FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_bio.c
1 /*-
2 * Copyright (c) 2004 Poul-Henning Kamp
3 * Copyright (c) 1994,1997 John S. Dyson
4 * Copyright (c) 2013 The FreeBSD Foundation
5 * All rights reserved.
6 *
7 * Portions of this software were developed by Konstantin Belousov
8 * under sponsorship from the FreeBSD Foundation.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * SUCH DAMAGE.
30 */
31
32 /*
33 * this file contains a new buffer I/O scheme implementing a coherent
34 * VM object and buffer cache scheme. Pains have been taken to make
35 * sure that the performance degradation associated with schemes such
36 * as this is not realized.
37 *
38 * Author: John S. Dyson
39 * Significant help during the development and debugging phases
40 * had been provided by David Greenman, also of the FreeBSD core team.
41 *
42 * see man buf(9) for more info.
43 */
44
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD: releng/10.1/sys/kern/vfs_bio.c 273776 2014-10-28 14:01:58Z mav $");
47
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/conf.h>
52 #include <sys/buf.h>
53 #include <sys/devicestat.h>
54 #include <sys/eventhandler.h>
55 #include <sys/fail.h>
56 #include <sys/limits.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/kernel.h>
62 #include <sys/kthread.h>
63 #include <sys/proc.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
66 #include <sys/sysctl.h>
67 #include <sys/vmem.h>
68 #include <sys/vmmeter.h>
69 #include <sys/vnode.h>
70 #include <geom/geom.h>
71 #include <vm/vm.h>
72 #include <vm/vm_param.h>
73 #include <vm/vm_kern.h>
74 #include <vm/vm_pageout.h>
75 #include <vm/vm_page.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/vm_map.h>
79 #include "opt_compat.h"
80 #include "opt_swap.h"
81
82 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
83
84 struct bio_ops bioops; /* I/O operation notification */
85
86 struct buf_ops buf_ops_bio = {
87 .bop_name = "buf_ops_bio",
88 .bop_write = bufwrite,
89 .bop_strategy = bufstrategy,
90 .bop_sync = bufsync,
91 .bop_bdflush = bufbdflush,
92 };
93
94 /*
95 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
96 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
97 */
98 struct buf *buf; /* buffer header pool */
99 caddr_t unmapped_buf;
100
101 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
102 struct proc *bufdaemonproc;
103
104 static int inmem(struct vnode *vp, daddr_t blkno);
105 static void vm_hold_free_pages(struct buf *bp, int newbsize);
106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
107 vm_offset_t to);
108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
110 vm_page_t m);
111 static void vfs_clean_pages_dirty_buf(struct buf *bp);
112 static void vfs_setdirty_locked_object(struct buf *bp);
113 static void vfs_vmio_release(struct buf *bp);
114 static int vfs_bio_clcheck(struct vnode *vp, int size,
115 daddr_t lblkno, daddr_t blkno);
116 static int buf_flush(int);
117 static int flushbufqueues(int, int);
118 static void buf_daemon(void);
119 static void bremfreel(struct buf *bp);
120 static __inline void bd_wakeup(void);
121 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
122 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
123 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
124 #endif
125
126 int vmiodirenable = TRUE;
127 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
128 "Use the VM system for directory writes");
129 long runningbufspace;
130 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
131 "Amount of presently outstanding async buffer io");
132 static long bufspace;
133 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
134 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
135 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
136 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
137 #else
138 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
139 "Virtual memory used for buffers");
140 #endif
141 static long unmapped_bufspace;
142 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD,
143 &unmapped_bufspace, 0,
144 "Amount of unmapped buffers, inclusive in the bufspace");
145 static long maxbufspace;
146 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
147 "Maximum allowed value of bufspace (including buf_daemon)");
148 static long bufmallocspace;
149 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
150 "Amount of malloced memory for buffers");
151 static long maxbufmallocspace;
152 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
153 "Maximum amount of malloced memory for buffers");
154 static long lobufspace;
155 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
156 "Minimum amount of buffers we want to have");
157 long hibufspace;
158 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
159 "Maximum allowed value of bufspace (excluding buf_daemon)");
160 static int bufreusecnt;
161 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
162 "Number of times we have reused a buffer");
163 static int buffreekvacnt;
164 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
165 "Number of times we have freed the KVA space from some buffer");
166 static int bufdefragcnt;
167 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
168 "Number of times we have had to repeat buffer allocation to defragment");
169 static long lorunningspace;
170 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
171 "Minimum preferred space used for in-progress I/O");
172 static long hirunningspace;
173 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
174 "Maximum amount of space to use for in-progress I/O");
175 int dirtybufferflushes;
176 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
177 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
178 int bdwriteskip;
179 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
180 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
181 int altbufferflushes;
182 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
183 0, "Number of fsync flushes to limit dirty buffers");
184 static int recursiveflushes;
185 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
186 0, "Number of flushes skipped due to being recursive");
187 static int numdirtybuffers;
188 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
189 "Number of buffers that are dirty (has unwritten changes) at the moment");
190 static int lodirtybuffers;
191 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
192 "How many buffers we want to have free before bufdaemon can sleep");
193 static int hidirtybuffers;
194 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
195 "When the number of dirty buffers is considered severe");
196 int dirtybufthresh;
197 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
198 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
199 static int numfreebuffers;
200 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
201 "Number of free buffers");
202 static int lofreebuffers;
203 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
204 "XXX Unused");
205 static int hifreebuffers;
206 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
207 "XXX Complicatedly unused");
208 static int getnewbufcalls;
209 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
210 "Number of calls to getnewbuf");
211 static int getnewbufrestarts;
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
213 "Number of times getnewbuf has had to restart a buffer aquisition");
214 static int mappingrestarts;
215 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
216 "Number of times getblk has had to restart a buffer mapping for "
217 "unmapped buffer");
218 static int flushbufqtarget = 100;
219 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
220 "Amount of work to do in flushbufqueues when helping bufdaemon");
221 static long notbufdflushes;
222 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0,
223 "Number of dirty buffer flushes done by the bufdaemon helpers");
224 static long barrierwrites;
225 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
226 "Number of barrier writes");
227 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
228 &unmapped_buf_allowed, 0,
229 "Permit the use of the unmapped i/o");
230
231 /*
232 * Lock for the non-dirty bufqueues
233 */
234 static struct mtx_padalign bqclean;
235
236 /*
237 * Lock for the dirty queue.
238 */
239 static struct mtx_padalign bqdirty;
240
241 /*
242 * This lock synchronizes access to bd_request.
243 */
244 static struct mtx_padalign bdlock;
245
246 /*
247 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
248 * waitrunningbufspace().
249 */
250 static struct mtx_padalign rbreqlock;
251
252 /*
253 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
254 */
255 static struct rwlock_padalign nblock;
256
257 /*
258 * Lock that protects bdirtywait.
259 */
260 static struct mtx_padalign bdirtylock;
261
262 /*
263 * Wakeup point for bufdaemon, as well as indicator of whether it is already
264 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
265 * is idling.
266 */
267 static int bd_request;
268
269 /*
270 * Request for the buf daemon to write more buffers than is indicated by
271 * lodirtybuf. This may be necessary to push out excess dependencies or
272 * defragment the address space where a simple count of the number of dirty
273 * buffers is insufficient to characterize the demand for flushing them.
274 */
275 static int bd_speedupreq;
276
277 /*
278 * bogus page -- for I/O to/from partially complete buffers
279 * this is a temporary solution to the problem, but it is not
280 * really that bad. it would be better to split the buffer
281 * for input in the case of buffers partially already in memory,
282 * but the code is intricate enough already.
283 */
284 vm_page_t bogus_page;
285
286 /*
287 * Synchronization (sleep/wakeup) variable for active buffer space requests.
288 * Set when wait starts, cleared prior to wakeup().
289 * Used in runningbufwakeup() and waitrunningbufspace().
290 */
291 static int runningbufreq;
292
293 /*
294 * Synchronization (sleep/wakeup) variable for buffer requests.
295 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
296 * by and/or.
297 * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
298 * getnewbuf(), and getblk().
299 */
300 static volatile int needsbuffer;
301
302 /*
303 * Synchronization for bwillwrite() waiters.
304 */
305 static int bdirtywait;
306
307 /*
308 * Definitions for the buffer free lists.
309 */
310 #define BUFFER_QUEUES 5 /* number of free buffer queues */
311
312 #define QUEUE_NONE 0 /* on no queue */
313 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
314 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
315 #define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */
316 #define QUEUE_EMPTY 4 /* empty buffer headers */
317 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */
318
319 /* Queues for free buffers with various properties */
320 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
321 #ifdef INVARIANTS
322 static int bq_len[BUFFER_QUEUES];
323 #endif
324
325 /*
326 * Single global constant for BUF_WMESG, to avoid getting multiple references.
327 * buf_wmesg is referred from macros.
328 */
329 const char *buf_wmesg = BUF_WMESG;
330
331 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
332 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
333 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
334
335 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
336 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
337 static int
338 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
339 {
340 long lvalue;
341 int ivalue;
342
343 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
344 return (sysctl_handle_long(oidp, arg1, arg2, req));
345 lvalue = *(long *)arg1;
346 if (lvalue > INT_MAX)
347 /* On overflow, still write out a long to trigger ENOMEM. */
348 return (sysctl_handle_long(oidp, &lvalue, 0, req));
349 ivalue = lvalue;
350 return (sysctl_handle_int(oidp, &ivalue, 0, req));
351 }
352 #endif
353
354 /*
355 * bqlock:
356 *
357 * Return the appropriate queue lock based on the index.
358 */
359 static inline struct mtx *
360 bqlock(int qindex)
361 {
362
363 if (qindex == QUEUE_DIRTY)
364 return (struct mtx *)(&bqdirty);
365 return (struct mtx *)(&bqclean);
366 }
367
368 /*
369 * bdirtywakeup:
370 *
371 * Wakeup any bwillwrite() waiters.
372 */
373 static void
374 bdirtywakeup(void)
375 {
376 mtx_lock(&bdirtylock);
377 if (bdirtywait) {
378 bdirtywait = 0;
379 wakeup(&bdirtywait);
380 }
381 mtx_unlock(&bdirtylock);
382 }
383
384 /*
385 * bdirtysub:
386 *
387 * Decrement the numdirtybuffers count by one and wakeup any
388 * threads blocked in bwillwrite().
389 */
390 static void
391 bdirtysub(void)
392 {
393
394 if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
395 (lodirtybuffers + hidirtybuffers) / 2)
396 bdirtywakeup();
397 }
398
399 /*
400 * bdirtyadd:
401 *
402 * Increment the numdirtybuffers count by one and wakeup the buf
403 * daemon if needed.
404 */
405 static void
406 bdirtyadd(void)
407 {
408
409 /*
410 * Only do the wakeup once as we cross the boundary. The
411 * buf daemon will keep running until the condition clears.
412 */
413 if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
414 (lodirtybuffers + hidirtybuffers) / 2)
415 bd_wakeup();
416 }
417
418 /*
419 * bufspacewakeup:
420 *
421 * Called when buffer space is potentially available for recovery.
422 * getnewbuf() will block on this flag when it is unable to free
423 * sufficient buffer space. Buffer space becomes recoverable when
424 * bp's get placed back in the queues.
425 */
426
427 static __inline void
428 bufspacewakeup(void)
429 {
430 int need_wakeup, on;
431
432 /*
433 * If someone is waiting for BUF space, wake them up. Even
434 * though we haven't freed the kva space yet, the waiting
435 * process will be able to now.
436 */
437 rw_rlock(&nblock);
438 for (;;) {
439 need_wakeup = 0;
440 on = needsbuffer;
441 if ((on & VFS_BIO_NEED_BUFSPACE) == 0)
442 break;
443 need_wakeup = 1;
444 if (atomic_cmpset_rel_int(&needsbuffer, on,
445 on & ~VFS_BIO_NEED_BUFSPACE))
446 break;
447 }
448 if (need_wakeup)
449 wakeup(__DEVOLATILE(void *, &needsbuffer));
450 rw_runlock(&nblock);
451 }
452
453 /*
454 * runningwakeup:
455 *
456 * Wake up processes that are waiting on asynchronous writes to fall
457 * below lorunningspace.
458 */
459 static void
460 runningwakeup(void)
461 {
462
463 mtx_lock(&rbreqlock);
464 if (runningbufreq) {
465 runningbufreq = 0;
466 wakeup(&runningbufreq);
467 }
468 mtx_unlock(&rbreqlock);
469 }
470
471 /*
472 * runningbufwakeup:
473 *
474 * Decrement the outstanding write count according.
475 */
476 void
477 runningbufwakeup(struct buf *bp)
478 {
479 long space, bspace;
480
481 bspace = bp->b_runningbufspace;
482 if (bspace == 0)
483 return;
484 space = atomic_fetchadd_long(&runningbufspace, -bspace);
485 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
486 space, bspace));
487 bp->b_runningbufspace = 0;
488 /*
489 * Only acquire the lock and wakeup on the transition from exceeding
490 * the threshold to falling below it.
491 */
492 if (space < lorunningspace)
493 return;
494 if (space - bspace > lorunningspace)
495 return;
496 runningwakeup();
497 }
498
499 /*
500 * bufcountadd:
501 *
502 * Called when a buffer has been added to one of the free queues to
503 * account for the buffer and to wakeup anyone waiting for free buffers.
504 * This typically occurs when large amounts of metadata are being handled
505 * by the buffer cache ( else buffer space runs out first, usually ).
506 */
507 static __inline void
508 bufcountadd(struct buf *bp)
509 {
510 int mask, need_wakeup, old, on;
511
512 KASSERT((bp->b_flags & B_INFREECNT) == 0,
513 ("buf %p already counted as free", bp));
514 bp->b_flags |= B_INFREECNT;
515 old = atomic_fetchadd_int(&numfreebuffers, 1);
516 KASSERT(old >= 0 && old < nbuf,
517 ("numfreebuffers climbed to %d", old + 1));
518 mask = VFS_BIO_NEED_ANY;
519 if (numfreebuffers >= hifreebuffers)
520 mask |= VFS_BIO_NEED_FREE;
521 rw_rlock(&nblock);
522 for (;;) {
523 need_wakeup = 0;
524 on = needsbuffer;
525 if (on == 0)
526 break;
527 need_wakeup = 1;
528 if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask))
529 break;
530 }
531 if (need_wakeup)
532 wakeup(__DEVOLATILE(void *, &needsbuffer));
533 rw_runlock(&nblock);
534 }
535
536 /*
537 * bufcountsub:
538 *
539 * Decrement the numfreebuffers count as needed.
540 */
541 static void
542 bufcountsub(struct buf *bp)
543 {
544 int old;
545
546 /*
547 * Fixup numfreebuffers count. If the buffer is invalid or not
548 * delayed-write, the buffer was free and we must decrement
549 * numfreebuffers.
550 */
551 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
552 KASSERT((bp->b_flags & B_INFREECNT) != 0,
553 ("buf %p not counted in numfreebuffers", bp));
554 bp->b_flags &= ~B_INFREECNT;
555 old = atomic_fetchadd_int(&numfreebuffers, -1);
556 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
557 }
558 }
559
560 /*
561 * waitrunningbufspace()
562 *
563 * runningbufspace is a measure of the amount of I/O currently
564 * running. This routine is used in async-write situations to
565 * prevent creating huge backups of pending writes to a device.
566 * Only asynchronous writes are governed by this function.
567 *
568 * This does NOT turn an async write into a sync write. It waits
569 * for earlier writes to complete and generally returns before the
570 * caller's write has reached the device.
571 */
572 void
573 waitrunningbufspace(void)
574 {
575
576 mtx_lock(&rbreqlock);
577 while (runningbufspace > hirunningspace) {
578 runningbufreq = 1;
579 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
580 }
581 mtx_unlock(&rbreqlock);
582 }
583
584
585 /*
586 * vfs_buf_test_cache:
587 *
588 * Called when a buffer is extended. This function clears the B_CACHE
589 * bit if the newly extended portion of the buffer does not contain
590 * valid data.
591 */
592 static __inline
593 void
594 vfs_buf_test_cache(struct buf *bp,
595 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
596 vm_page_t m)
597 {
598
599 VM_OBJECT_ASSERT_LOCKED(m->object);
600 if (bp->b_flags & B_CACHE) {
601 int base = (foff + off) & PAGE_MASK;
602 if (vm_page_is_valid(m, base, size) == 0)
603 bp->b_flags &= ~B_CACHE;
604 }
605 }
606
607 /* Wake up the buffer daemon if necessary */
608 static __inline void
609 bd_wakeup(void)
610 {
611
612 mtx_lock(&bdlock);
613 if (bd_request == 0) {
614 bd_request = 1;
615 wakeup(&bd_request);
616 }
617 mtx_unlock(&bdlock);
618 }
619
620 /*
621 * bd_speedup - speedup the buffer cache flushing code
622 */
623 void
624 bd_speedup(void)
625 {
626 int needwake;
627
628 mtx_lock(&bdlock);
629 needwake = 0;
630 if (bd_speedupreq == 0 || bd_request == 0)
631 needwake = 1;
632 bd_speedupreq = 1;
633 bd_request = 1;
634 if (needwake)
635 wakeup(&bd_request);
636 mtx_unlock(&bdlock);
637 }
638
639 #ifdef __i386__
640 #define TRANSIENT_DENOM 5
641 #else
642 #define TRANSIENT_DENOM 10
643 #endif
644
645 /*
646 * Calculating buffer cache scaling values and reserve space for buffer
647 * headers. This is called during low level kernel initialization and
648 * may be called more then once. We CANNOT write to the memory area
649 * being reserved at this time.
650 */
651 caddr_t
652 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
653 {
654 int tuned_nbuf;
655 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
656
657 /*
658 * physmem_est is in pages. Convert it to kilobytes (assumes
659 * PAGE_SIZE is >= 1K)
660 */
661 physmem_est = physmem_est * (PAGE_SIZE / 1024);
662
663 /*
664 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
665 * For the first 64MB of ram nominally allocate sufficient buffers to
666 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
667 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
668 * the buffer cache we limit the eventual kva reservation to
669 * maxbcache bytes.
670 *
671 * factor represents the 1/4 x ram conversion.
672 */
673 if (nbuf == 0) {
674 int factor = 4 * BKVASIZE / 1024;
675
676 nbuf = 50;
677 if (physmem_est > 4096)
678 nbuf += min((physmem_est - 4096) / factor,
679 65536 / factor);
680 if (physmem_est > 65536)
681 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
682 32 * 1024 * 1024 / (factor * 5));
683
684 if (maxbcache && nbuf > maxbcache / BKVASIZE)
685 nbuf = maxbcache / BKVASIZE;
686 tuned_nbuf = 1;
687 } else
688 tuned_nbuf = 0;
689
690 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
691 maxbuf = (LONG_MAX / 3) / BKVASIZE;
692 if (nbuf > maxbuf) {
693 if (!tuned_nbuf)
694 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
695 maxbuf);
696 nbuf = maxbuf;
697 }
698
699 /*
700 * Ideal allocation size for the transient bio submap if 10%
701 * of the maximal space buffer map. This roughly corresponds
702 * to the amount of the buffer mapped for typical UFS load.
703 *
704 * Clip the buffer map to reserve space for the transient
705 * BIOs, if its extent is bigger than 90% (80% on i386) of the
706 * maximum buffer map extent on the platform.
707 *
708 * The fall-back to the maxbuf in case of maxbcache unset,
709 * allows to not trim the buffer KVA for the architectures
710 * with ample KVA space.
711 */
712 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
713 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
714 buf_sz = (long)nbuf * BKVASIZE;
715 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
716 (TRANSIENT_DENOM - 1)) {
717 /*
718 * There is more KVA than memory. Do not
719 * adjust buffer map size, and assign the rest
720 * of maxbuf to transient map.
721 */
722 biotmap_sz = maxbuf_sz - buf_sz;
723 } else {
724 /*
725 * Buffer map spans all KVA we could afford on
726 * this platform. Give 10% (20% on i386) of
727 * the buffer map to the transient bio map.
728 */
729 biotmap_sz = buf_sz / TRANSIENT_DENOM;
730 buf_sz -= biotmap_sz;
731 }
732 if (biotmap_sz / INT_MAX > MAXPHYS)
733 bio_transient_maxcnt = INT_MAX;
734 else
735 bio_transient_maxcnt = biotmap_sz / MAXPHYS;
736 /*
737 * Artifically limit to 1024 simultaneous in-flight I/Os
738 * using the transient mapping.
739 */
740 if (bio_transient_maxcnt > 1024)
741 bio_transient_maxcnt = 1024;
742 if (tuned_nbuf)
743 nbuf = buf_sz / BKVASIZE;
744 }
745
746 /*
747 * swbufs are used as temporary holders for I/O, such as paging I/O.
748 * We have no less then 16 and no more then 256.
749 */
750 nswbuf = max(min(nbuf/4, 256), 16);
751 #ifdef NSWBUF_MIN
752 if (nswbuf < NSWBUF_MIN)
753 nswbuf = NSWBUF_MIN;
754 #endif
755
756 /*
757 * Reserve space for the buffer cache buffers
758 */
759 swbuf = (void *)v;
760 v = (caddr_t)(swbuf + nswbuf);
761 buf = (void *)v;
762 v = (caddr_t)(buf + nbuf);
763
764 return(v);
765 }
766
767 /* Initialize the buffer subsystem. Called before use of any buffers. */
768 void
769 bufinit(void)
770 {
771 struct buf *bp;
772 int i;
773
774 mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
775 mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
776 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
777 rw_init(&nblock, "needsbuffer lock");
778 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
779 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
780
781 /* next, make a null set of free lists */
782 for (i = 0; i < BUFFER_QUEUES; i++)
783 TAILQ_INIT(&bufqueues[i]);
784
785 /* finally, initialize each buffer header and stick on empty q */
786 for (i = 0; i < nbuf; i++) {
787 bp = &buf[i];
788 bzero(bp, sizeof *bp);
789 bp->b_flags = B_INVAL | B_INFREECNT;
790 bp->b_rcred = NOCRED;
791 bp->b_wcred = NOCRED;
792 bp->b_qindex = QUEUE_EMPTY;
793 bp->b_xflags = 0;
794 LIST_INIT(&bp->b_dep);
795 BUF_LOCKINIT(bp);
796 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
797 #ifdef INVARIANTS
798 bq_len[QUEUE_EMPTY]++;
799 #endif
800 }
801
802 /*
803 * maxbufspace is the absolute maximum amount of buffer space we are
804 * allowed to reserve in KVM and in real terms. The absolute maximum
805 * is nominally used by buf_daemon. hibufspace is the nominal maximum
806 * used by most other processes. The differential is required to
807 * ensure that buf_daemon is able to run when other processes might
808 * be blocked waiting for buffer space.
809 *
810 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
811 * this may result in KVM fragmentation which is not handled optimally
812 * by the system.
813 */
814 maxbufspace = (long)nbuf * BKVASIZE;
815 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
816 lobufspace = hibufspace - MAXBSIZE;
817
818 /*
819 * Note: The 16 MiB upper limit for hirunningspace was chosen
820 * arbitrarily and may need further tuning. It corresponds to
821 * 128 outstanding write IO requests (if IO size is 128 KiB),
822 * which fits with many RAID controllers' tagged queuing limits.
823 * The lower 1 MiB limit is the historical upper limit for
824 * hirunningspace.
825 */
826 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
827 16 * 1024 * 1024), 1024 * 1024);
828 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
829
830 /*
831 * Limit the amount of malloc memory since it is wired permanently into
832 * the kernel space. Even though this is accounted for in the buffer
833 * allocation, we don't want the malloced region to grow uncontrolled.
834 * The malloc scheme improves memory utilization significantly on average
835 * (small) directories.
836 */
837 maxbufmallocspace = hibufspace / 20;
838
839 /*
840 * Reduce the chance of a deadlock occuring by limiting the number
841 * of delayed-write dirty buffers we allow to stack up.
842 */
843 hidirtybuffers = nbuf / 4 + 20;
844 dirtybufthresh = hidirtybuffers * 9 / 10;
845 numdirtybuffers = 0;
846 /*
847 * To support extreme low-memory systems, make sure hidirtybuffers cannot
848 * eat up all available buffer space. This occurs when our minimum cannot
849 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
850 * BKVASIZE'd buffers.
851 */
852 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
853 hidirtybuffers >>= 1;
854 }
855 lodirtybuffers = hidirtybuffers / 2;
856
857 /*
858 * Try to keep the number of free buffers in the specified range,
859 * and give special processes (e.g. like buf_daemon) access to an
860 * emergency reserve.
861 */
862 lofreebuffers = nbuf / 18 + 5;
863 hifreebuffers = 2 * lofreebuffers;
864 numfreebuffers = nbuf;
865
866 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
867 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
868 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
869 }
870
871 #ifdef INVARIANTS
872 static inline void
873 vfs_buf_check_mapped(struct buf *bp)
874 {
875
876 KASSERT((bp->b_flags & B_UNMAPPED) == 0,
877 ("mapped buf %p %x", bp, bp->b_flags));
878 KASSERT(bp->b_kvabase != unmapped_buf,
879 ("mapped buf: b_kvabase was not updated %p", bp));
880 KASSERT(bp->b_data != unmapped_buf,
881 ("mapped buf: b_data was not updated %p", bp));
882 }
883
884 static inline void
885 vfs_buf_check_unmapped(struct buf *bp)
886 {
887
888 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
889 ("unmapped buf %p %x", bp, bp->b_flags));
890 KASSERT(bp->b_kvabase == unmapped_buf,
891 ("unmapped buf: corrupted b_kvabase %p", bp));
892 KASSERT(bp->b_data == unmapped_buf,
893 ("unmapped buf: corrupted b_data %p", bp));
894 }
895
896 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
897 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
898 #else
899 #define BUF_CHECK_MAPPED(bp) do {} while (0)
900 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
901 #endif
902
903 static void
904 bpmap_qenter(struct buf *bp)
905 {
906
907 BUF_CHECK_MAPPED(bp);
908
909 /*
910 * bp->b_data is relative to bp->b_offset, but
911 * bp->b_offset may be offset into the first page.
912 */
913 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
914 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
915 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
916 (vm_offset_t)(bp->b_offset & PAGE_MASK));
917 }
918
919 /*
920 * bfreekva() - free the kva allocation for a buffer.
921 *
922 * Since this call frees up buffer space, we call bufspacewakeup().
923 */
924 static void
925 bfreekva(struct buf *bp)
926 {
927
928 if (bp->b_kvasize == 0)
929 return;
930
931 atomic_add_int(&buffreekvacnt, 1);
932 atomic_subtract_long(&bufspace, bp->b_kvasize);
933 if ((bp->b_flags & B_UNMAPPED) == 0) {
934 BUF_CHECK_MAPPED(bp);
935 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
936 bp->b_kvasize);
937 } else {
938 BUF_CHECK_UNMAPPED(bp);
939 if ((bp->b_flags & B_KVAALLOC) != 0) {
940 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
941 bp->b_kvasize);
942 }
943 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
944 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
945 }
946 bp->b_kvasize = 0;
947 bufspacewakeup();
948 }
949
950 /*
951 * binsfree:
952 *
953 * Insert the buffer into the appropriate free list.
954 */
955 static void
956 binsfree(struct buf *bp, int qindex)
957 {
958 struct mtx *olock, *nlock;
959
960 BUF_ASSERT_XLOCKED(bp);
961
962 olock = bqlock(bp->b_qindex);
963 nlock = bqlock(qindex);
964 mtx_lock(olock);
965 /* Handle delayed bremfree() processing. */
966 if (bp->b_flags & B_REMFREE)
967 bremfreel(bp);
968
969 if (bp->b_qindex != QUEUE_NONE)
970 panic("binsfree: free buffer onto another queue???");
971
972 bp->b_qindex = qindex;
973 if (olock != nlock) {
974 mtx_unlock(olock);
975 mtx_lock(nlock);
976 }
977 if (bp->b_flags & B_AGE)
978 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
979 else
980 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
981 #ifdef INVARIANTS
982 bq_len[bp->b_qindex]++;
983 #endif
984 mtx_unlock(nlock);
985
986 /*
987 * Something we can maybe free or reuse.
988 */
989 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
990 bufspacewakeup();
991
992 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
993 bufcountadd(bp);
994 }
995
996 /*
997 * bremfree:
998 *
999 * Mark the buffer for removal from the appropriate free list.
1000 *
1001 */
1002 void
1003 bremfree(struct buf *bp)
1004 {
1005
1006 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1007 KASSERT((bp->b_flags & B_REMFREE) == 0,
1008 ("bremfree: buffer %p already marked for delayed removal.", bp));
1009 KASSERT(bp->b_qindex != QUEUE_NONE,
1010 ("bremfree: buffer %p not on a queue.", bp));
1011 BUF_ASSERT_XLOCKED(bp);
1012
1013 bp->b_flags |= B_REMFREE;
1014 bufcountsub(bp);
1015 }
1016
1017 /*
1018 * bremfreef:
1019 *
1020 * Force an immediate removal from a free list. Used only in nfs when
1021 * it abuses the b_freelist pointer.
1022 */
1023 void
1024 bremfreef(struct buf *bp)
1025 {
1026 struct mtx *qlock;
1027
1028 qlock = bqlock(bp->b_qindex);
1029 mtx_lock(qlock);
1030 bremfreel(bp);
1031 mtx_unlock(qlock);
1032 }
1033
1034 /*
1035 * bremfreel:
1036 *
1037 * Removes a buffer from the free list, must be called with the
1038 * correct qlock held.
1039 */
1040 static void
1041 bremfreel(struct buf *bp)
1042 {
1043
1044 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1045 bp, bp->b_vp, bp->b_flags);
1046 KASSERT(bp->b_qindex != QUEUE_NONE,
1047 ("bremfreel: buffer %p not on a queue.", bp));
1048 BUF_ASSERT_XLOCKED(bp);
1049 mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1050
1051 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1052 #ifdef INVARIANTS
1053 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1054 bp->b_qindex));
1055 bq_len[bp->b_qindex]--;
1056 #endif
1057 bp->b_qindex = QUEUE_NONE;
1058 /*
1059 * If this was a delayed bremfree() we only need to remove the buffer
1060 * from the queue and return the stats are already done.
1061 */
1062 if (bp->b_flags & B_REMFREE) {
1063 bp->b_flags &= ~B_REMFREE;
1064 return;
1065 }
1066 bufcountsub(bp);
1067 }
1068
1069 /*
1070 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
1071 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1072 * the buffer is valid and we do not have to do anything.
1073 */
1074 void
1075 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1076 int cnt, struct ucred * cred)
1077 {
1078 struct buf *rabp;
1079 int i;
1080
1081 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1082 if (inmem(vp, *rablkno))
1083 continue;
1084 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1085
1086 if ((rabp->b_flags & B_CACHE) == 0) {
1087 if (!TD_IS_IDLETHREAD(curthread))
1088 curthread->td_ru.ru_inblock++;
1089 rabp->b_flags |= B_ASYNC;
1090 rabp->b_flags &= ~B_INVAL;
1091 rabp->b_ioflags &= ~BIO_ERROR;
1092 rabp->b_iocmd = BIO_READ;
1093 if (rabp->b_rcred == NOCRED && cred != NOCRED)
1094 rabp->b_rcred = crhold(cred);
1095 vfs_busy_pages(rabp, 0);
1096 BUF_KERNPROC(rabp);
1097 rabp->b_iooffset = dbtob(rabp->b_blkno);
1098 bstrategy(rabp);
1099 } else {
1100 brelse(rabp);
1101 }
1102 }
1103 }
1104
1105 /*
1106 * Entry point for bread() and breadn() via #defines in sys/buf.h.
1107 *
1108 * Get a buffer with the specified data. Look in the cache first. We
1109 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
1110 * is set, the buffer is valid and we do not have to do anything, see
1111 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1112 */
1113 int
1114 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1115 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1116 {
1117 struct buf *bp;
1118 int rv = 0, readwait = 0;
1119
1120 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1121 /*
1122 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1123 */
1124 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1125 if (bp == NULL)
1126 return (EBUSY);
1127
1128 /* if not found in cache, do some I/O */
1129 if ((bp->b_flags & B_CACHE) == 0) {
1130 if (!TD_IS_IDLETHREAD(curthread))
1131 curthread->td_ru.ru_inblock++;
1132 bp->b_iocmd = BIO_READ;
1133 bp->b_flags &= ~B_INVAL;
1134 bp->b_ioflags &= ~BIO_ERROR;
1135 if (bp->b_rcred == NOCRED && cred != NOCRED)
1136 bp->b_rcred = crhold(cred);
1137 vfs_busy_pages(bp, 0);
1138 bp->b_iooffset = dbtob(bp->b_blkno);
1139 bstrategy(bp);
1140 ++readwait;
1141 }
1142
1143 breada(vp, rablkno, rabsize, cnt, cred);
1144
1145 if (readwait) {
1146 rv = bufwait(bp);
1147 }
1148 return (rv);
1149 }
1150
1151 /*
1152 * Write, release buffer on completion. (Done by iodone
1153 * if async). Do not bother writing anything if the buffer
1154 * is invalid.
1155 *
1156 * Note that we set B_CACHE here, indicating that buffer is
1157 * fully valid and thus cacheable. This is true even of NFS
1158 * now so we set it generally. This could be set either here
1159 * or in biodone() since the I/O is synchronous. We put it
1160 * here.
1161 */
1162 int
1163 bufwrite(struct buf *bp)
1164 {
1165 int oldflags;
1166 struct vnode *vp;
1167 long space;
1168 int vp_md;
1169
1170 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1171 if (bp->b_flags & B_INVAL) {
1172 brelse(bp);
1173 return (0);
1174 }
1175
1176 if (bp->b_flags & B_BARRIER)
1177 barrierwrites++;
1178
1179 oldflags = bp->b_flags;
1180
1181 BUF_ASSERT_HELD(bp);
1182
1183 if (bp->b_pin_count > 0)
1184 bunpin_wait(bp);
1185
1186 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1187 ("FFS background buffer should not get here %p", bp));
1188
1189 vp = bp->b_vp;
1190 if (vp)
1191 vp_md = vp->v_vflag & VV_MD;
1192 else
1193 vp_md = 0;
1194
1195 /*
1196 * Mark the buffer clean. Increment the bufobj write count
1197 * before bundirty() call, to prevent other thread from seeing
1198 * empty dirty list and zero counter for writes in progress,
1199 * falsely indicating that the bufobj is clean.
1200 */
1201 bufobj_wref(bp->b_bufobj);
1202 bundirty(bp);
1203
1204 bp->b_flags &= ~B_DONE;
1205 bp->b_ioflags &= ~BIO_ERROR;
1206 bp->b_flags |= B_CACHE;
1207 bp->b_iocmd = BIO_WRITE;
1208
1209 vfs_busy_pages(bp, 1);
1210
1211 /*
1212 * Normal bwrites pipeline writes
1213 */
1214 bp->b_runningbufspace = bp->b_bufsize;
1215 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1216
1217 if (!TD_IS_IDLETHREAD(curthread))
1218 curthread->td_ru.ru_oublock++;
1219 if (oldflags & B_ASYNC)
1220 BUF_KERNPROC(bp);
1221 bp->b_iooffset = dbtob(bp->b_blkno);
1222 bstrategy(bp);
1223
1224 if ((oldflags & B_ASYNC) == 0) {
1225 int rtval = bufwait(bp);
1226 brelse(bp);
1227 return (rtval);
1228 } else if (space > hirunningspace) {
1229 /*
1230 * don't allow the async write to saturate the I/O
1231 * system. We will not deadlock here because
1232 * we are blocking waiting for I/O that is already in-progress
1233 * to complete. We do not block here if it is the update
1234 * or syncer daemon trying to clean up as that can lead
1235 * to deadlock.
1236 */
1237 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1238 waitrunningbufspace();
1239 }
1240
1241 return (0);
1242 }
1243
1244 void
1245 bufbdflush(struct bufobj *bo, struct buf *bp)
1246 {
1247 struct buf *nbp;
1248
1249 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1250 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1251 altbufferflushes++;
1252 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1253 BO_LOCK(bo);
1254 /*
1255 * Try to find a buffer to flush.
1256 */
1257 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1258 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1259 BUF_LOCK(nbp,
1260 LK_EXCLUSIVE | LK_NOWAIT, NULL))
1261 continue;
1262 if (bp == nbp)
1263 panic("bdwrite: found ourselves");
1264 BO_UNLOCK(bo);
1265 /* Don't countdeps with the bo lock held. */
1266 if (buf_countdeps(nbp, 0)) {
1267 BO_LOCK(bo);
1268 BUF_UNLOCK(nbp);
1269 continue;
1270 }
1271 if (nbp->b_flags & B_CLUSTEROK) {
1272 vfs_bio_awrite(nbp);
1273 } else {
1274 bremfree(nbp);
1275 bawrite(nbp);
1276 }
1277 dirtybufferflushes++;
1278 break;
1279 }
1280 if (nbp == NULL)
1281 BO_UNLOCK(bo);
1282 }
1283 }
1284
1285 /*
1286 * Delayed write. (Buffer is marked dirty). Do not bother writing
1287 * anything if the buffer is marked invalid.
1288 *
1289 * Note that since the buffer must be completely valid, we can safely
1290 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1291 * biodone() in order to prevent getblk from writing the buffer
1292 * out synchronously.
1293 */
1294 void
1295 bdwrite(struct buf *bp)
1296 {
1297 struct thread *td = curthread;
1298 struct vnode *vp;
1299 struct bufobj *bo;
1300
1301 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1302 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1303 KASSERT((bp->b_flags & B_BARRIER) == 0,
1304 ("Barrier request in delayed write %p", bp));
1305 BUF_ASSERT_HELD(bp);
1306
1307 if (bp->b_flags & B_INVAL) {
1308 brelse(bp);
1309 return;
1310 }
1311
1312 /*
1313 * If we have too many dirty buffers, don't create any more.
1314 * If we are wildly over our limit, then force a complete
1315 * cleanup. Otherwise, just keep the situation from getting
1316 * out of control. Note that we have to avoid a recursive
1317 * disaster and not try to clean up after our own cleanup!
1318 */
1319 vp = bp->b_vp;
1320 bo = bp->b_bufobj;
1321 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1322 td->td_pflags |= TDP_INBDFLUSH;
1323 BO_BDFLUSH(bo, bp);
1324 td->td_pflags &= ~TDP_INBDFLUSH;
1325 } else
1326 recursiveflushes++;
1327
1328 bdirty(bp);
1329 /*
1330 * Set B_CACHE, indicating that the buffer is fully valid. This is
1331 * true even of NFS now.
1332 */
1333 bp->b_flags |= B_CACHE;
1334
1335 /*
1336 * This bmap keeps the system from needing to do the bmap later,
1337 * perhaps when the system is attempting to do a sync. Since it
1338 * is likely that the indirect block -- or whatever other datastructure
1339 * that the filesystem needs is still in memory now, it is a good
1340 * thing to do this. Note also, that if the pageout daemon is
1341 * requesting a sync -- there might not be enough memory to do
1342 * the bmap then... So, this is important to do.
1343 */
1344 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1345 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1346 }
1347
1348 /*
1349 * Set the *dirty* buffer range based upon the VM system dirty
1350 * pages.
1351 *
1352 * Mark the buffer pages as clean. We need to do this here to
1353 * satisfy the vnode_pager and the pageout daemon, so that it
1354 * thinks that the pages have been "cleaned". Note that since
1355 * the pages are in a delayed write buffer -- the VFS layer
1356 * "will" see that the pages get written out on the next sync,
1357 * or perhaps the cluster will be completed.
1358 */
1359 vfs_clean_pages_dirty_buf(bp);
1360 bqrelse(bp);
1361
1362 /*
1363 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1364 * due to the softdep code.
1365 */
1366 }
1367
1368 /*
1369 * bdirty:
1370 *
1371 * Turn buffer into delayed write request. We must clear BIO_READ and
1372 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1373 * itself to properly update it in the dirty/clean lists. We mark it
1374 * B_DONE to ensure that any asynchronization of the buffer properly
1375 * clears B_DONE ( else a panic will occur later ).
1376 *
1377 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1378 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1379 * should only be called if the buffer is known-good.
1380 *
1381 * Since the buffer is not on a queue, we do not update the numfreebuffers
1382 * count.
1383 *
1384 * The buffer must be on QUEUE_NONE.
1385 */
1386 void
1387 bdirty(struct buf *bp)
1388 {
1389
1390 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1391 bp, bp->b_vp, bp->b_flags);
1392 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1393 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1394 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1395 BUF_ASSERT_HELD(bp);
1396 bp->b_flags &= ~(B_RELBUF);
1397 bp->b_iocmd = BIO_WRITE;
1398
1399 if ((bp->b_flags & B_DELWRI) == 0) {
1400 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1401 reassignbuf(bp);
1402 bdirtyadd();
1403 }
1404 }
1405
1406 /*
1407 * bundirty:
1408 *
1409 * Clear B_DELWRI for buffer.
1410 *
1411 * Since the buffer is not on a queue, we do not update the numfreebuffers
1412 * count.
1413 *
1414 * The buffer must be on QUEUE_NONE.
1415 */
1416
1417 void
1418 bundirty(struct buf *bp)
1419 {
1420
1421 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1422 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1423 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1424 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1425 BUF_ASSERT_HELD(bp);
1426
1427 if (bp->b_flags & B_DELWRI) {
1428 bp->b_flags &= ~B_DELWRI;
1429 reassignbuf(bp);
1430 bdirtysub();
1431 }
1432 /*
1433 * Since it is now being written, we can clear its deferred write flag.
1434 */
1435 bp->b_flags &= ~B_DEFERRED;
1436 }
1437
1438 /*
1439 * bawrite:
1440 *
1441 * Asynchronous write. Start output on a buffer, but do not wait for
1442 * it to complete. The buffer is released when the output completes.
1443 *
1444 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1445 * B_INVAL buffers. Not us.
1446 */
1447 void
1448 bawrite(struct buf *bp)
1449 {
1450
1451 bp->b_flags |= B_ASYNC;
1452 (void) bwrite(bp);
1453 }
1454
1455 /*
1456 * babarrierwrite:
1457 *
1458 * Asynchronous barrier write. Start output on a buffer, but do not
1459 * wait for it to complete. Place a write barrier after this write so
1460 * that this buffer and all buffers written before it are committed to
1461 * the disk before any buffers written after this write are committed
1462 * to the disk. The buffer is released when the output completes.
1463 */
1464 void
1465 babarrierwrite(struct buf *bp)
1466 {
1467
1468 bp->b_flags |= B_ASYNC | B_BARRIER;
1469 (void) bwrite(bp);
1470 }
1471
1472 /*
1473 * bbarrierwrite:
1474 *
1475 * Synchronous barrier write. Start output on a buffer and wait for
1476 * it to complete. Place a write barrier after this write so that
1477 * this buffer and all buffers written before it are committed to
1478 * the disk before any buffers written after this write are committed
1479 * to the disk. The buffer is released when the output completes.
1480 */
1481 int
1482 bbarrierwrite(struct buf *bp)
1483 {
1484
1485 bp->b_flags |= B_BARRIER;
1486 return (bwrite(bp));
1487 }
1488
1489 /*
1490 * bwillwrite:
1491 *
1492 * Called prior to the locking of any vnodes when we are expecting to
1493 * write. We do not want to starve the buffer cache with too many
1494 * dirty buffers so we block here. By blocking prior to the locking
1495 * of any vnodes we attempt to avoid the situation where a locked vnode
1496 * prevents the various system daemons from flushing related buffers.
1497 */
1498 void
1499 bwillwrite(void)
1500 {
1501
1502 if (numdirtybuffers >= hidirtybuffers) {
1503 mtx_lock(&bdirtylock);
1504 while (numdirtybuffers >= hidirtybuffers) {
1505 bdirtywait = 1;
1506 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1507 "flswai", 0);
1508 }
1509 mtx_unlock(&bdirtylock);
1510 }
1511 }
1512
1513 /*
1514 * Return true if we have too many dirty buffers.
1515 */
1516 int
1517 buf_dirty_count_severe(void)
1518 {
1519
1520 return(numdirtybuffers >= hidirtybuffers);
1521 }
1522
1523 static __noinline int
1524 buf_vm_page_count_severe(void)
1525 {
1526
1527 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1528
1529 return vm_page_count_severe();
1530 }
1531
1532 /*
1533 * brelse:
1534 *
1535 * Release a busy buffer and, if requested, free its resources. The
1536 * buffer will be stashed in the appropriate bufqueue[] allowing it
1537 * to be accessed later as a cache entity or reused for other purposes.
1538 */
1539 void
1540 brelse(struct buf *bp)
1541 {
1542 int qindex;
1543
1544 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1545 bp, bp->b_vp, bp->b_flags);
1546 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1547 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1548
1549 if (BUF_LOCKRECURSED(bp)) {
1550 /*
1551 * Do not process, in particular, do not handle the
1552 * B_INVAL/B_RELBUF and do not release to free list.
1553 */
1554 BUF_UNLOCK(bp);
1555 return;
1556 }
1557
1558 if (bp->b_flags & B_MANAGED) {
1559 bqrelse(bp);
1560 return;
1561 }
1562
1563 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1564 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1565 /*
1566 * Failed write, redirty. Must clear BIO_ERROR to prevent
1567 * pages from being scrapped. If the error is anything
1568 * other than an I/O error (EIO), assume that retrying
1569 * is futile.
1570 */
1571 bp->b_ioflags &= ~BIO_ERROR;
1572 bdirty(bp);
1573 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1574 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1575 /*
1576 * Either a failed I/O or we were asked to free or not
1577 * cache the buffer.
1578 */
1579 bp->b_flags |= B_INVAL;
1580 if (!LIST_EMPTY(&bp->b_dep))
1581 buf_deallocate(bp);
1582 if (bp->b_flags & B_DELWRI)
1583 bdirtysub();
1584 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1585 if ((bp->b_flags & B_VMIO) == 0) {
1586 if (bp->b_bufsize)
1587 allocbuf(bp, 0);
1588 if (bp->b_vp)
1589 brelvp(bp);
1590 }
1591 }
1592
1593 /*
1594 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1595 * is called with B_DELWRI set, the underlying pages may wind up
1596 * getting freed causing a previous write (bdwrite()) to get 'lost'
1597 * because pages associated with a B_DELWRI bp are marked clean.
1598 *
1599 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1600 * if B_DELWRI is set.
1601 *
1602 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1603 * on pages to return pages to the VM page queues.
1604 */
1605 if (bp->b_flags & B_DELWRI)
1606 bp->b_flags &= ~B_RELBUF;
1607 else if (buf_vm_page_count_severe()) {
1608 /*
1609 * BKGRDINPROG can only be set with the buf and bufobj
1610 * locks both held. We tolerate a race to clear it here.
1611 */
1612 if (!(bp->b_vflags & BV_BKGRDINPROG))
1613 bp->b_flags |= B_RELBUF;
1614 }
1615
1616 /*
1617 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1618 * constituted, not even NFS buffers now. Two flags effect this. If
1619 * B_INVAL, the struct buf is invalidated but the VM object is kept
1620 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1621 *
1622 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1623 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1624 * buffer is also B_INVAL because it hits the re-dirtying code above.
1625 *
1626 * Normally we can do this whether a buffer is B_DELWRI or not. If
1627 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1628 * the commit state and we cannot afford to lose the buffer. If the
1629 * buffer has a background write in progress, we need to keep it
1630 * around to prevent it from being reconstituted and starting a second
1631 * background write.
1632 */
1633 if ((bp->b_flags & B_VMIO)
1634 && !(bp->b_vp->v_mount != NULL &&
1635 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1636 !vn_isdisk(bp->b_vp, NULL) &&
1637 (bp->b_flags & B_DELWRI))
1638 ) {
1639
1640 int i, j, resid;
1641 vm_page_t m;
1642 off_t foff;
1643 vm_pindex_t poff;
1644 vm_object_t obj;
1645
1646 obj = bp->b_bufobj->bo_object;
1647
1648 /*
1649 * Get the base offset and length of the buffer. Note that
1650 * in the VMIO case if the buffer block size is not
1651 * page-aligned then b_data pointer may not be page-aligned.
1652 * But our b_pages[] array *IS* page aligned.
1653 *
1654 * block sizes less then DEV_BSIZE (usually 512) are not
1655 * supported due to the page granularity bits (m->valid,
1656 * m->dirty, etc...).
1657 *
1658 * See man buf(9) for more information
1659 */
1660 resid = bp->b_bufsize;
1661 foff = bp->b_offset;
1662 for (i = 0; i < bp->b_npages; i++) {
1663 int had_bogus = 0;
1664
1665 m = bp->b_pages[i];
1666
1667 /*
1668 * If we hit a bogus page, fixup *all* the bogus pages
1669 * now.
1670 */
1671 if (m == bogus_page) {
1672 poff = OFF_TO_IDX(bp->b_offset);
1673 had_bogus = 1;
1674
1675 VM_OBJECT_RLOCK(obj);
1676 for (j = i; j < bp->b_npages; j++) {
1677 vm_page_t mtmp;
1678 mtmp = bp->b_pages[j];
1679 if (mtmp == bogus_page) {
1680 mtmp = vm_page_lookup(obj, poff + j);
1681 if (!mtmp) {
1682 panic("brelse: page missing\n");
1683 }
1684 bp->b_pages[j] = mtmp;
1685 }
1686 }
1687 VM_OBJECT_RUNLOCK(obj);
1688
1689 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1690 BUF_CHECK_MAPPED(bp);
1691 pmap_qenter(
1692 trunc_page((vm_offset_t)bp->b_data),
1693 bp->b_pages, bp->b_npages);
1694 }
1695 m = bp->b_pages[i];
1696 }
1697 if ((bp->b_flags & B_NOCACHE) ||
1698 (bp->b_ioflags & BIO_ERROR &&
1699 bp->b_iocmd == BIO_READ)) {
1700 int poffset = foff & PAGE_MASK;
1701 int presid = resid > (PAGE_SIZE - poffset) ?
1702 (PAGE_SIZE - poffset) : resid;
1703
1704 KASSERT(presid >= 0, ("brelse: extra page"));
1705 VM_OBJECT_WLOCK(obj);
1706 while (vm_page_xbusied(m)) {
1707 vm_page_lock(m);
1708 VM_OBJECT_WUNLOCK(obj);
1709 vm_page_busy_sleep(m, "mbncsh");
1710 VM_OBJECT_WLOCK(obj);
1711 }
1712 if (pmap_page_wired_mappings(m) == 0)
1713 vm_page_set_invalid(m, poffset, presid);
1714 VM_OBJECT_WUNLOCK(obj);
1715 if (had_bogus)
1716 printf("avoided corruption bug in bogus_page/brelse code\n");
1717 }
1718 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1719 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1720 }
1721 if (bp->b_flags & (B_INVAL | B_RELBUF))
1722 vfs_vmio_release(bp);
1723
1724 } else if (bp->b_flags & B_VMIO) {
1725
1726 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1727 vfs_vmio_release(bp);
1728 }
1729
1730 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1731 if (bp->b_bufsize != 0)
1732 allocbuf(bp, 0);
1733 if (bp->b_vp != NULL)
1734 brelvp(bp);
1735 }
1736
1737 /*
1738 * If the buffer has junk contents signal it and eventually
1739 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1740 * doesn't find it.
1741 */
1742 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1743 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1744 bp->b_flags |= B_INVAL;
1745 if (bp->b_flags & B_INVAL) {
1746 if (bp->b_flags & B_DELWRI)
1747 bundirty(bp);
1748 if (bp->b_vp)
1749 brelvp(bp);
1750 }
1751
1752 /* buffers with no memory */
1753 if (bp->b_bufsize == 0) {
1754 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1755 if (bp->b_vflags & BV_BKGRDINPROG)
1756 panic("losing buffer 1");
1757 if (bp->b_kvasize)
1758 qindex = QUEUE_EMPTYKVA;
1759 else
1760 qindex = QUEUE_EMPTY;
1761 bp->b_flags |= B_AGE;
1762 /* buffers with junk contents */
1763 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1764 (bp->b_ioflags & BIO_ERROR)) {
1765 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1766 if (bp->b_vflags & BV_BKGRDINPROG)
1767 panic("losing buffer 2");
1768 qindex = QUEUE_CLEAN;
1769 bp->b_flags |= B_AGE;
1770 /* remaining buffers */
1771 } else if (bp->b_flags & B_DELWRI)
1772 qindex = QUEUE_DIRTY;
1773 else
1774 qindex = QUEUE_CLEAN;
1775
1776 binsfree(bp, qindex);
1777
1778 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1779 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1780 panic("brelse: not dirty");
1781 /* unlock */
1782 BUF_UNLOCK(bp);
1783 }
1784
1785 /*
1786 * Release a buffer back to the appropriate queue but do not try to free
1787 * it. The buffer is expected to be used again soon.
1788 *
1789 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1790 * biodone() to requeue an async I/O on completion. It is also used when
1791 * known good buffers need to be requeued but we think we may need the data
1792 * again soon.
1793 *
1794 * XXX we should be able to leave the B_RELBUF hint set on completion.
1795 */
1796 void
1797 bqrelse(struct buf *bp)
1798 {
1799 int qindex;
1800
1801 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1802 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1803 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1804
1805 if (BUF_LOCKRECURSED(bp)) {
1806 /* do not release to free list */
1807 BUF_UNLOCK(bp);
1808 return;
1809 }
1810 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1811
1812 if (bp->b_flags & B_MANAGED) {
1813 if (bp->b_flags & B_REMFREE)
1814 bremfreef(bp);
1815 goto out;
1816 }
1817
1818 /* buffers with stale but valid contents */
1819 if (bp->b_flags & B_DELWRI) {
1820 qindex = QUEUE_DIRTY;
1821 } else {
1822 if ((bp->b_flags & B_DELWRI) == 0 &&
1823 (bp->b_xflags & BX_VNDIRTY))
1824 panic("bqrelse: not dirty");
1825 /*
1826 * BKGRDINPROG can only be set with the buf and bufobj
1827 * locks both held. We tolerate a race to clear it here.
1828 */
1829 if (buf_vm_page_count_severe() &&
1830 (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1831 /*
1832 * We are too low on memory, we have to try to free
1833 * the buffer (most importantly: the wired pages
1834 * making up its backing store) *now*.
1835 */
1836 brelse(bp);
1837 return;
1838 }
1839 qindex = QUEUE_CLEAN;
1840 }
1841 binsfree(bp, qindex);
1842
1843 out:
1844 /* unlock */
1845 BUF_UNLOCK(bp);
1846 }
1847
1848 /* Give pages used by the bp back to the VM system (where possible) */
1849 static void
1850 vfs_vmio_release(struct buf *bp)
1851 {
1852 int i;
1853 vm_page_t m;
1854
1855 if ((bp->b_flags & B_UNMAPPED) == 0) {
1856 BUF_CHECK_MAPPED(bp);
1857 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1858 } else
1859 BUF_CHECK_UNMAPPED(bp);
1860 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
1861 for (i = 0; i < bp->b_npages; i++) {
1862 m = bp->b_pages[i];
1863 bp->b_pages[i] = NULL;
1864 /*
1865 * In order to keep page LRU ordering consistent, put
1866 * everything on the inactive queue.
1867 */
1868 vm_page_lock(m);
1869 vm_page_unwire(m, 0);
1870
1871 /*
1872 * Might as well free the page if we can and it has
1873 * no valid data. We also free the page if the
1874 * buffer was used for direct I/O
1875 */
1876 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1877 if (m->wire_count == 0 && !vm_page_busied(m))
1878 vm_page_free(m);
1879 } else if (bp->b_flags & B_DIRECT)
1880 vm_page_try_to_free(m);
1881 else if (buf_vm_page_count_severe())
1882 vm_page_try_to_cache(m);
1883 vm_page_unlock(m);
1884 }
1885 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
1886
1887 if (bp->b_bufsize) {
1888 bufspacewakeup();
1889 bp->b_bufsize = 0;
1890 }
1891 bp->b_npages = 0;
1892 bp->b_flags &= ~B_VMIO;
1893 if (bp->b_vp)
1894 brelvp(bp);
1895 }
1896
1897 /*
1898 * Check to see if a block at a particular lbn is available for a clustered
1899 * write.
1900 */
1901 static int
1902 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1903 {
1904 struct buf *bpa;
1905 int match;
1906
1907 match = 0;
1908
1909 /* If the buf isn't in core skip it */
1910 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1911 return (0);
1912
1913 /* If the buf is busy we don't want to wait for it */
1914 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1915 return (0);
1916
1917 /* Only cluster with valid clusterable delayed write buffers */
1918 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1919 (B_DELWRI | B_CLUSTEROK))
1920 goto done;
1921
1922 if (bpa->b_bufsize != size)
1923 goto done;
1924
1925 /*
1926 * Check to see if it is in the expected place on disk and that the
1927 * block has been mapped.
1928 */
1929 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1930 match = 1;
1931 done:
1932 BUF_UNLOCK(bpa);
1933 return (match);
1934 }
1935
1936 /*
1937 * vfs_bio_awrite:
1938 *
1939 * Implement clustered async writes for clearing out B_DELWRI buffers.
1940 * This is much better then the old way of writing only one buffer at
1941 * a time. Note that we may not be presented with the buffers in the
1942 * correct order, so we search for the cluster in both directions.
1943 */
1944 int
1945 vfs_bio_awrite(struct buf *bp)
1946 {
1947 struct bufobj *bo;
1948 int i;
1949 int j;
1950 daddr_t lblkno = bp->b_lblkno;
1951 struct vnode *vp = bp->b_vp;
1952 int ncl;
1953 int nwritten;
1954 int size;
1955 int maxcl;
1956 int gbflags;
1957
1958 bo = &vp->v_bufobj;
1959 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1960 /*
1961 * right now we support clustered writing only to regular files. If
1962 * we find a clusterable block we could be in the middle of a cluster
1963 * rather then at the beginning.
1964 */
1965 if ((vp->v_type == VREG) &&
1966 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1967 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1968
1969 size = vp->v_mount->mnt_stat.f_iosize;
1970 maxcl = MAXPHYS / size;
1971
1972 BO_RLOCK(bo);
1973 for (i = 1; i < maxcl; i++)
1974 if (vfs_bio_clcheck(vp, size, lblkno + i,
1975 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1976 break;
1977
1978 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1979 if (vfs_bio_clcheck(vp, size, lblkno - j,
1980 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1981 break;
1982 BO_RUNLOCK(bo);
1983 --j;
1984 ncl = i + j;
1985 /*
1986 * this is a possible cluster write
1987 */
1988 if (ncl != 1) {
1989 BUF_UNLOCK(bp);
1990 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
1991 gbflags);
1992 return (nwritten);
1993 }
1994 }
1995 bremfree(bp);
1996 bp->b_flags |= B_ASYNC;
1997 /*
1998 * default (old) behavior, writing out only one block
1999 *
2000 * XXX returns b_bufsize instead of b_bcount for nwritten?
2001 */
2002 nwritten = bp->b_bufsize;
2003 (void) bwrite(bp);
2004
2005 return (nwritten);
2006 }
2007
2008 static void
2009 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2010 {
2011
2012 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2013 bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2014 if ((gbflags & GB_UNMAPPED) == 0) {
2015 bp->b_kvabase = (caddr_t)addr;
2016 } else if ((gbflags & GB_KVAALLOC) != 0) {
2017 KASSERT((gbflags & GB_UNMAPPED) != 0,
2018 ("GB_KVAALLOC without GB_UNMAPPED"));
2019 bp->b_kvaalloc = (caddr_t)addr;
2020 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2021 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2022 }
2023 bp->b_kvasize = maxsize;
2024 }
2025
2026 /*
2027 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2028 * needed.
2029 */
2030 static int
2031 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2032 {
2033 vm_offset_t addr;
2034
2035 bfreekva(bp);
2036 addr = 0;
2037
2038 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2039 /*
2040 * Buffer map is too fragmented. Request the caller
2041 * to defragment the map.
2042 */
2043 atomic_add_int(&bufdefragcnt, 1);
2044 return (1);
2045 }
2046 setbufkva(bp, addr, maxsize, gbflags);
2047 atomic_add_long(&bufspace, bp->b_kvasize);
2048 return (0);
2049 }
2050
2051 /*
2052 * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2053 * locked vnode is supplied.
2054 */
2055 static void
2056 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2057 int defrag)
2058 {
2059 struct thread *td;
2060 char *waitmsg;
2061 int cnt, error, flags, norunbuf, wait;
2062
2063 mtx_assert(&bqclean, MA_OWNED);
2064
2065 if (defrag) {
2066 flags = VFS_BIO_NEED_BUFSPACE;
2067 waitmsg = "nbufkv";
2068 } else if (bufspace >= hibufspace) {
2069 waitmsg = "nbufbs";
2070 flags = VFS_BIO_NEED_BUFSPACE;
2071 } else {
2072 waitmsg = "newbuf";
2073 flags = VFS_BIO_NEED_ANY;
2074 }
2075 atomic_set_int(&needsbuffer, flags);
2076 mtx_unlock(&bqclean);
2077
2078 bd_speedup(); /* heeeelp */
2079 if ((gbflags & GB_NOWAIT_BD) != 0)
2080 return;
2081
2082 td = curthread;
2083 cnt = 0;
2084 wait = MNT_NOWAIT;
2085 rw_wlock(&nblock);
2086 while ((needsbuffer & flags) != 0) {
2087 if (vp != NULL && vp->v_type != VCHR &&
2088 (td->td_pflags & TDP_BUFNEED) == 0) {
2089 rw_wunlock(&nblock);
2090 /*
2091 * getblk() is called with a vnode locked, and
2092 * some majority of the dirty buffers may as
2093 * well belong to the vnode. Flushing the
2094 * buffers there would make a progress that
2095 * cannot be achieved by the buf_daemon, that
2096 * cannot lock the vnode.
2097 */
2098 if (cnt++ > 2)
2099 wait = MNT_WAIT;
2100 ASSERT_VOP_LOCKED(vp, "bufd_helper");
2101 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2102 vn_lock(vp, LK_TRYUPGRADE);
2103 if (error == 0) {
2104 /* play bufdaemon */
2105 norunbuf = curthread_pflags_set(TDP_BUFNEED |
2106 TDP_NORUNNINGBUF);
2107 VOP_FSYNC(vp, wait, td);
2108 atomic_add_long(¬bufdflushes, 1);
2109 curthread_pflags_restore(norunbuf);
2110 }
2111 rw_wlock(&nblock);
2112 if ((needsbuffer & flags) == 0)
2113 break;
2114 }
2115 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
2116 (PRIBIO + 4) | slpflag, waitmsg, slptimeo);
2117 if (error != 0)
2118 break;
2119 }
2120 rw_wunlock(&nblock);
2121 }
2122
2123 static void
2124 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2125 {
2126
2127 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2128 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2129 bp->b_kvasize, bp->b_bufsize, qindex);
2130 mtx_assert(&bqclean, MA_NOTOWNED);
2131
2132 /*
2133 * Note: we no longer distinguish between VMIO and non-VMIO
2134 * buffers.
2135 */
2136 KASSERT((bp->b_flags & B_DELWRI) == 0,
2137 ("delwri buffer %p found in queue %d", bp, qindex));
2138
2139 if (qindex == QUEUE_CLEAN) {
2140 if (bp->b_flags & B_VMIO) {
2141 bp->b_flags &= ~B_ASYNC;
2142 vfs_vmio_release(bp);
2143 }
2144 if (bp->b_vp != NULL)
2145 brelvp(bp);
2146 }
2147
2148 /*
2149 * Get the rest of the buffer freed up. b_kva* is still valid
2150 * after this operation.
2151 */
2152
2153 if (bp->b_rcred != NOCRED) {
2154 crfree(bp->b_rcred);
2155 bp->b_rcred = NOCRED;
2156 }
2157 if (bp->b_wcred != NOCRED) {
2158 crfree(bp->b_wcred);
2159 bp->b_wcred = NOCRED;
2160 }
2161 if (!LIST_EMPTY(&bp->b_dep))
2162 buf_deallocate(bp);
2163 if (bp->b_vflags & BV_BKGRDINPROG)
2164 panic("losing buffer 3");
2165 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d",
2166 bp, bp->b_vp, qindex));
2167 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2168 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2169
2170 if (bp->b_bufsize)
2171 allocbuf(bp, 0);
2172
2173 bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2174 bp->b_ioflags = 0;
2175 bp->b_xflags = 0;
2176 KASSERT((bp->b_flags & B_INFREECNT) == 0,
2177 ("buf %p still counted as free?", bp));
2178 bp->b_vflags = 0;
2179 bp->b_vp = NULL;
2180 bp->b_blkno = bp->b_lblkno = 0;
2181 bp->b_offset = NOOFFSET;
2182 bp->b_iodone = 0;
2183 bp->b_error = 0;
2184 bp->b_resid = 0;
2185 bp->b_bcount = 0;
2186 bp->b_npages = 0;
2187 bp->b_dirtyoff = bp->b_dirtyend = 0;
2188 bp->b_bufobj = NULL;
2189 bp->b_pin_count = 0;
2190 bp->b_fsprivate1 = NULL;
2191 bp->b_fsprivate2 = NULL;
2192 bp->b_fsprivate3 = NULL;
2193
2194 LIST_INIT(&bp->b_dep);
2195 }
2196
2197 static int flushingbufs;
2198
2199 static struct buf *
2200 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2201 {
2202 struct buf *bp, *nbp;
2203 int nqindex, qindex, pass;
2204
2205 KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2206
2207 pass = 1;
2208 restart:
2209 atomic_add_int(&getnewbufrestarts, 1);
2210
2211 /*
2212 * Setup for scan. If we do not have enough free buffers,
2213 * we setup a degenerate case that immediately fails. Note
2214 * that if we are specially marked process, we are allowed to
2215 * dip into our reserves.
2216 *
2217 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2218 * for the allocation of the mapped buffer. For unmapped, the
2219 * easiest is to start with EMPTY outright.
2220 *
2221 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
2222 * However, there are a number of cases (defragging, reusing, ...)
2223 * where we cannot backup.
2224 */
2225 nbp = NULL;
2226 mtx_lock(&bqclean);
2227 if (!defrag && unmapped) {
2228 nqindex = QUEUE_EMPTY;
2229 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2230 }
2231 if (nbp == NULL) {
2232 nqindex = QUEUE_EMPTYKVA;
2233 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2234 }
2235
2236 /*
2237 * If no EMPTYKVA buffers and we are either defragging or
2238 * reusing, locate a CLEAN buffer to free or reuse. If
2239 * bufspace useage is low skip this step so we can allocate a
2240 * new buffer.
2241 */
2242 if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2243 nqindex = QUEUE_CLEAN;
2244 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2245 }
2246
2247 /*
2248 * If we could not find or were not allowed to reuse a CLEAN
2249 * buffer, check to see if it is ok to use an EMPTY buffer.
2250 * We can only use an EMPTY buffer if allocating its KVA would
2251 * not otherwise run us out of buffer space. No KVA is needed
2252 * for the unmapped allocation.
2253 */
2254 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2255 metadata)) {
2256 nqindex = QUEUE_EMPTY;
2257 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2258 }
2259
2260 /*
2261 * All available buffers might be clean, retry ignoring the
2262 * lobufspace as the last resort.
2263 */
2264 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2265 nqindex = QUEUE_CLEAN;
2266 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2267 }
2268
2269 /*
2270 * Run scan, possibly freeing data and/or kva mappings on the fly
2271 * depending.
2272 */
2273 while ((bp = nbp) != NULL) {
2274 qindex = nqindex;
2275
2276 /*
2277 * Calculate next bp (we can only use it if we do not
2278 * block or do other fancy things).
2279 */
2280 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2281 switch (qindex) {
2282 case QUEUE_EMPTY:
2283 nqindex = QUEUE_EMPTYKVA;
2284 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2285 if (nbp != NULL)
2286 break;
2287 /* FALLTHROUGH */
2288 case QUEUE_EMPTYKVA:
2289 nqindex = QUEUE_CLEAN;
2290 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2291 if (nbp != NULL)
2292 break;
2293 /* FALLTHROUGH */
2294 case QUEUE_CLEAN:
2295 if (metadata && pass == 1) {
2296 pass = 2;
2297 nqindex = QUEUE_EMPTY;
2298 nbp = TAILQ_FIRST(
2299 &bufqueues[QUEUE_EMPTY]);
2300 }
2301 /*
2302 * nbp is NULL.
2303 */
2304 break;
2305 }
2306 }
2307 /*
2308 * If we are defragging then we need a buffer with
2309 * b_kvasize != 0. XXX this situation should no longer
2310 * occur, if defrag is non-zero the buffer's b_kvasize
2311 * should also be non-zero at this point. XXX
2312 */
2313 if (defrag && bp->b_kvasize == 0) {
2314 printf("Warning: defrag empty buffer %p\n", bp);
2315 continue;
2316 }
2317
2318 /*
2319 * Start freeing the bp. This is somewhat involved. nbp
2320 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2321 */
2322 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2323 continue;
2324 /*
2325 * BKGRDINPROG can only be set with the buf and bufobj
2326 * locks both held. We tolerate a race to clear it here.
2327 */
2328 if (bp->b_vflags & BV_BKGRDINPROG) {
2329 BUF_UNLOCK(bp);
2330 continue;
2331 }
2332
2333 KASSERT(bp->b_qindex == qindex,
2334 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2335
2336 bremfreel(bp);
2337 mtx_unlock(&bqclean);
2338 /*
2339 * NOTE: nbp is now entirely invalid. We can only restart
2340 * the scan from this point on.
2341 */
2342
2343 getnewbuf_reuse_bp(bp, qindex);
2344 mtx_assert(&bqclean, MA_NOTOWNED);
2345
2346 /*
2347 * If we are defragging then free the buffer.
2348 */
2349 if (defrag) {
2350 bp->b_flags |= B_INVAL;
2351 bfreekva(bp);
2352 brelse(bp);
2353 defrag = 0;
2354 goto restart;
2355 }
2356
2357 /*
2358 * Notify any waiters for the buffer lock about
2359 * identity change by freeing the buffer.
2360 */
2361 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2362 bp->b_flags |= B_INVAL;
2363 bfreekva(bp);
2364 brelse(bp);
2365 goto restart;
2366 }
2367
2368 if (metadata)
2369 break;
2370
2371 /*
2372 * If we are overcomitted then recover the buffer and its
2373 * KVM space. This occurs in rare situations when multiple
2374 * processes are blocked in getnewbuf() or allocbuf().
2375 */
2376 if (bufspace >= hibufspace)
2377 flushingbufs = 1;
2378 if (flushingbufs && bp->b_kvasize != 0) {
2379 bp->b_flags |= B_INVAL;
2380 bfreekva(bp);
2381 brelse(bp);
2382 goto restart;
2383 }
2384 if (bufspace < lobufspace)
2385 flushingbufs = 0;
2386 break;
2387 }
2388 return (bp);
2389 }
2390
2391 /*
2392 * getnewbuf:
2393 *
2394 * Find and initialize a new buffer header, freeing up existing buffers
2395 * in the bufqueues as necessary. The new buffer is returned locked.
2396 *
2397 * Important: B_INVAL is not set. If the caller wishes to throw the
2398 * buffer away, the caller must set B_INVAL prior to calling brelse().
2399 *
2400 * We block if:
2401 * We have insufficient buffer headers
2402 * We have insufficient buffer space
2403 * buffer_arena is too fragmented ( space reservation fails )
2404 * If we have to flush dirty buffers ( but we try to avoid this )
2405 */
2406 static struct buf *
2407 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2408 int gbflags)
2409 {
2410 struct buf *bp;
2411 int defrag, metadata;
2412
2413 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2414 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2415 if (!unmapped_buf_allowed)
2416 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2417
2418 defrag = 0;
2419 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2420 vp->v_type == VCHR)
2421 metadata = 1;
2422 else
2423 metadata = 0;
2424 /*
2425 * We can't afford to block since we might be holding a vnode lock,
2426 * which may prevent system daemons from running. We deal with
2427 * low-memory situations by proactively returning memory and running
2428 * async I/O rather then sync I/O.
2429 */
2430 atomic_add_int(&getnewbufcalls, 1);
2431 atomic_subtract_int(&getnewbufrestarts, 1);
2432 restart:
2433 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2434 GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2435 if (bp != NULL)
2436 defrag = 0;
2437
2438 /*
2439 * If we exhausted our list, sleep as appropriate. We may have to
2440 * wakeup various daemons and write out some dirty buffers.
2441 *
2442 * Generally we are sleeping due to insufficient buffer space.
2443 */
2444 if (bp == NULL) {
2445 mtx_assert(&bqclean, MA_OWNED);
2446 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2447 mtx_assert(&bqclean, MA_NOTOWNED);
2448 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2449 mtx_assert(&bqclean, MA_NOTOWNED);
2450
2451 bfreekva(bp);
2452 bp->b_flags |= B_UNMAPPED;
2453 bp->b_kvabase = bp->b_data = unmapped_buf;
2454 bp->b_kvasize = maxsize;
2455 atomic_add_long(&bufspace, bp->b_kvasize);
2456 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2457 atomic_add_int(&bufreusecnt, 1);
2458 } else {
2459 mtx_assert(&bqclean, MA_NOTOWNED);
2460
2461 /*
2462 * We finally have a valid bp. We aren't quite out of the
2463 * woods, we still have to reserve kva space. In order
2464 * to keep fragmentation sane we only allocate kva in
2465 * BKVASIZE chunks.
2466 */
2467 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2468
2469 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2470 B_KVAALLOC)) == B_UNMAPPED) {
2471 if (allocbufkva(bp, maxsize, gbflags)) {
2472 defrag = 1;
2473 bp->b_flags |= B_INVAL;
2474 brelse(bp);
2475 goto restart;
2476 }
2477 atomic_add_int(&bufreusecnt, 1);
2478 } else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2479 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2480 /*
2481 * If the reused buffer has KVA allocated,
2482 * reassign b_kvaalloc to b_kvabase.
2483 */
2484 bp->b_kvabase = bp->b_kvaalloc;
2485 bp->b_flags &= ~B_KVAALLOC;
2486 atomic_subtract_long(&unmapped_bufspace,
2487 bp->b_kvasize);
2488 atomic_add_int(&bufreusecnt, 1);
2489 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2490 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2491 GB_KVAALLOC)) {
2492 /*
2493 * The case of reused buffer already have KVA
2494 * mapped, but the request is for unmapped
2495 * buffer with KVA allocated.
2496 */
2497 bp->b_kvaalloc = bp->b_kvabase;
2498 bp->b_data = bp->b_kvabase = unmapped_buf;
2499 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2500 atomic_add_long(&unmapped_bufspace,
2501 bp->b_kvasize);
2502 atomic_add_int(&bufreusecnt, 1);
2503 }
2504 if ((gbflags & GB_UNMAPPED) == 0) {
2505 bp->b_saveaddr = bp->b_kvabase;
2506 bp->b_data = bp->b_saveaddr;
2507 bp->b_flags &= ~B_UNMAPPED;
2508 BUF_CHECK_MAPPED(bp);
2509 }
2510 }
2511 return (bp);
2512 }
2513
2514 /*
2515 * buf_daemon:
2516 *
2517 * buffer flushing daemon. Buffers are normally flushed by the
2518 * update daemon but if it cannot keep up this process starts to
2519 * take the load in an attempt to prevent getnewbuf() from blocking.
2520 */
2521
2522 static struct kproc_desc buf_kp = {
2523 "bufdaemon",
2524 buf_daemon,
2525 &bufdaemonproc
2526 };
2527 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2528
2529 static int
2530 buf_flush(int target)
2531 {
2532 int flushed;
2533
2534 flushed = flushbufqueues(target, 0);
2535 if (flushed == 0) {
2536 /*
2537 * Could not find any buffers without rollback
2538 * dependencies, so just write the first one
2539 * in the hopes of eventually making progress.
2540 */
2541 flushed = flushbufqueues(target, 1);
2542 }
2543 return (flushed);
2544 }
2545
2546 static void
2547 buf_daemon()
2548 {
2549 int lodirty;
2550
2551 /*
2552 * This process needs to be suspended prior to shutdown sync.
2553 */
2554 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2555 SHUTDOWN_PRI_LAST);
2556
2557 /*
2558 * This process is allowed to take the buffer cache to the limit
2559 */
2560 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2561 mtx_lock(&bdlock);
2562 for (;;) {
2563 bd_request = 0;
2564 mtx_unlock(&bdlock);
2565
2566 kproc_suspend_check(bufdaemonproc);
2567 lodirty = lodirtybuffers;
2568 if (bd_speedupreq) {
2569 lodirty = numdirtybuffers / 2;
2570 bd_speedupreq = 0;
2571 }
2572 /*
2573 * Do the flush. Limit the amount of in-transit I/O we
2574 * allow to build up, otherwise we would completely saturate
2575 * the I/O system.
2576 */
2577 while (numdirtybuffers > lodirty) {
2578 if (buf_flush(numdirtybuffers - lodirty) == 0)
2579 break;
2580 kern_yield(PRI_USER);
2581 }
2582
2583 /*
2584 * Only clear bd_request if we have reached our low water
2585 * mark. The buf_daemon normally waits 1 second and
2586 * then incrementally flushes any dirty buffers that have
2587 * built up, within reason.
2588 *
2589 * If we were unable to hit our low water mark and couldn't
2590 * find any flushable buffers, we sleep for a short period
2591 * to avoid endless loops on unlockable buffers.
2592 */
2593 mtx_lock(&bdlock);
2594 if (numdirtybuffers <= lodirtybuffers) {
2595 /*
2596 * We reached our low water mark, reset the
2597 * request and sleep until we are needed again.
2598 * The sleep is just so the suspend code works.
2599 */
2600 bd_request = 0;
2601 /*
2602 * Do an extra wakeup in case dirty threshold
2603 * changed via sysctl and the explicit transition
2604 * out of shortfall was missed.
2605 */
2606 bdirtywakeup();
2607 if (runningbufspace <= lorunningspace)
2608 runningwakeup();
2609 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2610 } else {
2611 /*
2612 * We couldn't find any flushable dirty buffers but
2613 * still have too many dirty buffers, we
2614 * have to sleep and try again. (rare)
2615 */
2616 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2617 }
2618 }
2619 }
2620
2621 /*
2622 * flushbufqueues:
2623 *
2624 * Try to flush a buffer in the dirty queue. We must be careful to
2625 * free up B_INVAL buffers instead of write them, which NFS is
2626 * particularly sensitive to.
2627 */
2628 static int flushwithdeps = 0;
2629 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2630 0, "Number of buffers flushed with dependecies that require rollbacks");
2631
2632 static int
2633 flushbufqueues(int target, int flushdeps)
2634 {
2635 struct buf *sentinel;
2636 struct vnode *vp;
2637 struct mount *mp;
2638 struct buf *bp;
2639 int hasdeps;
2640 int flushed;
2641 int queue;
2642 int error;
2643
2644 flushed = 0;
2645 queue = QUEUE_DIRTY;
2646 bp = NULL;
2647 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2648 sentinel->b_qindex = QUEUE_SENTINEL;
2649 mtx_lock(&bqdirty);
2650 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2651 mtx_unlock(&bqdirty);
2652 while (flushed != target) {
2653 maybe_yield();
2654 mtx_lock(&bqdirty);
2655 bp = TAILQ_NEXT(sentinel, b_freelist);
2656 if (bp != NULL) {
2657 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2658 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2659 b_freelist);
2660 } else {
2661 mtx_unlock(&bqdirty);
2662 break;
2663 }
2664 KASSERT(bp->b_qindex != QUEUE_SENTINEL,
2665 ("parallel calls to flushbufqueues() bp %p", bp));
2666 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2667 mtx_unlock(&bqdirty);
2668 if (error != 0)
2669 continue;
2670 if (bp->b_pin_count > 0) {
2671 BUF_UNLOCK(bp);
2672 continue;
2673 }
2674 /*
2675 * BKGRDINPROG can only be set with the buf and bufobj
2676 * locks both held. We tolerate a race to clear it here.
2677 */
2678 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2679 (bp->b_flags & B_DELWRI) == 0) {
2680 BUF_UNLOCK(bp);
2681 continue;
2682 }
2683 if (bp->b_flags & B_INVAL) {
2684 bremfreef(bp);
2685 brelse(bp);
2686 flushed++;
2687 continue;
2688 }
2689
2690 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2691 if (flushdeps == 0) {
2692 BUF_UNLOCK(bp);
2693 continue;
2694 }
2695 hasdeps = 1;
2696 } else
2697 hasdeps = 0;
2698 /*
2699 * We must hold the lock on a vnode before writing
2700 * one of its buffers. Otherwise we may confuse, or
2701 * in the case of a snapshot vnode, deadlock the
2702 * system.
2703 *
2704 * The lock order here is the reverse of the normal
2705 * of vnode followed by buf lock. This is ok because
2706 * the NOWAIT will prevent deadlock.
2707 */
2708 vp = bp->b_vp;
2709 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2710 BUF_UNLOCK(bp);
2711 continue;
2712 }
2713 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2714 if (error == 0) {
2715 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2716 bp, bp->b_vp, bp->b_flags);
2717 vfs_bio_awrite(bp);
2718 vn_finished_write(mp);
2719 VOP_UNLOCK(vp, 0);
2720 flushwithdeps += hasdeps;
2721 flushed++;
2722 if (runningbufspace > hirunningspace)
2723 waitrunningbufspace();
2724 continue;
2725 }
2726 vn_finished_write(mp);
2727 BUF_UNLOCK(bp);
2728 }
2729 mtx_lock(&bqdirty);
2730 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2731 mtx_unlock(&bqdirty);
2732 free(sentinel, M_TEMP);
2733 return (flushed);
2734 }
2735
2736 /*
2737 * Check to see if a block is currently memory resident.
2738 */
2739 struct buf *
2740 incore(struct bufobj *bo, daddr_t blkno)
2741 {
2742 struct buf *bp;
2743
2744 BO_RLOCK(bo);
2745 bp = gbincore(bo, blkno);
2746 BO_RUNLOCK(bo);
2747 return (bp);
2748 }
2749
2750 /*
2751 * Returns true if no I/O is needed to access the
2752 * associated VM object. This is like incore except
2753 * it also hunts around in the VM system for the data.
2754 */
2755
2756 static int
2757 inmem(struct vnode * vp, daddr_t blkno)
2758 {
2759 vm_object_t obj;
2760 vm_offset_t toff, tinc, size;
2761 vm_page_t m;
2762 vm_ooffset_t off;
2763
2764 ASSERT_VOP_LOCKED(vp, "inmem");
2765
2766 if (incore(&vp->v_bufobj, blkno))
2767 return 1;
2768 if (vp->v_mount == NULL)
2769 return 0;
2770 obj = vp->v_object;
2771 if (obj == NULL)
2772 return (0);
2773
2774 size = PAGE_SIZE;
2775 if (size > vp->v_mount->mnt_stat.f_iosize)
2776 size = vp->v_mount->mnt_stat.f_iosize;
2777 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2778
2779 VM_OBJECT_RLOCK(obj);
2780 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2781 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2782 if (!m)
2783 goto notinmem;
2784 tinc = size;
2785 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2786 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2787 if (vm_page_is_valid(m,
2788 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2789 goto notinmem;
2790 }
2791 VM_OBJECT_RUNLOCK(obj);
2792 return 1;
2793
2794 notinmem:
2795 VM_OBJECT_RUNLOCK(obj);
2796 return (0);
2797 }
2798
2799 /*
2800 * Set the dirty range for a buffer based on the status of the dirty
2801 * bits in the pages comprising the buffer. The range is limited
2802 * to the size of the buffer.
2803 *
2804 * Tell the VM system that the pages associated with this buffer
2805 * are clean. This is used for delayed writes where the data is
2806 * going to go to disk eventually without additional VM intevention.
2807 *
2808 * Note that while we only really need to clean through to b_bcount, we
2809 * just go ahead and clean through to b_bufsize.
2810 */
2811 static void
2812 vfs_clean_pages_dirty_buf(struct buf *bp)
2813 {
2814 vm_ooffset_t foff, noff, eoff;
2815 vm_page_t m;
2816 int i;
2817
2818 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2819 return;
2820
2821 foff = bp->b_offset;
2822 KASSERT(bp->b_offset != NOOFFSET,
2823 ("vfs_clean_pages_dirty_buf: no buffer offset"));
2824
2825 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2826 vfs_drain_busy_pages(bp);
2827 vfs_setdirty_locked_object(bp);
2828 for (i = 0; i < bp->b_npages; i++) {
2829 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2830 eoff = noff;
2831 if (eoff > bp->b_offset + bp->b_bufsize)
2832 eoff = bp->b_offset + bp->b_bufsize;
2833 m = bp->b_pages[i];
2834 vfs_page_set_validclean(bp, foff, m);
2835 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2836 foff = noff;
2837 }
2838 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2839 }
2840
2841 static void
2842 vfs_setdirty_locked_object(struct buf *bp)
2843 {
2844 vm_object_t object;
2845 int i;
2846
2847 object = bp->b_bufobj->bo_object;
2848 VM_OBJECT_ASSERT_WLOCKED(object);
2849
2850 /*
2851 * We qualify the scan for modified pages on whether the
2852 * object has been flushed yet.
2853 */
2854 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2855 vm_offset_t boffset;
2856 vm_offset_t eoffset;
2857
2858 /*
2859 * test the pages to see if they have been modified directly
2860 * by users through the VM system.
2861 */
2862 for (i = 0; i < bp->b_npages; i++)
2863 vm_page_test_dirty(bp->b_pages[i]);
2864
2865 /*
2866 * Calculate the encompassing dirty range, boffset and eoffset,
2867 * (eoffset - boffset) bytes.
2868 */
2869
2870 for (i = 0; i < bp->b_npages; i++) {
2871 if (bp->b_pages[i]->dirty)
2872 break;
2873 }
2874 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2875
2876 for (i = bp->b_npages - 1; i >= 0; --i) {
2877 if (bp->b_pages[i]->dirty) {
2878 break;
2879 }
2880 }
2881 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2882
2883 /*
2884 * Fit it to the buffer.
2885 */
2886
2887 if (eoffset > bp->b_bcount)
2888 eoffset = bp->b_bcount;
2889
2890 /*
2891 * If we have a good dirty range, merge with the existing
2892 * dirty range.
2893 */
2894
2895 if (boffset < eoffset) {
2896 if (bp->b_dirtyoff > boffset)
2897 bp->b_dirtyoff = boffset;
2898 if (bp->b_dirtyend < eoffset)
2899 bp->b_dirtyend = eoffset;
2900 }
2901 }
2902 }
2903
2904 /*
2905 * Allocate the KVA mapping for an existing buffer. It handles the
2906 * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2907 * KVA which is not mapped (B_KVAALLOC).
2908 */
2909 static void
2910 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2911 {
2912 struct buf *scratch_bp;
2913 int bsize, maxsize, need_mapping, need_kva;
2914 off_t offset;
2915
2916 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2917 (gbflags & GB_UNMAPPED) == 0;
2918 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2919 (gbflags & GB_KVAALLOC) != 0;
2920 if (!need_mapping && !need_kva)
2921 return;
2922
2923 BUF_CHECK_UNMAPPED(bp);
2924
2925 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2926 /*
2927 * Buffer is not mapped, but the KVA was already
2928 * reserved at the time of the instantiation. Use the
2929 * allocated space.
2930 */
2931 bp->b_flags &= ~B_KVAALLOC;
2932 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2933 bp->b_kvabase = bp->b_kvaalloc;
2934 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
2935 goto has_addr;
2936 }
2937
2938 /*
2939 * Calculate the amount of the address space we would reserve
2940 * if the buffer was mapped.
2941 */
2942 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
2943 offset = blkno * bsize;
2944 maxsize = size + (offset & PAGE_MASK);
2945 maxsize = imax(maxsize, bsize);
2946
2947 mapping_loop:
2948 if (allocbufkva(bp, maxsize, gbflags)) {
2949 /*
2950 * Request defragmentation. getnewbuf() returns us the
2951 * allocated space by the scratch buffer KVA.
2952 */
2953 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
2954 (GB_UNMAPPED | GB_KVAALLOC));
2955 if (scratch_bp == NULL) {
2956 if ((gbflags & GB_NOWAIT_BD) != 0) {
2957 /*
2958 * XXXKIB: defragmentation cannot
2959 * succeed, not sure what else to do.
2960 */
2961 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
2962 }
2963 atomic_add_int(&mappingrestarts, 1);
2964 goto mapping_loop;
2965 }
2966 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
2967 ("scratch bp !B_KVAALLOC %p", scratch_bp));
2968 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
2969 scratch_bp->b_kvasize, gbflags);
2970
2971 /* Get rid of the scratch buffer. */
2972 scratch_bp->b_kvasize = 0;
2973 scratch_bp->b_flags |= B_INVAL;
2974 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
2975 brelse(scratch_bp);
2976 }
2977 if (!need_mapping)
2978 return;
2979
2980 has_addr:
2981 bp->b_saveaddr = bp->b_kvabase;
2982 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
2983 bp->b_flags &= ~B_UNMAPPED;
2984 BUF_CHECK_MAPPED(bp);
2985 bpmap_qenter(bp);
2986 }
2987
2988 /*
2989 * getblk:
2990 *
2991 * Get a block given a specified block and offset into a file/device.
2992 * The buffers B_DONE bit will be cleared on return, making it almost
2993 * ready for an I/O initiation. B_INVAL may or may not be set on
2994 * return. The caller should clear B_INVAL prior to initiating a
2995 * READ.
2996 *
2997 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2998 * an existing buffer.
2999 *
3000 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3001 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3002 * and then cleared based on the backing VM. If the previous buffer is
3003 * non-0-sized but invalid, B_CACHE will be cleared.
3004 *
3005 * If getblk() must create a new buffer, the new buffer is returned with
3006 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3007 * case it is returned with B_INVAL clear and B_CACHE set based on the
3008 * backing VM.
3009 *
3010 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
3011 * B_CACHE bit is clear.
3012 *
3013 * What this means, basically, is that the caller should use B_CACHE to
3014 * determine whether the buffer is fully valid or not and should clear
3015 * B_INVAL prior to issuing a read. If the caller intends to validate
3016 * the buffer by loading its data area with something, the caller needs
3017 * to clear B_INVAL. If the caller does this without issuing an I/O,
3018 * the caller should set B_CACHE ( as an optimization ), else the caller
3019 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3020 * a write attempt or if it was a successfull read. If the caller
3021 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3022 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3023 */
3024 struct buf *
3025 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3026 int flags)
3027 {
3028 struct buf *bp;
3029 struct bufobj *bo;
3030 int bsize, error, maxsize, vmio;
3031 off_t offset;
3032
3033 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3034 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3035 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3036 ASSERT_VOP_LOCKED(vp, "getblk");
3037 if (size > MAXBSIZE)
3038 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
3039 if (!unmapped_buf_allowed)
3040 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3041
3042 bo = &vp->v_bufobj;
3043 loop:
3044 BO_RLOCK(bo);
3045 bp = gbincore(bo, blkno);
3046 if (bp != NULL) {
3047 int lockflags;
3048 /*
3049 * Buffer is in-core. If the buffer is not busy nor managed,
3050 * it must be on a queue.
3051 */
3052 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3053
3054 if (flags & GB_LOCK_NOWAIT)
3055 lockflags |= LK_NOWAIT;
3056
3057 error = BUF_TIMELOCK(bp, lockflags,
3058 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3059
3060 /*
3061 * If we slept and got the lock we have to restart in case
3062 * the buffer changed identities.
3063 */
3064 if (error == ENOLCK)
3065 goto loop;
3066 /* We timed out or were interrupted. */
3067 else if (error)
3068 return (NULL);
3069 /* If recursed, assume caller knows the rules. */
3070 else if (BUF_LOCKRECURSED(bp))
3071 goto end;
3072
3073 /*
3074 * The buffer is locked. B_CACHE is cleared if the buffer is
3075 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3076 * and for a VMIO buffer B_CACHE is adjusted according to the
3077 * backing VM cache.
3078 */
3079 if (bp->b_flags & B_INVAL)
3080 bp->b_flags &= ~B_CACHE;
3081 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3082 bp->b_flags |= B_CACHE;
3083 if (bp->b_flags & B_MANAGED)
3084 MPASS(bp->b_qindex == QUEUE_NONE);
3085 else
3086 bremfree(bp);
3087
3088 /*
3089 * check for size inconsistencies for non-VMIO case.
3090 */
3091 if (bp->b_bcount != size) {
3092 if ((bp->b_flags & B_VMIO) == 0 ||
3093 (size > bp->b_kvasize)) {
3094 if (bp->b_flags & B_DELWRI) {
3095 /*
3096 * If buffer is pinned and caller does
3097 * not want sleep waiting for it to be
3098 * unpinned, bail out
3099 * */
3100 if (bp->b_pin_count > 0) {
3101 if (flags & GB_LOCK_NOWAIT) {
3102 bqrelse(bp);
3103 return (NULL);
3104 } else {
3105 bunpin_wait(bp);
3106 }
3107 }
3108 bp->b_flags |= B_NOCACHE;
3109 bwrite(bp);
3110 } else {
3111 if (LIST_EMPTY(&bp->b_dep)) {
3112 bp->b_flags |= B_RELBUF;
3113 brelse(bp);
3114 } else {
3115 bp->b_flags |= B_NOCACHE;
3116 bwrite(bp);
3117 }
3118 }
3119 goto loop;
3120 }
3121 }
3122
3123 /*
3124 * Handle the case of unmapped buffer which should
3125 * become mapped, or the buffer for which KVA
3126 * reservation is requested.
3127 */
3128 bp_unmapped_get_kva(bp, blkno, size, flags);
3129
3130 /*
3131 * If the size is inconsistant in the VMIO case, we can resize
3132 * the buffer. This might lead to B_CACHE getting set or
3133 * cleared. If the size has not changed, B_CACHE remains
3134 * unchanged from its previous state.
3135 */
3136 if (bp->b_bcount != size)
3137 allocbuf(bp, size);
3138
3139 KASSERT(bp->b_offset != NOOFFSET,
3140 ("getblk: no buffer offset"));
3141
3142 /*
3143 * A buffer with B_DELWRI set and B_CACHE clear must
3144 * be committed before we can return the buffer in
3145 * order to prevent the caller from issuing a read
3146 * ( due to B_CACHE not being set ) and overwriting
3147 * it.
3148 *
3149 * Most callers, including NFS and FFS, need this to
3150 * operate properly either because they assume they
3151 * can issue a read if B_CACHE is not set, or because
3152 * ( for example ) an uncached B_DELWRI might loop due
3153 * to softupdates re-dirtying the buffer. In the latter
3154 * case, B_CACHE is set after the first write completes,
3155 * preventing further loops.
3156 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3157 * above while extending the buffer, we cannot allow the
3158 * buffer to remain with B_CACHE set after the write
3159 * completes or it will represent a corrupt state. To
3160 * deal with this we set B_NOCACHE to scrap the buffer
3161 * after the write.
3162 *
3163 * We might be able to do something fancy, like setting
3164 * B_CACHE in bwrite() except if B_DELWRI is already set,
3165 * so the below call doesn't set B_CACHE, but that gets real
3166 * confusing. This is much easier.
3167 */
3168
3169 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3170 bp->b_flags |= B_NOCACHE;
3171 bwrite(bp);
3172 goto loop;
3173 }
3174 bp->b_flags &= ~B_DONE;
3175 } else {
3176 /*
3177 * Buffer is not in-core, create new buffer. The buffer
3178 * returned by getnewbuf() is locked. Note that the returned
3179 * buffer is also considered valid (not marked B_INVAL).
3180 */
3181 BO_RUNLOCK(bo);
3182 /*
3183 * If the user does not want us to create the buffer, bail out
3184 * here.
3185 */
3186 if (flags & GB_NOCREAT)
3187 return NULL;
3188 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3189 return NULL;
3190
3191 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3192 offset = blkno * bsize;
3193 vmio = vp->v_object != NULL;
3194 if (vmio) {
3195 maxsize = size + (offset & PAGE_MASK);
3196 } else {
3197 maxsize = size;
3198 /* Do not allow non-VMIO notmapped buffers. */
3199 flags &= ~GB_UNMAPPED;
3200 }
3201 maxsize = imax(maxsize, bsize);
3202
3203 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3204 if (bp == NULL) {
3205 if (slpflag || slptimeo)
3206 return NULL;
3207 goto loop;
3208 }
3209
3210 /*
3211 * This code is used to make sure that a buffer is not
3212 * created while the getnewbuf routine is blocked.
3213 * This can be a problem whether the vnode is locked or not.
3214 * If the buffer is created out from under us, we have to
3215 * throw away the one we just created.
3216 *
3217 * Note: this must occur before we associate the buffer
3218 * with the vp especially considering limitations in
3219 * the splay tree implementation when dealing with duplicate
3220 * lblkno's.
3221 */
3222 BO_LOCK(bo);
3223 if (gbincore(bo, blkno)) {
3224 BO_UNLOCK(bo);
3225 bp->b_flags |= B_INVAL;
3226 brelse(bp);
3227 goto loop;
3228 }
3229
3230 /*
3231 * Insert the buffer into the hash, so that it can
3232 * be found by incore.
3233 */
3234 bp->b_blkno = bp->b_lblkno = blkno;
3235 bp->b_offset = offset;
3236 bgetvp(vp, bp);
3237 BO_UNLOCK(bo);
3238
3239 /*
3240 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
3241 * buffer size starts out as 0, B_CACHE will be set by
3242 * allocbuf() for the VMIO case prior to it testing the
3243 * backing store for validity.
3244 */
3245
3246 if (vmio) {
3247 bp->b_flags |= B_VMIO;
3248 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3249 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3250 bp, vp->v_object, bp->b_bufobj->bo_object));
3251 } else {
3252 bp->b_flags &= ~B_VMIO;
3253 KASSERT(bp->b_bufobj->bo_object == NULL,
3254 ("ARGH! has b_bufobj->bo_object %p %p\n",
3255 bp, bp->b_bufobj->bo_object));
3256 BUF_CHECK_MAPPED(bp);
3257 }
3258
3259 allocbuf(bp, size);
3260 bp->b_flags &= ~B_DONE;
3261 }
3262 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3263 BUF_ASSERT_HELD(bp);
3264 end:
3265 KASSERT(bp->b_bufobj == bo,
3266 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3267 return (bp);
3268 }
3269
3270 /*
3271 * Get an empty, disassociated buffer of given size. The buffer is initially
3272 * set to B_INVAL.
3273 */
3274 struct buf *
3275 geteblk(int size, int flags)
3276 {
3277 struct buf *bp;
3278 int maxsize;
3279
3280 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3281 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3282 if ((flags & GB_NOWAIT_BD) &&
3283 (curthread->td_pflags & TDP_BUFNEED) != 0)
3284 return (NULL);
3285 }
3286 allocbuf(bp, size);
3287 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3288 BUF_ASSERT_HELD(bp);
3289 return (bp);
3290 }
3291
3292
3293 /*
3294 * This code constitutes the buffer memory from either anonymous system
3295 * memory (in the case of non-VMIO operations) or from an associated
3296 * VM object (in the case of VMIO operations). This code is able to
3297 * resize a buffer up or down.
3298 *
3299 * Note that this code is tricky, and has many complications to resolve
3300 * deadlock or inconsistant data situations. Tread lightly!!!
3301 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3302 * the caller. Calling this code willy nilly can result in the loss of data.
3303 *
3304 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3305 * B_CACHE for the non-VMIO case.
3306 */
3307
3308 int
3309 allocbuf(struct buf *bp, int size)
3310 {
3311 int newbsize, mbsize;
3312 int i;
3313
3314 BUF_ASSERT_HELD(bp);
3315
3316 if (bp->b_kvasize < size)
3317 panic("allocbuf: buffer too small");
3318
3319 if ((bp->b_flags & B_VMIO) == 0) {
3320 caddr_t origbuf;
3321 int origbufsize;
3322 /*
3323 * Just get anonymous memory from the kernel. Don't
3324 * mess with B_CACHE.
3325 */
3326 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3327 if (bp->b_flags & B_MALLOC)
3328 newbsize = mbsize;
3329 else
3330 newbsize = round_page(size);
3331
3332 if (newbsize < bp->b_bufsize) {
3333 /*
3334 * malloced buffers are not shrunk
3335 */
3336 if (bp->b_flags & B_MALLOC) {
3337 if (newbsize) {
3338 bp->b_bcount = size;
3339 } else {
3340 free(bp->b_data, M_BIOBUF);
3341 if (bp->b_bufsize) {
3342 atomic_subtract_long(
3343 &bufmallocspace,
3344 bp->b_bufsize);
3345 bufspacewakeup();
3346 bp->b_bufsize = 0;
3347 }
3348 bp->b_saveaddr = bp->b_kvabase;
3349 bp->b_data = bp->b_saveaddr;
3350 bp->b_bcount = 0;
3351 bp->b_flags &= ~B_MALLOC;
3352 }
3353 return 1;
3354 }
3355 vm_hold_free_pages(bp, newbsize);
3356 } else if (newbsize > bp->b_bufsize) {
3357 /*
3358 * We only use malloced memory on the first allocation.
3359 * and revert to page-allocated memory when the buffer
3360 * grows.
3361 */
3362 /*
3363 * There is a potential smp race here that could lead
3364 * to bufmallocspace slightly passing the max. It
3365 * is probably extremely rare and not worth worrying
3366 * over.
3367 */
3368 if ( (bufmallocspace < maxbufmallocspace) &&
3369 (bp->b_bufsize == 0) &&
3370 (mbsize <= PAGE_SIZE/2)) {
3371
3372 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3373 bp->b_bufsize = mbsize;
3374 bp->b_bcount = size;
3375 bp->b_flags |= B_MALLOC;
3376 atomic_add_long(&bufmallocspace, mbsize);
3377 return 1;
3378 }
3379 origbuf = NULL;
3380 origbufsize = 0;
3381 /*
3382 * If the buffer is growing on its other-than-first allocation,
3383 * then we revert to the page-allocation scheme.
3384 */
3385 if (bp->b_flags & B_MALLOC) {
3386 origbuf = bp->b_data;
3387 origbufsize = bp->b_bufsize;
3388 bp->b_data = bp->b_kvabase;
3389 if (bp->b_bufsize) {
3390 atomic_subtract_long(&bufmallocspace,
3391 bp->b_bufsize);
3392 bufspacewakeup();
3393 bp->b_bufsize = 0;
3394 }
3395 bp->b_flags &= ~B_MALLOC;
3396 newbsize = round_page(newbsize);
3397 }
3398 vm_hold_load_pages(
3399 bp,
3400 (vm_offset_t) bp->b_data + bp->b_bufsize,
3401 (vm_offset_t) bp->b_data + newbsize);
3402 if (origbuf) {
3403 bcopy(origbuf, bp->b_data, origbufsize);
3404 free(origbuf, M_BIOBUF);
3405 }
3406 }
3407 } else {
3408 int desiredpages;
3409
3410 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3411 desiredpages = (size == 0) ? 0 :
3412 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3413
3414 if (bp->b_flags & B_MALLOC)
3415 panic("allocbuf: VMIO buffer can't be malloced");
3416 /*
3417 * Set B_CACHE initially if buffer is 0 length or will become
3418 * 0-length.
3419 */
3420 if (size == 0 || bp->b_bufsize == 0)
3421 bp->b_flags |= B_CACHE;
3422
3423 if (newbsize < bp->b_bufsize) {
3424 /*
3425 * DEV_BSIZE aligned new buffer size is less then the
3426 * DEV_BSIZE aligned existing buffer size. Figure out
3427 * if we have to remove any pages.
3428 */
3429 if (desiredpages < bp->b_npages) {
3430 vm_page_t m;
3431
3432 if ((bp->b_flags & B_UNMAPPED) == 0) {
3433 BUF_CHECK_MAPPED(bp);
3434 pmap_qremove((vm_offset_t)trunc_page(
3435 (vm_offset_t)bp->b_data) +
3436 (desiredpages << PAGE_SHIFT),
3437 (bp->b_npages - desiredpages));
3438 } else
3439 BUF_CHECK_UNMAPPED(bp);
3440 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3441 for (i = desiredpages; i < bp->b_npages; i++) {
3442 /*
3443 * the page is not freed here -- it
3444 * is the responsibility of
3445 * vnode_pager_setsize
3446 */
3447 m = bp->b_pages[i];
3448 KASSERT(m != bogus_page,
3449 ("allocbuf: bogus page found"));
3450 while (vm_page_sleep_if_busy(m,
3451 "biodep"))
3452 continue;
3453
3454 bp->b_pages[i] = NULL;
3455 vm_page_lock(m);
3456 vm_page_unwire(m, 0);
3457 vm_page_unlock(m);
3458 }
3459 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3460 bp->b_npages = desiredpages;
3461 }
3462 } else if (size > bp->b_bcount) {
3463 /*
3464 * We are growing the buffer, possibly in a
3465 * byte-granular fashion.
3466 */
3467 vm_object_t obj;
3468 vm_offset_t toff;
3469 vm_offset_t tinc;
3470
3471 /*
3472 * Step 1, bring in the VM pages from the object,
3473 * allocating them if necessary. We must clear
3474 * B_CACHE if these pages are not valid for the
3475 * range covered by the buffer.
3476 */
3477
3478 obj = bp->b_bufobj->bo_object;
3479
3480 VM_OBJECT_WLOCK(obj);
3481 while (bp->b_npages < desiredpages) {
3482 vm_page_t m;
3483
3484 /*
3485 * We must allocate system pages since blocking
3486 * here could interfere with paging I/O, no
3487 * matter which process we are.
3488 *
3489 * Only exclusive busy can be tested here.
3490 * Blocking on shared busy might lead to
3491 * deadlocks once allocbuf() is called after
3492 * pages are vfs_busy_pages().
3493 */
3494 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3495 bp->b_npages, VM_ALLOC_NOBUSY |
3496 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3497 VM_ALLOC_IGN_SBUSY |
3498 VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3499 if (m->valid == 0)
3500 bp->b_flags &= ~B_CACHE;
3501 bp->b_pages[bp->b_npages] = m;
3502 ++bp->b_npages;
3503 }
3504
3505 /*
3506 * Step 2. We've loaded the pages into the buffer,
3507 * we have to figure out if we can still have B_CACHE
3508 * set. Note that B_CACHE is set according to the
3509 * byte-granular range ( bcount and size ), new the
3510 * aligned range ( newbsize ).
3511 *
3512 * The VM test is against m->valid, which is DEV_BSIZE
3513 * aligned. Needless to say, the validity of the data
3514 * needs to also be DEV_BSIZE aligned. Note that this
3515 * fails with NFS if the server or some other client
3516 * extends the file's EOF. If our buffer is resized,
3517 * B_CACHE may remain set! XXX
3518 */
3519
3520 toff = bp->b_bcount;
3521 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3522
3523 while ((bp->b_flags & B_CACHE) && toff < size) {
3524 vm_pindex_t pi;
3525
3526 if (tinc > (size - toff))
3527 tinc = size - toff;
3528
3529 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3530 PAGE_SHIFT;
3531
3532 vfs_buf_test_cache(
3533 bp,
3534 bp->b_offset,
3535 toff,
3536 tinc,
3537 bp->b_pages[pi]
3538 );
3539 toff += tinc;
3540 tinc = PAGE_SIZE;
3541 }
3542 VM_OBJECT_WUNLOCK(obj);
3543
3544 /*
3545 * Step 3, fixup the KVM pmap.
3546 */
3547 if ((bp->b_flags & B_UNMAPPED) == 0)
3548 bpmap_qenter(bp);
3549 else
3550 BUF_CHECK_UNMAPPED(bp);
3551 }
3552 }
3553 if (newbsize < bp->b_bufsize)
3554 bufspacewakeup();
3555 bp->b_bufsize = newbsize; /* actual buffer allocation */
3556 bp->b_bcount = size; /* requested buffer size */
3557 return 1;
3558 }
3559
3560 extern int inflight_transient_maps;
3561
3562 void
3563 biodone(struct bio *bp)
3564 {
3565 struct mtx *mtxp;
3566 void (*done)(struct bio *);
3567 vm_offset_t start, end;
3568
3569 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3570 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3571 bp->bio_flags |= BIO_UNMAPPED;
3572 start = trunc_page((vm_offset_t)bp->bio_data);
3573 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3574 pmap_qremove(start, OFF_TO_IDX(end - start));
3575 vmem_free(transient_arena, start, end - start);
3576 atomic_add_int(&inflight_transient_maps, -1);
3577 }
3578 done = bp->bio_done;
3579 if (done == NULL) {
3580 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3581 mtx_lock(mtxp);
3582 bp->bio_flags |= BIO_DONE;
3583 wakeup(bp);
3584 mtx_unlock(mtxp);
3585 } else {
3586 bp->bio_flags |= BIO_DONE;
3587 done(bp);
3588 }
3589 }
3590
3591 /*
3592 * Wait for a BIO to finish.
3593 */
3594 int
3595 biowait(struct bio *bp, const char *wchan)
3596 {
3597 struct mtx *mtxp;
3598
3599 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3600 mtx_lock(mtxp);
3601 while ((bp->bio_flags & BIO_DONE) == 0)
3602 msleep(bp, mtxp, PRIBIO, wchan, 0);
3603 mtx_unlock(mtxp);
3604 if (bp->bio_error != 0)
3605 return (bp->bio_error);
3606 if (!(bp->bio_flags & BIO_ERROR))
3607 return (0);
3608 return (EIO);
3609 }
3610
3611 void
3612 biofinish(struct bio *bp, struct devstat *stat, int error)
3613 {
3614
3615 if (error) {
3616 bp->bio_error = error;
3617 bp->bio_flags |= BIO_ERROR;
3618 }
3619 if (stat != NULL)
3620 devstat_end_transaction_bio(stat, bp);
3621 biodone(bp);
3622 }
3623
3624 /*
3625 * bufwait:
3626 *
3627 * Wait for buffer I/O completion, returning error status. The buffer
3628 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3629 * error and cleared.
3630 */
3631 int
3632 bufwait(struct buf *bp)
3633 {
3634 if (bp->b_iocmd == BIO_READ)
3635 bwait(bp, PRIBIO, "biord");
3636 else
3637 bwait(bp, PRIBIO, "biowr");
3638 if (bp->b_flags & B_EINTR) {
3639 bp->b_flags &= ~B_EINTR;
3640 return (EINTR);
3641 }
3642 if (bp->b_ioflags & BIO_ERROR) {
3643 return (bp->b_error ? bp->b_error : EIO);
3644 } else {
3645 return (0);
3646 }
3647 }
3648
3649 /*
3650 * Call back function from struct bio back up to struct buf.
3651 */
3652 static void
3653 bufdonebio(struct bio *bip)
3654 {
3655 struct buf *bp;
3656
3657 bp = bip->bio_caller2;
3658 bp->b_resid = bp->b_bcount - bip->bio_completed;
3659 bp->b_resid = bip->bio_resid; /* XXX: remove */
3660 bp->b_ioflags = bip->bio_flags;
3661 bp->b_error = bip->bio_error;
3662 if (bp->b_error)
3663 bp->b_ioflags |= BIO_ERROR;
3664 bufdone(bp);
3665 g_destroy_bio(bip);
3666 }
3667
3668 void
3669 dev_strategy(struct cdev *dev, struct buf *bp)
3670 {
3671 struct cdevsw *csw;
3672 int ref;
3673
3674 KASSERT(dev->si_refcount > 0,
3675 ("dev_strategy on un-referenced struct cdev *(%s) %p",
3676 devtoname(dev), dev));
3677
3678 csw = dev_refthread(dev, &ref);
3679 dev_strategy_csw(dev, csw, bp);
3680 dev_relthread(dev, ref);
3681 }
3682
3683 void
3684 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3685 {
3686 struct bio *bip;
3687
3688 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3689 ("b_iocmd botch"));
3690 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3691 dev->si_threadcount > 0,
3692 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3693 dev));
3694 if (csw == NULL) {
3695 bp->b_error = ENXIO;
3696 bp->b_ioflags = BIO_ERROR;
3697 bufdone(bp);
3698 return;
3699 }
3700 for (;;) {
3701 bip = g_new_bio();
3702 if (bip != NULL)
3703 break;
3704 /* Try again later */
3705 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3706 }
3707 bip->bio_cmd = bp->b_iocmd;
3708 bip->bio_offset = bp->b_iooffset;
3709 bip->bio_length = bp->b_bcount;
3710 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3711 bdata2bio(bp, bip);
3712 bip->bio_done = bufdonebio;
3713 bip->bio_caller2 = bp;
3714 bip->bio_dev = dev;
3715 (*csw->d_strategy)(bip);
3716 }
3717
3718 /*
3719 * bufdone:
3720 *
3721 * Finish I/O on a buffer, optionally calling a completion function.
3722 * This is usually called from an interrupt so process blocking is
3723 * not allowed.
3724 *
3725 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3726 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3727 * assuming B_INVAL is clear.
3728 *
3729 * For the VMIO case, we set B_CACHE if the op was a read and no
3730 * read error occured, or if the op was a write. B_CACHE is never
3731 * set if the buffer is invalid or otherwise uncacheable.
3732 *
3733 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3734 * initiator to leave B_INVAL set to brelse the buffer out of existance
3735 * in the biodone routine.
3736 */
3737 void
3738 bufdone(struct buf *bp)
3739 {
3740 struct bufobj *dropobj;
3741 void (*biodone)(struct buf *);
3742
3743 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3744 dropobj = NULL;
3745
3746 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3747 BUF_ASSERT_HELD(bp);
3748
3749 runningbufwakeup(bp);
3750 if (bp->b_iocmd == BIO_WRITE)
3751 dropobj = bp->b_bufobj;
3752 /* call optional completion function if requested */
3753 if (bp->b_iodone != NULL) {
3754 biodone = bp->b_iodone;
3755 bp->b_iodone = NULL;
3756 (*biodone) (bp);
3757 if (dropobj)
3758 bufobj_wdrop(dropobj);
3759 return;
3760 }
3761
3762 bufdone_finish(bp);
3763
3764 if (dropobj)
3765 bufobj_wdrop(dropobj);
3766 }
3767
3768 void
3769 bufdone_finish(struct buf *bp)
3770 {
3771 BUF_ASSERT_HELD(bp);
3772
3773 if (!LIST_EMPTY(&bp->b_dep))
3774 buf_complete(bp);
3775
3776 if (bp->b_flags & B_VMIO) {
3777 vm_ooffset_t foff;
3778 vm_page_t m;
3779 vm_object_t obj;
3780 struct vnode *vp;
3781 int bogus, i, iosize;
3782
3783 obj = bp->b_bufobj->bo_object;
3784 KASSERT(obj->paging_in_progress >= bp->b_npages,
3785 ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3786 obj->paging_in_progress, bp->b_npages));
3787
3788 vp = bp->b_vp;
3789 KASSERT(vp->v_holdcnt > 0,
3790 ("biodone_finish: vnode %p has zero hold count", vp));
3791 KASSERT(vp->v_object != NULL,
3792 ("biodone_finish: vnode %p has no vm_object", vp));
3793
3794 foff = bp->b_offset;
3795 KASSERT(bp->b_offset != NOOFFSET,
3796 ("biodone_finish: bp %p has no buffer offset", bp));
3797
3798 /*
3799 * Set B_CACHE if the op was a normal read and no error
3800 * occured. B_CACHE is set for writes in the b*write()
3801 * routines.
3802 */
3803 iosize = bp->b_bcount - bp->b_resid;
3804 if (bp->b_iocmd == BIO_READ &&
3805 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3806 !(bp->b_ioflags & BIO_ERROR)) {
3807 bp->b_flags |= B_CACHE;
3808 }
3809 bogus = 0;
3810 VM_OBJECT_WLOCK(obj);
3811 for (i = 0; i < bp->b_npages; i++) {
3812 int bogusflag = 0;
3813 int resid;
3814
3815 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3816 if (resid > iosize)
3817 resid = iosize;
3818
3819 /*
3820 * cleanup bogus pages, restoring the originals
3821 */
3822 m = bp->b_pages[i];
3823 if (m == bogus_page) {
3824 bogus = bogusflag = 1;
3825 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3826 if (m == NULL)
3827 panic("biodone: page disappeared!");
3828 bp->b_pages[i] = m;
3829 }
3830 KASSERT(OFF_TO_IDX(foff) == m->pindex,
3831 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3832 (intmax_t)foff, (uintmax_t)m->pindex));
3833
3834 /*
3835 * In the write case, the valid and clean bits are
3836 * already changed correctly ( see bdwrite() ), so we
3837 * only need to do this here in the read case.
3838 */
3839 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3840 KASSERT((m->dirty & vm_page_bits(foff &
3841 PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3842 " page %p has unexpected dirty bits", m));
3843 vfs_page_set_valid(bp, foff, m);
3844 }
3845
3846 vm_page_sunbusy(m);
3847 vm_object_pip_subtract(obj, 1);
3848 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3849 iosize -= resid;
3850 }
3851 vm_object_pip_wakeupn(obj, 0);
3852 VM_OBJECT_WUNLOCK(obj);
3853 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3854 BUF_CHECK_MAPPED(bp);
3855 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3856 bp->b_pages, bp->b_npages);
3857 }
3858 }
3859
3860 /*
3861 * For asynchronous completions, release the buffer now. The brelse
3862 * will do a wakeup there if necessary - so no need to do a wakeup
3863 * here in the async case. The sync case always needs to do a wakeup.
3864 */
3865
3866 if (bp->b_flags & B_ASYNC) {
3867 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3868 brelse(bp);
3869 else
3870 bqrelse(bp);
3871 } else
3872 bdone(bp);
3873 }
3874
3875 /*
3876 * This routine is called in lieu of iodone in the case of
3877 * incomplete I/O. This keeps the busy status for pages
3878 * consistant.
3879 */
3880 void
3881 vfs_unbusy_pages(struct buf *bp)
3882 {
3883 int i;
3884 vm_object_t obj;
3885 vm_page_t m;
3886
3887 runningbufwakeup(bp);
3888 if (!(bp->b_flags & B_VMIO))
3889 return;
3890
3891 obj = bp->b_bufobj->bo_object;
3892 VM_OBJECT_WLOCK(obj);
3893 for (i = 0; i < bp->b_npages; i++) {
3894 m = bp->b_pages[i];
3895 if (m == bogus_page) {
3896 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3897 if (!m)
3898 panic("vfs_unbusy_pages: page missing\n");
3899 bp->b_pages[i] = m;
3900 if ((bp->b_flags & B_UNMAPPED) == 0) {
3901 BUF_CHECK_MAPPED(bp);
3902 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3903 bp->b_pages, bp->b_npages);
3904 } else
3905 BUF_CHECK_UNMAPPED(bp);
3906 }
3907 vm_object_pip_subtract(obj, 1);
3908 vm_page_sunbusy(m);
3909 }
3910 vm_object_pip_wakeupn(obj, 0);
3911 VM_OBJECT_WUNLOCK(obj);
3912 }
3913
3914 /*
3915 * vfs_page_set_valid:
3916 *
3917 * Set the valid bits in a page based on the supplied offset. The
3918 * range is restricted to the buffer's size.
3919 *
3920 * This routine is typically called after a read completes.
3921 */
3922 static void
3923 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3924 {
3925 vm_ooffset_t eoff;
3926
3927 /*
3928 * Compute the end offset, eoff, such that [off, eoff) does not span a
3929 * page boundary and eoff is not greater than the end of the buffer.
3930 * The end of the buffer, in this case, is our file EOF, not the
3931 * allocation size of the buffer.
3932 */
3933 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3934 if (eoff > bp->b_offset + bp->b_bcount)
3935 eoff = bp->b_offset + bp->b_bcount;
3936
3937 /*
3938 * Set valid range. This is typically the entire buffer and thus the
3939 * entire page.
3940 */
3941 if (eoff > off)
3942 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3943 }
3944
3945 /*
3946 * vfs_page_set_validclean:
3947 *
3948 * Set the valid bits and clear the dirty bits in a page based on the
3949 * supplied offset. The range is restricted to the buffer's size.
3950 */
3951 static void
3952 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3953 {
3954 vm_ooffset_t soff, eoff;
3955
3956 /*
3957 * Start and end offsets in buffer. eoff - soff may not cross a
3958 * page boundry or cross the end of the buffer. The end of the
3959 * buffer, in this case, is our file EOF, not the allocation size
3960 * of the buffer.
3961 */
3962 soff = off;
3963 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3964 if (eoff > bp->b_offset + bp->b_bcount)
3965 eoff = bp->b_offset + bp->b_bcount;
3966
3967 /*
3968 * Set valid range. This is typically the entire buffer and thus the
3969 * entire page.
3970 */
3971 if (eoff > soff) {
3972 vm_page_set_validclean(
3973 m,
3974 (vm_offset_t) (soff & PAGE_MASK),
3975 (vm_offset_t) (eoff - soff)
3976 );
3977 }
3978 }
3979
3980 /*
3981 * Ensure that all buffer pages are not exclusive busied. If any page is
3982 * exclusive busy, drain it.
3983 */
3984 void
3985 vfs_drain_busy_pages(struct buf *bp)
3986 {
3987 vm_page_t m;
3988 int i, last_busied;
3989
3990 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
3991 last_busied = 0;
3992 for (i = 0; i < bp->b_npages; i++) {
3993 m = bp->b_pages[i];
3994 if (vm_page_xbusied(m)) {
3995 for (; last_busied < i; last_busied++)
3996 vm_page_sbusy(bp->b_pages[last_busied]);
3997 while (vm_page_xbusied(m)) {
3998 vm_page_lock(m);
3999 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4000 vm_page_busy_sleep(m, "vbpage");
4001 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4002 }
4003 }
4004 }
4005 for (i = 0; i < last_busied; i++)
4006 vm_page_sunbusy(bp->b_pages[i]);
4007 }
4008
4009 /*
4010 * This routine is called before a device strategy routine.
4011 * It is used to tell the VM system that paging I/O is in
4012 * progress, and treat the pages associated with the buffer
4013 * almost as being exclusive busy. Also the object paging_in_progress
4014 * flag is handled to make sure that the object doesn't become
4015 * inconsistant.
4016 *
4017 * Since I/O has not been initiated yet, certain buffer flags
4018 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4019 * and should be ignored.
4020 */
4021 void
4022 vfs_busy_pages(struct buf *bp, int clear_modify)
4023 {
4024 int i, bogus;
4025 vm_object_t obj;
4026 vm_ooffset_t foff;
4027 vm_page_t m;
4028
4029 if (!(bp->b_flags & B_VMIO))
4030 return;
4031
4032 obj = bp->b_bufobj->bo_object;
4033 foff = bp->b_offset;
4034 KASSERT(bp->b_offset != NOOFFSET,
4035 ("vfs_busy_pages: no buffer offset"));
4036 VM_OBJECT_WLOCK(obj);
4037 vfs_drain_busy_pages(bp);
4038 if (bp->b_bufsize != 0)
4039 vfs_setdirty_locked_object(bp);
4040 bogus = 0;
4041 for (i = 0; i < bp->b_npages; i++) {
4042 m = bp->b_pages[i];
4043
4044 if ((bp->b_flags & B_CLUSTER) == 0) {
4045 vm_object_pip_add(obj, 1);
4046 vm_page_sbusy(m);
4047 }
4048 /*
4049 * When readying a buffer for a read ( i.e
4050 * clear_modify == 0 ), it is important to do
4051 * bogus_page replacement for valid pages in
4052 * partially instantiated buffers. Partially
4053 * instantiated buffers can, in turn, occur when
4054 * reconstituting a buffer from its VM backing store
4055 * base. We only have to do this if B_CACHE is
4056 * clear ( which causes the I/O to occur in the
4057 * first place ). The replacement prevents the read
4058 * I/O from overwriting potentially dirty VM-backed
4059 * pages. XXX bogus page replacement is, uh, bogus.
4060 * It may not work properly with small-block devices.
4061 * We need to find a better way.
4062 */
4063 if (clear_modify) {
4064 pmap_remove_write(m);
4065 vfs_page_set_validclean(bp, foff, m);
4066 } else if (m->valid == VM_PAGE_BITS_ALL &&
4067 (bp->b_flags & B_CACHE) == 0) {
4068 bp->b_pages[i] = bogus_page;
4069 bogus++;
4070 }
4071 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4072 }
4073 VM_OBJECT_WUNLOCK(obj);
4074 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4075 BUF_CHECK_MAPPED(bp);
4076 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4077 bp->b_pages, bp->b_npages);
4078 }
4079 }
4080
4081 /*
4082 * vfs_bio_set_valid:
4083 *
4084 * Set the range within the buffer to valid. The range is
4085 * relative to the beginning of the buffer, b_offset. Note that
4086 * b_offset itself may be offset from the beginning of the first
4087 * page.
4088 */
4089 void
4090 vfs_bio_set_valid(struct buf *bp, int base, int size)
4091 {
4092 int i, n;
4093 vm_page_t m;
4094
4095 if (!(bp->b_flags & B_VMIO))
4096 return;
4097
4098 /*
4099 * Fixup base to be relative to beginning of first page.
4100 * Set initial n to be the maximum number of bytes in the
4101 * first page that can be validated.
4102 */
4103 base += (bp->b_offset & PAGE_MASK);
4104 n = PAGE_SIZE - (base & PAGE_MASK);
4105
4106 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4107 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4108 m = bp->b_pages[i];
4109 if (n > size)
4110 n = size;
4111 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4112 base += n;
4113 size -= n;
4114 n = PAGE_SIZE;
4115 }
4116 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4117 }
4118
4119 /*
4120 * vfs_bio_clrbuf:
4121 *
4122 * If the specified buffer is a non-VMIO buffer, clear the entire
4123 * buffer. If the specified buffer is a VMIO buffer, clear and
4124 * validate only the previously invalid portions of the buffer.
4125 * This routine essentially fakes an I/O, so we need to clear
4126 * BIO_ERROR and B_INVAL.
4127 *
4128 * Note that while we only theoretically need to clear through b_bcount,
4129 * we go ahead and clear through b_bufsize.
4130 */
4131 void
4132 vfs_bio_clrbuf(struct buf *bp)
4133 {
4134 int i, j, mask, sa, ea, slide;
4135
4136 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4137 clrbuf(bp);
4138 return;
4139 }
4140 bp->b_flags &= ~B_INVAL;
4141 bp->b_ioflags &= ~BIO_ERROR;
4142 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4143 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4144 (bp->b_offset & PAGE_MASK) == 0) {
4145 if (bp->b_pages[0] == bogus_page)
4146 goto unlock;
4147 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4148 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4149 if ((bp->b_pages[0]->valid & mask) == mask)
4150 goto unlock;
4151 if ((bp->b_pages[0]->valid & mask) == 0) {
4152 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4153 bp->b_pages[0]->valid |= mask;
4154 goto unlock;
4155 }
4156 }
4157 sa = bp->b_offset & PAGE_MASK;
4158 slide = 0;
4159 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4160 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4161 ea = slide & PAGE_MASK;
4162 if (ea == 0)
4163 ea = PAGE_SIZE;
4164 if (bp->b_pages[i] == bogus_page)
4165 continue;
4166 j = sa / DEV_BSIZE;
4167 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4168 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4169 if ((bp->b_pages[i]->valid & mask) == mask)
4170 continue;
4171 if ((bp->b_pages[i]->valid & mask) == 0)
4172 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4173 else {
4174 for (; sa < ea; sa += DEV_BSIZE, j++) {
4175 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4176 pmap_zero_page_area(bp->b_pages[i],
4177 sa, DEV_BSIZE);
4178 }
4179 }
4180 }
4181 bp->b_pages[i]->valid |= mask;
4182 }
4183 unlock:
4184 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4185 bp->b_resid = 0;
4186 }
4187
4188 void
4189 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4190 {
4191 vm_page_t m;
4192 int i, n;
4193
4194 if ((bp->b_flags & B_UNMAPPED) == 0) {
4195 BUF_CHECK_MAPPED(bp);
4196 bzero(bp->b_data + base, size);
4197 } else {
4198 BUF_CHECK_UNMAPPED(bp);
4199 n = PAGE_SIZE - (base & PAGE_MASK);
4200 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4201 m = bp->b_pages[i];
4202 if (n > size)
4203 n = size;
4204 pmap_zero_page_area(m, base & PAGE_MASK, n);
4205 base += n;
4206 size -= n;
4207 n = PAGE_SIZE;
4208 }
4209 }
4210 }
4211
4212 /*
4213 * vm_hold_load_pages and vm_hold_free_pages get pages into
4214 * a buffers address space. The pages are anonymous and are
4215 * not associated with a file object.
4216 */
4217 static void
4218 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4219 {
4220 vm_offset_t pg;
4221 vm_page_t p;
4222 int index;
4223
4224 BUF_CHECK_MAPPED(bp);
4225
4226 to = round_page(to);
4227 from = round_page(from);
4228 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4229
4230 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4231 tryagain:
4232 /*
4233 * note: must allocate system pages since blocking here
4234 * could interfere with paging I/O, no matter which
4235 * process we are.
4236 */
4237 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4238 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4239 if (p == NULL) {
4240 VM_WAIT;
4241 goto tryagain;
4242 }
4243 pmap_qenter(pg, &p, 1);
4244 bp->b_pages[index] = p;
4245 }
4246 bp->b_npages = index;
4247 }
4248
4249 /* Return pages associated with this buf to the vm system */
4250 static void
4251 vm_hold_free_pages(struct buf *bp, int newbsize)
4252 {
4253 vm_offset_t from;
4254 vm_page_t p;
4255 int index, newnpages;
4256
4257 BUF_CHECK_MAPPED(bp);
4258
4259 from = round_page((vm_offset_t)bp->b_data + newbsize);
4260 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4261 if (bp->b_npages > newnpages)
4262 pmap_qremove(from, bp->b_npages - newnpages);
4263 for (index = newnpages; index < bp->b_npages; index++) {
4264 p = bp->b_pages[index];
4265 bp->b_pages[index] = NULL;
4266 if (vm_page_sbusied(p))
4267 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4268 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4269 p->wire_count--;
4270 vm_page_free(p);
4271 atomic_subtract_int(&cnt.v_wire_count, 1);
4272 }
4273 bp->b_npages = newnpages;
4274 }
4275
4276 /*
4277 * Map an IO request into kernel virtual address space.
4278 *
4279 * All requests are (re)mapped into kernel VA space.
4280 * Notice that we use b_bufsize for the size of the buffer
4281 * to be mapped. b_bcount might be modified by the driver.
4282 *
4283 * Note that even if the caller determines that the address space should
4284 * be valid, a race or a smaller-file mapped into a larger space may
4285 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4286 * check the return value.
4287 */
4288 int
4289 vmapbuf(struct buf *bp, int mapbuf)
4290 {
4291 caddr_t kva;
4292 vm_prot_t prot;
4293 int pidx;
4294
4295 if (bp->b_bufsize < 0)
4296 return (-1);
4297 prot = VM_PROT_READ;
4298 if (bp->b_iocmd == BIO_READ)
4299 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4300 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4301 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4302 btoc(MAXPHYS))) < 0)
4303 return (-1);
4304 bp->b_npages = pidx;
4305 if (mapbuf || !unmapped_buf_allowed) {
4306 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4307 kva = bp->b_saveaddr;
4308 bp->b_saveaddr = bp->b_data;
4309 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4310 bp->b_flags &= ~B_UNMAPPED;
4311 } else {
4312 bp->b_flags |= B_UNMAPPED;
4313 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4314 bp->b_saveaddr = bp->b_data;
4315 bp->b_data = unmapped_buf;
4316 }
4317 return(0);
4318 }
4319
4320 /*
4321 * Free the io map PTEs associated with this IO operation.
4322 * We also invalidate the TLB entries and restore the original b_addr.
4323 */
4324 void
4325 vunmapbuf(struct buf *bp)
4326 {
4327 int npages;
4328
4329 npages = bp->b_npages;
4330 if (bp->b_flags & B_UNMAPPED)
4331 bp->b_flags &= ~B_UNMAPPED;
4332 else
4333 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4334 vm_page_unhold_pages(bp->b_pages, npages);
4335
4336 bp->b_data = bp->b_saveaddr;
4337 }
4338
4339 void
4340 bdone(struct buf *bp)
4341 {
4342 struct mtx *mtxp;
4343
4344 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4345 mtx_lock(mtxp);
4346 bp->b_flags |= B_DONE;
4347 wakeup(bp);
4348 mtx_unlock(mtxp);
4349 }
4350
4351 void
4352 bwait(struct buf *bp, u_char pri, const char *wchan)
4353 {
4354 struct mtx *mtxp;
4355
4356 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4357 mtx_lock(mtxp);
4358 while ((bp->b_flags & B_DONE) == 0)
4359 msleep(bp, mtxp, pri, wchan, 0);
4360 mtx_unlock(mtxp);
4361 }
4362
4363 int
4364 bufsync(struct bufobj *bo, int waitfor)
4365 {
4366
4367 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4368 }
4369
4370 void
4371 bufstrategy(struct bufobj *bo, struct buf *bp)
4372 {
4373 int i = 0;
4374 struct vnode *vp;
4375
4376 vp = bp->b_vp;
4377 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4378 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4379 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4380 i = VOP_STRATEGY(vp, bp);
4381 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4382 }
4383
4384 void
4385 bufobj_wrefl(struct bufobj *bo)
4386 {
4387
4388 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4389 ASSERT_BO_WLOCKED(bo);
4390 bo->bo_numoutput++;
4391 }
4392
4393 void
4394 bufobj_wref(struct bufobj *bo)
4395 {
4396
4397 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4398 BO_LOCK(bo);
4399 bo->bo_numoutput++;
4400 BO_UNLOCK(bo);
4401 }
4402
4403 void
4404 bufobj_wdrop(struct bufobj *bo)
4405 {
4406
4407 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4408 BO_LOCK(bo);
4409 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4410 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4411 bo->bo_flag &= ~BO_WWAIT;
4412 wakeup(&bo->bo_numoutput);
4413 }
4414 BO_UNLOCK(bo);
4415 }
4416
4417 int
4418 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4419 {
4420 int error;
4421
4422 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4423 ASSERT_BO_WLOCKED(bo);
4424 error = 0;
4425 while (bo->bo_numoutput) {
4426 bo->bo_flag |= BO_WWAIT;
4427 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4428 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4429 if (error)
4430 break;
4431 }
4432 return (error);
4433 }
4434
4435 void
4436 bpin(struct buf *bp)
4437 {
4438 struct mtx *mtxp;
4439
4440 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4441 mtx_lock(mtxp);
4442 bp->b_pin_count++;
4443 mtx_unlock(mtxp);
4444 }
4445
4446 void
4447 bunpin(struct buf *bp)
4448 {
4449 struct mtx *mtxp;
4450
4451 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4452 mtx_lock(mtxp);
4453 if (--bp->b_pin_count == 0)
4454 wakeup(bp);
4455 mtx_unlock(mtxp);
4456 }
4457
4458 void
4459 bunpin_wait(struct buf *bp)
4460 {
4461 struct mtx *mtxp;
4462
4463 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4464 mtx_lock(mtxp);
4465 while (bp->b_pin_count > 0)
4466 msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4467 mtx_unlock(mtxp);
4468 }
4469
4470 /*
4471 * Set bio_data or bio_ma for struct bio from the struct buf.
4472 */
4473 void
4474 bdata2bio(struct buf *bp, struct bio *bip)
4475 {
4476
4477 if ((bp->b_flags & B_UNMAPPED) != 0) {
4478 KASSERT(unmapped_buf_allowed, ("unmapped"));
4479 bip->bio_ma = bp->b_pages;
4480 bip->bio_ma_n = bp->b_npages;
4481 bip->bio_data = unmapped_buf;
4482 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4483 bip->bio_flags |= BIO_UNMAPPED;
4484 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4485 PAGE_SIZE == bp->b_npages,
4486 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4487 (long long)bip->bio_length, bip->bio_ma_n));
4488 } else {
4489 bip->bio_data = bp->b_data;
4490 bip->bio_ma = NULL;
4491 }
4492 }
4493
4494 #include "opt_ddb.h"
4495 #ifdef DDB
4496 #include <ddb/ddb.h>
4497
4498 /* DDB command to show buffer data */
4499 DB_SHOW_COMMAND(buffer, db_show_buffer)
4500 {
4501 /* get args */
4502 struct buf *bp = (struct buf *)addr;
4503
4504 if (!have_addr) {
4505 db_printf("usage: show buffer <addr>\n");
4506 return;
4507 }
4508
4509 db_printf("buf at %p\n", bp);
4510 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4511 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4512 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4513 db_printf(
4514 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4515 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4516 "b_dep = %p\n",
4517 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4518 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4519 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4520 if (bp->b_npages) {
4521 int i;
4522 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4523 for (i = 0; i < bp->b_npages; i++) {
4524 vm_page_t m;
4525 m = bp->b_pages[i];
4526 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4527 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4528 if ((i + 1) < bp->b_npages)
4529 db_printf(",");
4530 }
4531 db_printf("\n");
4532 }
4533 db_printf(" ");
4534 BUF_LOCKPRINTINFO(bp);
4535 }
4536
4537 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4538 {
4539 struct buf *bp;
4540 int i;
4541
4542 for (i = 0; i < nbuf; i++) {
4543 bp = &buf[i];
4544 if (BUF_ISLOCKED(bp)) {
4545 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4546 db_printf("\n");
4547 }
4548 }
4549 }
4550
4551 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4552 {
4553 struct vnode *vp;
4554 struct buf *bp;
4555
4556 if (!have_addr) {
4557 db_printf("usage: show vnodebufs <addr>\n");
4558 return;
4559 }
4560 vp = (struct vnode *)addr;
4561 db_printf("Clean buffers:\n");
4562 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4563 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4564 db_printf("\n");
4565 }
4566 db_printf("Dirty buffers:\n");
4567 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4568 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4569 db_printf("\n");
4570 }
4571 }
4572
4573 DB_COMMAND(countfreebufs, db_coundfreebufs)
4574 {
4575 struct buf *bp;
4576 int i, used = 0, nfree = 0;
4577
4578 if (have_addr) {
4579 db_printf("usage: countfreebufs\n");
4580 return;
4581 }
4582
4583 for (i = 0; i < nbuf; i++) {
4584 bp = &buf[i];
4585 if ((bp->b_flags & B_INFREECNT) != 0)
4586 nfree++;
4587 else
4588 used++;
4589 }
4590
4591 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4592 nfree + used);
4593 db_printf("numfreebuffers is %d\n", numfreebuffers);
4594 }
4595 #endif /* DDB */
Cache object: 80b925adee632018e1c7d1358b70936f
|