xref: /freebsd/sys/kern/vfs_bio.c (revision 3642298923e528d795e3a30ec165d2b469e28b40)
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, "BIO buffer", "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 flushdeps);
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  * Definitions for the buffer free lists.
241  */
242 #define BUFFER_QUEUES	5	/* number of free buffer queues */
243 
244 #define QUEUE_NONE	0	/* on no queue */
245 #define QUEUE_CLEAN	1	/* non-B_DELWRI buffers */
246 #define QUEUE_DIRTY	2	/* B_DELWRI buffers */
247 #define QUEUE_EMPTYKVA	3	/* empty buffer headers w/KVA assignment */
248 #define QUEUE_EMPTY	4	/* empty buffer headers */
249 
250 /* Queues for free buffers with various properties */
251 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
252 
253 /* Lock for the bufqueues */
254 static struct mtx bqlock;
255 
256 /*
257  * Single global constant for BUF_WMESG, to avoid getting multiple references.
258  * buf_wmesg is referred from macros.
259  */
260 const char *buf_wmesg = BUF_WMESG;
261 
262 #define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
263 #define VFS_BIO_NEED_DIRTYFLUSH	0x02	/* waiting for dirty buffer flush */
264 #define VFS_BIO_NEED_FREE	0x04	/* wait for free bufs, hi hysteresis */
265 #define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */
266 
267 #ifdef DIRECTIO
268 extern void ffs_rawread_setup(void);
269 #endif /* DIRECTIO */
270 /*
271  *	numdirtywakeup:
272  *
273  *	If someone is blocked due to there being too many dirty buffers,
274  *	and numdirtybuffers is now reasonable, wake them up.
275  */
276 
277 static __inline void
278 numdirtywakeup(int level)
279 {
280 
281 	if (numdirtybuffers <= level) {
282 		mtx_lock(&nblock);
283 		if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
284 			needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
285 			wakeup(&needsbuffer);
286 		}
287 		mtx_unlock(&nblock);
288 	}
289 }
290 
291 /*
292  *	bufspacewakeup:
293  *
294  *	Called when buffer space is potentially available for recovery.
295  *	getnewbuf() will block on this flag when it is unable to free
296  *	sufficient buffer space.  Buffer space becomes recoverable when
297  *	bp's get placed back in the queues.
298  */
299 
300 static __inline void
301 bufspacewakeup(void)
302 {
303 
304 	/*
305 	 * If someone is waiting for BUF space, wake them up.  Even
306 	 * though we haven't freed the kva space yet, the waiting
307 	 * process will be able to now.
308 	 */
309 	mtx_lock(&nblock);
310 	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
311 		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
312 		wakeup(&needsbuffer);
313 	}
314 	mtx_unlock(&nblock);
315 }
316 
317 /*
318  * runningbufwakeup() - in-progress I/O accounting.
319  *
320  */
321 static __inline void
322 runningbufwakeup(struct buf *bp)
323 {
324 
325 	if (bp->b_runningbufspace) {
326 		atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
327 		bp->b_runningbufspace = 0;
328 		mtx_lock(&rbreqlock);
329 		if (runningbufreq && runningbufspace <= lorunningspace) {
330 			runningbufreq = 0;
331 			wakeup(&runningbufreq);
332 		}
333 		mtx_unlock(&rbreqlock);
334 	}
335 }
336 
337 /*
338  *	bufcountwakeup:
339  *
340  *	Called when a buffer has been added to one of the free queues to
341  *	account for the buffer and to wakeup anyone waiting for free buffers.
342  *	This typically occurs when large amounts of metadata are being handled
343  *	by the buffer cache ( else buffer space runs out first, usually ).
344  */
345 
346 static __inline void
347 bufcountwakeup(void)
348 {
349 
350 	atomic_add_int(&numfreebuffers, 1);
351 	mtx_lock(&nblock);
352 	if (needsbuffer) {
353 		needsbuffer &= ~VFS_BIO_NEED_ANY;
354 		if (numfreebuffers >= hifreebuffers)
355 			needsbuffer &= ~VFS_BIO_NEED_FREE;
356 		wakeup(&needsbuffer);
357 	}
358 	mtx_unlock(&nblock);
359 }
360 
361 /*
362  *	waitrunningbufspace()
363  *
364  *	runningbufspace is a measure of the amount of I/O currently
365  *	running.  This routine is used in async-write situations to
366  *	prevent creating huge backups of pending writes to a device.
367  *	Only asynchronous writes are governed by this function.
368  *
369  *	Reads will adjust runningbufspace, but will not block based on it.
370  *	The read load has a side effect of reducing the allowed write load.
371  *
372  *	This does NOT turn an async write into a sync write.  It waits
373  *	for earlier writes to complete and generally returns before the
374  *	caller's write has reached the device.
375  */
376 static __inline void
377 waitrunningbufspace(void)
378 {
379 
380 	mtx_lock(&rbreqlock);
381 	while (runningbufspace > hirunningspace) {
382 		++runningbufreq;
383 		msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
384 	}
385 	mtx_unlock(&rbreqlock);
386 }
387 
388 
389 /*
390  *	vfs_buf_test_cache:
391  *
392  *	Called when a buffer is extended.  This function clears the B_CACHE
393  *	bit if the newly extended portion of the buffer does not contain
394  *	valid data.
395  */
396 static __inline
397 void
398 vfs_buf_test_cache(struct buf *bp,
399 		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
400 		  vm_page_t m)
401 {
402 
403 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
404 	if (bp->b_flags & B_CACHE) {
405 		int base = (foff + off) & PAGE_MASK;
406 		if (vm_page_is_valid(m, base, size) == 0)
407 			bp->b_flags &= ~B_CACHE;
408 	}
409 }
410 
411 /* Wake up the buffer deamon if necessary */
412 static __inline
413 void
414 bd_wakeup(int dirtybuflevel)
415 {
416 
417 	mtx_lock(&bdlock);
418 	if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
419 		bd_request = 1;
420 		wakeup(&bd_request);
421 	}
422 	mtx_unlock(&bdlock);
423 }
424 
425 /*
426  * bd_speedup - speedup the buffer cache flushing code
427  */
428 
429 static __inline
430 void
431 bd_speedup(void)
432 {
433 
434 	bd_wakeup(1);
435 }
436 
437 /*
438  * Calculating buffer cache scaling values and reserve space for buffer
439  * headers.  This is called during low level kernel initialization and
440  * may be called more then once.  We CANNOT write to the memory area
441  * being reserved at this time.
442  */
443 caddr_t
444 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
445 {
446 
447 	/*
448 	 * physmem_est is in pages.  Convert it to kilobytes (assumes
449 	 * PAGE_SIZE is >= 1K)
450 	 */
451 	physmem_est = physmem_est * (PAGE_SIZE / 1024);
452 
453 	/*
454 	 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
455 	 * For the first 64MB of ram nominally allocate sufficient buffers to
456 	 * cover 1/4 of our ram.  Beyond the first 64MB allocate additional
457 	 * buffers to cover 1/20 of our ram over 64MB.  When auto-sizing
458 	 * the buffer cache we limit the eventual kva reservation to
459 	 * maxbcache bytes.
460 	 *
461 	 * factor represents the 1/4 x ram conversion.
462 	 */
463 	if (nbuf == 0) {
464 		int factor = 4 * BKVASIZE / 1024;
465 
466 		nbuf = 50;
467 		if (physmem_est > 4096)
468 			nbuf += min((physmem_est - 4096) / factor,
469 			    65536 / factor);
470 		if (physmem_est > 65536)
471 			nbuf += (physmem_est - 65536) * 2 / (factor * 5);
472 
473 		if (maxbcache && nbuf > maxbcache / BKVASIZE)
474 			nbuf = maxbcache / BKVASIZE;
475 	}
476 
477 #if 0
478 	/*
479 	 * Do not allow the buffer_map to be more then 1/2 the size of the
480 	 * kernel_map.
481 	 */
482 	if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
483 	    (BKVASIZE * 2)) {
484 		nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
485 		    (BKVASIZE * 2);
486 		printf("Warning: nbufs capped at %d\n", nbuf);
487 	}
488 #endif
489 
490 	/*
491 	 * swbufs are used as temporary holders for I/O, such as paging I/O.
492 	 * We have no less then 16 and no more then 256.
493 	 */
494 	nswbuf = max(min(nbuf/4, 256), 16);
495 #ifdef NSWBUF_MIN
496 	if (nswbuf < NSWBUF_MIN)
497 		nswbuf = NSWBUF_MIN;
498 #endif
499 #ifdef DIRECTIO
500 	ffs_rawread_setup();
501 #endif
502 
503 	/*
504 	 * Reserve space for the buffer cache buffers
505 	 */
506 	swbuf = (void *)v;
507 	v = (caddr_t)(swbuf + nswbuf);
508 	buf = (void *)v;
509 	v = (caddr_t)(buf + nbuf);
510 
511 	return(v);
512 }
513 
514 /* Initialize the buffer subsystem.  Called before use of any buffers. */
515 void
516 bufinit(void)
517 {
518 	struct buf *bp;
519 	int i;
520 
521 	mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
522 	mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
523 	mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
524 	mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
525 	mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
526 
527 	/* next, make a null set of free lists */
528 	for (i = 0; i < BUFFER_QUEUES; i++)
529 		TAILQ_INIT(&bufqueues[i]);
530 
531 	/* finally, initialize each buffer header and stick on empty q */
532 	for (i = 0; i < nbuf; i++) {
533 		bp = &buf[i];
534 		bzero(bp, sizeof *bp);
535 		bp->b_flags = B_INVAL;	/* we're just an empty header */
536 		bp->b_rcred = NOCRED;
537 		bp->b_wcred = NOCRED;
538 		bp->b_qindex = QUEUE_EMPTY;
539 		bp->b_vflags = 0;
540 		bp->b_xflags = 0;
541 		LIST_INIT(&bp->b_dep);
542 		BUF_LOCKINIT(bp);
543 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
544 	}
545 
546 	/*
547 	 * maxbufspace is the absolute maximum amount of buffer space we are
548 	 * allowed to reserve in KVM and in real terms.  The absolute maximum
549 	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
550 	 * used by most other processes.  The differential is required to
551 	 * ensure that buf_daemon is able to run when other processes might
552 	 * be blocked waiting for buffer space.
553 	 *
554 	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
555 	 * this may result in KVM fragmentation which is not handled optimally
556 	 * by the system.
557 	 */
558 	maxbufspace = nbuf * BKVASIZE;
559 	hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
560 	lobufspace = hibufspace - MAXBSIZE;
561 
562 	lorunningspace = 512 * 1024;
563 	hirunningspace = 1024 * 1024;
564 
565 /*
566  * Limit the amount of malloc memory since it is wired permanently into
567  * the kernel space.  Even though this is accounted for in the buffer
568  * allocation, we don't want the malloced region to grow uncontrolled.
569  * The malloc scheme improves memory utilization significantly on average
570  * (small) directories.
571  */
572 	maxbufmallocspace = hibufspace / 20;
573 
574 /*
575  * Reduce the chance of a deadlock occuring by limiting the number
576  * of delayed-write dirty buffers we allow to stack up.
577  */
578 	hidirtybuffers = nbuf / 4 + 20;
579 	dirtybufthresh = hidirtybuffers * 9 / 10;
580 	numdirtybuffers = 0;
581 /*
582  * To support extreme low-memory systems, make sure hidirtybuffers cannot
583  * eat up all available buffer space.  This occurs when our minimum cannot
584  * be met.  We try to size hidirtybuffers to 3/4 our buffer space assuming
585  * BKVASIZE'd (8K) buffers.
586  */
587 	while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
588 		hidirtybuffers >>= 1;
589 	}
590 	lodirtybuffers = hidirtybuffers / 2;
591 
592 /*
593  * Try to keep the number of free buffers in the specified range,
594  * and give special processes (e.g. like buf_daemon) access to an
595  * emergency reserve.
596  */
597 	lofreebuffers = nbuf / 18 + 5;
598 	hifreebuffers = 2 * lofreebuffers;
599 	numfreebuffers = nbuf;
600 
601 /*
602  * Maximum number of async ops initiated per buf_daemon loop.  This is
603  * somewhat of a hack at the moment, we really need to limit ourselves
604  * based on the number of bytes of I/O in-transit that were initiated
605  * from buf_daemon.
606  */
607 
608 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
609 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
610 }
611 
612 /*
613  * bfreekva() - free the kva allocation for a buffer.
614  *
615  *	Since this call frees up buffer space, we call bufspacewakeup().
616  */
617 static void
618 bfreekva(struct buf *bp)
619 {
620 
621 	if (bp->b_kvasize) {
622 		atomic_add_int(&buffreekvacnt, 1);
623 		atomic_subtract_int(&bufspace, bp->b_kvasize);
624 		vm_map_lock(buffer_map);
625 		vm_map_delete(buffer_map,
626 		    (vm_offset_t) bp->b_kvabase,
627 		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize
628 		);
629 		vm_map_unlock(buffer_map);
630 		bp->b_kvasize = 0;
631 		bufspacewakeup();
632 	}
633 }
634 
635 /*
636  *	bremfree:
637  *
638  *	Mark the buffer for removal from the appropriate free list in brelse.
639  *
640  */
641 void
642 bremfree(struct buf *bp)
643 {
644 
645 	CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
646 	KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked."));
647 	KASSERT((bp->b_flags & B_REMFREE) == 0,
648 	    ("bremfree: buffer %p already marked for delayed removal.", bp));
649 	KASSERT(bp->b_qindex != QUEUE_NONE,
650 	    ("bremfree: buffer %p not on a queue.", bp));
651 
652 	bp->b_flags |= B_REMFREE;
653 	/* Fixup numfreebuffers count.  */
654 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
655 		atomic_subtract_int(&numfreebuffers, 1);
656 }
657 
658 /*
659  *	bremfreef:
660  *
661  *	Force an immediate removal from a free list.  Used only in nfs when
662  *	it abuses the b_freelist pointer.
663  */
664 void
665 bremfreef(struct buf *bp)
666 {
667 	mtx_lock(&bqlock);
668 	bremfreel(bp);
669 	mtx_unlock(&bqlock);
670 }
671 
672 /*
673  *	bremfreel:
674  *
675  *	Removes a buffer from the free list, must be called with the
676  *	bqlock held.
677  */
678 static void
679 bremfreel(struct buf *bp)
680 {
681 	CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
682 	    bp, bp->b_vp, bp->b_flags);
683 	KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp));
684 	KASSERT(bp->b_qindex != QUEUE_NONE,
685 	    ("bremfreel: buffer %p not on a queue.", bp));
686 	mtx_assert(&bqlock, MA_OWNED);
687 
688 	TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
689 	bp->b_qindex = QUEUE_NONE;
690 	/*
691 	 * If this was a delayed bremfree() we only need to remove the buffer
692 	 * from the queue and return the stats are already done.
693 	 */
694 	if (bp->b_flags & B_REMFREE) {
695 		bp->b_flags &= ~B_REMFREE;
696 		return;
697 	}
698 	/*
699 	 * Fixup numfreebuffers count.  If the buffer is invalid or not
700 	 * delayed-write, the buffer was free and we must decrement
701 	 * numfreebuffers.
702 	 */
703 	if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
704 		atomic_subtract_int(&numfreebuffers, 1);
705 }
706 
707 
708 /*
709  * Get a buffer with the specified data.  Look in the cache first.  We
710  * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
711  * is set, the buffer is valid and we do not have to do anything ( see
712  * getblk() ).  This is really just a special case of breadn().
713  */
714 int
715 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
716     struct buf **bpp)
717 {
718 
719 	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
720 }
721 
722 /*
723  * Operates like bread, but also starts asynchronous I/O on
724  * read-ahead blocks.  We must clear BIO_ERROR and B_INVAL prior
725  * to initiating I/O . If B_CACHE is set, the buffer is valid
726  * and we do not have to do anything.
727  */
728 int
729 breadn(struct vnode * vp, daddr_t blkno, int size,
730     daddr_t * rablkno, int *rabsize,
731     int cnt, struct ucred * cred, struct buf **bpp)
732 {
733 	struct buf *bp, *rabp;
734 	int i;
735 	int rv = 0, readwait = 0;
736 
737 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
738 	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
739 
740 	/* if not found in cache, do some I/O */
741 	if ((bp->b_flags & B_CACHE) == 0) {
742 		if (curthread != PCPU_GET(idlethread))
743 			curthread->td_proc->p_stats->p_ru.ru_inblock++;
744 		bp->b_iocmd = BIO_READ;
745 		bp->b_flags &= ~B_INVAL;
746 		bp->b_ioflags &= ~BIO_ERROR;
747 		if (bp->b_rcred == NOCRED && cred != NOCRED)
748 			bp->b_rcred = crhold(cred);
749 		vfs_busy_pages(bp, 0);
750 		bp->b_iooffset = dbtob(bp->b_blkno);
751 		bstrategy(bp);
752 		++readwait;
753 	}
754 
755 	for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
756 		if (inmem(vp, *rablkno))
757 			continue;
758 		rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
759 
760 		if ((rabp->b_flags & B_CACHE) == 0) {
761 			if (curthread != PCPU_GET(idlethread))
762 				curthread->td_proc->p_stats->p_ru.ru_inblock++;
763 			rabp->b_flags |= B_ASYNC;
764 			rabp->b_flags &= ~B_INVAL;
765 			rabp->b_ioflags &= ~BIO_ERROR;
766 			rabp->b_iocmd = BIO_READ;
767 			if (rabp->b_rcred == NOCRED && cred != NOCRED)
768 				rabp->b_rcred = crhold(cred);
769 			vfs_busy_pages(rabp, 0);
770 			BUF_KERNPROC(rabp);
771 			rabp->b_iooffset = dbtob(rabp->b_blkno);
772 			bstrategy(rabp);
773 		} else {
774 			brelse(rabp);
775 		}
776 	}
777 
778 	if (readwait) {
779 		rv = bufwait(bp);
780 	}
781 	return (rv);
782 }
783 
784 /*
785  * Write, release buffer on completion.  (Done by iodone
786  * if async).  Do not bother writing anything if the buffer
787  * is invalid.
788  *
789  * Note that we set B_CACHE here, indicating that buffer is
790  * fully valid and thus cacheable.  This is true even of NFS
791  * now so we set it generally.  This could be set either here
792  * or in biodone() since the I/O is synchronous.  We put it
793  * here.
794  */
795 int
796 bufwrite(struct buf *bp)
797 {
798 	int oldflags;
799 
800 	CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
801 	if (bp->b_flags & B_INVAL) {
802 		brelse(bp);
803 		return (0);
804 	}
805 
806 	oldflags = bp->b_flags;
807 
808 	if (BUF_REFCNT(bp) == 0)
809 		panic("bufwrite: buffer is not busy???");
810 	KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
811 	    ("FFS background buffer should not get here %p", bp));
812 
813 	/* Mark the buffer clean */
814 	bundirty(bp);
815 
816 	bp->b_flags &= ~B_DONE;
817 	bp->b_ioflags &= ~BIO_ERROR;
818 	bp->b_flags |= B_CACHE;
819 	bp->b_iocmd = BIO_WRITE;
820 
821 	bufobj_wref(bp->b_bufobj);
822 	vfs_busy_pages(bp, 1);
823 
824 	/*
825 	 * Normal bwrites pipeline writes
826 	 */
827 	bp->b_runningbufspace = bp->b_bufsize;
828 	atomic_add_int(&runningbufspace, bp->b_runningbufspace);
829 
830 	if (curthread != PCPU_GET(idlethread))
831 		curthread->td_proc->p_stats->p_ru.ru_oublock++;
832 	if (oldflags & B_ASYNC)
833 		BUF_KERNPROC(bp);
834 	bp->b_iooffset = dbtob(bp->b_blkno);
835 	bstrategy(bp);
836 
837 	if ((oldflags & B_ASYNC) == 0) {
838 		int rtval = bufwait(bp);
839 		brelse(bp);
840 		return (rtval);
841 	} else {
842 		/*
843 		 * don't allow the async write to saturate the I/O
844 		 * system.  We will not deadlock here because
845 		 * we are blocking waiting for I/O that is already in-progress
846 		 * to complete. We do not block here if it is the update
847 		 * or syncer daemon trying to clean up as that can lead
848 		 * to deadlock.
849 		 */
850 		if (curthread->td_proc != bufdaemonproc &&
851 		    curthread->td_proc != updateproc)
852 			waitrunningbufspace();
853 	}
854 
855 	return (0);
856 }
857 
858 /*
859  * Delayed write. (Buffer is marked dirty).  Do not bother writing
860  * anything if the buffer is marked invalid.
861  *
862  * Note that since the buffer must be completely valid, we can safely
863  * set B_CACHE.  In fact, we have to set B_CACHE here rather then in
864  * biodone() in order to prevent getblk from writing the buffer
865  * out synchronously.
866  */
867 void
868 bdwrite(struct buf *bp)
869 {
870 	struct thread *td = curthread;
871 	struct vnode *vp;
872 	struct buf *nbp;
873 	struct bufobj *bo;
874 
875 	CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
876 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
877 	KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy"));
878 
879 	if (bp->b_flags & B_INVAL) {
880 		brelse(bp);
881 		return;
882 	}
883 
884 	/*
885 	 * If we have too many dirty buffers, don't create any more.
886 	 * If we are wildly over our limit, then force a complete
887 	 * cleanup. Otherwise, just keep the situation from getting
888 	 * out of control. Note that we have to avoid a recursive
889 	 * disaster and not try to clean up after our own cleanup!
890 	 */
891 	vp = bp->b_vp;
892 	bo = bp->b_bufobj;
893 	if ((td->td_pflags & TDP_COWINPROGRESS) == 0) {
894 		BO_LOCK(bo);
895 		if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
896 			BO_UNLOCK(bo);
897 			(void) VOP_FSYNC(vp, MNT_NOWAIT, td);
898 			altbufferflushes++;
899 		} else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
900 			/*
901 			 * Try to find a buffer to flush.
902 			 */
903 			TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
904 				if ((nbp->b_vflags & BV_BKGRDINPROG) ||
905 				    BUF_LOCK(nbp,
906 				    LK_EXCLUSIVE | LK_NOWAIT, NULL))
907 					continue;
908 				if (bp == nbp)
909 					panic("bdwrite: found ourselves");
910 				BO_UNLOCK(bo);
911 				/* Don't countdeps with the bo lock held. */
912 				if (buf_countdeps(nbp, 0)) {
913 					BO_LOCK(bo);
914 					BUF_UNLOCK(nbp);
915 					continue;
916 				}
917 				if (nbp->b_flags & B_CLUSTEROK) {
918 					vfs_bio_awrite(nbp);
919 				} else {
920 					bremfree(nbp);
921 					bawrite(nbp);
922 				}
923 				dirtybufferflushes++;
924 				break;
925 			}
926 			if (nbp == NULL)
927 				BO_UNLOCK(bo);
928 		} else
929 			BO_UNLOCK(bo);
930 	} else
931 		recursiveflushes++;
932 
933 	bdirty(bp);
934 	/*
935 	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
936 	 * true even of NFS now.
937 	 */
938 	bp->b_flags |= B_CACHE;
939 
940 	/*
941 	 * This bmap keeps the system from needing to do the bmap later,
942 	 * perhaps when the system is attempting to do a sync.  Since it
943 	 * is likely that the indirect block -- or whatever other datastructure
944 	 * that the filesystem needs is still in memory now, it is a good
945 	 * thing to do this.  Note also, that if the pageout daemon is
946 	 * requesting a sync -- there might not be enough memory to do
947 	 * the bmap then...  So, this is important to do.
948 	 */
949 	if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
950 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
951 	}
952 
953 	/*
954 	 * Set the *dirty* buffer range based upon the VM system dirty pages.
955 	 */
956 	vfs_setdirty(bp);
957 
958 	/*
959 	 * We need to do this here to satisfy the vnode_pager and the
960 	 * pageout daemon, so that it thinks that the pages have been
961 	 * "cleaned".  Note that since the pages are in a delayed write
962 	 * buffer -- the VFS layer "will" see that the pages get written
963 	 * out on the next sync, or perhaps the cluster will be completed.
964 	 */
965 	vfs_clean_pages(bp);
966 	bqrelse(bp);
967 
968 	/*
969 	 * Wakeup the buffer flushing daemon if we have a lot of dirty
970 	 * buffers (midpoint between our recovery point and our stall
971 	 * point).
972 	 */
973 	bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
974 
975 	/*
976 	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
977 	 * due to the softdep code.
978 	 */
979 }
980 
981 /*
982  *	bdirty:
983  *
984  *	Turn buffer into delayed write request.  We must clear BIO_READ and
985  *	B_RELBUF, and we must set B_DELWRI.  We reassign the buffer to
986  *	itself to properly update it in the dirty/clean lists.  We mark it
987  *	B_DONE to ensure that any asynchronization of the buffer properly
988  *	clears B_DONE ( else a panic will occur later ).
989  *
990  *	bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
991  *	might have been set pre-getblk().  Unlike bwrite/bdwrite, bdirty()
992  *	should only be called if the buffer is known-good.
993  *
994  *	Since the buffer is not on a queue, we do not update the numfreebuffers
995  *	count.
996  *
997  *	The buffer must be on QUEUE_NONE.
998  */
999 void
1000 bdirty(struct buf *bp)
1001 {
1002 
1003 	CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1004 	    bp, bp->b_vp, bp->b_flags);
1005 	KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp));
1006 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1007 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1008 	    ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1009 	bp->b_flags &= ~(B_RELBUF);
1010 	bp->b_iocmd = BIO_WRITE;
1011 
1012 	if ((bp->b_flags & B_DELWRI) == 0) {
1013 		bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1014 		reassignbuf(bp);
1015 		atomic_add_int(&numdirtybuffers, 1);
1016 		bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1017 	}
1018 }
1019 
1020 /*
1021  *	bundirty:
1022  *
1023  *	Clear B_DELWRI for buffer.
1024  *
1025  *	Since the buffer is not on a queue, we do not update the numfreebuffers
1026  *	count.
1027  *
1028  *	The buffer must be on QUEUE_NONE.
1029  */
1030 
1031 void
1032 bundirty(struct buf *bp)
1033 {
1034 
1035 	CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1036 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1037 	KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1038 	    ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1039 	KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp));
1040 
1041 	if (bp->b_flags & B_DELWRI) {
1042 		bp->b_flags &= ~B_DELWRI;
1043 		reassignbuf(bp);
1044 		atomic_subtract_int(&numdirtybuffers, 1);
1045 		numdirtywakeup(lodirtybuffers);
1046 	}
1047 	/*
1048 	 * Since it is now being written, we can clear its deferred write flag.
1049 	 */
1050 	bp->b_flags &= ~B_DEFERRED;
1051 }
1052 
1053 /*
1054  *	bawrite:
1055  *
1056  *	Asynchronous write.  Start output on a buffer, but do not wait for
1057  *	it to complete.  The buffer is released when the output completes.
1058  *
1059  *	bwrite() ( or the VOP routine anyway ) is responsible for handling
1060  *	B_INVAL buffers.  Not us.
1061  */
1062 void
1063 bawrite(struct buf *bp)
1064 {
1065 
1066 	bp->b_flags |= B_ASYNC;
1067 	(void) bwrite(bp);
1068 }
1069 
1070 /*
1071  *	bwillwrite:
1072  *
1073  *	Called prior to the locking of any vnodes when we are expecting to
1074  *	write.  We do not want to starve the buffer cache with too many
1075  *	dirty buffers so we block here.  By blocking prior to the locking
1076  *	of any vnodes we attempt to avoid the situation where a locked vnode
1077  *	prevents the various system daemons from flushing related buffers.
1078  */
1079 
1080 void
1081 bwillwrite(void)
1082 {
1083 
1084 	if (numdirtybuffers >= hidirtybuffers) {
1085 		mtx_lock(&nblock);
1086 		while (numdirtybuffers >= hidirtybuffers) {
1087 			bd_wakeup(1);
1088 			needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1089 			msleep(&needsbuffer, &nblock,
1090 			    (PRIBIO + 4), "flswai", 0);
1091 		}
1092 		mtx_unlock(&nblock);
1093 	}
1094 }
1095 
1096 /*
1097  * Return true if we have too many dirty buffers.
1098  */
1099 int
1100 buf_dirty_count_severe(void)
1101 {
1102 
1103 	return(numdirtybuffers >= hidirtybuffers);
1104 }
1105 
1106 /*
1107  *	brelse:
1108  *
1109  *	Release a busy buffer and, if requested, free its resources.  The
1110  *	buffer will be stashed in the appropriate bufqueue[] allowing it
1111  *	to be accessed later as a cache entity or reused for other purposes.
1112  */
1113 void
1114 brelse(struct buf *bp)
1115 {
1116 	CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1117 	    bp, bp->b_vp, bp->b_flags);
1118 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1119 	    ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1120 
1121 	if (bp->b_iocmd == BIO_WRITE &&
1122 	    (bp->b_ioflags & BIO_ERROR) &&
1123 	    !(bp->b_flags & B_INVAL)) {
1124 		/*
1125 		 * Failed write, redirty.  Must clear BIO_ERROR to prevent
1126 		 * pages from being scrapped.  If B_INVAL is set then
1127 		 * this case is not run and the next case is run to
1128 		 * destroy the buffer.  B_INVAL can occur if the buffer
1129 		 * is outside the range supported by the underlying device.
1130 		 */
1131 		bp->b_ioflags &= ~BIO_ERROR;
1132 		bdirty(bp);
1133 	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1134 	    (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1135 		/*
1136 		 * Either a failed I/O or we were asked to free or not
1137 		 * cache the buffer.
1138 		 */
1139 		bp->b_flags |= B_INVAL;
1140 		if (LIST_FIRST(&bp->b_dep) != NULL)
1141 			buf_deallocate(bp);
1142 		if (bp->b_flags & B_DELWRI) {
1143 			atomic_subtract_int(&numdirtybuffers, 1);
1144 			numdirtywakeup(lodirtybuffers);
1145 		}
1146 		bp->b_flags &= ~(B_DELWRI | B_CACHE);
1147 		if ((bp->b_flags & B_VMIO) == 0) {
1148 			if (bp->b_bufsize)
1149 				allocbuf(bp, 0);
1150 			if (bp->b_vp)
1151 				brelvp(bp);
1152 		}
1153 	}
1154 
1155 	/*
1156 	 * We must clear B_RELBUF if B_DELWRI is set.  If vfs_vmio_release()
1157 	 * is called with B_DELWRI set, the underlying pages may wind up
1158 	 * getting freed causing a previous write (bdwrite()) to get 'lost'
1159 	 * because pages associated with a B_DELWRI bp are marked clean.
1160 	 *
1161 	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1162 	 * if B_DELWRI is set.
1163 	 *
1164 	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1165 	 * on pages to return pages to the VM page queues.
1166 	 */
1167 	if (bp->b_flags & B_DELWRI)
1168 		bp->b_flags &= ~B_RELBUF;
1169 	else if (vm_page_count_severe()) {
1170 		/*
1171 		 * XXX This lock may not be necessary since BKGRDINPROG
1172 		 * cannot be set while we hold the buf lock, it can only be
1173 		 * cleared if it is already pending.
1174 		 */
1175 		if (bp->b_vp) {
1176 			BO_LOCK(bp->b_bufobj);
1177 			if (!(bp->b_vflags & BV_BKGRDINPROG))
1178 				bp->b_flags |= B_RELBUF;
1179 			BO_UNLOCK(bp->b_bufobj);
1180 		} else
1181 			bp->b_flags |= B_RELBUF;
1182 	}
1183 
1184 	/*
1185 	 * VMIO buffer rundown.  It is not very necessary to keep a VMIO buffer
1186 	 * constituted, not even NFS buffers now.  Two flags effect this.  If
1187 	 * B_INVAL, the struct buf is invalidated but the VM object is kept
1188 	 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1189 	 *
1190 	 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1191 	 * invalidated.  BIO_ERROR cannot be set for a failed write unless the
1192 	 * buffer is also B_INVAL because it hits the re-dirtying code above.
1193 	 *
1194 	 * Normally we can do this whether a buffer is B_DELWRI or not.  If
1195 	 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1196 	 * the commit state and we cannot afford to lose the buffer. If the
1197 	 * buffer has a background write in progress, we need to keep it
1198 	 * around to prevent it from being reconstituted and starting a second
1199 	 * background write.
1200 	 */
1201 	if ((bp->b_flags & B_VMIO)
1202 	    && !(bp->b_vp->v_mount != NULL &&
1203 		 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1204 		 !vn_isdisk(bp->b_vp, NULL) &&
1205 		 (bp->b_flags & B_DELWRI))
1206 	    ) {
1207 
1208 		int i, j, resid;
1209 		vm_page_t m;
1210 		off_t foff;
1211 		vm_pindex_t poff;
1212 		vm_object_t obj;
1213 
1214 		obj = bp->b_bufobj->bo_object;
1215 
1216 		/*
1217 		 * Get the base offset and length of the buffer.  Note that
1218 		 * in the VMIO case if the buffer block size is not
1219 		 * page-aligned then b_data pointer may not be page-aligned.
1220 		 * But our b_pages[] array *IS* page aligned.
1221 		 *
1222 		 * block sizes less then DEV_BSIZE (usually 512) are not
1223 		 * supported due to the page granularity bits (m->valid,
1224 		 * m->dirty, etc...).
1225 		 *
1226 		 * See man buf(9) for more information
1227 		 */
1228 		resid = bp->b_bufsize;
1229 		foff = bp->b_offset;
1230 		VM_OBJECT_LOCK(obj);
1231 		for (i = 0; i < bp->b_npages; i++) {
1232 			int had_bogus = 0;
1233 
1234 			m = bp->b_pages[i];
1235 
1236 			/*
1237 			 * If we hit a bogus page, fixup *all* the bogus pages
1238 			 * now.
1239 			 */
1240 			if (m == bogus_page) {
1241 				poff = OFF_TO_IDX(bp->b_offset);
1242 				had_bogus = 1;
1243 
1244 				for (j = i; j < bp->b_npages; j++) {
1245 					vm_page_t mtmp;
1246 					mtmp = bp->b_pages[j];
1247 					if (mtmp == bogus_page) {
1248 						mtmp = vm_page_lookup(obj, poff + j);
1249 						if (!mtmp) {
1250 							panic("brelse: page missing\n");
1251 						}
1252 						bp->b_pages[j] = mtmp;
1253 					}
1254 				}
1255 
1256 				if ((bp->b_flags & B_INVAL) == 0) {
1257 					pmap_qenter(
1258 					    trunc_page((vm_offset_t)bp->b_data),
1259 					    bp->b_pages, bp->b_npages);
1260 				}
1261 				m = bp->b_pages[i];
1262 			}
1263 			if ((bp->b_flags & B_NOCACHE) ||
1264 			    (bp->b_ioflags & BIO_ERROR)) {
1265 				int poffset = foff & PAGE_MASK;
1266 				int presid = resid > (PAGE_SIZE - poffset) ?
1267 					(PAGE_SIZE - poffset) : resid;
1268 
1269 				KASSERT(presid >= 0, ("brelse: extra page"));
1270 				vm_page_lock_queues();
1271 				vm_page_set_invalid(m, poffset, presid);
1272 				vm_page_unlock_queues();
1273 				if (had_bogus)
1274 					printf("avoided corruption bug in bogus_page/brelse code\n");
1275 			}
1276 			resid -= PAGE_SIZE - (foff & PAGE_MASK);
1277 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1278 		}
1279 		VM_OBJECT_UNLOCK(obj);
1280 		if (bp->b_flags & (B_INVAL | B_RELBUF))
1281 			vfs_vmio_release(bp);
1282 
1283 	} else if (bp->b_flags & B_VMIO) {
1284 
1285 		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1286 			vfs_vmio_release(bp);
1287 		}
1288 
1289 	}
1290 
1291 	if (BUF_REFCNT(bp) > 1) {
1292 		/* do not release to free list */
1293 		BUF_UNLOCK(bp);
1294 		return;
1295 	}
1296 
1297 	/* enqueue */
1298 	mtx_lock(&bqlock);
1299 	/* Handle delayed bremfree() processing. */
1300 	if (bp->b_flags & B_REMFREE)
1301 		bremfreel(bp);
1302 	if (bp->b_qindex != QUEUE_NONE)
1303 		panic("brelse: free buffer onto another queue???");
1304 
1305 	/* buffers with no memory */
1306 	if (bp->b_bufsize == 0) {
1307 		bp->b_flags |= B_INVAL;
1308 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1309 		if (bp->b_vflags & BV_BKGRDINPROG)
1310 			panic("losing buffer 1");
1311 		if (bp->b_kvasize) {
1312 			bp->b_qindex = QUEUE_EMPTYKVA;
1313 		} else {
1314 			bp->b_qindex = QUEUE_EMPTY;
1315 		}
1316 		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1317 	/* buffers with junk contents */
1318 	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1319 	    (bp->b_ioflags & BIO_ERROR)) {
1320 		bp->b_flags |= B_INVAL;
1321 		bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1322 		if (bp->b_vflags & BV_BKGRDINPROG)
1323 			panic("losing buffer 2");
1324 		bp->b_qindex = QUEUE_CLEAN;
1325 		TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1326 	/* remaining buffers */
1327 	} else {
1328 		if (bp->b_flags & B_DELWRI)
1329 			bp->b_qindex = QUEUE_DIRTY;
1330 		else
1331 			bp->b_qindex = QUEUE_CLEAN;
1332 		if (bp->b_flags & B_AGE)
1333 			TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1334 		else
1335 			TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1336 	}
1337 	mtx_unlock(&bqlock);
1338 
1339 	/*
1340 	 * If B_INVAL and B_DELWRI is set, clear B_DELWRI.  We have already
1341 	 * placed the buffer on the correct queue.  We must also disassociate
1342 	 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1343 	 * find it.
1344 	 */
1345 	if (bp->b_flags & B_INVAL) {
1346 		if (bp->b_flags & B_DELWRI)
1347 			bundirty(bp);
1348 		if (bp->b_vp)
1349 			brelvp(bp);
1350 	}
1351 
1352 	/*
1353 	 * Fixup numfreebuffers count.  The bp is on an appropriate queue
1354 	 * unless locked.  We then bump numfreebuffers if it is not B_DELWRI.
1355 	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1356 	 * if B_INVAL is set ).
1357 	 */
1358 
1359 	if (!(bp->b_flags & B_DELWRI))
1360 		bufcountwakeup();
1361 
1362 	/*
1363 	 * Something we can maybe free or reuse
1364 	 */
1365 	if (bp->b_bufsize || bp->b_kvasize)
1366 		bufspacewakeup();
1367 
1368 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1369 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1370 		panic("brelse: not dirty");
1371 	/* unlock */
1372 	BUF_UNLOCK(bp);
1373 }
1374 
1375 /*
1376  * Release a buffer back to the appropriate queue but do not try to free
1377  * it.  The buffer is expected to be used again soon.
1378  *
1379  * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1380  * biodone() to requeue an async I/O on completion.  It is also used when
1381  * known good buffers need to be requeued but we think we may need the data
1382  * again soon.
1383  *
1384  * XXX we should be able to leave the B_RELBUF hint set on completion.
1385  */
1386 void
1387 bqrelse(struct buf *bp)
1388 {
1389 	CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1390 	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1391 	    ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1392 
1393 	if (BUF_REFCNT(bp) > 1) {
1394 		/* do not release to free list */
1395 		BUF_UNLOCK(bp);
1396 		return;
1397 	}
1398 	mtx_lock(&bqlock);
1399 	/* Handle delayed bremfree() processing. */
1400 	if (bp->b_flags & B_REMFREE)
1401 		bremfreel(bp);
1402 	if (bp->b_qindex != QUEUE_NONE)
1403 		panic("bqrelse: free buffer onto another queue???");
1404 	/* buffers with stale but valid contents */
1405 	if (bp->b_flags & B_DELWRI) {
1406 		bp->b_qindex = QUEUE_DIRTY;
1407 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1408 	} else {
1409 		/*
1410 		 * XXX This lock may not be necessary since BKGRDINPROG
1411 		 * cannot be set while we hold the buf lock, it can only be
1412 		 * cleared if it is already pending.
1413 		 */
1414 		BO_LOCK(bp->b_bufobj);
1415 		if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) {
1416 			BO_UNLOCK(bp->b_bufobj);
1417 			bp->b_qindex = QUEUE_CLEAN;
1418 			TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1419 			    b_freelist);
1420 		} else {
1421 			/*
1422 			 * We are too low on memory, we have to try to free
1423 			 * the buffer (most importantly: the wired pages
1424 			 * making up its backing store) *now*.
1425 			 */
1426 			BO_UNLOCK(bp->b_bufobj);
1427 			mtx_unlock(&bqlock);
1428 			brelse(bp);
1429 			return;
1430 		}
1431 	}
1432 	mtx_unlock(&bqlock);
1433 
1434 	if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1435 		bufcountwakeup();
1436 
1437 	/*
1438 	 * Something we can maybe free or reuse.
1439 	 */
1440 	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1441 		bufspacewakeup();
1442 
1443 	bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1444 	if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1445 		panic("bqrelse: not dirty");
1446 	/* unlock */
1447 	BUF_UNLOCK(bp);
1448 }
1449 
1450 /* Give pages used by the bp back to the VM system (where possible) */
1451 static void
1452 vfs_vmio_release(struct buf *bp)
1453 {
1454 	int i;
1455 	vm_page_t m;
1456 
1457 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1458 	vm_page_lock_queues();
1459 	for (i = 0; i < bp->b_npages; i++) {
1460 		m = bp->b_pages[i];
1461 		bp->b_pages[i] = NULL;
1462 		/*
1463 		 * In order to keep page LRU ordering consistent, put
1464 		 * everything on the inactive queue.
1465 		 */
1466 		vm_page_unwire(m, 0);
1467 		/*
1468 		 * We don't mess with busy pages, it is
1469 		 * the responsibility of the process that
1470 		 * busied the pages to deal with them.
1471 		 */
1472 		if ((m->flags & PG_BUSY) || (m->busy != 0))
1473 			continue;
1474 
1475 		if (m->wire_count == 0) {
1476 			/*
1477 			 * Might as well free the page if we can and it has
1478 			 * no valid data.  We also free the page if the
1479 			 * buffer was used for direct I/O
1480 			 */
1481 			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1482 			    m->hold_count == 0) {
1483 				pmap_remove_all(m);
1484 				vm_page_free(m);
1485 			} else if (bp->b_flags & B_DIRECT) {
1486 				vm_page_try_to_free(m);
1487 			} else if (vm_page_count_severe()) {
1488 				vm_page_try_to_cache(m);
1489 			}
1490 		}
1491 	}
1492 	vm_page_unlock_queues();
1493 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1494 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1495 
1496 	if (bp->b_bufsize) {
1497 		bufspacewakeup();
1498 		bp->b_bufsize = 0;
1499 	}
1500 	bp->b_npages = 0;
1501 	bp->b_flags &= ~B_VMIO;
1502 	if (bp->b_vp)
1503 		brelvp(bp);
1504 }
1505 
1506 /*
1507  * Check to see if a block at a particular lbn is available for a clustered
1508  * write.
1509  */
1510 static int
1511 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1512 {
1513 	struct buf *bpa;
1514 	int match;
1515 
1516 	match = 0;
1517 
1518 	/* If the buf isn't in core skip it */
1519 	if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1520 		return (0);
1521 
1522 	/* If the buf is busy we don't want to wait for it */
1523 	if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1524 		return (0);
1525 
1526 	/* Only cluster with valid clusterable delayed write buffers */
1527 	if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1528 	    (B_DELWRI | B_CLUSTEROK))
1529 		goto done;
1530 
1531 	if (bpa->b_bufsize != size)
1532 		goto done;
1533 
1534 	/*
1535 	 * Check to see if it is in the expected place on disk and that the
1536 	 * block has been mapped.
1537 	 */
1538 	if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1539 		match = 1;
1540 done:
1541 	BUF_UNLOCK(bpa);
1542 	return (match);
1543 }
1544 
1545 /*
1546  *	vfs_bio_awrite:
1547  *
1548  *	Implement clustered async writes for clearing out B_DELWRI buffers.
1549  *	This is much better then the old way of writing only one buffer at
1550  *	a time.  Note that we may not be presented with the buffers in the
1551  *	correct order, so we search for the cluster in both directions.
1552  */
1553 int
1554 vfs_bio_awrite(struct buf *bp)
1555 {
1556 	int i;
1557 	int j;
1558 	daddr_t lblkno = bp->b_lblkno;
1559 	struct vnode *vp = bp->b_vp;
1560 	int ncl;
1561 	int nwritten;
1562 	int size;
1563 	int maxcl;
1564 
1565 	/*
1566 	 * right now we support clustered writing only to regular files.  If
1567 	 * we find a clusterable block we could be in the middle of a cluster
1568 	 * rather then at the beginning.
1569 	 */
1570 	if ((vp->v_type == VREG) &&
1571 	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
1572 	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1573 
1574 		size = vp->v_mount->mnt_stat.f_iosize;
1575 		maxcl = MAXPHYS / size;
1576 
1577 		VI_LOCK(vp);
1578 		for (i = 1; i < maxcl; i++)
1579 			if (vfs_bio_clcheck(vp, size, lblkno + i,
1580 			    bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1581 				break;
1582 
1583 		for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1584 			if (vfs_bio_clcheck(vp, size, lblkno - j,
1585 			    bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1586 				break;
1587 
1588 		VI_UNLOCK(vp);
1589 		--j;
1590 		ncl = i + j;
1591 		/*
1592 		 * this is a possible cluster write
1593 		 */
1594 		if (ncl != 1) {
1595 			BUF_UNLOCK(bp);
1596 			nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1597 			return nwritten;
1598 		}
1599 	}
1600 	bremfree(bp);
1601 	bp->b_flags |= B_ASYNC;
1602 	/*
1603 	 * default (old) behavior, writing out only one block
1604 	 *
1605 	 * XXX returns b_bufsize instead of b_bcount for nwritten?
1606 	 */
1607 	nwritten = bp->b_bufsize;
1608 	(void) bwrite(bp);
1609 
1610 	return nwritten;
1611 }
1612 
1613 /*
1614  *	getnewbuf:
1615  *
1616  *	Find and initialize a new buffer header, freeing up existing buffers
1617  *	in the bufqueues as necessary.  The new buffer is returned locked.
1618  *
1619  *	Important:  B_INVAL is not set.  If the caller wishes to throw the
1620  *	buffer away, the caller must set B_INVAL prior to calling brelse().
1621  *
1622  *	We block if:
1623  *		We have insufficient buffer headers
1624  *		We have insufficient buffer space
1625  *		buffer_map is too fragmented ( space reservation fails )
1626  *		If we have to flush dirty buffers ( but we try to avoid this )
1627  *
1628  *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1629  *	Instead we ask the buf daemon to do it for us.  We attempt to
1630  *	avoid piecemeal wakeups of the pageout daemon.
1631  */
1632 
1633 static struct buf *
1634 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1635 {
1636 	struct buf *bp;
1637 	struct buf *nbp;
1638 	int defrag = 0;
1639 	int nqindex;
1640 	static int flushingbufs;
1641 
1642 	/*
1643 	 * We can't afford to block since we might be holding a vnode lock,
1644 	 * which may prevent system daemons from running.  We deal with
1645 	 * low-memory situations by proactively returning memory and running
1646 	 * async I/O rather then sync I/O.
1647 	 */
1648 
1649 	atomic_add_int(&getnewbufcalls, 1);
1650 	atomic_subtract_int(&getnewbufrestarts, 1);
1651 restart:
1652 	atomic_add_int(&getnewbufrestarts, 1);
1653 
1654 	/*
1655 	 * Setup for scan.  If we do not have enough free buffers,
1656 	 * we setup a degenerate case that immediately fails.  Note
1657 	 * that if we are specially marked process, we are allowed to
1658 	 * dip into our reserves.
1659 	 *
1660 	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
1661 	 *
1662 	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
1663 	 * However, there are a number of cases (defragging, reusing, ...)
1664 	 * where we cannot backup.
1665 	 */
1666 	mtx_lock(&bqlock);
1667 	nqindex = QUEUE_EMPTYKVA;
1668 	nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1669 
1670 	if (nbp == NULL) {
1671 		/*
1672 		 * If no EMPTYKVA buffers and we are either
1673 		 * defragging or reusing, locate a CLEAN buffer
1674 		 * to free or reuse.  If bufspace useage is low
1675 		 * skip this step so we can allocate a new buffer.
1676 		 */
1677 		if (defrag || bufspace >= lobufspace) {
1678 			nqindex = QUEUE_CLEAN;
1679 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1680 		}
1681 
1682 		/*
1683 		 * If we could not find or were not allowed to reuse a
1684 		 * CLEAN buffer, check to see if it is ok to use an EMPTY
1685 		 * buffer.  We can only use an EMPTY buffer if allocating
1686 		 * its KVA would not otherwise run us out of buffer space.
1687 		 */
1688 		if (nbp == NULL && defrag == 0 &&
1689 		    bufspace + maxsize < hibufspace) {
1690 			nqindex = QUEUE_EMPTY;
1691 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1692 		}
1693 	}
1694 
1695 	/*
1696 	 * Run scan, possibly freeing data and/or kva mappings on the fly
1697 	 * depending.
1698 	 */
1699 
1700 	while ((bp = nbp) != NULL) {
1701 		int qindex = nqindex;
1702 
1703 		/*
1704 		 * Calculate next bp ( we can only use it if we do not block
1705 		 * or do other fancy things ).
1706 		 */
1707 		if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1708 			switch(qindex) {
1709 			case QUEUE_EMPTY:
1710 				nqindex = QUEUE_EMPTYKVA;
1711 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1712 					break;
1713 				/* FALLTHROUGH */
1714 			case QUEUE_EMPTYKVA:
1715 				nqindex = QUEUE_CLEAN;
1716 				if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1717 					break;
1718 				/* FALLTHROUGH */
1719 			case QUEUE_CLEAN:
1720 				/*
1721 				 * nbp is NULL.
1722 				 */
1723 				break;
1724 			}
1725 		}
1726 		/*
1727 		 * If we are defragging then we need a buffer with
1728 		 * b_kvasize != 0.  XXX this situation should no longer
1729 		 * occur, if defrag is non-zero the buffer's b_kvasize
1730 		 * should also be non-zero at this point.  XXX
1731 		 */
1732 		if (defrag && bp->b_kvasize == 0) {
1733 			printf("Warning: defrag empty buffer %p\n", bp);
1734 			continue;
1735 		}
1736 
1737 		/*
1738 		 * Start freeing the bp.  This is somewhat involved.  nbp
1739 		 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1740 		 */
1741 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1742 			continue;
1743 		if (bp->b_vp) {
1744 			BO_LOCK(bp->b_bufobj);
1745 			if (bp->b_vflags & BV_BKGRDINPROG) {
1746 				BO_UNLOCK(bp->b_bufobj);
1747 				BUF_UNLOCK(bp);
1748 				continue;
1749 			}
1750 			BO_UNLOCK(bp->b_bufobj);
1751 		}
1752 		CTR6(KTR_BUF,
1753 		    "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1754 		    "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1755 		    bp->b_kvasize, bp->b_bufsize, qindex);
1756 
1757 		/*
1758 		 * Sanity Checks
1759 		 */
1760 		KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1761 
1762 		/*
1763 		 * Note: we no longer distinguish between VMIO and non-VMIO
1764 		 * buffers.
1765 		 */
1766 
1767 		KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1768 
1769 		bremfreel(bp);
1770 		mtx_unlock(&bqlock);
1771 
1772 		if (qindex == QUEUE_CLEAN) {
1773 			if (bp->b_flags & B_VMIO) {
1774 				bp->b_flags &= ~B_ASYNC;
1775 				vfs_vmio_release(bp);
1776 			}
1777 			if (bp->b_vp)
1778 				brelvp(bp);
1779 		}
1780 
1781 		/*
1782 		 * NOTE:  nbp is now entirely invalid.  We can only restart
1783 		 * the scan from this point on.
1784 		 *
1785 		 * Get the rest of the buffer freed up.  b_kva* is still
1786 		 * valid after this operation.
1787 		 */
1788 
1789 		if (bp->b_rcred != NOCRED) {
1790 			crfree(bp->b_rcred);
1791 			bp->b_rcred = NOCRED;
1792 		}
1793 		if (bp->b_wcred != NOCRED) {
1794 			crfree(bp->b_wcred);
1795 			bp->b_wcred = NOCRED;
1796 		}
1797 		if (LIST_FIRST(&bp->b_dep) != NULL)
1798 			buf_deallocate(bp);
1799 		if (bp->b_vflags & BV_BKGRDINPROG)
1800 			panic("losing buffer 3");
1801 		KASSERT(bp->b_vp == NULL,
1802 		    ("bp: %p still has vnode %p.  qindex: %d",
1803 		    bp, bp->b_vp, qindex));
1804 		KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1805 		   ("bp: %p still on a buffer list. xflags %X",
1806 		    bp, bp->b_xflags));
1807 
1808 		if (bp->b_bufsize)
1809 			allocbuf(bp, 0);
1810 
1811 		bp->b_flags = 0;
1812 		bp->b_ioflags = 0;
1813 		bp->b_xflags = 0;
1814 		bp->b_vflags = 0;
1815 		bp->b_vp = NULL;
1816 		bp->b_blkno = bp->b_lblkno = 0;
1817 		bp->b_offset = NOOFFSET;
1818 		bp->b_iodone = 0;
1819 		bp->b_error = 0;
1820 		bp->b_resid = 0;
1821 		bp->b_bcount = 0;
1822 		bp->b_npages = 0;
1823 		bp->b_dirtyoff = bp->b_dirtyend = 0;
1824 		bp->b_bufobj = NULL;
1825 
1826 		LIST_INIT(&bp->b_dep);
1827 
1828 		/*
1829 		 * If we are defragging then free the buffer.
1830 		 */
1831 		if (defrag) {
1832 			bp->b_flags |= B_INVAL;
1833 			bfreekva(bp);
1834 			brelse(bp);
1835 			defrag = 0;
1836 			goto restart;
1837 		}
1838 
1839 		/*
1840 		 * If we are overcomitted then recover the buffer and its
1841 		 * KVM space.  This occurs in rare situations when multiple
1842 		 * processes are blocked in getnewbuf() or allocbuf().
1843 		 */
1844 		if (bufspace >= hibufspace)
1845 			flushingbufs = 1;
1846 		if (flushingbufs && bp->b_kvasize != 0) {
1847 			bp->b_flags |= B_INVAL;
1848 			bfreekva(bp);
1849 			brelse(bp);
1850 			goto restart;
1851 		}
1852 		if (bufspace < lobufspace)
1853 			flushingbufs = 0;
1854 		break;
1855 	}
1856 
1857 	/*
1858 	 * If we exhausted our list, sleep as appropriate.  We may have to
1859 	 * wakeup various daemons and write out some dirty buffers.
1860 	 *
1861 	 * Generally we are sleeping due to insufficient buffer space.
1862 	 */
1863 
1864 	if (bp == NULL) {
1865 		int flags;
1866 		char *waitmsg;
1867 
1868 		mtx_unlock(&bqlock);
1869 		if (defrag) {
1870 			flags = VFS_BIO_NEED_BUFSPACE;
1871 			waitmsg = "nbufkv";
1872 		} else if (bufspace >= hibufspace) {
1873 			waitmsg = "nbufbs";
1874 			flags = VFS_BIO_NEED_BUFSPACE;
1875 		} else {
1876 			waitmsg = "newbuf";
1877 			flags = VFS_BIO_NEED_ANY;
1878 		}
1879 
1880 		bd_speedup();	/* heeeelp */
1881 
1882 		mtx_lock(&nblock);
1883 		needsbuffer |= flags;
1884 		while (needsbuffer & flags) {
1885 			if (msleep(&needsbuffer, &nblock,
1886 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1887 				mtx_unlock(&nblock);
1888 				return (NULL);
1889 			}
1890 		}
1891 		mtx_unlock(&nblock);
1892 	} else {
1893 		/*
1894 		 * We finally have a valid bp.  We aren't quite out of the
1895 		 * woods, we still have to reserve kva space.  In order
1896 		 * to keep fragmentation sane we only allocate kva in
1897 		 * BKVASIZE chunks.
1898 		 */
1899 		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1900 
1901 		if (maxsize != bp->b_kvasize) {
1902 			vm_offset_t addr = 0;
1903 
1904 			bfreekva(bp);
1905 
1906 			vm_map_lock(buffer_map);
1907 			if (vm_map_findspace(buffer_map,
1908 				vm_map_min(buffer_map), maxsize, &addr)) {
1909 				/*
1910 				 * Uh oh.  Buffer map is to fragmented.  We
1911 				 * must defragment the map.
1912 				 */
1913 				atomic_add_int(&bufdefragcnt, 1);
1914 				vm_map_unlock(buffer_map);
1915 				defrag = 1;
1916 				bp->b_flags |= B_INVAL;
1917 				brelse(bp);
1918 				goto restart;
1919 			}
1920 			if (addr) {
1921 				vm_map_insert(buffer_map, NULL, 0,
1922 					addr, addr + maxsize,
1923 					VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1924 
1925 				bp->b_kvabase = (caddr_t) addr;
1926 				bp->b_kvasize = maxsize;
1927 				atomic_add_int(&bufspace, bp->b_kvasize);
1928 				atomic_add_int(&bufreusecnt, 1);
1929 			}
1930 			vm_map_unlock(buffer_map);
1931 		}
1932 		bp->b_saveaddr = bp->b_kvabase;
1933 		bp->b_data = bp->b_saveaddr;
1934 	}
1935 	return(bp);
1936 }
1937 
1938 /*
1939  *	buf_daemon:
1940  *
1941  *	buffer flushing daemon.  Buffers are normally flushed by the
1942  *	update daemon but if it cannot keep up this process starts to
1943  *	take the load in an attempt to prevent getnewbuf() from blocking.
1944  */
1945 
1946 static struct kproc_desc buf_kp = {
1947 	"bufdaemon",
1948 	buf_daemon,
1949 	&bufdaemonproc
1950 };
1951 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1952 
1953 static void
1954 buf_daemon()
1955 {
1956 	mtx_lock(&Giant);
1957 
1958 	/*
1959 	 * This process needs to be suspended prior to shutdown sync.
1960 	 */
1961 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
1962 	    SHUTDOWN_PRI_LAST);
1963 
1964 	/*
1965 	 * This process is allowed to take the buffer cache to the limit
1966 	 */
1967 	mtx_lock(&bdlock);
1968 	for (;;) {
1969 		bd_request = 0;
1970 		mtx_unlock(&bdlock);
1971 
1972 		kthread_suspend_check(bufdaemonproc);
1973 
1974 		/*
1975 		 * Do the flush.  Limit the amount of in-transit I/O we
1976 		 * allow to build up, otherwise we would completely saturate
1977 		 * the I/O system.  Wakeup any waiting processes before we
1978 		 * normally would so they can run in parallel with our drain.
1979 		 */
1980 		while (numdirtybuffers > lodirtybuffers) {
1981 			if (flushbufqueues(0) == 0) {
1982 				/*
1983 				 * Could not find any buffers without rollback
1984 				 * dependencies, so just write the first one
1985 				 * in the hopes of eventually making progress.
1986 				 */
1987 				flushbufqueues(1);
1988 				break;
1989 			}
1990 			uio_yield();
1991 		}
1992 
1993 		/*
1994 		 * Only clear bd_request if we have reached our low water
1995 		 * mark.  The buf_daemon normally waits 1 second and
1996 		 * then incrementally flushes any dirty buffers that have
1997 		 * built up, within reason.
1998 		 *
1999 		 * If we were unable to hit our low water mark and couldn't
2000 		 * find any flushable buffers, we sleep half a second.
2001 		 * Otherwise we loop immediately.
2002 		 */
2003 		mtx_lock(&bdlock);
2004 		if (numdirtybuffers <= lodirtybuffers) {
2005 			/*
2006 			 * We reached our low water mark, reset the
2007 			 * request and sleep until we are needed again.
2008 			 * The sleep is just so the suspend code works.
2009 			 */
2010 			bd_request = 0;
2011 			msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2012 		} else {
2013 			/*
2014 			 * We couldn't find any flushable dirty buffers but
2015 			 * still have too many dirty buffers, we
2016 			 * have to sleep and try again.  (rare)
2017 			 */
2018 			msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2019 		}
2020 	}
2021 }
2022 
2023 /*
2024  *	flushbufqueues:
2025  *
2026  *	Try to flush a buffer in the dirty queue.  We must be careful to
2027  *	free up B_INVAL buffers instead of write them, which NFS is
2028  *	particularly sensitive to.
2029  */
2030 static int flushwithdeps = 0;
2031 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2032     0, "Number of buffers flushed with dependecies that require rollbacks");
2033 
2034 static int
2035 flushbufqueues(int flushdeps)
2036 {
2037 	struct thread *td = curthread;
2038 	struct buf sentinel;
2039 	struct vnode *vp;
2040 	struct mount *mp;
2041 	struct buf *bp;
2042 	int hasdeps;
2043 	int flushed;
2044 	int target;
2045 
2046 	target = numdirtybuffers - lodirtybuffers;
2047 	if (flushdeps && target > 2)
2048 		target /= 2;
2049 	flushed = 0;
2050 	bp = NULL;
2051 	mtx_lock(&bqlock);
2052 	TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], &sentinel, b_freelist);
2053 	while (flushed != target) {
2054 		bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
2055 		if (bp == &sentinel)
2056 			break;
2057 		TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
2058 		TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
2059 
2060 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2061 			continue;
2062 		BO_LOCK(bp->b_bufobj);
2063 		if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2064 		    (bp->b_flags & B_DELWRI) == 0) {
2065 			BO_UNLOCK(bp->b_bufobj);
2066 			BUF_UNLOCK(bp);
2067 			continue;
2068 		}
2069 		BO_UNLOCK(bp->b_bufobj);
2070 		if (bp->b_flags & B_INVAL) {
2071 			bremfreel(bp);
2072 			mtx_unlock(&bqlock);
2073 			brelse(bp);
2074 			flushed++;
2075 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2076 			mtx_lock(&bqlock);
2077 			continue;
2078 		}
2079 
2080 		if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) {
2081 			if (flushdeps == 0) {
2082 				BUF_UNLOCK(bp);
2083 				continue;
2084 			}
2085 			hasdeps = 1;
2086 		} else
2087 			hasdeps = 0;
2088 		/*
2089 		 * We must hold the lock on a vnode before writing
2090 		 * one of its buffers. Otherwise we may confuse, or
2091 		 * in the case of a snapshot vnode, deadlock the
2092 		 * system.
2093 		 *
2094 		 * The lock order here is the reverse of the normal
2095 		 * of vnode followed by buf lock.  This is ok because
2096 		 * the NOWAIT will prevent deadlock.
2097 		 */
2098 		vp = bp->b_vp;
2099 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2100 			BUF_UNLOCK(bp);
2101 			continue;
2102 		}
2103 		if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2104 			mtx_unlock(&bqlock);
2105 			CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2106 			    bp, bp->b_vp, bp->b_flags);
2107 			vfs_bio_awrite(bp);
2108 			vn_finished_write(mp);
2109 			VOP_UNLOCK(vp, 0, td);
2110 			flushwithdeps += hasdeps;
2111 			flushed++;
2112 			waitrunningbufspace();
2113 			numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2114 			mtx_lock(&bqlock);
2115 			continue;
2116 		}
2117 		vn_finished_write(mp);
2118 		BUF_UNLOCK(bp);
2119 	}
2120 	TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], &sentinel, b_freelist);
2121 	mtx_unlock(&bqlock);
2122 	return (flushed);
2123 }
2124 
2125 /*
2126  * Check to see if a block is currently memory resident.
2127  */
2128 struct buf *
2129 incore(struct bufobj *bo, daddr_t blkno)
2130 {
2131 	struct buf *bp;
2132 
2133 	BO_LOCK(bo);
2134 	bp = gbincore(bo, blkno);
2135 	BO_UNLOCK(bo);
2136 	return (bp);
2137 }
2138 
2139 /*
2140  * Returns true if no I/O is needed to access the
2141  * associated VM object.  This is like incore except
2142  * it also hunts around in the VM system for the data.
2143  */
2144 
2145 static int
2146 inmem(struct vnode * vp, daddr_t blkno)
2147 {
2148 	vm_object_t obj;
2149 	vm_offset_t toff, tinc, size;
2150 	vm_page_t m;
2151 	vm_ooffset_t off;
2152 
2153 	ASSERT_VOP_LOCKED(vp, "inmem");
2154 
2155 	if (incore(&vp->v_bufobj, blkno))
2156 		return 1;
2157 	if (vp->v_mount == NULL)
2158 		return 0;
2159 	obj = vp->v_object;
2160 	if (obj == NULL)
2161 		return (0);
2162 
2163 	size = PAGE_SIZE;
2164 	if (size > vp->v_mount->mnt_stat.f_iosize)
2165 		size = vp->v_mount->mnt_stat.f_iosize;
2166 	off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2167 
2168 	VM_OBJECT_LOCK(obj);
2169 	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2170 		m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2171 		if (!m)
2172 			goto notinmem;
2173 		tinc = size;
2174 		if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2175 			tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2176 		if (vm_page_is_valid(m,
2177 		    (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2178 			goto notinmem;
2179 	}
2180 	VM_OBJECT_UNLOCK(obj);
2181 	return 1;
2182 
2183 notinmem:
2184 	VM_OBJECT_UNLOCK(obj);
2185 	return (0);
2186 }
2187 
2188 /*
2189  *	vfs_setdirty:
2190  *
2191  *	Sets the dirty range for a buffer based on the status of the dirty
2192  *	bits in the pages comprising the buffer.
2193  *
2194  *	The range is limited to the size of the buffer.
2195  *
2196  *	This routine is primarily used by NFS, but is generalized for the
2197  *	B_VMIO case.
2198  */
2199 static void
2200 vfs_setdirty(struct buf *bp)
2201 {
2202 	int i;
2203 	vm_object_t object;
2204 
2205 	/*
2206 	 * Degenerate case - empty buffer
2207 	 */
2208 
2209 	if (bp->b_bufsize == 0)
2210 		return;
2211 
2212 	/*
2213 	 * We qualify the scan for modified pages on whether the
2214 	 * object has been flushed yet.  The OBJ_WRITEABLE flag
2215 	 * is not cleared simply by protecting pages off.
2216 	 */
2217 
2218 	if ((bp->b_flags & B_VMIO) == 0)
2219 		return;
2220 
2221 	object = bp->b_pages[0]->object;
2222 	VM_OBJECT_LOCK(object);
2223 	if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2224 		printf("Warning: object %p writeable but not mightbedirty\n", object);
2225 	if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2226 		printf("Warning: object %p mightbedirty but not writeable\n", object);
2227 
2228 	if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2229 		vm_offset_t boffset;
2230 		vm_offset_t eoffset;
2231 
2232 		vm_page_lock_queues();
2233 		/*
2234 		 * test the pages to see if they have been modified directly
2235 		 * by users through the VM system.
2236 		 */
2237 		for (i = 0; i < bp->b_npages; i++)
2238 			vm_page_test_dirty(bp->b_pages[i]);
2239 
2240 		/*
2241 		 * Calculate the encompassing dirty range, boffset and eoffset,
2242 		 * (eoffset - boffset) bytes.
2243 		 */
2244 
2245 		for (i = 0; i < bp->b_npages; i++) {
2246 			if (bp->b_pages[i]->dirty)
2247 				break;
2248 		}
2249 		boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2250 
2251 		for (i = bp->b_npages - 1; i >= 0; --i) {
2252 			if (bp->b_pages[i]->dirty) {
2253 				break;
2254 			}
2255 		}
2256 		eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2257 
2258 		vm_page_unlock_queues();
2259 		/*
2260 		 * Fit it to the buffer.
2261 		 */
2262 
2263 		if (eoffset > bp->b_bcount)
2264 			eoffset = bp->b_bcount;
2265 
2266 		/*
2267 		 * If we have a good dirty range, merge with the existing
2268 		 * dirty range.
2269 		 */
2270 
2271 		if (boffset < eoffset) {
2272 			if (bp->b_dirtyoff > boffset)
2273 				bp->b_dirtyoff = boffset;
2274 			if (bp->b_dirtyend < eoffset)
2275 				bp->b_dirtyend = eoffset;
2276 		}
2277 	}
2278 	VM_OBJECT_UNLOCK(object);
2279 }
2280 
2281 /*
2282  *	getblk:
2283  *
2284  *	Get a block given a specified block and offset into a file/device.
2285  *	The buffers B_DONE bit will be cleared on return, making it almost
2286  * 	ready for an I/O initiation.  B_INVAL may or may not be set on
2287  *	return.  The caller should clear B_INVAL prior to initiating a
2288  *	READ.
2289  *
2290  *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2291  *	an existing buffer.
2292  *
2293  *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
2294  *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2295  *	and then cleared based on the backing VM.  If the previous buffer is
2296  *	non-0-sized but invalid, B_CACHE will be cleared.
2297  *
2298  *	If getblk() must create a new buffer, the new buffer is returned with
2299  *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2300  *	case it is returned with B_INVAL clear and B_CACHE set based on the
2301  *	backing VM.
2302  *
2303  *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
2304  *	B_CACHE bit is clear.
2305  *
2306  *	What this means, basically, is that the caller should use B_CACHE to
2307  *	determine whether the buffer is fully valid or not and should clear
2308  *	B_INVAL prior to issuing a read.  If the caller intends to validate
2309  *	the buffer by loading its data area with something, the caller needs
2310  *	to clear B_INVAL.  If the caller does this without issuing an I/O,
2311  *	the caller should set B_CACHE ( as an optimization ), else the caller
2312  *	should issue the I/O and biodone() will set B_CACHE if the I/O was
2313  *	a write attempt or if it was a successfull read.  If the caller
2314  *	intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2315  *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
2316  */
2317 struct buf *
2318 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2319     int flags)
2320 {
2321 	struct buf *bp;
2322 	struct bufobj *bo;
2323 	int error;
2324 
2325 	CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2326 	ASSERT_VOP_LOCKED(vp, "getblk");
2327 	if (size > MAXBSIZE)
2328 		panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2329 
2330 	bo = &vp->v_bufobj;
2331 loop:
2332 	/*
2333 	 * Block if we are low on buffers.   Certain processes are allowed
2334 	 * to completely exhaust the buffer cache.
2335          *
2336          * If this check ever becomes a bottleneck it may be better to
2337          * move it into the else, when gbincore() fails.  At the moment
2338          * it isn't a problem.
2339 	 *
2340 	 * XXX remove if 0 sections (clean this up after its proven)
2341          */
2342 	if (numfreebuffers == 0) {
2343 		if (curthread == PCPU_GET(idlethread))
2344 			return NULL;
2345 		mtx_lock(&nblock);
2346 		needsbuffer |= VFS_BIO_NEED_ANY;
2347 		mtx_unlock(&nblock);
2348 	}
2349 
2350 	VI_LOCK(vp);
2351 	bp = gbincore(bo, blkno);
2352 	if (bp != NULL) {
2353 		int lockflags;
2354 		/*
2355 		 * Buffer is in-core.  If the buffer is not busy, it must
2356 		 * be on a queue.
2357 		 */
2358 		lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2359 
2360 		if (flags & GB_LOCK_NOWAIT)
2361 			lockflags |= LK_NOWAIT;
2362 
2363 		error = BUF_TIMELOCK(bp, lockflags,
2364 		    VI_MTX(vp), "getblk", slpflag, slptimeo);
2365 
2366 		/*
2367 		 * If we slept and got the lock we have to restart in case
2368 		 * the buffer changed identities.
2369 		 */
2370 		if (error == ENOLCK)
2371 			goto loop;
2372 		/* We timed out or were interrupted. */
2373 		else if (error)
2374 			return (NULL);
2375 
2376 		/*
2377 		 * The buffer is locked.  B_CACHE is cleared if the buffer is
2378 		 * invalid.  Otherwise, for a non-VMIO buffer, B_CACHE is set
2379 		 * and for a VMIO buffer B_CACHE is adjusted according to the
2380 		 * backing VM cache.
2381 		 */
2382 		if (bp->b_flags & B_INVAL)
2383 			bp->b_flags &= ~B_CACHE;
2384 		else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2385 			bp->b_flags |= B_CACHE;
2386 		bremfree(bp);
2387 
2388 		/*
2389 		 * check for size inconsistancies for non-VMIO case.
2390 		 */
2391 
2392 		if (bp->b_bcount != size) {
2393 			if ((bp->b_flags & B_VMIO) == 0 ||
2394 			    (size > bp->b_kvasize)) {
2395 				if (bp->b_flags & B_DELWRI) {
2396 					bp->b_flags |= B_NOCACHE;
2397 					bwrite(bp);
2398 				} else {
2399 					if ((bp->b_flags & B_VMIO) &&
2400 					   (LIST_FIRST(&bp->b_dep) == NULL)) {
2401 						bp->b_flags |= B_RELBUF;
2402 						brelse(bp);
2403 					} else {
2404 						bp->b_flags |= B_NOCACHE;
2405 						bwrite(bp);
2406 					}
2407 				}
2408 				goto loop;
2409 			}
2410 		}
2411 
2412 		/*
2413 		 * If the size is inconsistant in the VMIO case, we can resize
2414 		 * the buffer.  This might lead to B_CACHE getting set or
2415 		 * cleared.  If the size has not changed, B_CACHE remains
2416 		 * unchanged from its previous state.
2417 		 */
2418 
2419 		if (bp->b_bcount != size)
2420 			allocbuf(bp, size);
2421 
2422 		KASSERT(bp->b_offset != NOOFFSET,
2423 		    ("getblk: no buffer offset"));
2424 
2425 		/*
2426 		 * A buffer with B_DELWRI set and B_CACHE clear must
2427 		 * be committed before we can return the buffer in
2428 		 * order to prevent the caller from issuing a read
2429 		 * ( due to B_CACHE not being set ) and overwriting
2430 		 * it.
2431 		 *
2432 		 * Most callers, including NFS and FFS, need this to
2433 		 * operate properly either because they assume they
2434 		 * can issue a read if B_CACHE is not set, or because
2435 		 * ( for example ) an uncached B_DELWRI might loop due
2436 		 * to softupdates re-dirtying the buffer.  In the latter
2437 		 * case, B_CACHE is set after the first write completes,
2438 		 * preventing further loops.
2439 		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
2440 		 * above while extending the buffer, we cannot allow the
2441 		 * buffer to remain with B_CACHE set after the write
2442 		 * completes or it will represent a corrupt state.  To
2443 		 * deal with this we set B_NOCACHE to scrap the buffer
2444 		 * after the write.
2445 		 *
2446 		 * We might be able to do something fancy, like setting
2447 		 * B_CACHE in bwrite() except if B_DELWRI is already set,
2448 		 * so the below call doesn't set B_CACHE, but that gets real
2449 		 * confusing.  This is much easier.
2450 		 */
2451 
2452 		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2453 			bp->b_flags |= B_NOCACHE;
2454 			bwrite(bp);
2455 			goto loop;
2456 		}
2457 		bp->b_flags &= ~B_DONE;
2458 	} else {
2459 		int bsize, maxsize, vmio;
2460 		off_t offset;
2461 
2462 		/*
2463 		 * Buffer is not in-core, create new buffer.  The buffer
2464 		 * returned by getnewbuf() is locked.  Note that the returned
2465 		 * buffer is also considered valid (not marked B_INVAL).
2466 		 */
2467 		VI_UNLOCK(vp);
2468 		/*
2469 		 * If the user does not want us to create the buffer, bail out
2470 		 * here.
2471 		 */
2472 		if (flags & GB_NOCREAT)
2473 			return NULL;
2474 		bsize = bo->bo_bsize;
2475 		offset = blkno * bsize;
2476 		vmio = vp->v_object != NULL;
2477 		maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2478 		maxsize = imax(maxsize, bsize);
2479 
2480 		bp = getnewbuf(slpflag, slptimeo, size, maxsize);
2481 		if (bp == NULL) {
2482 			if (slpflag || slptimeo)
2483 				return NULL;
2484 			goto loop;
2485 		}
2486 
2487 		/*
2488 		 * This code is used to make sure that a buffer is not
2489 		 * created while the getnewbuf routine is blocked.
2490 		 * This can be a problem whether the vnode is locked or not.
2491 		 * If the buffer is created out from under us, we have to
2492 		 * throw away the one we just created.
2493 		 *
2494 		 * Note: this must occur before we associate the buffer
2495 		 * with the vp especially considering limitations in
2496 		 * the splay tree implementation when dealing with duplicate
2497 		 * lblkno's.
2498 		 */
2499 		BO_LOCK(bo);
2500 		if (gbincore(bo, blkno)) {
2501 			BO_UNLOCK(bo);
2502 			bp->b_flags |= B_INVAL;
2503 			brelse(bp);
2504 			goto loop;
2505 		}
2506 
2507 		/*
2508 		 * Insert the buffer into the hash, so that it can
2509 		 * be found by incore.
2510 		 */
2511 		bp->b_blkno = bp->b_lblkno = blkno;
2512 		bp->b_offset = offset;
2513 
2514 		bgetvp(vp, bp);
2515 		BO_UNLOCK(bo);
2516 
2517 		/*
2518 		 * set B_VMIO bit.  allocbuf() the buffer bigger.  Since the
2519 		 * buffer size starts out as 0, B_CACHE will be set by
2520 		 * allocbuf() for the VMIO case prior to it testing the
2521 		 * backing store for validity.
2522 		 */
2523 
2524 		if (vmio) {
2525 			bp->b_flags |= B_VMIO;
2526 #if defined(VFS_BIO_DEBUG)
2527 			if (vn_canvmio(vp) != TRUE)
2528 				printf("getblk: VMIO on vnode type %d\n",
2529 					vp->v_type);
2530 #endif
2531 			KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2532 			    ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2533 			    bp, vp->v_object, bp->b_bufobj->bo_object));
2534 		} else {
2535 			bp->b_flags &= ~B_VMIO;
2536 			KASSERT(bp->b_bufobj->bo_object == NULL,
2537 			    ("ARGH! has b_bufobj->bo_object %p %p\n",
2538 			    bp, bp->b_bufobj->bo_object));
2539 		}
2540 
2541 		allocbuf(bp, size);
2542 		bp->b_flags &= ~B_DONE;
2543 	}
2544 	CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2545 	KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2546 	KASSERT(bp->b_bufobj == bo,
2547 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2548 	return (bp);
2549 }
2550 
2551 /*
2552  * Get an empty, disassociated buffer of given size.  The buffer is initially
2553  * set to B_INVAL.
2554  */
2555 struct buf *
2556 geteblk(int size)
2557 {
2558 	struct buf *bp;
2559 	int maxsize;
2560 
2561 	maxsize = (size + BKVAMASK) & ~BKVAMASK;
2562 	while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2563 		continue;
2564 	allocbuf(bp, size);
2565 	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
2566 	KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2567 	return (bp);
2568 }
2569 
2570 
2571 /*
2572  * This code constitutes the buffer memory from either anonymous system
2573  * memory (in the case of non-VMIO operations) or from an associated
2574  * VM object (in the case of VMIO operations).  This code is able to
2575  * resize a buffer up or down.
2576  *
2577  * Note that this code is tricky, and has many complications to resolve
2578  * deadlock or inconsistant data situations.  Tread lightly!!!
2579  * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2580  * the caller.  Calling this code willy nilly can result in the loss of data.
2581  *
2582  * allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
2583  * B_CACHE for the non-VMIO case.
2584  */
2585 
2586 int
2587 allocbuf(struct buf *bp, int size)
2588 {
2589 	int newbsize, mbsize;
2590 	int i;
2591 
2592 	if (BUF_REFCNT(bp) == 0)
2593 		panic("allocbuf: buffer not busy");
2594 
2595 	if (bp->b_kvasize < size)
2596 		panic("allocbuf: buffer too small");
2597 
2598 	if ((bp->b_flags & B_VMIO) == 0) {
2599 		caddr_t origbuf;
2600 		int origbufsize;
2601 		/*
2602 		 * Just get anonymous memory from the kernel.  Don't
2603 		 * mess with B_CACHE.
2604 		 */
2605 		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2606 		if (bp->b_flags & B_MALLOC)
2607 			newbsize = mbsize;
2608 		else
2609 			newbsize = round_page(size);
2610 
2611 		if (newbsize < bp->b_bufsize) {
2612 			/*
2613 			 * malloced buffers are not shrunk
2614 			 */
2615 			if (bp->b_flags & B_MALLOC) {
2616 				if (newbsize) {
2617 					bp->b_bcount = size;
2618 				} else {
2619 					free(bp->b_data, M_BIOBUF);
2620 					if (bp->b_bufsize) {
2621 						atomic_subtract_int(
2622 						    &bufmallocspace,
2623 						    bp->b_bufsize);
2624 						bufspacewakeup();
2625 						bp->b_bufsize = 0;
2626 					}
2627 					bp->b_saveaddr = bp->b_kvabase;
2628 					bp->b_data = bp->b_saveaddr;
2629 					bp->b_bcount = 0;
2630 					bp->b_flags &= ~B_MALLOC;
2631 				}
2632 				return 1;
2633 			}
2634 			vm_hold_free_pages(
2635 			    bp,
2636 			    (vm_offset_t) bp->b_data + newbsize,
2637 			    (vm_offset_t) bp->b_data + bp->b_bufsize);
2638 		} else if (newbsize > bp->b_bufsize) {
2639 			/*
2640 			 * We only use malloced memory on the first allocation.
2641 			 * and revert to page-allocated memory when the buffer
2642 			 * grows.
2643 			 */
2644 			/*
2645 			 * There is a potential smp race here that could lead
2646 			 * to bufmallocspace slightly passing the max.  It
2647 			 * is probably extremely rare and not worth worrying
2648 			 * over.
2649 			 */
2650 			if ( (bufmallocspace < maxbufmallocspace) &&
2651 				(bp->b_bufsize == 0) &&
2652 				(mbsize <= PAGE_SIZE/2)) {
2653 
2654 				bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2655 				bp->b_bufsize = mbsize;
2656 				bp->b_bcount = size;
2657 				bp->b_flags |= B_MALLOC;
2658 				atomic_add_int(&bufmallocspace, mbsize);
2659 				return 1;
2660 			}
2661 			origbuf = NULL;
2662 			origbufsize = 0;
2663 			/*
2664 			 * If the buffer is growing on its other-than-first allocation,
2665 			 * then we revert to the page-allocation scheme.
2666 			 */
2667 			if (bp->b_flags & B_MALLOC) {
2668 				origbuf = bp->b_data;
2669 				origbufsize = bp->b_bufsize;
2670 				bp->b_data = bp->b_kvabase;
2671 				if (bp->b_bufsize) {
2672 					atomic_subtract_int(&bufmallocspace,
2673 					    bp->b_bufsize);
2674 					bufspacewakeup();
2675 					bp->b_bufsize = 0;
2676 				}
2677 				bp->b_flags &= ~B_MALLOC;
2678 				newbsize = round_page(newbsize);
2679 			}
2680 			vm_hold_load_pages(
2681 			    bp,
2682 			    (vm_offset_t) bp->b_data + bp->b_bufsize,
2683 			    (vm_offset_t) bp->b_data + newbsize);
2684 			if (origbuf) {
2685 				bcopy(origbuf, bp->b_data, origbufsize);
2686 				free(origbuf, M_BIOBUF);
2687 			}
2688 		}
2689 	} else {
2690 		int desiredpages;
2691 
2692 		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2693 		desiredpages = (size == 0) ? 0 :
2694 			num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2695 
2696 		if (bp->b_flags & B_MALLOC)
2697 			panic("allocbuf: VMIO buffer can't be malloced");
2698 		/*
2699 		 * Set B_CACHE initially if buffer is 0 length or will become
2700 		 * 0-length.
2701 		 */
2702 		if (size == 0 || bp->b_bufsize == 0)
2703 			bp->b_flags |= B_CACHE;
2704 
2705 		if (newbsize < bp->b_bufsize) {
2706 			/*
2707 			 * DEV_BSIZE aligned new buffer size is less then the
2708 			 * DEV_BSIZE aligned existing buffer size.  Figure out
2709 			 * if we have to remove any pages.
2710 			 */
2711 			if (desiredpages < bp->b_npages) {
2712 				vm_page_t m;
2713 
2714 				VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2715 				vm_page_lock_queues();
2716 				for (i = desiredpages; i < bp->b_npages; i++) {
2717 					/*
2718 					 * the page is not freed here -- it
2719 					 * is the responsibility of
2720 					 * vnode_pager_setsize
2721 					 */
2722 					m = bp->b_pages[i];
2723 					KASSERT(m != bogus_page,
2724 					    ("allocbuf: bogus page found"));
2725 					while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2726 						vm_page_lock_queues();
2727 
2728 					bp->b_pages[i] = NULL;
2729 					vm_page_unwire(m, 0);
2730 				}
2731 				vm_page_unlock_queues();
2732 				VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2733 				pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2734 				    (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2735 				bp->b_npages = desiredpages;
2736 			}
2737 		} else if (size > bp->b_bcount) {
2738 			/*
2739 			 * We are growing the buffer, possibly in a
2740 			 * byte-granular fashion.
2741 			 */
2742 			struct vnode *vp;
2743 			vm_object_t obj;
2744 			vm_offset_t toff;
2745 			vm_offset_t tinc;
2746 
2747 			/*
2748 			 * Step 1, bring in the VM pages from the object,
2749 			 * allocating them if necessary.  We must clear
2750 			 * B_CACHE if these pages are not valid for the
2751 			 * range covered by the buffer.
2752 			 */
2753 
2754 			vp = bp->b_vp;
2755 			obj = bp->b_bufobj->bo_object;
2756 
2757 			VM_OBJECT_LOCK(obj);
2758 			while (bp->b_npages < desiredpages) {
2759 				vm_page_t m;
2760 				vm_pindex_t pi;
2761 
2762 				pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2763 				if ((m = vm_page_lookup(obj, pi)) == NULL) {
2764 					/*
2765 					 * note: must allocate system pages
2766 					 * since blocking here could intefere
2767 					 * with paging I/O, no matter which
2768 					 * process we are.
2769 					 */
2770 					m = vm_page_alloc(obj, pi,
2771 					    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2772 					    VM_ALLOC_WIRED);
2773 					if (m == NULL) {
2774 						atomic_add_int(&vm_pageout_deficit,
2775 						    desiredpages - bp->b_npages);
2776 						VM_OBJECT_UNLOCK(obj);
2777 						VM_WAIT;
2778 						VM_OBJECT_LOCK(obj);
2779 					} else {
2780 						bp->b_flags &= ~B_CACHE;
2781 						bp->b_pages[bp->b_npages] = m;
2782 						++bp->b_npages;
2783 					}
2784 					continue;
2785 				}
2786 
2787 				/*
2788 				 * We found a page.  If we have to sleep on it,
2789 				 * retry because it might have gotten freed out
2790 				 * from under us.
2791 				 *
2792 				 * We can only test PG_BUSY here.  Blocking on
2793 				 * m->busy might lead to a deadlock:
2794 				 *
2795 				 *  vm_fault->getpages->cluster_read->allocbuf
2796 				 *
2797 				 */
2798 				vm_page_lock_queues();
2799 				if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2800 					continue;
2801 
2802 				/*
2803 				 * We have a good page.  Should we wakeup the
2804 				 * page daemon?
2805 				 */
2806 				if ((curproc != pageproc) &&
2807 				    ((m->queue - m->pc) == PQ_CACHE) &&
2808 				    ((cnt.v_free_count + cnt.v_cache_count) <
2809 					(cnt.v_free_min + cnt.v_cache_min))) {
2810 					pagedaemon_wakeup();
2811 				}
2812 				vm_page_wire(m);
2813 				vm_page_unlock_queues();
2814 				bp->b_pages[bp->b_npages] = m;
2815 				++bp->b_npages;
2816 			}
2817 
2818 			/*
2819 			 * Step 2.  We've loaded the pages into the buffer,
2820 			 * we have to figure out if we can still have B_CACHE
2821 			 * set.  Note that B_CACHE is set according to the
2822 			 * byte-granular range ( bcount and size ), new the
2823 			 * aligned range ( newbsize ).
2824 			 *
2825 			 * The VM test is against m->valid, which is DEV_BSIZE
2826 			 * aligned.  Needless to say, the validity of the data
2827 			 * needs to also be DEV_BSIZE aligned.  Note that this
2828 			 * fails with NFS if the server or some other client
2829 			 * extends the file's EOF.  If our buffer is resized,
2830 			 * B_CACHE may remain set! XXX
2831 			 */
2832 
2833 			toff = bp->b_bcount;
2834 			tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2835 
2836 			while ((bp->b_flags & B_CACHE) && toff < size) {
2837 				vm_pindex_t pi;
2838 
2839 				if (tinc > (size - toff))
2840 					tinc = size - toff;
2841 
2842 				pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2843 				    PAGE_SHIFT;
2844 
2845 				vfs_buf_test_cache(
2846 				    bp,
2847 				    bp->b_offset,
2848 				    toff,
2849 				    tinc,
2850 				    bp->b_pages[pi]
2851 				);
2852 				toff += tinc;
2853 				tinc = PAGE_SIZE;
2854 			}
2855 			VM_OBJECT_UNLOCK(obj);
2856 
2857 			/*
2858 			 * Step 3, fixup the KVM pmap.  Remember that
2859 			 * bp->b_data is relative to bp->b_offset, but
2860 			 * bp->b_offset may be offset into the first page.
2861 			 */
2862 
2863 			bp->b_data = (caddr_t)
2864 			    trunc_page((vm_offset_t)bp->b_data);
2865 			pmap_qenter(
2866 			    (vm_offset_t)bp->b_data,
2867 			    bp->b_pages,
2868 			    bp->b_npages
2869 			);
2870 
2871 			bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2872 			    (vm_offset_t)(bp->b_offset & PAGE_MASK));
2873 		}
2874 	}
2875 	if (newbsize < bp->b_bufsize)
2876 		bufspacewakeup();
2877 	bp->b_bufsize = newbsize;	/* actual buffer allocation	*/
2878 	bp->b_bcount = size;		/* requested buffer size	*/
2879 	return 1;
2880 }
2881 
2882 void
2883 biodone(struct bio *bp)
2884 {
2885 
2886 	mtx_lock(&bdonelock);
2887 	bp->bio_flags |= BIO_DONE;
2888 	if (bp->bio_done == NULL)
2889 		wakeup(bp);
2890 	mtx_unlock(&bdonelock);
2891 	if (bp->bio_done != NULL)
2892 		bp->bio_done(bp);
2893 }
2894 
2895 /*
2896  * Wait for a BIO to finish.
2897  *
2898  * XXX: resort to a timeout for now.  The optimal locking (if any) for this
2899  * case is not yet clear.
2900  */
2901 int
2902 biowait(struct bio *bp, const char *wchan)
2903 {
2904 
2905 	mtx_lock(&bdonelock);
2906 	while ((bp->bio_flags & BIO_DONE) == 0)
2907 		msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
2908 	mtx_unlock(&bdonelock);
2909 	if (bp->bio_error != 0)
2910 		return (bp->bio_error);
2911 	if (!(bp->bio_flags & BIO_ERROR))
2912 		return (0);
2913 	return (EIO);
2914 }
2915 
2916 void
2917 biofinish(struct bio *bp, struct devstat *stat, int error)
2918 {
2919 
2920 	if (error) {
2921 		bp->bio_error = error;
2922 		bp->bio_flags |= BIO_ERROR;
2923 	}
2924 	if (stat != NULL)
2925 		devstat_end_transaction_bio(stat, bp);
2926 	biodone(bp);
2927 }
2928 
2929 /*
2930  *	bufwait:
2931  *
2932  *	Wait for buffer I/O completion, returning error status.  The buffer
2933  *	is left locked and B_DONE on return.  B_EINTR is converted into an EINTR
2934  *	error and cleared.
2935  */
2936 int
2937 bufwait(struct buf *bp)
2938 {
2939 	if (bp->b_iocmd == BIO_READ)
2940 		bwait(bp, PRIBIO, "biord");
2941 	else
2942 		bwait(bp, PRIBIO, "biowr");
2943 	if (bp->b_flags & B_EINTR) {
2944 		bp->b_flags &= ~B_EINTR;
2945 		return (EINTR);
2946 	}
2947 	if (bp->b_ioflags & BIO_ERROR) {
2948 		return (bp->b_error ? bp->b_error : EIO);
2949 	} else {
2950 		return (0);
2951 	}
2952 }
2953 
2954  /*
2955   * Call back function from struct bio back up to struct buf.
2956   */
2957 static void
2958 bufdonebio(struct bio *bip)
2959 {
2960 	struct buf *bp;
2961 
2962 	bp = bip->bio_caller2;
2963 	bp->b_resid = bp->b_bcount - bip->bio_completed;
2964 	bp->b_resid = bip->bio_resid;	/* XXX: remove */
2965 	bp->b_ioflags = bip->bio_flags;
2966 	bp->b_error = bip->bio_error;
2967 	if (bp->b_error)
2968 		bp->b_ioflags |= BIO_ERROR;
2969 	bufdone(bp);
2970 	g_destroy_bio(bip);
2971 }
2972 
2973 void
2974 dev_strategy(struct cdev *dev, struct buf *bp)
2975 {
2976 	struct cdevsw *csw;
2977 	struct bio *bip;
2978 
2979 	if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
2980 		panic("b_iocmd botch");
2981 	for (;;) {
2982 		bip = g_new_bio();
2983 		if (bip != NULL)
2984 			break;
2985 		/* Try again later */
2986 		tsleep(&bp, PRIBIO, "dev_strat", hz/10);
2987 	}
2988 	bip->bio_cmd = bp->b_iocmd;
2989 	bip->bio_offset = bp->b_iooffset;
2990 	bip->bio_length = bp->b_bcount;
2991 	bip->bio_bcount = bp->b_bcount;	/* XXX: remove */
2992 	bip->bio_data = bp->b_data;
2993 	bip->bio_done = bufdonebio;
2994 	bip->bio_caller2 = bp;
2995 	bip->bio_dev = dev;
2996 	KASSERT(dev->si_refcount > 0,
2997 	    ("dev_strategy on un-referenced struct cdev *(%s)",
2998 	    devtoname(dev)));
2999 	csw = dev_refthread(dev);
3000 	if (csw == NULL) {
3001 		bp->b_error = ENXIO;
3002 		bp->b_ioflags = BIO_ERROR;
3003 		bufdone(bp);
3004 		return;
3005 	}
3006 	(*csw->d_strategy)(bip);
3007 	dev_relthread(dev);
3008 }
3009 
3010 /*
3011  *	bufdone:
3012  *
3013  *	Finish I/O on a buffer, optionally calling a completion function.
3014  *	This is usually called from an interrupt so process blocking is
3015  *	not allowed.
3016  *
3017  *	biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3018  *	In a non-VMIO bp, B_CACHE will be set on the next getblk()
3019  *	assuming B_INVAL is clear.
3020  *
3021  *	For the VMIO case, we set B_CACHE if the op was a read and no
3022  *	read error occured, or if the op was a write.  B_CACHE is never
3023  *	set if the buffer is invalid or otherwise uncacheable.
3024  *
3025  *	biodone does not mess with B_INVAL, allowing the I/O routine or the
3026  *	initiator to leave B_INVAL set to brelse the buffer out of existance
3027  *	in the biodone routine.
3028  */
3029 void
3030 bufdone(struct buf *bp)
3031 {
3032 	struct bufobj *dropobj;
3033 	void    (*biodone)(struct buf *);
3034 
3035 
3036 	CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3037 	dropobj = NULL;
3038 
3039 	KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
3040 	KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3041 
3042 	runningbufwakeup(bp);
3043 	if (bp->b_iocmd == BIO_WRITE)
3044 		dropobj = bp->b_bufobj;
3045 	/* call optional completion function if requested */
3046 	if (bp->b_iodone != NULL) {
3047 		biodone = bp->b_iodone;
3048 		bp->b_iodone = NULL;
3049 		(*biodone) (bp);
3050 		if (dropobj)
3051 			bufobj_wdrop(dropobj);
3052 		return;
3053 	}
3054 	if (LIST_FIRST(&bp->b_dep) != NULL)
3055 		buf_complete(bp);
3056 
3057 	if (bp->b_flags & B_VMIO) {
3058 		int i;
3059 		vm_ooffset_t foff;
3060 		vm_page_t m;
3061 		vm_object_t obj;
3062 		int iosize;
3063 		struct vnode *vp = bp->b_vp;
3064 
3065 		obj = bp->b_bufobj->bo_object;
3066 
3067 #if defined(VFS_BIO_DEBUG)
3068 		mp_fixme("usecount and vflag accessed without locks.");
3069 		if (vp->v_usecount == 0) {
3070 			panic("biodone: zero vnode ref count");
3071 		}
3072 
3073 		KASSERT(vp->v_object != NULL,
3074 			("biodone: vnode %p has no vm_object", vp));
3075 #endif
3076 
3077 		foff = bp->b_offset;
3078 		KASSERT(bp->b_offset != NOOFFSET,
3079 		    ("biodone: no buffer offset"));
3080 
3081 		VM_OBJECT_LOCK(obj);
3082 #if defined(VFS_BIO_DEBUG)
3083 		if (obj->paging_in_progress < bp->b_npages) {
3084 			printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3085 			    obj->paging_in_progress, bp->b_npages);
3086 		}
3087 #endif
3088 
3089 		/*
3090 		 * Set B_CACHE if the op was a normal read and no error
3091 		 * occured.  B_CACHE is set for writes in the b*write()
3092 		 * routines.
3093 		 */
3094 		iosize = bp->b_bcount - bp->b_resid;
3095 		if (bp->b_iocmd == BIO_READ &&
3096 		    !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3097 		    !(bp->b_ioflags & BIO_ERROR)) {
3098 			bp->b_flags |= B_CACHE;
3099 		}
3100 		vm_page_lock_queues();
3101 		for (i = 0; i < bp->b_npages; i++) {
3102 			int bogusflag = 0;
3103 			int resid;
3104 
3105 			resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3106 			if (resid > iosize)
3107 				resid = iosize;
3108 
3109 			/*
3110 			 * cleanup bogus pages, restoring the originals
3111 			 */
3112 			m = bp->b_pages[i];
3113 			if (m == bogus_page) {
3114 				bogusflag = 1;
3115 				m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3116 				if (m == NULL)
3117 					panic("biodone: page disappeared!");
3118 				bp->b_pages[i] = m;
3119 				pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3120 			}
3121 #if defined(VFS_BIO_DEBUG)
3122 			if (OFF_TO_IDX(foff) != m->pindex) {
3123 				printf(
3124 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3125 				    (intmax_t)foff, (uintmax_t)m->pindex);
3126 			}
3127 #endif
3128 
3129 			/*
3130 			 * In the write case, the valid and clean bits are
3131 			 * already changed correctly ( see bdwrite() ), so we
3132 			 * only need to do this here in the read case.
3133 			 */
3134 			if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3135 				vfs_page_set_valid(bp, foff, i, m);
3136 			}
3137 
3138 			/*
3139 			 * when debugging new filesystems or buffer I/O methods, this
3140 			 * is the most common error that pops up.  if you see this, you
3141 			 * have not set the page busy flag correctly!!!
3142 			 */
3143 			if (m->busy == 0) {
3144 				printf("biodone: page busy < 0, "
3145 				    "pindex: %d, foff: 0x(%x,%x), "
3146 				    "resid: %d, index: %d\n",
3147 				    (int) m->pindex, (int)(foff >> 32),
3148 						(int) foff & 0xffffffff, resid, i);
3149 				if (!vn_isdisk(vp, NULL))
3150 					printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3151 					    (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3152 					    (intmax_t) bp->b_lblkno,
3153 					    bp->b_flags, bp->b_npages);
3154 				else
3155 					printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3156 					    (intmax_t) bp->b_lblkno,
3157 					    bp->b_flags, bp->b_npages);
3158 				printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3159 				    (u_long)m->valid, (u_long)m->dirty,
3160 				    m->wire_count);
3161 				panic("biodone: page busy < 0\n");
3162 			}
3163 			vm_page_io_finish(m);
3164 			vm_object_pip_subtract(obj, 1);
3165 			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3166 			iosize -= resid;
3167 		}
3168 		vm_page_unlock_queues();
3169 		vm_object_pip_wakeupn(obj, 0);
3170 		VM_OBJECT_UNLOCK(obj);
3171 	}
3172 
3173 	/*
3174 	 * For asynchronous completions, release the buffer now. The brelse
3175 	 * will do a wakeup there if necessary - so no need to do a wakeup
3176 	 * here in the async case. The sync case always needs to do a wakeup.
3177 	 */
3178 
3179 	if (bp->b_flags & B_ASYNC) {
3180 		if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3181 			brelse(bp);
3182 		else
3183 			bqrelse(bp);
3184 	} else
3185 		bdone(bp);
3186 	if (dropobj)
3187 		bufobj_wdrop(dropobj);
3188 }
3189 
3190 /*
3191  * This routine is called in lieu of iodone in the case of
3192  * incomplete I/O.  This keeps the busy status for pages
3193  * consistant.
3194  */
3195 void
3196 vfs_unbusy_pages(struct buf *bp)
3197 {
3198 	int i;
3199 	vm_object_t obj;
3200 	vm_page_t m;
3201 
3202 	runningbufwakeup(bp);
3203 	if (!(bp->b_flags & B_VMIO))
3204 		return;
3205 
3206 	obj = bp->b_bufobj->bo_object;
3207 	VM_OBJECT_LOCK(obj);
3208 	vm_page_lock_queues();
3209 	for (i = 0; i < bp->b_npages; i++) {
3210 		m = bp->b_pages[i];
3211 		if (m == bogus_page) {
3212 			m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3213 			if (!m)
3214 				panic("vfs_unbusy_pages: page missing\n");
3215 			bp->b_pages[i] = m;
3216 			pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3217 			    bp->b_pages, bp->b_npages);
3218 		}
3219 		vm_object_pip_subtract(obj, 1);
3220 		vm_page_io_finish(m);
3221 	}
3222 	vm_page_unlock_queues();
3223 	vm_object_pip_wakeupn(obj, 0);
3224 	VM_OBJECT_UNLOCK(obj);
3225 }
3226 
3227 /*
3228  * vfs_page_set_valid:
3229  *
3230  *	Set the valid bits in a page based on the supplied offset.   The
3231  *	range is restricted to the buffer's size.
3232  *
3233  *	This routine is typically called after a read completes.
3234  */
3235 static void
3236 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3237 {
3238 	vm_ooffset_t soff, eoff;
3239 
3240 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3241 	/*
3242 	 * Start and end offsets in buffer.  eoff - soff may not cross a
3243 	 * page boundry or cross the end of the buffer.  The end of the
3244 	 * buffer, in this case, is our file EOF, not the allocation size
3245 	 * of the buffer.
3246 	 */
3247 	soff = off;
3248 	eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3249 	if (eoff > bp->b_offset + bp->b_bcount)
3250 		eoff = bp->b_offset + bp->b_bcount;
3251 
3252 	/*
3253 	 * Set valid range.  This is typically the entire buffer and thus the
3254 	 * entire page.
3255 	 */
3256 	if (eoff > soff) {
3257 		vm_page_set_validclean(
3258 		    m,
3259 		   (vm_offset_t) (soff & PAGE_MASK),
3260 		   (vm_offset_t) (eoff - soff)
3261 		);
3262 	}
3263 }
3264 
3265 /*
3266  * This routine is called before a device strategy routine.
3267  * It is used to tell the VM system that paging I/O is in
3268  * progress, and treat the pages associated with the buffer
3269  * almost as being PG_BUSY.  Also the object paging_in_progress
3270  * flag is handled to make sure that the object doesn't become
3271  * inconsistant.
3272  *
3273  * Since I/O has not been initiated yet, certain buffer flags
3274  * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3275  * and should be ignored.
3276  */
3277 void
3278 vfs_busy_pages(struct buf *bp, int clear_modify)
3279 {
3280 	int i, bogus;
3281 	vm_object_t obj;
3282 	vm_ooffset_t foff;
3283 	vm_page_t m;
3284 
3285 	if (!(bp->b_flags & B_VMIO))
3286 		return;
3287 
3288 	obj = bp->b_bufobj->bo_object;
3289 	foff = bp->b_offset;
3290 	KASSERT(bp->b_offset != NOOFFSET,
3291 	    ("vfs_busy_pages: no buffer offset"));
3292 	vfs_setdirty(bp);
3293 	VM_OBJECT_LOCK(obj);
3294 retry:
3295 	vm_page_lock_queues();
3296 	for (i = 0; i < bp->b_npages; i++) {
3297 		m = bp->b_pages[i];
3298 
3299 		if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3300 			goto retry;
3301 	}
3302 	bogus = 0;
3303 	for (i = 0; i < bp->b_npages; i++) {
3304 		m = bp->b_pages[i];
3305 
3306 		if ((bp->b_flags & B_CLUSTER) == 0) {
3307 			vm_object_pip_add(obj, 1);
3308 			vm_page_io_start(m);
3309 		}
3310 		/*
3311 		 * When readying a buffer for a read ( i.e
3312 		 * clear_modify == 0 ), it is important to do
3313 		 * bogus_page replacement for valid pages in
3314 		 * partially instantiated buffers.  Partially
3315 		 * instantiated buffers can, in turn, occur when
3316 		 * reconstituting a buffer from its VM backing store
3317 		 * base.  We only have to do this if B_CACHE is
3318 		 * clear ( which causes the I/O to occur in the
3319 		 * first place ).  The replacement prevents the read
3320 		 * I/O from overwriting potentially dirty VM-backed
3321 		 * pages.  XXX bogus page replacement is, uh, bogus.
3322 		 * It may not work properly with small-block devices.
3323 		 * We need to find a better way.
3324 		 */
3325 		pmap_remove_all(m);
3326 		if (clear_modify)
3327 			vfs_page_set_valid(bp, foff, i, m);
3328 		else if (m->valid == VM_PAGE_BITS_ALL &&
3329 		    (bp->b_flags & B_CACHE) == 0) {
3330 			bp->b_pages[i] = bogus_page;
3331 			bogus++;
3332 		}
3333 		foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3334 	}
3335 	vm_page_unlock_queues();
3336 	VM_OBJECT_UNLOCK(obj);
3337 	if (bogus)
3338 		pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3339 		    bp->b_pages, bp->b_npages);
3340 }
3341 
3342 /*
3343  * Tell the VM system that the pages associated with this buffer
3344  * are clean.  This is used for delayed writes where the data is
3345  * going to go to disk eventually without additional VM intevention.
3346  *
3347  * Note that while we only really need to clean through to b_bcount, we
3348  * just go ahead and clean through to b_bufsize.
3349  */
3350 static void
3351 vfs_clean_pages(struct buf *bp)
3352 {
3353 	int i;
3354 	vm_ooffset_t foff, noff, eoff;
3355 	vm_page_t m;
3356 
3357 	if (!(bp->b_flags & B_VMIO))
3358 		return;
3359 
3360 	foff = bp->b_offset;
3361 	KASSERT(bp->b_offset != NOOFFSET,
3362 	    ("vfs_clean_pages: no buffer offset"));
3363 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3364 	vm_page_lock_queues();
3365 	for (i = 0; i < bp->b_npages; i++) {
3366 		m = bp->b_pages[i];
3367 		noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3368 		eoff = noff;
3369 
3370 		if (eoff > bp->b_offset + bp->b_bufsize)
3371 			eoff = bp->b_offset + bp->b_bufsize;
3372 		vfs_page_set_valid(bp, foff, i, m);
3373 		/* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3374 		foff = noff;
3375 	}
3376 	vm_page_unlock_queues();
3377 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3378 }
3379 
3380 /*
3381  *	vfs_bio_set_validclean:
3382  *
3383  *	Set the range within the buffer to valid and clean.  The range is
3384  *	relative to the beginning of the buffer, b_offset.  Note that b_offset
3385  *	itself may be offset from the beginning of the first page.
3386  *
3387  */
3388 
3389 void
3390 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3391 {
3392 	int i, n;
3393 	vm_page_t m;
3394 
3395 	if (!(bp->b_flags & B_VMIO))
3396 		return;
3397 	/*
3398 	 * Fixup base to be relative to beginning of first page.
3399 	 * Set initial n to be the maximum number of bytes in the
3400 	 * first page that can be validated.
3401 	 */
3402 
3403 	base += (bp->b_offset & PAGE_MASK);
3404 	n = PAGE_SIZE - (base & PAGE_MASK);
3405 
3406 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3407 	vm_page_lock_queues();
3408 	for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3409 		m = bp->b_pages[i];
3410 		if (n > size)
3411 			n = size;
3412 		vm_page_set_validclean(m, base & PAGE_MASK, n);
3413 		base += n;
3414 		size -= n;
3415 		n = PAGE_SIZE;
3416 	}
3417 	vm_page_unlock_queues();
3418 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3419 }
3420 
3421 /*
3422  *	vfs_bio_clrbuf:
3423  *
3424  *	clear a buffer.  This routine essentially fakes an I/O, so we need
3425  *	to clear BIO_ERROR and B_INVAL.
3426  *
3427  *	Note that while we only theoretically need to clear through b_bcount,
3428  *	we go ahead and clear through b_bufsize.
3429  */
3430 
3431 void
3432 vfs_bio_clrbuf(struct buf *bp)
3433 {
3434 	int i, j, mask = 0;
3435 	caddr_t sa, ea;
3436 
3437 	if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3438 		clrbuf(bp);
3439 		return;
3440 	}
3441 
3442 	bp->b_flags &= ~B_INVAL;
3443 	bp->b_ioflags &= ~BIO_ERROR;
3444 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3445 	if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3446 	    (bp->b_offset & PAGE_MASK) == 0) {
3447 		if (bp->b_pages[0] == bogus_page)
3448 			goto unlock;
3449 		mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3450 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3451 		if ((bp->b_pages[0]->valid & mask) == mask)
3452 			goto unlock;
3453 		if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3454 		    ((bp->b_pages[0]->valid & mask) == 0)) {
3455 			bzero(bp->b_data, bp->b_bufsize);
3456 			bp->b_pages[0]->valid |= mask;
3457 			goto unlock;
3458 		}
3459 	}
3460 	ea = sa = bp->b_data;
3461 	for(i = 0; i < bp->b_npages; i++, sa = ea) {
3462 		ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3463 		ea = (caddr_t)(vm_offset_t)ulmin(
3464 		    (u_long)(vm_offset_t)ea,
3465 		    (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3466 		if (bp->b_pages[i] == bogus_page)
3467 			continue;
3468 		j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3469 		mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3470 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3471 		if ((bp->b_pages[i]->valid & mask) == mask)
3472 			continue;
3473 		if ((bp->b_pages[i]->valid & mask) == 0) {
3474 			if ((bp->b_pages[i]->flags & PG_ZERO) == 0)
3475 				bzero(sa, ea - sa);
3476 		} else {
3477 			for (; sa < ea; sa += DEV_BSIZE, j++) {
3478 				if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3479 				    (bp->b_pages[i]->valid & (1 << j)) == 0)
3480 					bzero(sa, DEV_BSIZE);
3481 			}
3482 		}
3483 		bp->b_pages[i]->valid |= mask;
3484 	}
3485 unlock:
3486 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3487 	bp->b_resid = 0;
3488 }
3489 
3490 /*
3491  * vm_hold_load_pages and vm_hold_free_pages get pages into
3492  * a buffers address space.  The pages are anonymous and are
3493  * not associated with a file object.
3494  */
3495 static void
3496 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3497 {
3498 	vm_offset_t pg;
3499 	vm_page_t p;
3500 	int index;
3501 
3502 	to = round_page(to);
3503 	from = round_page(from);
3504 	index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3505 
3506 	VM_OBJECT_LOCK(kernel_object);
3507 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3508 tryagain:
3509 		/*
3510 		 * note: must allocate system pages since blocking here
3511 		 * could intefere with paging I/O, no matter which
3512 		 * process we are.
3513 		 */
3514 		p = vm_page_alloc(kernel_object,
3515 			((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3516 		    VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3517 		if (!p) {
3518 			atomic_add_int(&vm_pageout_deficit,
3519 			    (to - pg) >> PAGE_SHIFT);
3520 			VM_OBJECT_UNLOCK(kernel_object);
3521 			VM_WAIT;
3522 			VM_OBJECT_LOCK(kernel_object);
3523 			goto tryagain;
3524 		}
3525 		p->valid = VM_PAGE_BITS_ALL;
3526 		pmap_qenter(pg, &p, 1);
3527 		bp->b_pages[index] = p;
3528 	}
3529 	VM_OBJECT_UNLOCK(kernel_object);
3530 	bp->b_npages = index;
3531 }
3532 
3533 /* Return pages associated with this buf to the vm system */
3534 static void
3535 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3536 {
3537 	vm_offset_t pg;
3538 	vm_page_t p;
3539 	int index, newnpages;
3540 
3541 	from = round_page(from);
3542 	to = round_page(to);
3543 	newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3544 
3545 	VM_OBJECT_LOCK(kernel_object);
3546 	for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3547 		p = bp->b_pages[index];
3548 		if (p && (index < bp->b_npages)) {
3549 			if (p->busy) {
3550 				printf(
3551 			    "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3552 				    (intmax_t)bp->b_blkno,
3553 				    (intmax_t)bp->b_lblkno);
3554 			}
3555 			bp->b_pages[index] = NULL;
3556 			pmap_qremove(pg, 1);
3557 			vm_page_lock_queues();
3558 			vm_page_unwire(p, 0);
3559 			vm_page_free(p);
3560 			vm_page_unlock_queues();
3561 		}
3562 	}
3563 	VM_OBJECT_UNLOCK(kernel_object);
3564 	bp->b_npages = newnpages;
3565 }
3566 
3567 /*
3568  * Map an IO request into kernel virtual address space.
3569  *
3570  * All requests are (re)mapped into kernel VA space.
3571  * Notice that we use b_bufsize for the size of the buffer
3572  * to be mapped.  b_bcount might be modified by the driver.
3573  *
3574  * Note that even if the caller determines that the address space should
3575  * be valid, a race or a smaller-file mapped into a larger space may
3576  * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3577  * check the return value.
3578  */
3579 int
3580 vmapbuf(struct buf *bp)
3581 {
3582 	caddr_t addr, kva;
3583 	vm_prot_t prot;
3584 	int pidx, i;
3585 	struct vm_page *m;
3586 	struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3587 
3588 	if (bp->b_bufsize < 0)
3589 		return (-1);
3590 	prot = VM_PROT_READ;
3591 	if (bp->b_iocmd == BIO_READ)
3592 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
3593 	for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3594 	     addr < bp->b_data + bp->b_bufsize;
3595 	     addr += PAGE_SIZE, pidx++) {
3596 		/*
3597 		 * Do the vm_fault if needed; do the copy-on-write thing
3598 		 * when reading stuff off device into memory.
3599 		 *
3600 		 * NOTE! Must use pmap_extract() because addr may be in
3601 		 * the userland address space, and kextract is only guarenteed
3602 		 * to work for the kernland address space (see: sparc64 port).
3603 		 */
3604 retry:
3605 		if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3606 		    prot) < 0) {
3607 			vm_page_lock_queues();
3608 			for (i = 0; i < pidx; ++i) {
3609 				vm_page_unhold(bp->b_pages[i]);
3610 				bp->b_pages[i] = NULL;
3611 			}
3612 			vm_page_unlock_queues();
3613 			return(-1);
3614 		}
3615 		m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3616 		if (m == NULL)
3617 			goto retry;
3618 		bp->b_pages[pidx] = m;
3619 	}
3620 	if (pidx > btoc(MAXPHYS))
3621 		panic("vmapbuf: mapped more than MAXPHYS");
3622 	pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3623 
3624 	kva = bp->b_saveaddr;
3625 	bp->b_npages = pidx;
3626 	bp->b_saveaddr = bp->b_data;
3627 	bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3628 	return(0);
3629 }
3630 
3631 /*
3632  * Free the io map PTEs associated with this IO operation.
3633  * We also invalidate the TLB entries and restore the original b_addr.
3634  */
3635 void
3636 vunmapbuf(struct buf *bp)
3637 {
3638 	int pidx;
3639 	int npages;
3640 
3641 	npages = bp->b_npages;
3642 	pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3643 	vm_page_lock_queues();
3644 	for (pidx = 0; pidx < npages; pidx++)
3645 		vm_page_unhold(bp->b_pages[pidx]);
3646 	vm_page_unlock_queues();
3647 
3648 	bp->b_data = bp->b_saveaddr;
3649 }
3650 
3651 void
3652 bdone(struct buf *bp)
3653 {
3654 
3655 	mtx_lock(&bdonelock);
3656 	bp->b_flags |= B_DONE;
3657 	wakeup(bp);
3658 	mtx_unlock(&bdonelock);
3659 }
3660 
3661 void
3662 bwait(struct buf *bp, u_char pri, const char *wchan)
3663 {
3664 
3665 	mtx_lock(&bdonelock);
3666 	while ((bp->b_flags & B_DONE) == 0)
3667 		msleep(bp, &bdonelock, pri, wchan, 0);
3668 	mtx_unlock(&bdonelock);
3669 }
3670 
3671 int
3672 bufsync(struct bufobj *bo, int waitfor, struct thread *td)
3673 {
3674 
3675 	return (VOP_FSYNC(bo->__bo_vnode, waitfor, td));
3676 }
3677 
3678 void
3679 bufstrategy(struct bufobj *bo, struct buf *bp)
3680 {
3681 	int i = 0;
3682 	struct vnode *vp;
3683 
3684 	vp = bp->b_vp;
3685 	KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3686 	KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3687 	    ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3688 	i = VOP_STRATEGY(vp, bp);
3689 	KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3690 }
3691 
3692 void
3693 bufobj_wrefl(struct bufobj *bo)
3694 {
3695 
3696 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3697 	ASSERT_BO_LOCKED(bo);
3698 	bo->bo_numoutput++;
3699 }
3700 
3701 void
3702 bufobj_wref(struct bufobj *bo)
3703 {
3704 
3705 	KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3706 	BO_LOCK(bo);
3707 	bo->bo_numoutput++;
3708 	BO_UNLOCK(bo);
3709 }
3710 
3711 void
3712 bufobj_wdrop(struct bufobj *bo)
3713 {
3714 
3715 	KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3716 	BO_LOCK(bo);
3717 	KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3718 	if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3719 		bo->bo_flag &= ~BO_WWAIT;
3720 		wakeup(&bo->bo_numoutput);
3721 	}
3722 	BO_UNLOCK(bo);
3723 }
3724 
3725 int
3726 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3727 {
3728 	int error;
3729 
3730 	KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3731 	ASSERT_BO_LOCKED(bo);
3732 	error = 0;
3733 	while (bo->bo_numoutput) {
3734 		bo->bo_flag |= BO_WWAIT;
3735 		error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3736 		    slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3737 		if (error)
3738 			break;
3739 	}
3740 	return (error);
3741 }
3742 
3743 #include "opt_ddb.h"
3744 #ifdef DDB
3745 #include <ddb/ddb.h>
3746 
3747 /* DDB command to show buffer data */
3748 DB_SHOW_COMMAND(buffer, db_show_buffer)
3749 {
3750 	/* get args */
3751 	struct buf *bp = (struct buf *)addr;
3752 
3753 	if (!have_addr) {
3754 		db_printf("usage: show buffer <addr>\n");
3755 		return;
3756 	}
3757 
3758 	db_printf("buf at %p\n", bp);
3759 	db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3760 	db_printf(
3761 	    "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3762 	    "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n",
3763 	    bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3764 	    bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno);
3765 	if (bp->b_npages) {
3766 		int i;
3767 		db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3768 		for (i = 0; i < bp->b_npages; i++) {
3769 			vm_page_t m;
3770 			m = bp->b_pages[i];
3771 			db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3772 			    (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3773 			if ((i + 1) < bp->b_npages)
3774 				db_printf(",");
3775 		}
3776 		db_printf("\n");
3777 	}
3778 	lockmgr_printinfo(&bp->b_lock);
3779 }
3780 
3781 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
3782 {
3783 	struct buf *bp;
3784 	int i;
3785 
3786 	for (i = 0; i < nbuf; i++) {
3787 		bp = &buf[i];
3788 		if (lockcount(&bp->b_lock)) {
3789 			db_show_buffer((uintptr_t)bp, 1, 0, NULL);
3790 			db_printf("\n");
3791 		}
3792 	}
3793 }
3794 #endif /* DDB */
3795