2913 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2914 
2915 			while ((bp->b_flags & B_CACHE) && toff < size) {
2916 				vm_pindex_t pi;
2917 
2918 				if (tinc > (size - toff))
2919 					tinc = size - toff;
2920 
2921 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2922 				    PAGE_SHIFT;
2923 
2924 				vfs_buf_test_cache(
2925 				    bp,
2926 				    bp->b_offset,
2927 				    toff,
2928 				    tinc,
2929 				    bp->b_pages[pi]
2930 				);
2931 				toff += tinc;
2932 				tinc = PAGE_SIZE;
2933 			}
2934 			VM_OBJECT_UNLOCK(obj);
2935 
2936 			/*
2937 			 * Step 3, fixup the KVM pmap.  Remember that
2938 			 * bp->b_data is relative to bp->b_offset, but
2939 			 * bp->b_offset may be offset into the first page.
2940 			 */
2941 
2942 			bp->b_data = (caddr_t)
2943 			    trunc_page((vm_offset_t)bp->b_data);
2944 			pmap_qenter(
2945 			    (vm_offset_t)bp->b_data,
2946 			    bp->b_pages,
2947 			    bp->b_npages
2948 			);
2949 
2950 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2951 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
2952 		}
2953 	}
2954 	if (newbsize < bp->b_bufsize)
2955 		bufspacewakeup();
2956 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
2957 	bp->b_bcount = size;		/* requested buffer size	*/
2958 	return 1;
2959 }
2960 
2961 void
2962 biodone(struct bio *bp)
2963 {
2964 	void (*done)(struct bio *);
2965 
2966 	mtx_lock(&bdonelock);
2967 	bp->bio_flags |= BIO_DONE;
2968 	done = bp->bio_done;
2969 	if (done == NULL)
2970 		wakeup(bp);
2971 	mtx_unlock(&bdonelock);
2972 	if (done != NULL)
2973 		done(bp);
2974 }
2975 
2976 /*
2977  * Wait for a BIO to finish.
2978  *
2979  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
2980  * case is not yet clear.
2981  */
2982 int
2983 biowait(struct bio *bp, const char *wchan)
2984 {
2985 
2986 	mtx_lock(&bdonelock);
2987 	while ((bp->bio_flags & BIO_DONE) == 0)
2988 		msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
2989 	mtx_unlock(&bdonelock);
2990 	if (bp->bio_error != 0)
2991 		return (bp->bio_error);
2992 	if (!(bp->bio_flags & BIO_ERROR))
2993 		return (0);
2994 	return (EIO);
2995 }
2996 
2997 void
2998 biofinish(struct bio *bp, struct devstat *stat, int error)
2999 {
3000 
3001 	if (error) {
3002 		bp->bio_error = error;
3003 		bp->bio_flags |= BIO_ERROR;
3004 	}
3005 	if (stat != NULL)
3006 		devstat_end_transaction_bio(stat, bp);
3007 	biodone(bp);
3008 }
3009 
3010 /*
3011  *	bufwait:
3012  *
3013  *	Wait for buffer I/O completion, returning error status.  The buffer
3014  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
3015  *	error and cleared.
3016  */
3017 int
3018 bufwait(struct buf *bp)
3019 {
3020 	if (bp->b_iocmd == BIO_READ)
3021 		bwait(bp, PRIBIO, "biord");
3022 	else
3023 		bwait(bp, PRIBIO, "biowr");
3024 	if (bp->b_flags & B_EINTR) {
3025 		bp->b_flags &= ~B_EINTR;
3026 		return (EINTR);
3027 	}
3028 	if (bp->b_ioflags & BIO_ERROR) {
3029 		return (bp->b_error ? bp->b_error : EIO);
3030 	} else {
3031 		return (0);
3032 	}
3033 }
3034 
3035  /*
3036   * Call back function from struct bio back up to struct buf.
3037   */
3038 static void
3039 bufdonebio(struct bio *bip)
3040 {
3041 	struct buf *bp;
3042 
3043 	bp = bip->bio_caller2;
3044 	bp->b_resid = bp->b_bcount - bip->bio_completed;
3045 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
3046 	bp->b_ioflags = bip->bio_flags;
3047 	bp->b_error = bip->bio_error;
3048 	if (bp->b_error)
3049 		bp->b_ioflags |= BIO_ERROR;
3050 	bufdone(bp);
3051 	g_destroy_bio(bip);
3052 }
3053 
3054 void
3055 dev_strategy(struct cdev *dev, struct buf *bp)
3056 {
3057 	struct cdevsw *csw;
3058 	struct bio *bip;
3059 
3060 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3061 		panic("b_iocmd botch");
3062 	for (;;) {
3063 		bip = g_new_bio();
3064 		if (bip != NULL)
3065 			break;
3066 		/* Try again later */
3067 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3068 	}
3069 	bip->bio_cmd = bp->b_iocmd;
3070 	bip->bio_offset = bp->b_iooffset;
3071 	bip->bio_length = bp->b_bcount;
3072 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
3073 	bip->bio_data = bp->b_data;
3074 	bip->bio_done = bufdonebio;
3075 	bip->bio_caller2 = bp;
3076 	bip->bio_dev = dev;
3077 	KASSERT(dev->si_refcount > 0,
3078 	    ("dev_strategy on un-referenced struct cdev *(%s)",
3079 	    devtoname(dev)));
3080 	csw = dev_refthread(dev);
3081 	if (csw == NULL) {
3082 		g_destroy_bio(bip);
3083 		bp->b_error = ENXIO;
3084 		bp->b_ioflags = BIO_ERROR;
3085 		bufdone(bp);
3086 		return;
3087 	}
3088 	(*csw->d_strategy)(bip);
3089 	dev_relthread(dev);
3090 }
3091 
3092 /*
3093  *	bufdone:
3094  *
3095  *	Finish I/O on a buffer, optionally calling a completion function.
3096  *	This is usually called from an interrupt so process blocking is
3097  *	not allowed.
3098  *
3099  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3100  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3101  *	assuming B_INVAL is clear.
3102  *
3103  *	For the VMIO case, we set B_CACHE if the op was a read and no
3104  *	read error occured, or if the op was a write.  B_CACHE is never
3105  *	set if the buffer is invalid or otherwise uncacheable.
3106  *
3107  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3108  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3109  *	in the biodone routine.
3110  */
3111 void
3112 bufdone(struct buf *bp)
3113 {
3114 	struct bufobj *dropobj;
3115 	void    (*biodone)(struct buf *);
3116 
3117 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3118 	dropobj = NULL;
3119 
3120 	KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp,
3121 	    BUF_REFCNT(bp)));
3122 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3123 
3124 	runningbufwakeup(bp);
3125 	if (bp->b_iocmd == BIO_WRITE)
3126 		dropobj = bp->b_bufobj;
3127 	/* call optional completion function if requested */
3128 	if (bp->b_iodone != NULL) {
3129 		biodone = bp->b_iodone;
3130 		bp->b_iodone = NULL;
3131 		(*biodone) (bp);
3132 		if (dropobj)
3133 			bufobj_wdrop(dropobj);
3134 		return;
3135 	}
3136 
3137 	bufdone_finish(bp);
3138 
3139 	if (dropobj)
3140 		bufobj_wdrop(dropobj);
3141 }
3142 
3143 void
3144 bufdone_finish(struct buf *bp)
3145 {
3146 	KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp,
3147 	    BUF_REFCNT(bp)));
3148 
3149 	if (LIST_FIRST(&bp->b_dep) != NULL)
3150 		buf_complete(bp);
3151 
3152 	if (bp->b_flags & B_VMIO) {
3153 		int i;
3154 		vm_ooffset_t foff;
3155 		vm_page_t m;
3156 		vm_object_t obj;
3157 		int iosize;
3158 		struct vnode *vp = bp->b_vp;
3159 
3160 		obj = bp->b_bufobj->bo_object;
3161 
3162 #if defined(VFS_BIO_DEBUG)
3163 		mp_fixme("usecount and vflag accessed without locks.");
3164 		if (vp->v_usecount == 0) {
3165 			panic("biodone: zero vnode ref count");
3166 		}
3167 
3168 		KASSERT(vp->v_object != NULL,
3169 			("biodone: vnode %p has no vm_object", vp));
3170 #endif
3171 
3172 		foff = bp->b_offset;
3173 		KASSERT(bp->b_offset != NOOFFSET,
3174 		    ("biodone: no buffer offset"));
3175 
3176 		VM_OBJECT_LOCK(obj);
3177 #if defined(VFS_BIO_DEBUG)
3178 		if (obj->paging_in_progress < bp->b_npages) {
3179 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3180 			    obj->paging_in_progress, bp->b_npages);
3181 		}
3182 #endif
3183 
3184 		/*
3185 		 * Set B_CACHE if the op was a normal read and no error
3186 		 * occured.  B_CACHE is set for writes in the b*write()
3187 		 * routines.
3188 		 */
3189 		iosize = bp->b_bcount - bp->b_resid;
3190 		if (bp->b_iocmd == BIO_READ &&
3191 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3192 		    !(bp->b_ioflags & BIO_ERROR)) {
3193 			bp->b_flags |= B_CACHE;
3194 		}
3195 		vm_page_lock_queues();
3196 		for (i = 0; i < bp->b_npages; i++) {
3197 			int bogusflag = 0;
3198 			int resid;
3199 
3200 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3201 			if (resid > iosize)
3202 				resid = iosize;
3203 
3204 			/*
3205 			 * cleanup bogus pages, restoring the originals
3206 			 */
3207 			m = bp->b_pages[i];
3208 			if (m == bogus_page) {
3209 				bogusflag = 1;
3210 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3211 				if (m == NULL)
3212 					panic("biodone: page disappeared!");
3213 				bp->b_pages[i] = m;
3214 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3215 				    bp->b_pages, bp->b_npages);
3216 			}
3217 #if defined(VFS_BIO_DEBUG)
3218 			if (OFF_TO_IDX(foff) != m->pindex) {
3219 				printf(
3220 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3221 				    (intmax_t)foff, (uintmax_t)m->pindex);
3222 			}
3223 #endif
3224 
3225 			/*
3226 			 * In the write case, the valid and clean bits are
3227 			 * already changed correctly ( see bdwrite() ), so we
3228 			 * only need to do this here in the read case.
3229 			 */
3230 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3231 				vfs_page_set_valid(bp, foff, i, m);
3232 			}
3233 
3234 			/*
3235 			 * when debugging new filesystems or buffer I/O methods, this
3236 			 * is the most common error that pops up.  if you see this, you
3237 			 * have not set the page busy flag correctly!!!
3238 			 */
3239 			if (m->busy == 0) {
3240 				printf("biodone: page busy < 0, "
3241 				    "pindex: %d, foff: 0x(%x,%x), "
3242 				    "resid: %d, index: %d\n",
3243 				    (int) m->pindex, (int)(foff >> 32),
3244 						(int) foff & 0xffffffff, resid, i);
3245 				if (!vn_isdisk(vp, NULL))
3246 					printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3247 					    (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3248 					    (intmax_t) bp->b_lblkno,
3249 					    bp->b_flags, bp->b_npages);
3250 				else
3251 					printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3252 					    (intmax_t) bp->b_lblkno,
3253 					    bp->b_flags, bp->b_npages);
3254 				printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3255 				    (u_long)m->valid, (u_long)m->dirty,
3256 				    m->wire_count);
3257 				panic("biodone: page busy < 0\n");
3258 			}
3259 			vm_page_io_finish(m);
3260 			vm_object_pip_subtract(obj, 1);
3261 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3262 			iosize -= resid;
3263 		}
3264 		vm_page_unlock_queues();
3265 		vm_object_pip_wakeupn(obj, 0);
3266 		VM_OBJECT_UNLOCK(obj);
3267 	}
3268 
3269 	/*
3270 	 * For asynchronous completions, release the buffer now. The brelse
3271 	 * will do a wakeup there if necessary - so no need to do a wakeup
3272 	 * here in the async case. The sync case always needs to do a wakeup.
3273 	 */
3274 
3275 	if (bp->b_flags & B_ASYNC) {
3276 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3277 			brelse(bp);
3278 		else
3279 			bqrelse(bp);
3280 	} else
3281 		bdone(bp);
3282 }
3283 
3284 /*
3285  * This routine is called in lieu of iodone in the case of
3286  * incomplete I/O.  This keeps the busy status for pages
3287  * consistant.
3288  */
3289 void
3290 vfs_unbusy_pages(struct buf *bp)
3291 {
3292 	int i;
3293 	vm_object_t obj;
3294 	vm_page_t m;
3295 
3296 	runningbufwakeup(bp);
3297 	if (!(bp->b_flags & B_VMIO))
3298 		return;
3299 
3300 	obj = bp->b_bufobj->bo_object;
3301 	VM_OBJECT_LOCK(obj);
3302 	vm_page_lock_queues();
3303 	for (i = 0; i < bp->b_npages; i++) {
3304 		m = bp->b_pages[i];
3305 		if (m == bogus_page) {
3306 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3307 			if (!m)
3308 				panic("vfs_unbusy_pages: page missing\n");
3309 			bp->b_pages[i] = m;
3310 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3311 			    bp->b_pages, bp->b_npages);
3312 		}
3313 		vm_object_pip_subtract(obj, 1);
3314 		vm_page_io_finish(m);
3315 	}
3316 	vm_page_unlock_queues();
3317 	vm_object_pip_wakeupn(obj, 0);
3318 	VM_OBJECT_UNLOCK(obj);
3319 }
3320 
3321 /*
3322  * vfs_page_set_valid:
3323  *
3324  *	Set the valid bits in a page based on the supplied offset.   The
3325  *	range is restricted to the buffer's size.
3326  *
3327  *	This routine is typically called after a read completes.
3328  */
3329 static void
3330 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3331 {
3332 	vm_ooffset_t soff, eoff;
3333 
3334 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3335 	/*
3336 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3337 	 * page boundry or cross the end of the buffer.  The end of the
3338 	 * buffer, in this case, is our file EOF, not the allocation size
3339 	 * of the buffer.
3340 	 */
3341 	soff = off;
3342 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3343 	if (eoff > bp->b_offset + bp->b_bcount)
3344 		eoff = bp->b_offset + bp->b_bcount;
3345 
3346 	/*
3347 	 * Set valid range.  This is typically the entire buffer and thus the
3348 	 * entire page.
3349 	 */
3350 	if (eoff > soff) {
3351 		vm_page_set_validclean(
3352 		    m,
3353 		   (vm_offset_t) (soff & PAGE_MASK),
3354 		   (vm_offset_t) (eoff - soff)
3355 		);
3356 	}
3357 }
3358 
3359 /*
3360  * This routine is called before a device strategy routine.
3361  * It is used to tell the VM system that paging I/O is in
3362  * progress, and treat the pages associated with the buffer
3363  * almost as being PG_BUSY.  Also the object paging_in_progress
3364  * flag is handled to make sure that the object doesn't become
3365  * inconsistant.
3366  *
3367  * Since I/O has not been initiated yet, certain buffer flags
3368  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3369  * and should be ignored.
3370  */
3371 void
3372 vfs_busy_pages(struct buf *bp, int clear_modify)
3373 {
3374 	int i, bogus;
3375 	vm_object_t obj;
3376 	vm_ooffset_t foff;
3377 	vm_page_t m;
3378 
3379 	if (!(bp->b_flags & B_VMIO))
3380 		return;
3381 
3382 	obj = bp->b_bufobj->bo_object;
3383 	foff = bp->b_offset;
3384 	KASSERT(bp->b_offset != NOOFFSET,
3385 	    ("vfs_busy_pages: no buffer offset"));
3386 	vfs_setdirty(bp);
3387 	VM_OBJECT_LOCK(obj);
3388 retry:
3389 	vm_page_lock_queues();
3390 	for (i = 0; i < bp->b_npages; i++) {
3391 		m = bp->b_pages[i];
3392 
3393 		if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3394 			goto retry;
3395 	}
3396 	bogus = 0;
3397 	for (i = 0; i < bp->b_npages; i++) {
3398 		m = bp->b_pages[i];
3399 
3400 		if ((bp->b_flags & B_CLUSTER) == 0) {
3401 			vm_object_pip_add(obj, 1);
3402 			vm_page_io_start(m);
3403 		}
3404 		/*
3405 		 * When readying a buffer for a read ( i.e
3406 		 * clear_modify == 0 ), it is important to do
3407 		 * bogus_page replacement for valid pages in
3408 		 * partially instantiated buffers.  Partially
3409 		 * instantiated buffers can, in turn, occur when
3410 		 * reconstituting a buffer from its VM backing store
3411 		 * base.  We only have to do this if B_CACHE is
3412 		 * clear ( which causes the I/O to occur in the
3413 		 * first place ).  The replacement prevents the read
3414 		 * I/O from overwriting potentially dirty VM-backed
3415 		 * pages.  XXX bogus page replacement is, uh, bogus.
3416 		 * It may not work properly with small-block devices.
3417 		 * We need to find a better way.
3418 		 */
3419 		pmap_remove_all(m);
3420 		if (clear_modify)
3421 			vfs_page_set_valid(bp, foff, i, m);
3422 		else if (m->valid == VM_PAGE_BITS_ALL &&
3423 		    (bp->b_flags & B_CACHE) == 0) {
3424 			bp->b_pages[i] = bogus_page;
3425 			bogus++;
3426 		}
3427 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3428 	}
3429 	vm_page_unlock_queues();
3430 	VM_OBJECT_UNLOCK(obj);
3431 	if (bogus)
3432 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3433 		    bp->b_pages, bp->b_npages);
3434 }
3435 
3436 /*
3437  * Tell the VM system that the pages associated with this buffer
3438  * are clean.  This is used for delayed writes where the data is
3439  * going to go to disk eventually without additional VM intevention.
3440  *
3441  * Note that while we only really need to clean through to b_bcount, we
3442  * just go ahead and clean through to b_bufsize.
3443  */
3444 static void
3445 vfs_clean_pages(struct buf *bp)
3446 {
3447 	int i;
3448 	vm_ooffset_t foff, noff, eoff;
3449 	vm_page_t m;
3450 
3451 	if (!(bp->b_flags & B_VMIO))
3452 		return;
3453 
3454 	foff = bp->b_offset;
3455 	KASSERT(bp->b_offset != NOOFFSET,
3456 	    ("vfs_clean_pages: no buffer offset"));
3457 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3458 	vm_page_lock_queues();
3459 	for (i = 0; i < bp->b_npages; i++) {
3460 		m = bp->b_pages[i];
3461 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3462 		eoff = noff;
3463 
3464 		if (eoff > bp->b_offset + bp->b_bufsize)
3465 			eoff = bp->b_offset + bp->b_bufsize;
3466 		vfs_page_set_valid(bp, foff, i, m);
3467 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3468 		foff = noff;
3469 	}
3470 	vm_page_unlock_queues();
3471 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3472 }
3473 
3474 /*
3475  *	vfs_bio_set_validclean:
3476  *
3477  *	Set the range within the buffer to valid and clean.  The range is
3478  *	relative to the beginning of the buffer, b_offset.  Note that b_offset
3479  *	itself may be offset from the beginning of the first page.
3480  *
3481  */
3482 
3483 void
3484 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3485 {
3486 	int i, n;
3487 	vm_page_t m;
3488 
3489 	if (!(bp->b_flags & B_VMIO))
3490 		return;
3491 	/*
3492 	 * Fixup base to be relative to beginning of first page.
3493 	 * Set initial n to be the maximum number of bytes in the
3494 	 * first page that can be validated.
3495 	 */
3496 
3497 	base += (bp->b_offset & PAGE_MASK);
3498 	n = PAGE_SIZE - (base & PAGE_MASK);
3499 
3500 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3501 	vm_page_lock_queues();
3502 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3503 		m = bp->b_pages[i];
3504 		if (n > size)
3505 			n = size;
3506 		vm_page_set_validclean(m, base & PAGE_MASK, n);
3507 		base += n;
3508 		size -= n;
3509 		n = PAGE_SIZE;
3510 	}
3511 	vm_page_unlock_queues();
3512 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3513 }
3514 
3515 /*
3516  *	vfs_bio_clrbuf:
3517  *
3518  *	clear a buffer.  This routine essentially fakes an I/O, so we need
3519  *	to clear BIO_ERROR and B_INVAL.
3520  *
3521  *	Note that while we only theoretically need to clear through b_bcount,
3522  *	we go ahead and clear through b_bufsize.
3523  */
3524 
3525 void
3526 vfs_bio_clrbuf(struct buf *bp)
3527 {
3528 	int i, j, mask = 0;
3529 	caddr_t sa, ea;
3530 
3531 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3532 		clrbuf(bp);
3533 		return;
3534 	}
3535 
3536 	bp->b_flags &= ~B_INVAL;
3537 	bp->b_ioflags &= ~BIO_ERROR;
3538 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3539 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3540 	    (bp->b_offset & PAGE_MASK) == 0) {
3541 		if (bp->b_pages[0] == bogus_page)
3542 			goto unlock;
3543 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3544 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3545 		if ((bp->b_pages[0]->valid & mask) == mask)
3546 			goto unlock;
3547 		if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3548 		    ((bp->b_pages[0]->valid & mask) == 0)) {
3549 			bzero(bp->b_data, bp->b_bufsize);
3550 			bp->b_pages[0]->valid |= mask;
3551 			goto unlock;
3552 		}
3553 	}
3554 	ea = sa = bp->b_data;
3555 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3556 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3557 		ea = (caddr_t)(vm_offset_t)ulmin(
3558 		    (u_long)(vm_offset_t)ea,
3559 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3560 		if (bp->b_pages[i] == bogus_page)
3561 			continue;
3562 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3563 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3564 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3565 		if ((bp->b_pages[i]->valid & mask) == mask)
3566 			continue;
3567 		if ((bp->b_pages[i]->valid & mask) == 0) {
3568 			if ((bp->b_pages[i]->flags & PG_ZERO) == 0)
3569 				bzero(sa, ea - sa);
3570 		} else {
3571 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3572 				if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3573 				    (bp->b_pages[i]->valid & (1 << j)) == 0)
3574 					bzero(sa, DEV_BSIZE);
3575 			}
3576 		}
3577 		bp->b_pages[i]->valid |= mask;
3578 	}
3579 unlock:
3580 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3581 	bp->b_resid = 0;
3582 }
3583 
3584 /*
3585  * vm_hold_load_pages and vm_hold_free_pages get pages into
3586  * a buffers address space.  The pages are anonymous and are
3587  * not associated with a file object.
3588  */
3589 static void
3590 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3591 {
3592 	vm_offset_t pg;
3593 	vm_page_t p;
3594 	int index;
3595 
3596 	to = round_page(to);
3597 	from = round_page(from);
3598 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3599 
3600 	VM_OBJECT_LOCK(kernel_object);
3601 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3602 tryagain:
3603 		/*
3604 		 * note: must allocate system pages since blocking here
3605 		 * could intefere with paging I/O, no matter which
3606 		 * process we are.
3607 		 */
3608 		p = vm_page_alloc(kernel_object,
3609 			((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3610 		    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3611 		if (!p) {
3612 			atomic_add_int(&vm_pageout_deficit,
3613 			    (to - pg) >> PAGE_SHIFT);
3614 			VM_OBJECT_UNLOCK(kernel_object);
3615 			VM_WAIT;
3616 			VM_OBJECT_LOCK(kernel_object);
3617 			goto tryagain;
3618 		}
3619 		p->valid = VM_PAGE_BITS_ALL;
3620 		pmap_qenter(pg, &p, 1);
3621 		bp->b_pages[index] = p;
3622 	}
3623 	VM_OBJECT_UNLOCK(kernel_object);
3624 	bp->b_npages = index;
3625 }
3626 
3627 /* Return pages associated with this buf to the vm system */
3628 static void
3629 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3630 {
3631 	vm_offset_t pg;
3632 	vm_page_t p;
3633 	int index, newnpages;
3634 
3635 	from = round_page(from);
3636 	to = round_page(to);
3637 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3638 
3639 	VM_OBJECT_LOCK(kernel_object);
3640 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3641 		p = bp->b_pages[index];
3642 		if (p && (index < bp->b_npages)) {
3643 			if (p->busy) {
3644 				printf(
3645 			    "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3646 				    (intmax_t)bp->b_blkno,
3647 				    (intmax_t)bp->b_lblkno);
3648 			}
3649 			bp->b_pages[index] = NULL;
3650 			pmap_qremove(pg, 1);
3651 			vm_page_lock_queues();
3652 			vm_page_unwire(p, 0);
3653 			vm_page_free(p);
3654 			vm_page_unlock_queues();
3655 		}
3656 	}
3657 	VM_OBJECT_UNLOCK(kernel_object);
3658 	bp->b_npages = newnpages;
3659 }
3660 
3661 /*
3662  * Map an IO request into kernel virtual address space.
3663  *
3664  * All requests are (re)mapped into kernel VA space.
3665  * Notice that we use b_bufsize for the size of the buffer
3666  * to be mapped.  b_bcount might be modified by the driver.
3667  *
3668  * Note that even if the caller determines that the address space should
3669  * be valid, a race or a smaller-file mapped into a larger space may
3670  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3671  * check the return value.
3672  */
3673 int
3674 vmapbuf(struct buf *bp)
3675 {
3676 	caddr_t addr, kva;
3677 	vm_prot_t prot;
3678 	int pidx, i;
3679 	struct vm_page *m;
3680 	struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3681 
3682 	if (bp->b_bufsize < 0)
3683 		return (-1);
3684 	prot = VM_PROT_READ;
3685 	if (bp->b_iocmd == BIO_READ)
3686 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3687 	for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3688 	     addr < bp->b_data + bp->b_bufsize;
3689 	     addr += PAGE_SIZE, pidx++) {
3690 		/*
3691 		 * Do the vm_fault if needed; do the copy-on-write thing
3692 		 * when reading stuff off device into memory.
3693 		 *
3694 		 * NOTE! Must use pmap_extract() because addr may be in
3695 		 * the userland address space, and kextract is only guarenteed
3696 		 * to work for the kernland address space (see: sparc64 port).
3697 		 */
3698 retry:
3699 		if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3700 		    prot) < 0) {
3701 			vm_page_lock_queues();
3702 			for (i = 0; i < pidx; ++i) {
3703 				vm_page_unhold(bp->b_pages[i]);
3704 				bp->b_pages[i] = NULL;
3705 			}
3706 			vm_page_unlock_queues();
3707 			return(-1);
3708 		}
3709 		m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3710 		if (m == NULL)
3711 			goto retry;
3712 		bp->b_pages[pidx] = m;
3713 	}
3714 	if (pidx > btoc(MAXPHYS))
3715 		panic("vmapbuf: mapped more than MAXPHYS");
3716 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3717 
3718 	kva = bp->b_saveaddr;
3719 	bp->b_npages = pidx;
3720 	bp->b_saveaddr = bp->b_data;
3721 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3722 	return(0);
3723 }
3724 
3725 /*
3726  * Free the io map PTEs associated with this IO operation.
3727  * We also invalidate the TLB entries and restore the original b_addr.
3728  */
3729 void
3730 vunmapbuf(struct buf *bp)
3731 {
3732 	int pidx;
3733 	int npages;
3734 
3735 	npages = bp->b_npages;
3736 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3737 	vm_page_lock_queues();
3738 	for (pidx = 0; pidx < npages; pidx++)
3739 		vm_page_unhold(bp->b_pages[pidx]);
3740 	vm_page_unlock_queues();
3741 
3742 	bp->b_data = bp->b_saveaddr;
3743 }
3744 
3745 void
3746 bdone(struct buf *bp)
3747 {
3748 
3749 	mtx_lock(&bdonelock);
3750 	bp->b_flags |= B_DONE;
3751 	wakeup(bp);
3752 	mtx_unlock(&bdonelock);
3753 }
3754 
3755 void
3756 bwait(struct buf *bp, u_char pri, const char *wchan)
3757 {
3758 
3759 	mtx_lock(&bdonelock);
3760 	while ((bp->b_flags & B_DONE) == 0)
3761 		msleep(bp, &bdonelock, pri, wchan, 0);
3762 	mtx_unlock(&bdonelock);
3763 }
3764 
3765 int
3766 bufsync(struct bufobj *bo, int waitfor, struct thread *td)
3767 {
3768 
3769 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, td));
3770 }
3771 
3772 void
3773 bufstrategy(struct bufobj *bo, struct buf *bp)
3774 {
3775 	int i = 0;
3776 	struct vnode *vp;
3777 
3778 	vp = bp->b_vp;
3779 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3780 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3781 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3782 	i = VOP_STRATEGY(vp, bp);
3783 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3784 }
3785 
3786 void
3787 bufobj_wrefl(struct bufobj *bo)
3788 {
3789 
3790 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3791 	ASSERT_BO_LOCKED(bo);
3792 	bo->bo_numoutput++;
3793 }
3794 
3795 void
3796 bufobj_wref(struct bufobj *bo)
3797 {
3798 
3799 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3800 	BO_LOCK(bo);
3801 	bo->bo_numoutput++;
3802 	BO_UNLOCK(bo);
3803 }
3804 
3805 void
3806 bufobj_wdrop(struct bufobj *bo)
3807 {
3808 
3809 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3810 	BO_LOCK(bo);
3811 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3812 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3813 		bo->bo_flag &= ~BO_WWAIT;
3814 		wakeup(&bo->bo_numoutput);
3815 	}
3816 	BO_UNLOCK(bo);
3817 }
3818 
3819 int
3820 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3821 {
3822 	int error;
3823 
3824 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3825 	ASSERT_BO_LOCKED(bo);
3826 	error = 0;
3827 	while (bo->bo_numoutput) {
3828 		bo->bo_flag |= BO_WWAIT;
3829 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3830 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3831 		if (error)
3832 			break;
3833 	}
3834 	return (error);
3835 }
3836 
3837 void
3838 bpin(struct buf *bp)
3839 {
3840 	mtx_lock(&bpinlock);
3841 	bp->b_pin_count++;
3842 	mtx_unlock(&bpinlock);
3843 }
3844 
3845 void
3846 bunpin(struct buf *bp)
3847 {
3848 	mtx_lock(&bpinlock);
3849 	if (--bp->b_pin_count == 0)
3850 		wakeup(bp);
3851 	mtx_unlock(&bpinlock);
3852 }
3853 
3854 void
3855 bunpin_wait(struct buf *bp)
3856 {
3857 	mtx_lock(&bpinlock);
3858 	while (bp->b_pin_count > 0)
3859 		msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0);
3860 	mtx_unlock(&bpinlock);
3861 }
3862 
3863 #include "opt_ddb.h"
3864 #ifdef DDB
3865 #include <ddb/ddb.h>
3866 
3867 /* DDB command to show buffer data */
3868 DB_SHOW_COMMAND(buffer, db_show_buffer)
3869 {
3870 	/* get args */
3871 	struct buf *bp = (struct buf *)addr;
3872 
3873 	if (!have_addr) {
3874 		db_printf("usage: show buffer <addr>\n");
3875 		return;
3876 	}
3877 
3878 	db_printf("buf at %p\n", bp);
3879 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3880 	db_printf(
3881 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3882 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n",
3883 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3884 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno);
3885 	if (bp->b_npages) {
3886 		int i;
3887 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3888 		for (i = 0; i < bp->b_npages; i++) {
3889 			vm_page_t m;
3890 			m = bp->b_pages[i];
3891 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3892 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3893 			if ((i + 1) < bp->b_npages)
3894 				db_printf(",");
3895 		}
3896 		db_printf("\n");
3897 	}
3898 	lockmgr_printinfo(&bp->b_lock);
3899 }
3900 
3901 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
3902 {
3903 	struct buf *bp;
3904 	int i;
3905 
3906 	for (i = 0; i < nbuf; i++) {
3907 		bp = &buf[i];
3908 		if (lockcount(&bp->b_lock)) {
3909 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
3910 			db_printf("\n");
3911 		}
3912 	}
3913 }
3914 #endif /* DDB */
3915