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 * All rights reserved.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * SUCH DAMAGE.
26 */
27
28 /*
29 * this file contains a new buffer I/O scheme implementing a coherent
30 * VM object and buffer cache scheme. Pains have been taken to make
31 * sure that the performance degradation associated with schemes such
32 * as this is not realized.
33 *
34 * Author: John S. Dyson
35 * Significant help during the development and debugging phases
36 * had been provided by David Greenman, also of the FreeBSD core team.
37 *
38 * see man buf(9) for more info.
39 */
40
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bio.h>
47 #include <sys/conf.h>
48 #include <sys/buf.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/malloc.h>
54 #include <sys/mount.h>
55 #include <sys/mutex.h>
56 #include <sys/kernel.h>
57 #include <sys/kthread.h>
58 #include <sys/proc.h>
59 #include <sys/resourcevar.h>
60 #include <sys/sysctl.h>
61 #include <sys/vmmeter.h>
62 #include <sys/vnode.h>
63 #include <geom/geom.h>
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_pageout.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_map.h>
72 #include "opt_directio.h"
73 #include "opt_swap.h"
74
75 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
76
77 struct bio_ops bioops; /* I/O operation notification */
78
79 struct buf_ops buf_ops_bio = {
80 .bop_name = "buf_ops_bio",
81 .bop_write = bufwrite,
82 .bop_strategy = bufstrategy,
83 .bop_sync = bufsync,
84 .bop_bdflush = bufbdflush,
85 };
86
87 /*
88 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
89 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
90 */
91 struct buf *buf; /* buffer header pool */
92
93 static struct proc *bufdaemonproc;
94
95 static int inmem(struct vnode *vp, daddr_t blkno);
96 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from,
97 vm_offset_t to);
98 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
99 vm_offset_t to);
100 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
101 int pageno, vm_page_t m);
102 static void vfs_clean_pages(struct buf *bp);
103 static void vfs_setdirty(struct buf *bp);
104 static void vfs_setdirty_locked_object(struct buf *bp);
105 static void vfs_vmio_release(struct buf *bp);
106 static int vfs_bio_clcheck(struct vnode *vp, int size,
107 daddr_t lblkno, daddr_t blkno);
108 static int flushbufqueues(int, int);
109 static void buf_daemon(void);
110 static void bremfreel(struct buf *bp);
111
112 int vmiodirenable = TRUE;
113 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
114 "Use the VM system for directory writes");
115 int runningbufspace;
116 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
117 "Amount of presently outstanding async buffer io");
118 static int bufspace;
119 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
120 "KVA memory used for bufs");
121 static int maxbufspace;
122 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
123 "Maximum allowed value of bufspace (including buf_daemon)");
124 static int bufmallocspace;
125 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
126 "Amount of malloced memory for buffers");
127 static int maxbufmallocspace;
128 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
129 "Maximum amount of malloced memory for buffers");
130 static int lobufspace;
131 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
132 "Minimum amount of buffers we want to have");
133 int hibufspace;
134 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
135 "Maximum allowed value of bufspace (excluding buf_daemon)");
136 static int bufreusecnt;
137 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
138 "Number of times we have reused a buffer");
139 static int buffreekvacnt;
140 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
141 "Number of times we have freed the KVA space from some buffer");
142 static int bufdefragcnt;
143 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
144 "Number of times we have had to repeat buffer allocation to defragment");
145 static int lorunningspace;
146 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
147 "Minimum preferred space used for in-progress I/O");
148 static int hirunningspace;
149 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
150 "Maximum amount of space to use for in-progress I/O");
151 int dirtybufferflushes;
152 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
153 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
154 int bdwriteskip;
155 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
156 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
157 int altbufferflushes;
158 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
159 0, "Number of fsync flushes to limit dirty buffers");
160 static int recursiveflushes;
161 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
162 0, "Number of flushes skipped due to being recursive");
163 static int numdirtybuffers;
164 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
165 "Number of buffers that are dirty (has unwritten changes) at the moment");
166 static int lodirtybuffers;
167 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
168 "How many buffers we want to have free before bufdaemon can sleep");
169 static int hidirtybuffers;
170 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
171 "When the number of dirty buffers is considered severe");
172 int dirtybufthresh;
173 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
174 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
175 static int numfreebuffers;
176 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
177 "Number of free buffers");
178 static int lofreebuffers;
179 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
180 "XXX Unused");
181 static int hifreebuffers;
182 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
183 "XXX Complicatedly unused");
184 static int getnewbufcalls;
185 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
186 "Number of calls to getnewbuf");
187 static int getnewbufrestarts;
188 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
189 "Number of times getnewbuf has had to restart a buffer aquisition");
190
191 /*
192 * Wakeup point for bufdaemon, as well as indicator of whether it is already
193 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
194 * is idling.
195 */
196 static int bd_request;
197
198 /*
199 * This lock synchronizes access to bd_request.
200 */
201 static struct mtx bdlock;
202
203 /*
204 * bogus page -- for I/O to/from partially complete buffers
205 * this is a temporary solution to the problem, but it is not
206 * really that bad. it would be better to split the buffer
207 * for input in the case of buffers partially already in memory,
208 * but the code is intricate enough already.
209 */
210 vm_page_t bogus_page;
211
212 /*
213 * Synchronization (sleep/wakeup) variable for active buffer space requests.
214 * Set when wait starts, cleared prior to wakeup().
215 * Used in runningbufwakeup() and waitrunningbufspace().
216 */
217 static int runningbufreq;
218
219 /*
220 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
221 * waitrunningbufspace().
222 */
223 static struct mtx rbreqlock;
224
225 /*
226 * Synchronization (sleep/wakeup) variable for buffer requests.
227 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
228 * by and/or.
229 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
230 * getnewbuf(), and getblk().
231 */
232 static int needsbuffer;
233
234 /*
235 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
236 */
237 static struct mtx nblock;
238
239 /*
240 * Lock that protects against bwait()/bdone()/B_DONE races.
241 */
242
243 static struct mtx bdonelock;
244
245 /*
246 * Lock that protects against bwait()/bdone()/B_DONE races.
247 */
248 static struct mtx bpinlock;
249
250 /*
251 * Definitions for the buffer free lists.
252 */
253 #define BUFFER_QUEUES 6 /* number of free buffer queues */
254
255 #define QUEUE_NONE 0 /* on no queue */
256 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
257 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
258 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */
259 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */
260 #define QUEUE_EMPTY 5 /* empty buffer headers */
261
262 /* Queues for free buffers with various properties */
263 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
264
265 /* Lock for the bufqueues */
266 static struct mtx bqlock;
267
268 /*
269 * Single global constant for BUF_WMESG, to avoid getting multiple references.
270 * buf_wmesg is referred from macros.
271 */
272 const char *buf_wmesg = BUF_WMESG;
273
274 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
275 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
276 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
277 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
278
279 #ifdef DIRECTIO
280 extern void ffs_rawread_setup(void);
281 #endif /* DIRECTIO */
282 /*
283 * numdirtywakeup:
284 *
285 * If someone is blocked due to there being too many dirty buffers,
286 * and numdirtybuffers is now reasonable, wake them up.
287 */
288
289 static __inline void
290 numdirtywakeup(int level)
291 {
292
293 if (numdirtybuffers <= level) {
294 mtx_lock(&nblock);
295 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
296 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
297 wakeup(&needsbuffer);
298 }
299 mtx_unlock(&nblock);
300 }
301 }
302
303 /*
304 * bufspacewakeup:
305 *
306 * Called when buffer space is potentially available for recovery.
307 * getnewbuf() will block on this flag when it is unable to free
308 * sufficient buffer space. Buffer space becomes recoverable when
309 * bp's get placed back in the queues.
310 */
311
312 static __inline void
313 bufspacewakeup(void)
314 {
315
316 /*
317 * If someone is waiting for BUF space, wake them up. Even
318 * though we haven't freed the kva space yet, the waiting
319 * process will be able to now.
320 */
321 mtx_lock(&nblock);
322 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
323 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
324 wakeup(&needsbuffer);
325 }
326 mtx_unlock(&nblock);
327 }
328
329 /*
330 * runningbufwakeup() - in-progress I/O accounting.
331 *
332 */
333 void
334 runningbufwakeup(struct buf *bp)
335 {
336
337 if (bp->b_runningbufspace) {
338 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
339 bp->b_runningbufspace = 0;
340 mtx_lock(&rbreqlock);
341 if (runningbufreq && runningbufspace <= lorunningspace) {
342 runningbufreq = 0;
343 wakeup(&runningbufreq);
344 }
345 mtx_unlock(&rbreqlock);
346 }
347 }
348
349 /*
350 * bufcountwakeup:
351 *
352 * Called when a buffer has been added to one of the free queues to
353 * account for the buffer and to wakeup anyone waiting for free buffers.
354 * This typically occurs when large amounts of metadata are being handled
355 * by the buffer cache ( else buffer space runs out first, usually ).
356 */
357
358 static __inline void
359 bufcountwakeup(void)
360 {
361
362 atomic_add_int(&numfreebuffers, 1);
363 mtx_lock(&nblock);
364 if (needsbuffer) {
365 needsbuffer &= ~VFS_BIO_NEED_ANY;
366 if (numfreebuffers >= hifreebuffers)
367 needsbuffer &= ~VFS_BIO_NEED_FREE;
368 wakeup(&needsbuffer);
369 }
370 mtx_unlock(&nblock);
371 }
372
373 /*
374 * waitrunningbufspace()
375 *
376 * runningbufspace is a measure of the amount of I/O currently
377 * running. This routine is used in async-write situations to
378 * prevent creating huge backups of pending writes to a device.
379 * Only asynchronous writes are governed by this function.
380 *
381 * Reads will adjust runningbufspace, but will not block based on it.
382 * The read load has a side effect of reducing the allowed write load.
383 *
384 * This does NOT turn an async write into a sync write. It waits
385 * for earlier writes to complete and generally returns before the
386 * caller's write has reached the device.
387 */
388 void
389 waitrunningbufspace(void)
390 {
391
392 mtx_lock(&rbreqlock);
393 while (runningbufspace > hirunningspace) {
394 ++runningbufreq;
395 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
396 }
397 mtx_unlock(&rbreqlock);
398 }
399
400
401 /*
402 * vfs_buf_test_cache:
403 *
404 * Called when a buffer is extended. This function clears the B_CACHE
405 * bit if the newly extended portion of the buffer does not contain
406 * valid data.
407 */
408 static __inline
409 void
410 vfs_buf_test_cache(struct buf *bp,
411 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
412 vm_page_t m)
413 {
414
415 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
416 if (bp->b_flags & B_CACHE) {
417 int base = (foff + off) & PAGE_MASK;
418 if (vm_page_is_valid(m, base, size) == 0)
419 bp->b_flags &= ~B_CACHE;
420 }
421 }
422
423 /* Wake up the buffer daemon if necessary */
424 static __inline
425 void
426 bd_wakeup(int dirtybuflevel)
427 {
428
429 mtx_lock(&bdlock);
430 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
431 bd_request = 1;
432 wakeup(&bd_request);
433 }
434 mtx_unlock(&bdlock);
435 }
436
437 /*
438 * bd_speedup - speedup the buffer cache flushing code
439 */
440
441 static __inline
442 void
443 bd_speedup(void)
444 {
445
446 bd_wakeup(1);
447 }
448
449 /*
450 * Calculating buffer cache scaling values and reserve space for buffer
451 * headers. This is called during low level kernel initialization and
452 * may be called more then once. We CANNOT write to the memory area
453 * being reserved at this time.
454 */
455 caddr_t
456 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
457 {
458 int maxbuf;
459
460 /*
461 * physmem_est is in pages. Convert it to kilobytes (assumes
462 * PAGE_SIZE is >= 1K)
463 */
464 physmem_est = physmem_est * (PAGE_SIZE / 1024);
465
466 /*
467 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
468 * For the first 64MB of ram nominally allocate sufficient buffers to
469 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
470 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
471 * the buffer cache we limit the eventual kva reservation to
472 * maxbcache bytes.
473 *
474 * factor represents the 1/4 x ram conversion.
475 */
476 if (nbuf == 0) {
477 int factor = 4 * BKVASIZE / 1024;
478
479 nbuf = 50;
480 if (physmem_est > 4096)
481 nbuf += min((physmem_est - 4096) / factor,
482 65536 / factor);
483 if (physmem_est > 65536)
484 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
485
486 if (maxbcache && nbuf > maxbcache / BKVASIZE)
487 nbuf = maxbcache / BKVASIZE;
488
489 /* XXX Avoid integer overflows later on with maxbufspace. */
490 maxbuf = (INT_MAX / 3) / BKVASIZE;
491 if (nbuf > maxbuf)
492 nbuf = maxbuf;
493 }
494
495 #if 0
496 /*
497 * Do not allow the buffer_map to be more then 1/2 the size of the
498 * kernel_map.
499 */
500 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
501 (BKVASIZE * 2)) {
502 nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
503 (BKVASIZE * 2);
504 printf("Warning: nbufs capped at %d\n", nbuf);
505 }
506 #endif
507
508 /*
509 * swbufs are used as temporary holders for I/O, such as paging I/O.
510 * We have no less then 16 and no more then 256.
511 */
512 nswbuf = max(min(nbuf/4, 256), 16);
513 #ifdef NSWBUF_MIN
514 if (nswbuf < NSWBUF_MIN)
515 nswbuf = NSWBUF_MIN;
516 #endif
517 #ifdef DIRECTIO
518 ffs_rawread_setup();
519 #endif
520
521 /*
522 * Reserve space for the buffer cache buffers
523 */
524 swbuf = (void *)v;
525 v = (caddr_t)(swbuf + nswbuf);
526 buf = (void *)v;
527 v = (caddr_t)(buf + nbuf);
528
529 return(v);
530 }
531
532 /* Initialize the buffer subsystem. Called before use of any buffers. */
533 void
534 bufinit(void)
535 {
536 struct buf *bp;
537 int i;
538
539 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
540 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
541 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
542 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
543 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
544 mtx_init(&bpinlock, "bpin lock", NULL, MTX_DEF);
545
546 /* next, make a null set of free lists */
547 for (i = 0; i < BUFFER_QUEUES; i++)
548 TAILQ_INIT(&bufqueues[i]);
549
550 /* finally, initialize each buffer header and stick on empty q */
551 for (i = 0; i < nbuf; i++) {
552 bp = &buf[i];
553 bzero(bp, sizeof *bp);
554 bp->b_flags = B_INVAL; /* we're just an empty header */
555 bp->b_rcred = NOCRED;
556 bp->b_wcred = NOCRED;
557 bp->b_qindex = QUEUE_EMPTY;
558 bp->b_vflags = 0;
559 bp->b_xflags = 0;
560 LIST_INIT(&bp->b_dep);
561 BUF_LOCKINIT(bp);
562 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
563 }
564
565 /*
566 * maxbufspace is the absolute maximum amount of buffer space we are
567 * allowed to reserve in KVM and in real terms. The absolute maximum
568 * is nominally used by buf_daemon. hibufspace is the nominal maximum
569 * used by most other processes. The differential is required to
570 * ensure that buf_daemon is able to run when other processes might
571 * be blocked waiting for buffer space.
572 *
573 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
574 * this may result in KVM fragmentation which is not handled optimally
575 * by the system.
576 */
577 maxbufspace = nbuf * BKVASIZE;
578 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
579 lobufspace = hibufspace - MAXBSIZE;
580
581 lorunningspace = 512 * 1024;
582 hirunningspace = 1024 * 1024;
583
584 /*
585 * Limit the amount of malloc memory since it is wired permanently into
586 * the kernel space. Even though this is accounted for in the buffer
587 * allocation, we don't want the malloced region to grow uncontrolled.
588 * The malloc scheme improves memory utilization significantly on average
589 * (small) directories.
590 */
591 maxbufmallocspace = hibufspace / 20;
592
593 /*
594 * Reduce the chance of a deadlock occuring by limiting the number
595 * of delayed-write dirty buffers we allow to stack up.
596 */
597 hidirtybuffers = nbuf / 4 + 20;
598 dirtybufthresh = hidirtybuffers * 9 / 10;
599 numdirtybuffers = 0;
600 /*
601 * To support extreme low-memory systems, make sure hidirtybuffers cannot
602 * eat up all available buffer space. This occurs when our minimum cannot
603 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
604 * BKVASIZE'd (8K) buffers.
605 */
606 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
607 hidirtybuffers >>= 1;
608 }
609 lodirtybuffers = hidirtybuffers / 2;
610
611 /*
612 * Try to keep the number of free buffers in the specified range,
613 * and give special processes (e.g. like buf_daemon) access to an
614 * emergency reserve.
615 */
616 lofreebuffers = nbuf / 18 + 5;
617 hifreebuffers = 2 * lofreebuffers;
618 numfreebuffers = nbuf;
619
620 /*
621 * Maximum number of async ops initiated per buf_daemon loop. This is
622 * somewhat of a hack at the moment, we really need to limit ourselves
623 * based on the number of bytes of I/O in-transit that were initiated
624 * from buf_daemon.
625 */
626
627 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
628 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
629 }
630
631 /*
632 * bfreekva() - free the kva allocation for a buffer.
633 *
634 * Since this call frees up buffer space, we call bufspacewakeup().
635 */
636 static void
637 bfreekva(struct buf *bp)
638 {
639
640 if (bp->b_kvasize) {
641 atomic_add_int(&buffreekvacnt, 1);
642 atomic_subtract_int(&bufspace, bp->b_kvasize);
643 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
644 (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
645 bp->b_kvasize = 0;
646 bufspacewakeup();
647 }
648 }
649
650 /*
651 * bremfree:
652 *
653 * Mark the buffer for removal from the appropriate free list in brelse.
654 *
655 */
656 void
657 bremfree(struct buf *bp)
658 {
659
660 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
661 KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked."));
662 KASSERT((bp->b_flags & B_REMFREE) == 0,
663 ("bremfree: buffer %p already marked for delayed removal.", bp));
664 KASSERT(bp->b_qindex != QUEUE_NONE,
665 ("bremfree: buffer %p not on a queue.", bp));
666
667 bp->b_flags |= B_REMFREE;
668 /* Fixup numfreebuffers count. */
669 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
670 atomic_subtract_int(&numfreebuffers, 1);
671 }
672
673 /*
674 * bremfreef:
675 *
676 * Force an immediate removal from a free list. Used only in nfs when
677 * it abuses the b_freelist pointer.
678 */
679 void
680 bremfreef(struct buf *bp)
681 {
682 mtx_lock(&bqlock);
683 bremfreel(bp);
684 mtx_unlock(&bqlock);
685 }
686
687 /*
688 * bremfreel:
689 *
690 * Removes a buffer from the free list, must be called with the
691 * bqlock held.
692 */
693 static void
694 bremfreel(struct buf *bp)
695 {
696 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
697 bp, bp->b_vp, bp->b_flags);
698 KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp));
699 KASSERT(bp->b_qindex != QUEUE_NONE,
700 ("bremfreel: buffer %p not on a queue.", bp));
701 mtx_assert(&bqlock, MA_OWNED);
702
703 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
704 bp->b_qindex = QUEUE_NONE;
705 /*
706 * If this was a delayed bremfree() we only need to remove the buffer
707 * from the queue and return the stats are already done.
708 */
709 if (bp->b_flags & B_REMFREE) {
710 bp->b_flags &= ~B_REMFREE;
711 return;
712 }
713 /*
714 * Fixup numfreebuffers count. If the buffer is invalid or not
715 * delayed-write, the buffer was free and we must decrement
716 * numfreebuffers.
717 */
718 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
719 atomic_subtract_int(&numfreebuffers, 1);
720 }
721
722
723 /*
724 * Get a buffer with the specified data. Look in the cache first. We
725 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
726 * is set, the buffer is valid and we do not have to do anything ( see
727 * getblk() ). This is really just a special case of breadn().
728 */
729 int
730 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
731 struct buf **bpp)
732 {
733
734 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
735 }
736
737 /*
738 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
739 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
740 * the buffer is valid and we do not have to do anything.
741 */
742 void
743 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
744 int cnt, struct ucred * cred)
745 {
746 struct buf *rabp;
747 int i;
748
749 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
750 if (inmem(vp, *rablkno))
751 continue;
752 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
753
754 if ((rabp->b_flags & B_CACHE) == 0) {
755 if (!TD_IS_IDLETHREAD(curthread))
756 curthread->td_ru.ru_inblock++;
757 rabp->b_flags |= B_ASYNC;
758 rabp->b_flags &= ~B_INVAL;
759 rabp->b_ioflags &= ~BIO_ERROR;
760 rabp->b_iocmd = BIO_READ;
761 if (rabp->b_rcred == NOCRED && cred != NOCRED)
762 rabp->b_rcred = crhold(cred);
763 vfs_busy_pages(rabp, 0);
764 BUF_KERNPROC(rabp);
765 rabp->b_iooffset = dbtob(rabp->b_blkno);
766 bstrategy(rabp);
767 } else {
768 brelse(rabp);
769 }
770 }
771 }
772
773 /*
774 * Operates like bread, but also starts asynchronous I/O on
775 * read-ahead blocks.
776 */
777 int
778 breadn(struct vnode * vp, daddr_t blkno, int size,
779 daddr_t * rablkno, int *rabsize,
780 int cnt, struct ucred * cred, struct buf **bpp)
781 {
782 struct buf *bp;
783 int rv = 0, readwait = 0;
784
785 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
786 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
787
788 /* if not found in cache, do some I/O */
789 if ((bp->b_flags & B_CACHE) == 0) {
790 if (!TD_IS_IDLETHREAD(curthread))
791 curthread->td_ru.ru_inblock++;
792 bp->b_iocmd = BIO_READ;
793 bp->b_flags &= ~B_INVAL;
794 bp->b_ioflags &= ~BIO_ERROR;
795 if (bp->b_rcred == NOCRED && cred != NOCRED)
796 bp->b_rcred = crhold(cred);
797 vfs_busy_pages(bp, 0);
798 bp->b_iooffset = dbtob(bp->b_blkno);
799 bstrategy(bp);
800 ++readwait;
801 }
802
803 breada(vp, rablkno, rabsize, cnt, cred);
804
805 if (readwait) {
806 rv = bufwait(bp);
807 }
808 return (rv);
809 }
810
811 /*
812 * Write, release buffer on completion. (Done by iodone
813 * if async). Do not bother writing anything if the buffer
814 * is invalid.
815 *
816 * Note that we set B_CACHE here, indicating that buffer is
817 * fully valid and thus cacheable. This is true even of NFS
818 * now so we set it generally. This could be set either here
819 * or in biodone() since the I/O is synchronous. We put it
820 * here.
821 */
822 int
823 bufwrite(struct buf *bp)
824 {
825 int oldflags;
826 struct vnode *vp;
827 int vp_md;
828
829 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
830 if (bp->b_flags & B_INVAL) {
831 brelse(bp);
832 return (0);
833 }
834
835 oldflags = bp->b_flags;
836
837 if (BUF_REFCNT(bp) == 0)
838 panic("bufwrite: buffer is not busy???");
839
840 if (bp->b_pin_count > 0)
841 bunpin_wait(bp);
842
843 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
844 ("FFS background buffer should not get here %p", bp));
845
846 vp = bp->b_vp;
847 if (vp)
848 vp_md = vp->v_vflag & VV_MD;
849 else
850 vp_md = 0;
851
852 /* Mark the buffer clean */
853 bundirty(bp);
854
855 bp->b_flags &= ~B_DONE;
856 bp->b_ioflags &= ~BIO_ERROR;
857 bp->b_flags |= B_CACHE;
858 bp->b_iocmd = BIO_WRITE;
859
860 bufobj_wref(bp->b_bufobj);
861 vfs_busy_pages(bp, 1);
862
863 /*
864 * Normal bwrites pipeline writes
865 */
866 bp->b_runningbufspace = bp->b_bufsize;
867 atomic_add_int(&runningbufspace, bp->b_runningbufspace);
868
869 if (!TD_IS_IDLETHREAD(curthread))
870 curthread->td_ru.ru_oublock++;
871 if (oldflags & B_ASYNC)
872 BUF_KERNPROC(bp);
873 bp->b_iooffset = dbtob(bp->b_blkno);
874 bstrategy(bp);
875
876 if ((oldflags & B_ASYNC) == 0) {
877 int rtval = bufwait(bp);
878 brelse(bp);
879 return (rtval);
880 } else {
881 /*
882 * don't allow the async write to saturate the I/O
883 * system. We will not deadlock here because
884 * we are blocking waiting for I/O that is already in-progress
885 * to complete. We do not block here if it is the update
886 * or syncer daemon trying to clean up as that can lead
887 * to deadlock.
888 */
889 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
890 waitrunningbufspace();
891 }
892
893 return (0);
894 }
895
896 void
897 bufbdflush(struct bufobj *bo, struct buf *bp)
898 {
899 struct buf *nbp;
900
901 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
902 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
903 altbufferflushes++;
904 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
905 BO_LOCK(bo);
906 /*
907 * Try to find a buffer to flush.
908 */
909 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
910 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
911 BUF_LOCK(nbp,
912 LK_EXCLUSIVE | LK_NOWAIT, NULL))
913 continue;
914 if (bp == nbp)
915 panic("bdwrite: found ourselves");
916 BO_UNLOCK(bo);
917 /* Don't countdeps with the bo lock held. */
918 if (buf_countdeps(nbp, 0)) {
919 BO_LOCK(bo);
920 BUF_UNLOCK(nbp);
921 continue;
922 }
923 if (nbp->b_flags & B_CLUSTEROK) {
924 vfs_bio_awrite(nbp);
925 } else {
926 bremfree(nbp);
927 bawrite(nbp);
928 }
929 dirtybufferflushes++;
930 break;
931 }
932 if (nbp == NULL)
933 BO_UNLOCK(bo);
934 }
935 }
936
937 /*
938 * Delayed write. (Buffer is marked dirty). Do not bother writing
939 * anything if the buffer is marked invalid.
940 *
941 * Note that since the buffer must be completely valid, we can safely
942 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
943 * biodone() in order to prevent getblk from writing the buffer
944 * out synchronously.
945 */
946 void
947 bdwrite(struct buf *bp)
948 {
949 struct thread *td = curthread;
950 struct vnode *vp;
951 struct bufobj *bo;
952
953 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
954 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
955 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy"));
956
957 if (bp->b_flags & B_INVAL) {
958 brelse(bp);
959 return;
960 }
961
962 /*
963 * If we have too many dirty buffers, don't create any more.
964 * If we are wildly over our limit, then force a complete
965 * cleanup. Otherwise, just keep the situation from getting
966 * out of control. Note that we have to avoid a recursive
967 * disaster and not try to clean up after our own cleanup!
968 */
969 vp = bp->b_vp;
970 bo = bp->b_bufobj;
971 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
972 td->td_pflags |= TDP_INBDFLUSH;
973 BO_BDFLUSH(bo, bp);
974 td->td_pflags &= ~TDP_INBDFLUSH;
975 } else
976 recursiveflushes++;
977
978 bdirty(bp);
979 /*
980 * Set B_CACHE, indicating that the buffer is fully valid. This is
981 * true even of NFS now.
982 */
983 bp->b_flags |= B_CACHE;
984
985 /*
986 * This bmap keeps the system from needing to do the bmap later,
987 * perhaps when the system is attempting to do a sync. Since it
988 * is likely that the indirect block -- or whatever other datastructure
989 * that the filesystem needs is still in memory now, it is a good
990 * thing to do this. Note also, that if the pageout daemon is
991 * requesting a sync -- there might not be enough memory to do
992 * the bmap then... So, this is important to do.
993 */
994 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
995 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
996 }
997
998 /*
999 * Set the *dirty* buffer range based upon the VM system dirty pages.
1000 */
1001 vfs_setdirty(bp);
1002
1003 /*
1004 * We need to do this here to satisfy the vnode_pager and the
1005 * pageout daemon, so that it thinks that the pages have been
1006 * "cleaned". Note that since the pages are in a delayed write
1007 * buffer -- the VFS layer "will" see that the pages get written
1008 * out on the next sync, or perhaps the cluster will be completed.
1009 */
1010 vfs_clean_pages(bp);
1011 bqrelse(bp);
1012
1013 /*
1014 * Wakeup the buffer flushing daemon if we have a lot of dirty
1015 * buffers (midpoint between our recovery point and our stall
1016 * point).
1017 */
1018 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1019
1020 /*
1021 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1022 * due to the softdep code.
1023 */
1024 }
1025
1026 /*
1027 * bdirty:
1028 *
1029 * Turn buffer into delayed write request. We must clear BIO_READ and
1030 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1031 * itself to properly update it in the dirty/clean lists. We mark it
1032 * B_DONE to ensure that any asynchronization of the buffer properly
1033 * clears B_DONE ( else a panic will occur later ).
1034 *
1035 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1036 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1037 * should only be called if the buffer is known-good.
1038 *
1039 * Since the buffer is not on a queue, we do not update the numfreebuffers
1040 * count.
1041 *
1042 * The buffer must be on QUEUE_NONE.
1043 */
1044 void
1045 bdirty(struct buf *bp)
1046 {
1047
1048 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1049 bp, bp->b_vp, bp->b_flags);
1050 KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp));
1051 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1052 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1053 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1054 bp->b_flags &= ~(B_RELBUF);
1055 bp->b_iocmd = BIO_WRITE;
1056
1057 if ((bp->b_flags & B_DELWRI) == 0) {
1058 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1059 reassignbuf(bp);
1060 atomic_add_int(&numdirtybuffers, 1);
1061 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1062 }
1063 }
1064
1065 /*
1066 * bundirty:
1067 *
1068 * Clear B_DELWRI for buffer.
1069 *
1070 * Since the buffer is not on a queue, we do not update the numfreebuffers
1071 * count.
1072 *
1073 * The buffer must be on QUEUE_NONE.
1074 */
1075
1076 void
1077 bundirty(struct buf *bp)
1078 {
1079
1080 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1081 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1082 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1083 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1084 KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp));
1085
1086 if (bp->b_flags & B_DELWRI) {
1087 bp->b_flags &= ~B_DELWRI;
1088 reassignbuf(bp);
1089 atomic_subtract_int(&numdirtybuffers, 1);
1090 numdirtywakeup(lodirtybuffers);
1091 }
1092 /*
1093 * Since it is now being written, we can clear its deferred write flag.
1094 */
1095 bp->b_flags &= ~B_DEFERRED;
1096 }
1097
1098 /*
1099 * bawrite:
1100 *
1101 * Asynchronous write. Start output on a buffer, but do not wait for
1102 * it to complete. The buffer is released when the output completes.
1103 *
1104 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1105 * B_INVAL buffers. Not us.
1106 */
1107 void
1108 bawrite(struct buf *bp)
1109 {
1110
1111 bp->b_flags |= B_ASYNC;
1112 (void) bwrite(bp);
1113 }
1114
1115 /*
1116 * bwillwrite:
1117 *
1118 * Called prior to the locking of any vnodes when we are expecting to
1119 * write. We do not want to starve the buffer cache with too many
1120 * dirty buffers so we block here. By blocking prior to the locking
1121 * of any vnodes we attempt to avoid the situation where a locked vnode
1122 * prevents the various system daemons from flushing related buffers.
1123 */
1124
1125 void
1126 bwillwrite(void)
1127 {
1128
1129 if (numdirtybuffers >= hidirtybuffers) {
1130 mtx_lock(&nblock);
1131 while (numdirtybuffers >= hidirtybuffers) {
1132 bd_wakeup(1);
1133 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1134 msleep(&needsbuffer, &nblock,
1135 (PRIBIO + 4), "flswai", 0);
1136 }
1137 mtx_unlock(&nblock);
1138 }
1139 }
1140
1141 /*
1142 * Return true if we have too many dirty buffers.
1143 */
1144 int
1145 buf_dirty_count_severe(void)
1146 {
1147
1148 return(numdirtybuffers >= hidirtybuffers);
1149 }
1150
1151 /*
1152 * brelse:
1153 *
1154 * Release a busy buffer and, if requested, free its resources. The
1155 * buffer will be stashed in the appropriate bufqueue[] allowing it
1156 * to be accessed later as a cache entity or reused for other purposes.
1157 */
1158 void
1159 brelse(struct buf *bp)
1160 {
1161 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1162 bp, bp->b_vp, bp->b_flags);
1163 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1164 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1165
1166 if (bp->b_flags & B_MANAGED) {
1167 bqrelse(bp);
1168 return;
1169 }
1170
1171 if (bp->b_iocmd == BIO_WRITE &&
1172 (bp->b_ioflags & BIO_ERROR) &&
1173 !(bp->b_flags & B_INVAL)) {
1174 /*
1175 * Failed write, redirty. Must clear BIO_ERROR to prevent
1176 * pages from being scrapped. If B_INVAL is set then
1177 * this case is not run and the next case is run to
1178 * destroy the buffer. B_INVAL can occur if the buffer
1179 * is outside the range supported by the underlying device.
1180 */
1181 bp->b_ioflags &= ~BIO_ERROR;
1182 bdirty(bp);
1183 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1184 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1185 /*
1186 * Either a failed I/O or we were asked to free or not
1187 * cache the buffer.
1188 */
1189 bp->b_flags |= B_INVAL;
1190 if (!LIST_EMPTY(&bp->b_dep))
1191 buf_deallocate(bp);
1192 if (bp->b_flags & B_DELWRI) {
1193 atomic_subtract_int(&numdirtybuffers, 1);
1194 numdirtywakeup(lodirtybuffers);
1195 }
1196 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1197 if ((bp->b_flags & B_VMIO) == 0) {
1198 if (bp->b_bufsize)
1199 allocbuf(bp, 0);
1200 if (bp->b_vp)
1201 brelvp(bp);
1202 }
1203 }
1204
1205 /*
1206 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1207 * is called with B_DELWRI set, the underlying pages may wind up
1208 * getting freed causing a previous write (bdwrite()) to get 'lost'
1209 * because pages associated with a B_DELWRI bp are marked clean.
1210 *
1211 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1212 * if B_DELWRI is set.
1213 *
1214 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1215 * on pages to return pages to the VM page queues.
1216 */
1217 if (bp->b_flags & B_DELWRI)
1218 bp->b_flags &= ~B_RELBUF;
1219 else if (vm_page_count_severe()) {
1220 /*
1221 * XXX This lock may not be necessary since BKGRDINPROG
1222 * cannot be set while we hold the buf lock, it can only be
1223 * cleared if it is already pending.
1224 */
1225 if (bp->b_vp) {
1226 BO_LOCK(bp->b_bufobj);
1227 if (!(bp->b_vflags & BV_BKGRDINPROG))
1228 bp->b_flags |= B_RELBUF;
1229 BO_UNLOCK(bp->b_bufobj);
1230 } else
1231 bp->b_flags |= B_RELBUF;
1232 }
1233
1234 /*
1235 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1236 * constituted, not even NFS buffers now. Two flags effect this. If
1237 * B_INVAL, the struct buf is invalidated but the VM object is kept
1238 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1239 *
1240 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1241 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1242 * buffer is also B_INVAL because it hits the re-dirtying code above.
1243 *
1244 * Normally we can do this whether a buffer is B_DELWRI or not. If
1245 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1246 * the commit state and we cannot afford to lose the buffer. If the
1247 * buffer has a background write in progress, we need to keep it
1248 * around to prevent it from being reconstituted and starting a second
1249 * background write.
1250 */
1251 if ((bp->b_flags & B_VMIO)
1252 && !(bp->b_vp->v_mount != NULL &&
1253 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1254 !vn_isdisk(bp->b_vp, NULL) &&
1255 (bp->b_flags & B_DELWRI))
1256 ) {
1257
1258 int i, j, resid;
1259 vm_page_t m;
1260 off_t foff;
1261 vm_pindex_t poff;
1262 vm_object_t obj;
1263
1264 obj = bp->b_bufobj->bo_object;
1265
1266 /*
1267 * Get the base offset and length of the buffer. Note that
1268 * in the VMIO case if the buffer block size is not
1269 * page-aligned then b_data pointer may not be page-aligned.
1270 * But our b_pages[] array *IS* page aligned.
1271 *
1272 * block sizes less then DEV_BSIZE (usually 512) are not
1273 * supported due to the page granularity bits (m->valid,
1274 * m->dirty, etc...).
1275 *
1276 * See man buf(9) for more information
1277 */
1278 resid = bp->b_bufsize;
1279 foff = bp->b_offset;
1280 VM_OBJECT_LOCK(obj);
1281 for (i = 0; i < bp->b_npages; i++) {
1282 int had_bogus = 0;
1283
1284 m = bp->b_pages[i];
1285
1286 /*
1287 * If we hit a bogus page, fixup *all* the bogus pages
1288 * now.
1289 */
1290 if (m == bogus_page) {
1291 poff = OFF_TO_IDX(bp->b_offset);
1292 had_bogus = 1;
1293
1294 for (j = i; j < bp->b_npages; j++) {
1295 vm_page_t mtmp;
1296 mtmp = bp->b_pages[j];
1297 if (mtmp == bogus_page) {
1298 mtmp = vm_page_lookup(obj, poff + j);
1299 if (!mtmp) {
1300 panic("brelse: page missing\n");
1301 }
1302 bp->b_pages[j] = mtmp;
1303 }
1304 }
1305
1306 if ((bp->b_flags & B_INVAL) == 0) {
1307 pmap_qenter(
1308 trunc_page((vm_offset_t)bp->b_data),
1309 bp->b_pages, bp->b_npages);
1310 }
1311 m = bp->b_pages[i];
1312 }
1313 if ((bp->b_flags & B_NOCACHE) ||
1314 (bp->b_ioflags & BIO_ERROR)) {
1315 int poffset = foff & PAGE_MASK;
1316 int presid = resid > (PAGE_SIZE - poffset) ?
1317 (PAGE_SIZE - poffset) : resid;
1318
1319 KASSERT(presid >= 0, ("brelse: extra page"));
1320 vm_page_lock_queues();
1321 vm_page_set_invalid(m, poffset, presid);
1322 vm_page_unlock_queues();
1323 if (had_bogus)
1324 printf("avoided corruption bug in bogus_page/brelse code\n");
1325 }
1326 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1327 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1328 }
1329 VM_OBJECT_UNLOCK(obj);
1330 if (bp->b_flags & (B_INVAL | B_RELBUF))
1331 vfs_vmio_release(bp);
1332
1333 } else if (bp->b_flags & B_VMIO) {
1334
1335 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1336 vfs_vmio_release(bp);
1337 }
1338
1339 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1340 if (bp->b_bufsize != 0)
1341 allocbuf(bp, 0);
1342 if (bp->b_vp != NULL)
1343 brelvp(bp);
1344 }
1345
1346 if (BUF_REFCNT(bp) > 1) {
1347 /* do not release to free list */
1348 BUF_UNLOCK(bp);
1349 return;
1350 }
1351
1352 /* enqueue */
1353 mtx_lock(&bqlock);
1354 /* Handle delayed bremfree() processing. */
1355 if (bp->b_flags & B_REMFREE)
1356 bremfreel(bp);
1357 if (bp->b_qindex != QUEUE_NONE)
1358 panic("brelse: free buffer onto another queue???");
1359
1360 /* buffers with no memory */
1361 if (bp->b_bufsize == 0) {
1362 bp->b_flags |= B_INVAL;
1363 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1364 if (bp->b_vflags & BV_BKGRDINPROG)
1365 panic("losing buffer 1");
1366 if (bp->b_kvasize) {
1367 bp->b_qindex = QUEUE_EMPTYKVA;
1368 } else {
1369 bp->b_qindex = QUEUE_EMPTY;
1370 }
1371 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1372 /* buffers with junk contents */
1373 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1374 (bp->b_ioflags & BIO_ERROR)) {
1375 bp->b_flags |= B_INVAL;
1376 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1377 if (bp->b_vflags & BV_BKGRDINPROG)
1378 panic("losing buffer 2");
1379 bp->b_qindex = QUEUE_CLEAN;
1380 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1381 /* remaining buffers */
1382 } else {
1383 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1384 (B_DELWRI|B_NEEDSGIANT))
1385 bp->b_qindex = QUEUE_DIRTY_GIANT;
1386 if (bp->b_flags & B_DELWRI)
1387 bp->b_qindex = QUEUE_DIRTY;
1388 else
1389 bp->b_qindex = QUEUE_CLEAN;
1390 if (bp->b_flags & B_AGE)
1391 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1392 else
1393 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1394 }
1395 mtx_unlock(&bqlock);
1396
1397 /*
1398 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already
1399 * placed the buffer on the correct queue. We must also disassociate
1400 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1401 * find it.
1402 */
1403 if (bp->b_flags & B_INVAL) {
1404 if (bp->b_flags & B_DELWRI)
1405 bundirty(bp);
1406 if (bp->b_vp)
1407 brelvp(bp);
1408 }
1409
1410 /*
1411 * Fixup numfreebuffers count. The bp is on an appropriate queue
1412 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1413 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1414 * if B_INVAL is set ).
1415 */
1416
1417 if (!(bp->b_flags & B_DELWRI))
1418 bufcountwakeup();
1419
1420 /*
1421 * Something we can maybe free or reuse
1422 */
1423 if (bp->b_bufsize || bp->b_kvasize)
1424 bufspacewakeup();
1425
1426 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1427 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1428 panic("brelse: not dirty");
1429 /* unlock */
1430 BUF_UNLOCK(bp);
1431 }
1432
1433 /*
1434 * Release a buffer back to the appropriate queue but do not try to free
1435 * it. The buffer is expected to be used again soon.
1436 *
1437 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1438 * biodone() to requeue an async I/O on completion. It is also used when
1439 * known good buffers need to be requeued but we think we may need the data
1440 * again soon.
1441 *
1442 * XXX we should be able to leave the B_RELBUF hint set on completion.
1443 */
1444 void
1445 bqrelse(struct buf *bp)
1446 {
1447 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1448 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1449 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1450
1451 if (BUF_REFCNT(bp) > 1) {
1452 /* do not release to free list */
1453 BUF_UNLOCK(bp);
1454 return;
1455 }
1456
1457 if (bp->b_flags & B_MANAGED) {
1458 if (bp->b_flags & B_REMFREE) {
1459 mtx_lock(&bqlock);
1460 bremfreel(bp);
1461 mtx_unlock(&bqlock);
1462 }
1463 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1464 BUF_UNLOCK(bp);
1465 return;
1466 }
1467
1468 mtx_lock(&bqlock);
1469 /* Handle delayed bremfree() processing. */
1470 if (bp->b_flags & B_REMFREE)
1471 bremfreel(bp);
1472 if (bp->b_qindex != QUEUE_NONE)
1473 panic("bqrelse: free buffer onto another queue???");
1474 /* buffers with stale but valid contents */
1475 if (bp->b_flags & B_DELWRI) {
1476 if (bp->b_flags & B_NEEDSGIANT)
1477 bp->b_qindex = QUEUE_DIRTY_GIANT;
1478 else
1479 bp->b_qindex = QUEUE_DIRTY;
1480 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1481 } else {
1482 /*
1483 * XXX This lock may not be necessary since BKGRDINPROG
1484 * cannot be set while we hold the buf lock, it can only be
1485 * cleared if it is already pending.
1486 */
1487 BO_LOCK(bp->b_bufobj);
1488 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) {
1489 BO_UNLOCK(bp->b_bufobj);
1490 bp->b_qindex = QUEUE_CLEAN;
1491 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1492 b_freelist);
1493 } else {
1494 /*
1495 * We are too low on memory, we have to try to free
1496 * the buffer (most importantly: the wired pages
1497 * making up its backing store) *now*.
1498 */
1499 BO_UNLOCK(bp->b_bufobj);
1500 mtx_unlock(&bqlock);
1501 brelse(bp);
1502 return;
1503 }
1504 }
1505 mtx_unlock(&bqlock);
1506
1507 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1508 bufcountwakeup();
1509
1510 /*
1511 * Something we can maybe free or reuse.
1512 */
1513 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1514 bufspacewakeup();
1515
1516 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1517 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1518 panic("bqrelse: not dirty");
1519 /* unlock */
1520 BUF_UNLOCK(bp);
1521 }
1522
1523 /* Give pages used by the bp back to the VM system (where possible) */
1524 static void
1525 vfs_vmio_release(struct buf *bp)
1526 {
1527 int i;
1528 vm_page_t m;
1529
1530 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1531 vm_page_lock_queues();
1532 for (i = 0; i < bp->b_npages; i++) {
1533 m = bp->b_pages[i];
1534 bp->b_pages[i] = NULL;
1535 /*
1536 * In order to keep page LRU ordering consistent, put
1537 * everything on the inactive queue.
1538 */
1539 vm_page_unwire(m, 0);
1540 /*
1541 * We don't mess with busy pages, it is
1542 * the responsibility of the process that
1543 * busied the pages to deal with them.
1544 */
1545 if ((m->oflags & VPO_BUSY) || (m->busy != 0))
1546 continue;
1547
1548 if (m->wire_count == 0) {
1549 /*
1550 * Might as well free the page if we can and it has
1551 * no valid data. We also free the page if the
1552 * buffer was used for direct I/O
1553 */
1554 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1555 m->hold_count == 0) {
1556 vm_page_free(m);
1557 } else if (bp->b_flags & B_DIRECT) {
1558 vm_page_try_to_free(m);
1559 } else if (vm_page_count_severe()) {
1560 vm_page_try_to_cache(m);
1561 }
1562 }
1563 }
1564 vm_page_unlock_queues();
1565 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1566 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1567
1568 if (bp->b_bufsize) {
1569 bufspacewakeup();
1570 bp->b_bufsize = 0;
1571 }
1572 bp->b_npages = 0;
1573 bp->b_flags &= ~B_VMIO;
1574 if (bp->b_vp)
1575 brelvp(bp);
1576 }
1577
1578 /*
1579 * Check to see if a block at a particular lbn is available for a clustered
1580 * write.
1581 */
1582 static int
1583 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1584 {
1585 struct buf *bpa;
1586 int match;
1587
1588 match = 0;
1589
1590 /* If the buf isn't in core skip it */
1591 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1592 return (0);
1593
1594 /* If the buf is busy we don't want to wait for it */
1595 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1596 return (0);
1597
1598 /* Only cluster with valid clusterable delayed write buffers */
1599 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1600 (B_DELWRI | B_CLUSTEROK))
1601 goto done;
1602
1603 if (bpa->b_bufsize != size)
1604 goto done;
1605
1606 /*
1607 * Check to see if it is in the expected place on disk and that the
1608 * block has been mapped.
1609 */
1610 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1611 match = 1;
1612 done:
1613 BUF_UNLOCK(bpa);
1614 return (match);
1615 }
1616
1617 /*
1618 * vfs_bio_awrite:
1619 *
1620 * Implement clustered async writes for clearing out B_DELWRI buffers.
1621 * This is much better then the old way of writing only one buffer at
1622 * a time. Note that we may not be presented with the buffers in the
1623 * correct order, so we search for the cluster in both directions.
1624 */
1625 int
1626 vfs_bio_awrite(struct buf *bp)
1627 {
1628 int i;
1629 int j;
1630 daddr_t lblkno = bp->b_lblkno;
1631 struct vnode *vp = bp->b_vp;
1632 int ncl;
1633 int nwritten;
1634 int size;
1635 int maxcl;
1636
1637 /*
1638 * right now we support clustered writing only to regular files. If
1639 * we find a clusterable block we could be in the middle of a cluster
1640 * rather then at the beginning.
1641 */
1642 if ((vp->v_type == VREG) &&
1643 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1644 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1645
1646 size = vp->v_mount->mnt_stat.f_iosize;
1647 maxcl = MAXPHYS / size;
1648
1649 VI_LOCK(vp);
1650 for (i = 1; i < maxcl; i++)
1651 if (vfs_bio_clcheck(vp, size, lblkno + i,
1652 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1653 break;
1654
1655 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1656 if (vfs_bio_clcheck(vp, size, lblkno - j,
1657 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1658 break;
1659
1660 VI_UNLOCK(vp);
1661 --j;
1662 ncl = i + j;
1663 /*
1664 * this is a possible cluster write
1665 */
1666 if (ncl != 1) {
1667 BUF_UNLOCK(bp);
1668 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1669 return nwritten;
1670 }
1671 }
1672 bremfree(bp);
1673 bp->b_flags |= B_ASYNC;
1674 /*
1675 * default (old) behavior, writing out only one block
1676 *
1677 * XXX returns b_bufsize instead of b_bcount for nwritten?
1678 */
1679 nwritten = bp->b_bufsize;
1680 (void) bwrite(bp);
1681
1682 return nwritten;
1683 }
1684
1685 /*
1686 * getnewbuf:
1687 *
1688 * Find and initialize a new buffer header, freeing up existing buffers
1689 * in the bufqueues as necessary. The new buffer is returned locked.
1690 *
1691 * Important: B_INVAL is not set. If the caller wishes to throw the
1692 * buffer away, the caller must set B_INVAL prior to calling brelse().
1693 *
1694 * We block if:
1695 * We have insufficient buffer headers
1696 * We have insufficient buffer space
1697 * buffer_map is too fragmented ( space reservation fails )
1698 * If we have to flush dirty buffers ( but we try to avoid this )
1699 *
1700 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1701 * Instead we ask the buf daemon to do it for us. We attempt to
1702 * avoid piecemeal wakeups of the pageout daemon.
1703 */
1704
1705 static struct buf *
1706 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1707 {
1708 struct buf *bp;
1709 struct buf *nbp;
1710 int defrag = 0;
1711 int nqindex;
1712 static int flushingbufs;
1713
1714 /*
1715 * We can't afford to block since we might be holding a vnode lock,
1716 * which may prevent system daemons from running. We deal with
1717 * low-memory situations by proactively returning memory and running
1718 * async I/O rather then sync I/O.
1719 */
1720
1721 atomic_add_int(&getnewbufcalls, 1);
1722 atomic_subtract_int(&getnewbufrestarts, 1);
1723 restart:
1724 atomic_add_int(&getnewbufrestarts, 1);
1725
1726 /*
1727 * Setup for scan. If we do not have enough free buffers,
1728 * we setup a degenerate case that immediately fails. Note
1729 * that if we are specially marked process, we are allowed to
1730 * dip into our reserves.
1731 *
1732 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1733 *
1734 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1735 * However, there are a number of cases (defragging, reusing, ...)
1736 * where we cannot backup.
1737 */
1738 mtx_lock(&bqlock);
1739 nqindex = QUEUE_EMPTYKVA;
1740 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1741
1742 if (nbp == NULL) {
1743 /*
1744 * If no EMPTYKVA buffers and we are either
1745 * defragging or reusing, locate a CLEAN buffer
1746 * to free or reuse. If bufspace useage is low
1747 * skip this step so we can allocate a new buffer.
1748 */
1749 if (defrag || bufspace >= lobufspace) {
1750 nqindex = QUEUE_CLEAN;
1751 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1752 }
1753
1754 /*
1755 * If we could not find or were not allowed to reuse a
1756 * CLEAN buffer, check to see if it is ok to use an EMPTY
1757 * buffer. We can only use an EMPTY buffer if allocating
1758 * its KVA would not otherwise run us out of buffer space.
1759 */
1760 if (nbp == NULL && defrag == 0 &&
1761 bufspace + maxsize < hibufspace) {
1762 nqindex = QUEUE_EMPTY;
1763 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1764 }
1765 }
1766
1767 /*
1768 * Run scan, possibly freeing data and/or kva mappings on the fly
1769 * depending.
1770 */
1771
1772 while ((bp = nbp) != NULL) {
1773 int qindex = nqindex;
1774
1775 /*
1776 * Calculate next bp ( we can only use it if we do not block
1777 * or do other fancy things ).
1778 */
1779 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1780 switch(qindex) {
1781 case QUEUE_EMPTY:
1782 nqindex = QUEUE_EMPTYKVA;
1783 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1784 break;
1785 /* FALLTHROUGH */
1786 case QUEUE_EMPTYKVA:
1787 nqindex = QUEUE_CLEAN;
1788 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1789 break;
1790 /* FALLTHROUGH */
1791 case QUEUE_CLEAN:
1792 /*
1793 * nbp is NULL.
1794 */
1795 break;
1796 }
1797 }
1798 /*
1799 * If we are defragging then we need a buffer with
1800 * b_kvasize != 0. XXX this situation should no longer
1801 * occur, if defrag is non-zero the buffer's b_kvasize
1802 * should also be non-zero at this point. XXX
1803 */
1804 if (defrag && bp->b_kvasize == 0) {
1805 printf("Warning: defrag empty buffer %p\n", bp);
1806 continue;
1807 }
1808
1809 /*
1810 * Start freeing the bp. This is somewhat involved. nbp
1811 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1812 */
1813 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1814 continue;
1815 if (bp->b_vp) {
1816 BO_LOCK(bp->b_bufobj);
1817 if (bp->b_vflags & BV_BKGRDINPROG) {
1818 BO_UNLOCK(bp->b_bufobj);
1819 BUF_UNLOCK(bp);
1820 continue;
1821 }
1822 BO_UNLOCK(bp->b_bufobj);
1823 }
1824 CTR6(KTR_BUF,
1825 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1826 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1827 bp->b_kvasize, bp->b_bufsize, qindex);
1828
1829 /*
1830 * Sanity Checks
1831 */
1832 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1833
1834 /*
1835 * Note: we no longer distinguish between VMIO and non-VMIO
1836 * buffers.
1837 */
1838
1839 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1840
1841 bremfreel(bp);
1842 mtx_unlock(&bqlock);
1843
1844 if (qindex == QUEUE_CLEAN) {
1845 if (bp->b_flags & B_VMIO) {
1846 bp->b_flags &= ~B_ASYNC;
1847 vfs_vmio_release(bp);
1848 }
1849 if (bp->b_vp)
1850 brelvp(bp);
1851 }
1852
1853 /*
1854 * NOTE: nbp is now entirely invalid. We can only restart
1855 * the scan from this point on.
1856 *
1857 * Get the rest of the buffer freed up. b_kva* is still
1858 * valid after this operation.
1859 */
1860
1861 if (bp->b_rcred != NOCRED) {
1862 crfree(bp->b_rcred);
1863 bp->b_rcred = NOCRED;
1864 }
1865 if (bp->b_wcred != NOCRED) {
1866 crfree(bp->b_wcred);
1867 bp->b_wcred = NOCRED;
1868 }
1869 if (!LIST_EMPTY(&bp->b_dep))
1870 buf_deallocate(bp);
1871 if (bp->b_vflags & BV_BKGRDINPROG)
1872 panic("losing buffer 3");
1873 KASSERT(bp->b_vp == NULL,
1874 ("bp: %p still has vnode %p. qindex: %d",
1875 bp, bp->b_vp, qindex));
1876 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1877 ("bp: %p still on a buffer list. xflags %X",
1878 bp, bp->b_xflags));
1879
1880 if (bp->b_bufsize)
1881 allocbuf(bp, 0);
1882
1883 bp->b_flags = 0;
1884 bp->b_ioflags = 0;
1885 bp->b_xflags = 0;
1886 bp->b_vflags = 0;
1887 bp->b_vp = NULL;
1888 bp->b_blkno = bp->b_lblkno = 0;
1889 bp->b_offset = NOOFFSET;
1890 bp->b_iodone = 0;
1891 bp->b_error = 0;
1892 bp->b_resid = 0;
1893 bp->b_bcount = 0;
1894 bp->b_npages = 0;
1895 bp->b_dirtyoff = bp->b_dirtyend = 0;
1896 bp->b_bufobj = NULL;
1897 bp->b_pin_count = 0;
1898 bp->b_fsprivate1 = NULL;
1899 bp->b_fsprivate2 = NULL;
1900 bp->b_fsprivate3 = NULL;
1901
1902 LIST_INIT(&bp->b_dep);
1903
1904 /*
1905 * If we are defragging then free the buffer.
1906 */
1907 if (defrag) {
1908 bp->b_flags |= B_INVAL;
1909 bfreekva(bp);
1910 brelse(bp);
1911 defrag = 0;
1912 goto restart;
1913 }
1914
1915 /*
1916 * Notify any waiters for the buffer lock about
1917 * identity change by freeing the buffer.
1918 */
1919 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp) > 0) {
1920 bp->b_flags |= B_INVAL;
1921 bfreekva(bp);
1922 brelse(bp);
1923 goto restart;
1924 }
1925
1926 /*
1927 * If we are overcomitted then recover the buffer and its
1928 * KVM space. This occurs in rare situations when multiple
1929 * processes are blocked in getnewbuf() or allocbuf().
1930 */
1931 if (bufspace >= hibufspace)
1932 flushingbufs = 1;
1933 if (flushingbufs && bp->b_kvasize != 0) {
1934 bp->b_flags |= B_INVAL;
1935 bfreekva(bp);
1936 brelse(bp);
1937 goto restart;
1938 }
1939 if (bufspace < lobufspace)
1940 flushingbufs = 0;
1941 break;
1942 }
1943
1944 /*
1945 * If we exhausted our list, sleep as appropriate. We may have to
1946 * wakeup various daemons and write out some dirty buffers.
1947 *
1948 * Generally we are sleeping due to insufficient buffer space.
1949 */
1950
1951 if (bp == NULL) {
1952 int flags;
1953 char *waitmsg;
1954
1955 if (defrag) {
1956 flags = VFS_BIO_NEED_BUFSPACE;
1957 waitmsg = "nbufkv";
1958 } else if (bufspace >= hibufspace) {
1959 waitmsg = "nbufbs";
1960 flags = VFS_BIO_NEED_BUFSPACE;
1961 } else {
1962 waitmsg = "newbuf";
1963 flags = VFS_BIO_NEED_ANY;
1964 }
1965 mtx_lock(&nblock);
1966 needsbuffer |= flags;
1967 mtx_unlock(&nblock);
1968 mtx_unlock(&bqlock);
1969
1970 bd_speedup(); /* heeeelp */
1971
1972 mtx_lock(&nblock);
1973 while (needsbuffer & flags) {
1974 if (msleep(&needsbuffer, &nblock,
1975 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1976 mtx_unlock(&nblock);
1977 return (NULL);
1978 }
1979 }
1980 mtx_unlock(&nblock);
1981 } else {
1982 /*
1983 * We finally have a valid bp. We aren't quite out of the
1984 * woods, we still have to reserve kva space. In order
1985 * to keep fragmentation sane we only allocate kva in
1986 * BKVASIZE chunks.
1987 */
1988 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1989
1990 if (maxsize != bp->b_kvasize) {
1991 vm_offset_t addr = 0;
1992
1993 bfreekva(bp);
1994
1995 vm_map_lock(buffer_map);
1996 if (vm_map_findspace(buffer_map,
1997 vm_map_min(buffer_map), maxsize, &addr)) {
1998 /*
1999 * Uh oh. Buffer map is to fragmented. We
2000 * must defragment the map.
2001 */
2002 atomic_add_int(&bufdefragcnt, 1);
2003 vm_map_unlock(buffer_map);
2004 defrag = 1;
2005 bp->b_flags |= B_INVAL;
2006 brelse(bp);
2007 goto restart;
2008 }
2009 if (addr) {
2010 vm_map_insert(buffer_map, NULL, 0,
2011 addr, addr + maxsize,
2012 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2013
2014 bp->b_kvabase = (caddr_t) addr;
2015 bp->b_kvasize = maxsize;
2016 atomic_add_int(&bufspace, bp->b_kvasize);
2017 atomic_add_int(&bufreusecnt, 1);
2018 }
2019 vm_map_unlock(buffer_map);
2020 }
2021 bp->b_saveaddr = bp->b_kvabase;
2022 bp->b_data = bp->b_saveaddr;
2023 }
2024 return(bp);
2025 }
2026
2027 /*
2028 * buf_daemon:
2029 *
2030 * buffer flushing daemon. Buffers are normally flushed by the
2031 * update daemon but if it cannot keep up this process starts to
2032 * take the load in an attempt to prevent getnewbuf() from blocking.
2033 */
2034
2035 static struct kproc_desc buf_kp = {
2036 "bufdaemon",
2037 buf_daemon,
2038 &bufdaemonproc
2039 };
2040 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
2041
2042 static void
2043 buf_daemon()
2044 {
2045
2046 /*
2047 * This process needs to be suspended prior to shutdown sync.
2048 */
2049 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2050 SHUTDOWN_PRI_LAST);
2051
2052 /*
2053 * This process is allowed to take the buffer cache to the limit
2054 */
2055 curthread->td_pflags |= TDP_NORUNNINGBUF;
2056 mtx_lock(&bdlock);
2057 for (;;) {
2058 bd_request = 0;
2059 mtx_unlock(&bdlock);
2060
2061 kthread_suspend_check(bufdaemonproc);
2062
2063 /*
2064 * Do the flush. Limit the amount of in-transit I/O we
2065 * allow to build up, otherwise we would completely saturate
2066 * the I/O system. Wakeup any waiting processes before we
2067 * normally would so they can run in parallel with our drain.
2068 */
2069 while (numdirtybuffers > lodirtybuffers) {
2070 int flushed;
2071
2072 flushed = flushbufqueues(QUEUE_DIRTY, 0);
2073 /* The list empty check here is slightly racy */
2074 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2075 mtx_lock(&Giant);
2076 flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0);
2077 mtx_unlock(&Giant);
2078 }
2079 if (flushed == 0) {
2080 /*
2081 * Could not find any buffers without rollback
2082 * dependencies, so just write the first one
2083 * in the hopes of eventually making progress.
2084 */
2085 flushbufqueues(QUEUE_DIRTY, 1);
2086 if (!TAILQ_EMPTY(
2087 &bufqueues[QUEUE_DIRTY_GIANT])) {
2088 mtx_lock(&Giant);
2089 flushbufqueues(QUEUE_DIRTY_GIANT, 1);
2090 mtx_unlock(&Giant);
2091 }
2092 break;
2093 }
2094 uio_yield();
2095 }
2096
2097 /*
2098 * Only clear bd_request if we have reached our low water
2099 * mark. The buf_daemon normally waits 1 second and
2100 * then incrementally flushes any dirty buffers that have
2101 * built up, within reason.
2102 *
2103 * If we were unable to hit our low water mark and couldn't
2104 * find any flushable buffers, we sleep half a second.
2105 * Otherwise we loop immediately.
2106 */
2107 mtx_lock(&bdlock);
2108 if (numdirtybuffers <= lodirtybuffers) {
2109 /*
2110 * We reached our low water mark, reset the
2111 * request and sleep until we are needed again.
2112 * The sleep is just so the suspend code works.
2113 */
2114 bd_request = 0;
2115 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2116 } else {
2117 /*
2118 * We couldn't find any flushable dirty buffers but
2119 * still have too many dirty buffers, we
2120 * have to sleep and try again. (rare)
2121 */
2122 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2123 }
2124 }
2125 }
2126
2127 /*
2128 * flushbufqueues:
2129 *
2130 * Try to flush a buffer in the dirty queue. We must be careful to
2131 * free up B_INVAL buffers instead of write them, which NFS is
2132 * particularly sensitive to.
2133 */
2134 static int flushwithdeps = 0;
2135 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2136 0, "Number of buffers flushed with dependecies that require rollbacks");
2137
2138 static int
2139 flushbufqueues(int queue, int flushdeps)
2140 {
2141 struct thread *td = curthread;
2142 struct buf sentinel;
2143 struct vnode *vp;
2144 struct mount *mp;
2145 struct buf *bp;
2146 int hasdeps;
2147 int flushed;
2148 int target;
2149
2150 target = numdirtybuffers - lodirtybuffers;
2151 if (flushdeps && target > 2)
2152 target /= 2;
2153 flushed = 0;
2154 bp = NULL;
2155 mtx_lock(&bqlock);
2156 TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist);
2157 while (flushed != target) {
2158 bp = TAILQ_FIRST(&bufqueues[queue]);
2159 if (bp == &sentinel)
2160 break;
2161 TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist);
2162 TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist);
2163
2164 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2165 continue;
2166 if (bp->b_pin_count > 0) {
2167 BUF_UNLOCK(bp);
2168 continue;
2169 }
2170 BO_LOCK(bp->b_bufobj);
2171 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2172 (bp->b_flags & B_DELWRI) == 0) {
2173 BO_UNLOCK(bp->b_bufobj);
2174 BUF_UNLOCK(bp);
2175 continue;
2176 }
2177 BO_UNLOCK(bp->b_bufobj);
2178 if (bp->b_flags & B_INVAL) {
2179 bremfreel(bp);
2180 mtx_unlock(&bqlock);
2181 brelse(bp);
2182 flushed++;
2183 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2184 mtx_lock(&bqlock);
2185 continue;
2186 }
2187
2188 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2189 if (flushdeps == 0) {
2190 BUF_UNLOCK(bp);
2191 continue;
2192 }
2193 hasdeps = 1;
2194 } else
2195 hasdeps = 0;
2196 /*
2197 * We must hold the lock on a vnode before writing
2198 * one of its buffers. Otherwise we may confuse, or
2199 * in the case of a snapshot vnode, deadlock the
2200 * system.
2201 *
2202 * The lock order here is the reverse of the normal
2203 * of vnode followed by buf lock. This is ok because
2204 * the NOWAIT will prevent deadlock.
2205 */
2206 vp = bp->b_vp;
2207 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2208 BUF_UNLOCK(bp);
2209 continue;
2210 }
2211 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2212 mtx_unlock(&bqlock);
2213 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2214 bp, bp->b_vp, bp->b_flags);
2215 vfs_bio_awrite(bp);
2216 vn_finished_write(mp);
2217 VOP_UNLOCK(vp, 0, td);
2218 flushwithdeps += hasdeps;
2219 flushed++;
2220 waitrunningbufspace();
2221 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2222 mtx_lock(&bqlock);
2223 continue;
2224 }
2225 vn_finished_write(mp);
2226 BUF_UNLOCK(bp);
2227 }
2228 TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist);
2229 mtx_unlock(&bqlock);
2230 return (flushed);
2231 }
2232
2233 /*
2234 * Check to see if a block is currently memory resident.
2235 */
2236 struct buf *
2237 incore(struct bufobj *bo, daddr_t blkno)
2238 {
2239 struct buf *bp;
2240
2241 BO_LOCK(bo);
2242 bp = gbincore(bo, blkno);
2243 BO_UNLOCK(bo);
2244 return (bp);
2245 }
2246
2247 /*
2248 * Returns true if no I/O is needed to access the
2249 * associated VM object. This is like incore except
2250 * it also hunts around in the VM system for the data.
2251 */
2252
2253 static int
2254 inmem(struct vnode * vp, daddr_t blkno)
2255 {
2256 vm_object_t obj;
2257 vm_offset_t toff, tinc, size;
2258 vm_page_t m;
2259 vm_ooffset_t off;
2260
2261 ASSERT_VOP_LOCKED(vp, "inmem");
2262
2263 if (incore(&vp->v_bufobj, blkno))
2264 return 1;
2265 if (vp->v_mount == NULL)
2266 return 0;
2267 obj = vp->v_object;
2268 if (obj == NULL)
2269 return (0);
2270
2271 size = PAGE_SIZE;
2272 if (size > vp->v_mount->mnt_stat.f_iosize)
2273 size = vp->v_mount->mnt_stat.f_iosize;
2274 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2275
2276 VM_OBJECT_LOCK(obj);
2277 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2278 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2279 if (!m)
2280 goto notinmem;
2281 tinc = size;
2282 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2283 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2284 if (vm_page_is_valid(m,
2285 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2286 goto notinmem;
2287 }
2288 VM_OBJECT_UNLOCK(obj);
2289 return 1;
2290
2291 notinmem:
2292 VM_OBJECT_UNLOCK(obj);
2293 return (0);
2294 }
2295
2296 /*
2297 * vfs_setdirty:
2298 *
2299 * Sets the dirty range for a buffer based on the status of the dirty
2300 * bits in the pages comprising the buffer.
2301 *
2302 * The range is limited to the size of the buffer.
2303 *
2304 * This routine is primarily used by NFS, but is generalized for the
2305 * B_VMIO case.
2306 */
2307 static void
2308 vfs_setdirty(struct buf *bp)
2309 {
2310
2311 /*
2312 * Degenerate case - empty buffer
2313 */
2314
2315 if (bp->b_bufsize == 0)
2316 return;
2317
2318 /*
2319 * We qualify the scan for modified pages on whether the
2320 * object has been flushed yet.
2321 */
2322
2323 if ((bp->b_flags & B_VMIO) == 0)
2324 return;
2325
2326 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2327 vfs_setdirty_locked_object(bp);
2328 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2329 }
2330
2331 static void
2332 vfs_setdirty_locked_object(struct buf *bp)
2333 {
2334 vm_object_t object;
2335 int i;
2336
2337 object = bp->b_bufobj->bo_object;
2338 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2339 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2340 vm_offset_t boffset;
2341 vm_offset_t eoffset;
2342
2343 vm_page_lock_queues();
2344 /*
2345 * test the pages to see if they have been modified directly
2346 * by users through the VM system.
2347 */
2348 for (i = 0; i < bp->b_npages; i++)
2349 vm_page_test_dirty(bp->b_pages[i]);
2350
2351 /*
2352 * Calculate the encompassing dirty range, boffset and eoffset,
2353 * (eoffset - boffset) bytes.
2354 */
2355
2356 for (i = 0; i < bp->b_npages; i++) {
2357 if (bp->b_pages[i]->dirty)
2358 break;
2359 }
2360 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2361
2362 for (i = bp->b_npages - 1; i >= 0; --i) {
2363 if (bp->b_pages[i]->dirty) {
2364 break;
2365 }
2366 }
2367 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2368
2369 vm_page_unlock_queues();
2370 /*
2371 * Fit it to the buffer.
2372 */
2373
2374 if (eoffset > bp->b_bcount)
2375 eoffset = bp->b_bcount;
2376
2377 /*
2378 * If we have a good dirty range, merge with the existing
2379 * dirty range.
2380 */
2381
2382 if (boffset < eoffset) {
2383 if (bp->b_dirtyoff > boffset)
2384 bp->b_dirtyoff = boffset;
2385 if (bp->b_dirtyend < eoffset)
2386 bp->b_dirtyend = eoffset;
2387 }
2388 }
2389 }
2390
2391 /*
2392 * getblk:
2393 *
2394 * Get a block given a specified block and offset into a file/device.
2395 * The buffers B_DONE bit will be cleared on return, making it almost
2396 * ready for an I/O initiation. B_INVAL may or may not be set on
2397 * return. The caller should clear B_INVAL prior to initiating a
2398 * READ.
2399 *
2400 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2401 * an existing buffer.
2402 *
2403 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2404 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2405 * and then cleared based on the backing VM. If the previous buffer is
2406 * non-0-sized but invalid, B_CACHE will be cleared.
2407 *
2408 * If getblk() must create a new buffer, the new buffer is returned with
2409 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2410 * case it is returned with B_INVAL clear and B_CACHE set based on the
2411 * backing VM.
2412 *
2413 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2414 * B_CACHE bit is clear.
2415 *
2416 * What this means, basically, is that the caller should use B_CACHE to
2417 * determine whether the buffer is fully valid or not and should clear
2418 * B_INVAL prior to issuing a read. If the caller intends to validate
2419 * the buffer by loading its data area with something, the caller needs
2420 * to clear B_INVAL. If the caller does this without issuing an I/O,
2421 * the caller should set B_CACHE ( as an optimization ), else the caller
2422 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2423 * a write attempt or if it was a successfull read. If the caller
2424 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2425 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2426 */
2427 struct buf *
2428 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2429 int flags)
2430 {
2431 struct buf *bp;
2432 struct bufobj *bo;
2433 int error;
2434
2435 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2436 ASSERT_VOP_LOCKED(vp, "getblk");
2437 if (size > MAXBSIZE)
2438 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2439
2440 bo = &vp->v_bufobj;
2441 loop:
2442 /*
2443 * Block if we are low on buffers. Certain processes are allowed
2444 * to completely exhaust the buffer cache.
2445 *
2446 * If this check ever becomes a bottleneck it may be better to
2447 * move it into the else, when gbincore() fails. At the moment
2448 * it isn't a problem.
2449 *
2450 * XXX remove if 0 sections (clean this up after its proven)
2451 */
2452 if (numfreebuffers == 0) {
2453 if (TD_IS_IDLETHREAD(curthread))
2454 return NULL;
2455 mtx_lock(&nblock);
2456 needsbuffer |= VFS_BIO_NEED_ANY;
2457 mtx_unlock(&nblock);
2458 }
2459
2460 BO_LOCK(bo);
2461 bp = gbincore(bo, blkno);
2462 if (bp != NULL) {
2463 int lockflags;
2464 /*
2465 * Buffer is in-core. If the buffer is not busy, it must
2466 * be on a queue.
2467 */
2468 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2469
2470 if (flags & GB_LOCK_NOWAIT)
2471 lockflags |= LK_NOWAIT;
2472
2473 error = BUF_TIMELOCK(bp, lockflags,
2474 VI_MTX(vp), "getblk", slpflag, slptimeo);
2475
2476 /*
2477 * If we slept and got the lock we have to restart in case
2478 * the buffer changed identities.
2479 */
2480 if (error == ENOLCK)
2481 goto loop;
2482 /* We timed out or were interrupted. */
2483 else if (error)
2484 return (NULL);
2485
2486 /*
2487 * The buffer is locked. B_CACHE is cleared if the buffer is
2488 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2489 * and for a VMIO buffer B_CACHE is adjusted according to the
2490 * backing VM cache.
2491 */
2492 if (bp->b_flags & B_INVAL)
2493 bp->b_flags &= ~B_CACHE;
2494 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2495 bp->b_flags |= B_CACHE;
2496 bremfree(bp);
2497
2498 /*
2499 * check for size inconsistancies for non-VMIO case.
2500 */
2501
2502 if (bp->b_bcount != size) {
2503 if ((bp->b_flags & B_VMIO) == 0 ||
2504 (size > bp->b_kvasize)) {
2505 if (bp->b_flags & B_DELWRI) {
2506 /*
2507 * If buffer is pinned and caller does
2508 * not want sleep waiting for it to be
2509 * unpinned, bail out
2510 * */
2511 if (bp->b_pin_count > 0) {
2512 if (flags & GB_LOCK_NOWAIT) {
2513 bqrelse(bp);
2514 return (NULL);
2515 } else {
2516 bunpin_wait(bp);
2517 }
2518 }
2519 bp->b_flags |= B_NOCACHE;
2520 bwrite(bp);
2521 } else {
2522 if (LIST_EMPTY(&bp->b_dep)) {
2523 bp->b_flags |= B_RELBUF;
2524 brelse(bp);
2525 } else {
2526 bp->b_flags |= B_NOCACHE;
2527 bwrite(bp);
2528 }
2529 }
2530 goto loop;
2531 }
2532 }
2533
2534 /*
2535 * If the size is inconsistant in the VMIO case, we can resize
2536 * the buffer. This might lead to B_CACHE getting set or
2537 * cleared. If the size has not changed, B_CACHE remains
2538 * unchanged from its previous state.
2539 */
2540
2541 if (bp->b_bcount != size)
2542 allocbuf(bp, size);
2543
2544 KASSERT(bp->b_offset != NOOFFSET,
2545 ("getblk: no buffer offset"));
2546
2547 /*
2548 * A buffer with B_DELWRI set and B_CACHE clear must
2549 * be committed before we can return the buffer in
2550 * order to prevent the caller from issuing a read
2551 * ( due to B_CACHE not being set ) and overwriting
2552 * it.
2553 *
2554 * Most callers, including NFS and FFS, need this to
2555 * operate properly either because they assume they
2556 * can issue a read if B_CACHE is not set, or because
2557 * ( for example ) an uncached B_DELWRI might loop due
2558 * to softupdates re-dirtying the buffer. In the latter
2559 * case, B_CACHE is set after the first write completes,
2560 * preventing further loops.
2561 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2562 * above while extending the buffer, we cannot allow the
2563 * buffer to remain with B_CACHE set after the write
2564 * completes or it will represent a corrupt state. To
2565 * deal with this we set B_NOCACHE to scrap the buffer
2566 * after the write.
2567 *
2568 * We might be able to do something fancy, like setting
2569 * B_CACHE in bwrite() except if B_DELWRI is already set,
2570 * so the below call doesn't set B_CACHE, but that gets real
2571 * confusing. This is much easier.
2572 */
2573
2574 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2575 bp->b_flags |= B_NOCACHE;
2576 bwrite(bp);
2577 goto loop;
2578 }
2579 bp->b_flags &= ~B_DONE;
2580 } else {
2581 int bsize, maxsize, vmio;
2582 off_t offset;
2583
2584 /*
2585 * Buffer is not in-core, create new buffer. The buffer
2586 * returned by getnewbuf() is locked. Note that the returned
2587 * buffer is also considered valid (not marked B_INVAL).
2588 */
2589 BO_UNLOCK(bo);
2590 /*
2591 * If the user does not want us to create the buffer, bail out
2592 * here.
2593 */
2594 if (flags & GB_NOCREAT)
2595 return NULL;
2596 bsize = bo->bo_bsize;
2597 offset = blkno * bsize;
2598 vmio = vp->v_object != NULL;
2599 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2600 maxsize = imax(maxsize, bsize);
2601
2602 bp = getnewbuf(slpflag, slptimeo, size, maxsize);
2603 if (bp == NULL) {
2604 if (slpflag || slptimeo)
2605 return NULL;
2606 goto loop;
2607 }
2608
2609 /*
2610 * This code is used to make sure that a buffer is not
2611 * created while the getnewbuf routine is blocked.
2612 * This can be a problem whether the vnode is locked or not.
2613 * If the buffer is created out from under us, we have to
2614 * throw away the one we just created.
2615 *
2616 * Note: this must occur before we associate the buffer
2617 * with the vp especially considering limitations in
2618 * the splay tree implementation when dealing with duplicate
2619 * lblkno's.
2620 */
2621 BO_LOCK(bo);
2622 if (gbincore(bo, blkno)) {
2623 BO_UNLOCK(bo);
2624 bp->b_flags |= B_INVAL;
2625 brelse(bp);
2626 goto loop;
2627 }
2628
2629 /*
2630 * Insert the buffer into the hash, so that it can
2631 * be found by incore.
2632 */
2633 bp->b_blkno = bp->b_lblkno = blkno;
2634 bp->b_offset = offset;
2635 bgetvp(vp, bp);
2636 BO_UNLOCK(bo);
2637
2638 /*
2639 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2640 * buffer size starts out as 0, B_CACHE will be set by
2641 * allocbuf() for the VMIO case prior to it testing the
2642 * backing store for validity.
2643 */
2644
2645 if (vmio) {
2646 bp->b_flags |= B_VMIO;
2647 #if defined(VFS_BIO_DEBUG)
2648 if (vn_canvmio(vp) != TRUE)
2649 printf("getblk: VMIO on vnode type %d\n",
2650 vp->v_type);
2651 #endif
2652 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2653 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2654 bp, vp->v_object, bp->b_bufobj->bo_object));
2655 } else {
2656 bp->b_flags &= ~B_VMIO;
2657 KASSERT(bp->b_bufobj->bo_object == NULL,
2658 ("ARGH! has b_bufobj->bo_object %p %p\n",
2659 bp, bp->b_bufobj->bo_object));
2660 }
2661
2662 allocbuf(bp, size);
2663 bp->b_flags &= ~B_DONE;
2664 }
2665 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2666 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2667 KASSERT(bp->b_bufobj == bo,
2668 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2669 return (bp);
2670 }
2671
2672 /*
2673 * Get an empty, disassociated buffer of given size. The buffer is initially
2674 * set to B_INVAL.
2675 */
2676 struct buf *
2677 geteblk(int size)
2678 {
2679 struct buf *bp;
2680 int maxsize;
2681
2682 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2683 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2684 continue;
2685 allocbuf(bp, size);
2686 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2687 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2688 return (bp);
2689 }
2690
2691
2692 /*
2693 * This code constitutes the buffer memory from either anonymous system
2694 * memory (in the case of non-VMIO operations) or from an associated
2695 * VM object (in the case of VMIO operations). This code is able to
2696 * resize a buffer up or down.
2697 *
2698 * Note that this code is tricky, and has many complications to resolve
2699 * deadlock or inconsistant data situations. Tread lightly!!!
2700 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2701 * the caller. Calling this code willy nilly can result in the loss of data.
2702 *
2703 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2704 * B_CACHE for the non-VMIO case.
2705 */
2706
2707 int
2708 allocbuf(struct buf *bp, int size)
2709 {
2710 int newbsize, mbsize;
2711 int i;
2712
2713 if (BUF_REFCNT(bp) == 0)
2714 panic("allocbuf: buffer not busy");
2715
2716 if (bp->b_kvasize < size)
2717 panic("allocbuf: buffer too small");
2718
2719 if ((bp->b_flags & B_VMIO) == 0) {
2720 caddr_t origbuf;
2721 int origbufsize;
2722 /*
2723 * Just get anonymous memory from the kernel. Don't
2724 * mess with B_CACHE.
2725 */
2726 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2727 if (bp->b_flags & B_MALLOC)
2728 newbsize = mbsize;
2729 else
2730 newbsize = round_page(size);
2731
2732 if (newbsize < bp->b_bufsize) {
2733 /*
2734 * malloced buffers are not shrunk
2735 */
2736 if (bp->b_flags & B_MALLOC) {
2737 if (newbsize) {
2738 bp->b_bcount = size;
2739 } else {
2740 free(bp->b_data, M_BIOBUF);
2741 if (bp->b_bufsize) {
2742 atomic_subtract_int(
2743 &bufmallocspace,
2744 bp->b_bufsize);
2745 bufspacewakeup();
2746 bp->b_bufsize = 0;
2747 }
2748 bp->b_saveaddr = bp->b_kvabase;
2749 bp->b_data = bp->b_saveaddr;
2750 bp->b_bcount = 0;
2751 bp->b_flags &= ~B_MALLOC;
2752 }
2753 return 1;
2754 }
2755 vm_hold_free_pages(
2756 bp,
2757 (vm_offset_t) bp->b_data + newbsize,
2758 (vm_offset_t) bp->b_data + bp->b_bufsize);
2759 } else if (newbsize > bp->b_bufsize) {
2760 /*
2761 * We only use malloced memory on the first allocation.
2762 * and revert to page-allocated memory when the buffer
2763 * grows.
2764 */
2765 /*
2766 * There is a potential smp race here that could lead
2767 * to bufmallocspace slightly passing the max. It
2768 * is probably extremely rare and not worth worrying
2769 * over.
2770 */
2771 if ( (bufmallocspace < maxbufmallocspace) &&
2772 (bp->b_bufsize == 0) &&
2773 (mbsize <= PAGE_SIZE/2)) {
2774
2775 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2776 bp->b_bufsize = mbsize;
2777 bp->b_bcount = size;
2778 bp->b_flags |= B_MALLOC;
2779 atomic_add_int(&bufmallocspace, mbsize);
2780 return 1;
2781 }
2782 origbuf = NULL;
2783 origbufsize = 0;
2784 /*
2785 * If the buffer is growing on its other-than-first allocation,
2786 * then we revert to the page-allocation scheme.
2787 */
2788 if (bp->b_flags & B_MALLOC) {
2789 origbuf = bp->b_data;
2790 origbufsize = bp->b_bufsize;
2791 bp->b_data = bp->b_kvabase;
2792 if (bp->b_bufsize) {
2793 atomic_subtract_int(&bufmallocspace,
2794 bp->b_bufsize);
2795 bufspacewakeup();
2796 bp->b_bufsize = 0;
2797 }
2798 bp->b_flags &= ~B_MALLOC;
2799 newbsize = round_page(newbsize);
2800 }
2801 vm_hold_load_pages(
2802 bp,
2803 (vm_offset_t) bp->b_data + bp->b_bufsize,
2804 (vm_offset_t) bp->b_data + newbsize);
2805 if (origbuf) {
2806 bcopy(origbuf, bp->b_data, origbufsize);
2807 free(origbuf, M_BIOBUF);
2808 }
2809 }
2810 } else {
2811 int desiredpages;
2812
2813 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2814 desiredpages = (size == 0) ? 0 :
2815 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2816
2817 if (bp->b_flags & B_MALLOC)
2818 panic("allocbuf: VMIO buffer can't be malloced");
2819 /*
2820 * Set B_CACHE initially if buffer is 0 length or will become
2821 * 0-length.
2822 */
2823 if (size == 0 || bp->b_bufsize == 0)
2824 bp->b_flags |= B_CACHE;
2825
2826 if (newbsize < bp->b_bufsize) {
2827 /*
2828 * DEV_BSIZE aligned new buffer size is less then the
2829 * DEV_BSIZE aligned existing buffer size. Figure out
2830 * if we have to remove any pages.
2831 */
2832 if (desiredpages < bp->b_npages) {
2833 vm_page_t m;
2834
2835 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2836 vm_page_lock_queues();
2837 for (i = desiredpages; i < bp->b_npages; i++) {
2838 /*
2839 * the page is not freed here -- it
2840 * is the responsibility of
2841 * vnode_pager_setsize
2842 */
2843 m = bp->b_pages[i];
2844 KASSERT(m != bogus_page,
2845 ("allocbuf: bogus page found"));
2846 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2847 vm_page_lock_queues();
2848
2849 bp->b_pages[i] = NULL;
2850 vm_page_unwire(m, 0);
2851 }
2852 vm_page_unlock_queues();
2853 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2854 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2855 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2856 bp->b_npages = desiredpages;
2857 }
2858 } else if (size > bp->b_bcount) {
2859 /*
2860 * We are growing the buffer, possibly in a
2861 * byte-granular fashion.
2862 */
2863 struct vnode *vp;
2864 vm_object_t obj;
2865 vm_offset_t toff;
2866 vm_offset_t tinc;
2867
2868 /*
2869 * Step 1, bring in the VM pages from the object,
2870 * allocating them if necessary. We must clear
2871 * B_CACHE if these pages are not valid for the
2872 * range covered by the buffer.
2873 */
2874
2875 vp = bp->b_vp;
2876 obj = bp->b_bufobj->bo_object;
2877
2878 VM_OBJECT_LOCK(obj);
2879 while (bp->b_npages < desiredpages) {
2880 vm_page_t m;
2881 vm_pindex_t pi;
2882
2883 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2884 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2885 /*
2886 * note: must allocate system pages
2887 * since blocking here could intefere
2888 * with paging I/O, no matter which
2889 * process we are.
2890 */
2891 m = vm_page_alloc(obj, pi,
2892 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2893 VM_ALLOC_WIRED);
2894 if (m == NULL) {
2895 atomic_add_int(&vm_pageout_deficit,
2896 desiredpages - bp->b_npages);
2897 VM_OBJECT_UNLOCK(obj);
2898 VM_WAIT;
2899 VM_OBJECT_LOCK(obj);
2900 } else {
2901 if (m->valid == 0)
2902 bp->b_flags &= ~B_CACHE;
2903 bp->b_pages[bp->b_npages] = m;
2904 ++bp->b_npages;
2905 }
2906 continue;
2907 }
2908
2909 /*
2910 * We found a page. If we have to sleep on it,
2911 * retry because it might have gotten freed out
2912 * from under us.
2913 *
2914 * We can only test VPO_BUSY here. Blocking on
2915 * m->busy might lead to a deadlock:
2916 *
2917 * vm_fault->getpages->cluster_read->allocbuf
2918 *
2919 */
2920 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2921 continue;
2922
2923 /*
2924 * We have a good page.
2925 */
2926 vm_page_lock_queues();
2927 vm_page_wire(m);
2928 vm_page_unlock_queues();
2929 bp->b_pages[bp->b_npages] = m;
2930 ++bp->b_npages;
2931 }
2932
2933 /*
2934 * Step 2. We've loaded the pages into the buffer,
2935 * we have to figure out if we can still have B_CACHE
2936 * set. Note that B_CACHE is set according to the
2937 * byte-granular range ( bcount and size ), new the
2938 * aligned range ( newbsize ).
2939 *
2940 * The VM test is against m->valid, which is DEV_BSIZE
2941 * aligned. Needless to say, the validity of the data
2942 * needs to also be DEV_BSIZE aligned. Note that this
2943 * fails with NFS if the server or some other client
2944 * extends the file's EOF. If our buffer is resized,
2945 * B_CACHE may remain set! XXX
2946 */
2947
2948 toff = bp->b_bcount;
2949 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2950
2951 while ((bp->b_flags & B_CACHE) && toff < size) {
2952 vm_pindex_t pi;
2953
2954 if (tinc > (size - toff))
2955 tinc = size - toff;
2956
2957 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2958 PAGE_SHIFT;
2959
2960 vfs_buf_test_cache(
2961 bp,
2962 bp->b_offset,
2963 toff,
2964 tinc,
2965 bp->b_pages[pi]
2966 );
2967 toff += tinc;
2968 tinc = PAGE_SIZE;
2969 }
2970 VM_OBJECT_UNLOCK(obj);
2971
2972 /*
2973 * Step 3, fixup the KVM pmap. Remember that
2974 * bp->b_data is relative to bp->b_offset, but
2975 * bp->b_offset may be offset into the first page.
2976 */
2977
2978 bp->b_data = (caddr_t)
2979 trunc_page((vm_offset_t)bp->b_data);
2980 pmap_qenter(
2981 (vm_offset_t)bp->b_data,
2982 bp->b_pages,
2983 bp->b_npages
2984 );
2985
2986 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2987 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2988 }
2989 }
2990 if (newbsize < bp->b_bufsize)
2991 bufspacewakeup();
2992 bp->b_bufsize = newbsize; /* actual buffer allocation */
2993 bp->b_bcount = size; /* requested buffer size */
2994 return 1;
2995 }
2996
2997 void
2998 biodone(struct bio *bp)
2999 {
3000 void (*done)(struct bio *);
3001
3002 mtx_lock(&bdonelock);
3003 bp->bio_flags |= BIO_DONE;
3004 done = bp->bio_done;
3005 if (done == NULL)
3006 wakeup(bp);
3007 mtx_unlock(&bdonelock);
3008 if (done != NULL)
3009 done(bp);
3010 }
3011
3012 /*
3013 * Wait for a BIO to finish.
3014 *
3015 * XXX: resort to a timeout for now. The optimal locking (if any) for this
3016 * case is not yet clear.
3017 */
3018 int
3019 biowait(struct bio *bp, const char *wchan)
3020 {
3021
3022 mtx_lock(&bdonelock);
3023 while ((bp->bio_flags & BIO_DONE) == 0)
3024 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
3025 mtx_unlock(&bdonelock);
3026 if (bp->bio_error != 0)
3027 return (bp->bio_error);
3028 if (!(bp->bio_flags & BIO_ERROR))
3029 return (0);
3030 return (EIO);
3031 }
3032
3033 void
3034 biofinish(struct bio *bp, struct devstat *stat, int error)
3035 {
3036
3037 if (error) {
3038 bp->bio_error = error;
3039 bp->bio_flags |= BIO_ERROR;
3040 }
3041 if (stat != NULL)
3042 devstat_end_transaction_bio(stat, bp);
3043 biodone(bp);
3044 }
3045
3046 /*
3047 * bufwait:
3048 *
3049 * Wait for buffer I/O completion, returning error status. The buffer
3050 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3051 * error and cleared.
3052 */
3053 int
3054 bufwait(struct buf *bp)
3055 {
3056 if (bp->b_iocmd == BIO_READ)
3057 bwait(bp, PRIBIO, "biord");
3058 else
3059 bwait(bp, PRIBIO, "biowr");
3060 if (bp->b_flags & B_EINTR) {
3061 bp->b_flags &= ~B_EINTR;
3062 return (EINTR);
3063 }
3064 if (bp->b_ioflags & BIO_ERROR) {
3065 return (bp->b_error ? bp->b_error : EIO);
3066 } else {
3067 return (0);
3068 }
3069 }
3070
3071 /*
3072 * Call back function from struct bio back up to struct buf.
3073 */
3074 static void
3075 bufdonebio(struct bio *bip)
3076 {
3077 struct buf *bp;
3078
3079 bp = bip->bio_caller2;
3080 bp->b_resid = bp->b_bcount - bip->bio_completed;
3081 bp->b_resid = bip->bio_resid; /* XXX: remove */
3082 bp->b_ioflags = bip->bio_flags;
3083 bp->b_error = bip->bio_error;
3084 if (bp->b_error)
3085 bp->b_ioflags |= BIO_ERROR;
3086 bufdone(bp);
3087 g_destroy_bio(bip);
3088 }
3089
3090 void
3091 dev_strategy(struct cdev *dev, struct buf *bp)
3092 {
3093 struct cdevsw *csw;
3094 struct bio *bip;
3095
3096 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3097 panic("b_iocmd botch");
3098 for (;;) {
3099 bip = g_new_bio();
3100 if (bip != NULL)
3101 break;
3102 /* Try again later */
3103 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3104 }
3105 bip->bio_cmd = bp->b_iocmd;
3106 bip->bio_offset = bp->b_iooffset;
3107 bip->bio_length = bp->b_bcount;
3108 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3109 bip->bio_data = bp->b_data;
3110 bip->bio_done = bufdonebio;
3111 bip->bio_caller2 = bp;
3112 bip->bio_dev = dev;
3113 KASSERT(dev->si_refcount > 0,
3114 ("dev_strategy on un-referenced struct cdev *(%s)",
3115 devtoname(dev)));
3116 csw = dev_refthread(dev);
3117 if (csw == NULL) {
3118 g_destroy_bio(bip);
3119 bp->b_error = ENXIO;
3120 bp->b_ioflags = BIO_ERROR;
3121 bufdone(bp);
3122 return;
3123 }
3124 (*csw->d_strategy)(bip);
3125 dev_relthread(dev);
3126 }
3127
3128 /*
3129 * bufdone:
3130 *
3131 * Finish I/O on a buffer, optionally calling a completion function.
3132 * This is usually called from an interrupt so process blocking is
3133 * not allowed.
3134 *
3135 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3136 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3137 * assuming B_INVAL is clear.
3138 *
3139 * For the VMIO case, we set B_CACHE if the op was a read and no
3140 * read error occured, or if the op was a write. B_CACHE is never
3141 * set if the buffer is invalid or otherwise uncacheable.
3142 *
3143 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3144 * initiator to leave B_INVAL set to brelse the buffer out of existance
3145 * in the biodone routine.
3146 */
3147 void
3148 bufdone(struct buf *bp)
3149 {
3150 struct bufobj *dropobj;
3151 void (*biodone)(struct buf *);
3152
3153 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3154 dropobj = NULL;
3155
3156 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp,
3157 BUF_REFCNT(bp)));
3158 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3159
3160 runningbufwakeup(bp);
3161 if (bp->b_iocmd == BIO_WRITE)
3162 dropobj = bp->b_bufobj;
3163 /* call optional completion function if requested */
3164 if (bp->b_iodone != NULL) {
3165 biodone = bp->b_iodone;
3166 bp->b_iodone = NULL;
3167 (*biodone) (bp);
3168 if (dropobj)
3169 bufobj_wdrop(dropobj);
3170 return;
3171 }
3172
3173 bufdone_finish(bp);
3174
3175 if (dropobj)
3176 bufobj_wdrop(dropobj);
3177 }
3178
3179 void
3180 bufdone_finish(struct buf *bp)
3181 {
3182 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp,
3183 BUF_REFCNT(bp)));
3184
3185 if (!LIST_EMPTY(&bp->b_dep))
3186 buf_complete(bp);
3187
3188 if (bp->b_flags & B_VMIO) {
3189 int i;
3190 vm_ooffset_t foff;
3191 vm_page_t m;
3192 vm_object_t obj;
3193 int iosize;
3194 struct vnode *vp = bp->b_vp;
3195 boolean_t are_queues_locked;
3196
3197 obj = bp->b_bufobj->bo_object;
3198
3199 #if defined(VFS_BIO_DEBUG)
3200 mp_fixme("usecount and vflag accessed without locks.");
3201 if (vp->v_usecount == 0) {
3202 panic("biodone: zero vnode ref count");
3203 }
3204
3205 KASSERT(vp->v_object != NULL,
3206 ("biodone: vnode %p has no vm_object", vp));
3207 #endif
3208
3209 foff = bp->b_offset;
3210 KASSERT(bp->b_offset != NOOFFSET,
3211 ("biodone: no buffer offset"));
3212
3213 VM_OBJECT_LOCK(obj);
3214 #if defined(VFS_BIO_DEBUG)
3215 if (obj->paging_in_progress < bp->b_npages) {
3216 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3217 obj->paging_in_progress, bp->b_npages);
3218 }
3219 #endif
3220
3221 /*
3222 * Set B_CACHE if the op was a normal read and no error
3223 * occured. B_CACHE is set for writes in the b*write()
3224 * routines.
3225 */
3226 iosize = bp->b_bcount - bp->b_resid;
3227 if (bp->b_iocmd == BIO_READ &&
3228 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3229 !(bp->b_ioflags & BIO_ERROR)) {
3230 bp->b_flags |= B_CACHE;
3231 }
3232 if (bp->b_iocmd == BIO_READ) {
3233 vm_page_lock_queues();
3234 are_queues_locked = TRUE;
3235 } else
3236 are_queues_locked = FALSE;
3237 for (i = 0; i < bp->b_npages; i++) {
3238 int bogusflag = 0;
3239 int resid;
3240
3241 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3242 if (resid > iosize)
3243 resid = iosize;
3244
3245 /*
3246 * cleanup bogus pages, restoring the originals
3247 */
3248 m = bp->b_pages[i];
3249 if (m == bogus_page) {
3250 bogusflag = 1;
3251 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3252 if (m == NULL)
3253 panic("biodone: page disappeared!");
3254 bp->b_pages[i] = m;
3255 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3256 bp->b_pages, bp->b_npages);
3257 }
3258 #if defined(VFS_BIO_DEBUG)
3259 if (OFF_TO_IDX(foff) != m->pindex) {
3260 printf(
3261 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3262 (intmax_t)foff, (uintmax_t)m->pindex);
3263 }
3264 #endif
3265
3266 /*
3267 * In the write case, the valid and clean bits are
3268 * already changed correctly ( see bdwrite() ), so we
3269 * only need to do this here in the read case.
3270 */
3271 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3272 vfs_page_set_valid(bp, foff, i, m);
3273 }
3274
3275 /*
3276 * when debugging new filesystems or buffer I/O methods, this
3277 * is the most common error that pops up. if you see this, you
3278 * have not set the page busy flag correctly!!!
3279 */
3280 if (m->busy == 0) {
3281 printf("biodone: page busy < 0, "
3282 "pindex: %d, foff: 0x(%x,%x), "
3283 "resid: %d, index: %d\n",
3284 (int) m->pindex, (int)(foff >> 32),
3285 (int) foff & 0xffffffff, resid, i);
3286 if (!vn_isdisk(vp, NULL))
3287 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3288 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3289 (intmax_t) bp->b_lblkno,
3290 bp->b_flags, bp->b_npages);
3291 else
3292 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3293 (intmax_t) bp->b_lblkno,
3294 bp->b_flags, bp->b_npages);
3295 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3296 (u_long)m->valid, (u_long)m->dirty,
3297 m->wire_count);
3298 panic("biodone: page busy < 0\n");
3299 }
3300 vm_page_io_finish(m);
3301 vm_object_pip_subtract(obj, 1);
3302 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3303 iosize -= resid;
3304 }
3305 if (are_queues_locked)
3306 vm_page_unlock_queues();
3307 vm_object_pip_wakeupn(obj, 0);
3308 VM_OBJECT_UNLOCK(obj);
3309 }
3310
3311 /*
3312 * For asynchronous completions, release the buffer now. The brelse
3313 * will do a wakeup there if necessary - so no need to do a wakeup
3314 * here in the async case. The sync case always needs to do a wakeup.
3315 */
3316
3317 if (bp->b_flags & B_ASYNC) {
3318 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3319 brelse(bp);
3320 else
3321 bqrelse(bp);
3322 } else
3323 bdone(bp);
3324 }
3325
3326 /*
3327 * This routine is called in lieu of iodone in the case of
3328 * incomplete I/O. This keeps the busy status for pages
3329 * consistant.
3330 */
3331 void
3332 vfs_unbusy_pages(struct buf *bp)
3333 {
3334 int i;
3335 vm_object_t obj;
3336 vm_page_t m;
3337
3338 runningbufwakeup(bp);
3339 if (!(bp->b_flags & B_VMIO))
3340 return;
3341
3342 obj = bp->b_bufobj->bo_object;
3343 VM_OBJECT_LOCK(obj);
3344 for (i = 0; i < bp->b_npages; i++) {
3345 m = bp->b_pages[i];
3346 if (m == bogus_page) {
3347 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3348 if (!m)
3349 panic("vfs_unbusy_pages: page missing\n");
3350 bp->b_pages[i] = m;
3351 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3352 bp->b_pages, bp->b_npages);
3353 }
3354 vm_object_pip_subtract(obj, 1);
3355 vm_page_io_finish(m);
3356 }
3357 vm_object_pip_wakeupn(obj, 0);
3358 VM_OBJECT_UNLOCK(obj);
3359 }
3360
3361 /*
3362 * vfs_page_set_valid:
3363 *
3364 * Set the valid bits in a page based on the supplied offset. The
3365 * range is restricted to the buffer's size.
3366 *
3367 * This routine is typically called after a read completes.
3368 */
3369 static void
3370 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3371 {
3372 vm_ooffset_t soff, eoff;
3373
3374 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3375 /*
3376 * Start and end offsets in buffer. eoff - soff may not cross a
3377 * page boundry or cross the end of the buffer. The end of the
3378 * buffer, in this case, is our file EOF, not the allocation size
3379 * of the buffer.
3380 */
3381 soff = off;
3382 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3383 if (eoff > bp->b_offset + bp->b_bcount)
3384 eoff = bp->b_offset + bp->b_bcount;
3385
3386 /*
3387 * Set valid range. This is typically the entire buffer and thus the
3388 * entire page.
3389 */
3390 if (eoff > soff) {
3391 vm_page_set_validclean(
3392 m,
3393 (vm_offset_t) (soff & PAGE_MASK),
3394 (vm_offset_t) (eoff - soff)
3395 );
3396 }
3397 }
3398
3399 /*
3400 * This routine is called before a device strategy routine.
3401 * It is used to tell the VM system that paging I/O is in
3402 * progress, and treat the pages associated with the buffer
3403 * almost as being VPO_BUSY. Also the object paging_in_progress
3404 * flag is handled to make sure that the object doesn't become
3405 * inconsistant.
3406 *
3407 * Since I/O has not been initiated yet, certain buffer flags
3408 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3409 * and should be ignored.
3410 */
3411 void
3412 vfs_busy_pages(struct buf *bp, int clear_modify)
3413 {
3414 int i, bogus;
3415 vm_object_t obj;
3416 vm_ooffset_t foff;
3417 vm_page_t m;
3418
3419 if (!(bp->b_flags & B_VMIO))
3420 return;
3421
3422 obj = bp->b_bufobj->bo_object;
3423 foff = bp->b_offset;
3424 KASSERT(bp->b_offset != NOOFFSET,
3425 ("vfs_busy_pages: no buffer offset"));
3426 VM_OBJECT_LOCK(obj);
3427 if (bp->b_bufsize != 0)
3428 vfs_setdirty_locked_object(bp);
3429 retry:
3430 for (i = 0; i < bp->b_npages; i++) {
3431 m = bp->b_pages[i];
3432
3433 if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3434 goto retry;
3435 }
3436 bogus = 0;
3437 vm_page_lock_queues();
3438 for (i = 0; i < bp->b_npages; i++) {
3439 m = bp->b_pages[i];
3440
3441 if ((bp->b_flags & B_CLUSTER) == 0) {
3442 vm_object_pip_add(obj, 1);
3443 vm_page_io_start(m);
3444 }
3445 /*
3446 * When readying a buffer for a read ( i.e
3447 * clear_modify == 0 ), it is important to do
3448 * bogus_page replacement for valid pages in
3449 * partially instantiated buffers. Partially
3450 * instantiated buffers can, in turn, occur when
3451 * reconstituting a buffer from its VM backing store
3452 * base. We only have to do this if B_CACHE is
3453 * clear ( which causes the I/O to occur in the
3454 * first place ). The replacement prevents the read
3455 * I/O from overwriting potentially dirty VM-backed
3456 * pages. XXX bogus page replacement is, uh, bogus.
3457 * It may not work properly with small-block devices.
3458 * We need to find a better way.
3459 */
3460 pmap_remove_all(m);
3461 if (clear_modify)
3462 vfs_page_set_valid(bp, foff, i, m);
3463 else if (m->valid == VM_PAGE_BITS_ALL &&
3464 (bp->b_flags & B_CACHE) == 0) {
3465 bp->b_pages[i] = bogus_page;
3466 bogus++;
3467 }
3468 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3469 }
3470 vm_page_unlock_queues();
3471 VM_OBJECT_UNLOCK(obj);
3472 if (bogus)
3473 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3474 bp->b_pages, bp->b_npages);
3475 }
3476
3477 /*
3478 * Tell the VM system that the pages associated with this buffer
3479 * are clean. This is used for delayed writes where the data is
3480 * going to go to disk eventually without additional VM intevention.
3481 *
3482 * Note that while we only really need to clean through to b_bcount, we
3483 * just go ahead and clean through to b_bufsize.
3484 */
3485 static void
3486 vfs_clean_pages(struct buf *bp)
3487 {
3488 int i;
3489 vm_ooffset_t foff, noff, eoff;
3490 vm_page_t m;
3491
3492 if (!(bp->b_flags & B_VMIO))
3493 return;
3494
3495 foff = bp->b_offset;
3496 KASSERT(bp->b_offset != NOOFFSET,
3497 ("vfs_clean_pages: no buffer offset"));
3498 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3499 vm_page_lock_queues();
3500 for (i = 0; i < bp->b_npages; i++) {
3501 m = bp->b_pages[i];
3502 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3503 eoff = noff;
3504
3505 if (eoff > bp->b_offset + bp->b_bufsize)
3506 eoff = bp->b_offset + bp->b_bufsize;
3507 vfs_page_set_valid(bp, foff, i, m);
3508 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3509 foff = noff;
3510 }
3511 vm_page_unlock_queues();
3512 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3513 }
3514
3515 /*
3516 * vfs_bio_set_validclean:
3517 *
3518 * Set the range within the buffer to valid and clean. The range is
3519 * relative to the beginning of the buffer, b_offset. Note that b_offset
3520 * itself may be offset from the beginning of the first page.
3521 *
3522 */
3523
3524 void
3525 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3526 {
3527 int i, n;
3528 vm_page_t m;
3529
3530 if (!(bp->b_flags & B_VMIO))
3531 return;
3532 /*
3533 * Fixup base to be relative to beginning of first page.
3534 * Set initial n to be the maximum number of bytes in the
3535 * first page that can be validated.
3536 */
3537
3538 base += (bp->b_offset & PAGE_MASK);
3539 n = PAGE_SIZE - (base & PAGE_MASK);
3540
3541 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3542 vm_page_lock_queues();
3543 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3544 m = bp->b_pages[i];
3545 if (n > size)
3546 n = size;
3547 vm_page_set_validclean(m, base & PAGE_MASK, n);
3548 base += n;
3549 size -= n;
3550 n = PAGE_SIZE;
3551 }
3552 vm_page_unlock_queues();
3553 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3554 }
3555
3556 /*
3557 * vfs_bio_clrbuf:
3558 *
3559 * clear a buffer. This routine essentially fakes an I/O, so we need
3560 * to clear BIO_ERROR and B_INVAL.
3561 *
3562 * Note that while we only theoretically need to clear through b_bcount,
3563 * we go ahead and clear through b_bufsize.
3564 */
3565
3566 void
3567 vfs_bio_clrbuf(struct buf *bp)
3568 {
3569 int i, j, mask = 0;
3570 caddr_t sa, ea;
3571
3572 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3573 clrbuf(bp);
3574 return;
3575 }
3576
3577 bp->b_flags &= ~B_INVAL;
3578 bp->b_ioflags &= ~BIO_ERROR;
3579 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3580 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3581 (bp->b_offset & PAGE_MASK) == 0) {
3582 if (bp->b_pages[0] == bogus_page)
3583 goto unlock;
3584 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3585 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3586 if ((bp->b_pages[0]->valid & mask) == mask)
3587 goto unlock;
3588 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3589 ((bp->b_pages[0]->valid & mask) == 0)) {
3590 bzero(bp->b_data, bp->b_bufsize);
3591 bp->b_pages[0]->valid |= mask;
3592 goto unlock;
3593 }
3594 }
3595 ea = sa = bp->b_data;
3596 for(i = 0; i < bp->b_npages; i++, sa = ea) {
3597 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3598 ea = (caddr_t)(vm_offset_t)ulmin(
3599 (u_long)(vm_offset_t)ea,
3600 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3601 if (bp->b_pages[i] == bogus_page)
3602 continue;
3603 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3604 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3605 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3606 if ((bp->b_pages[i]->valid & mask) == mask)
3607 continue;
3608 if ((bp->b_pages[i]->valid & mask) == 0) {
3609 if ((bp->b_pages[i]->flags & PG_ZERO) == 0)
3610 bzero(sa, ea - sa);
3611 } else {
3612 for (; sa < ea; sa += DEV_BSIZE, j++) {
3613 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3614 (bp->b_pages[i]->valid & (1 << j)) == 0)
3615 bzero(sa, DEV_BSIZE);
3616 }
3617 }
3618 bp->b_pages[i]->valid |= mask;
3619 }
3620 unlock:
3621 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3622 bp->b_resid = 0;
3623 }
3624
3625 /*
3626 * vm_hold_load_pages and vm_hold_free_pages get pages into
3627 * a buffers address space. The pages are anonymous and are
3628 * not associated with a file object.
3629 */
3630 static void
3631 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3632 {
3633 vm_offset_t pg;
3634 vm_page_t p;
3635 int index;
3636
3637 to = round_page(to);
3638 from = round_page(from);
3639 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3640
3641 VM_OBJECT_LOCK(kernel_object);
3642 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3643 tryagain:
3644 /*
3645 * note: must allocate system pages since blocking here
3646 * could intefere with paging I/O, no matter which
3647 * process we are.
3648 */
3649 p = vm_page_alloc(kernel_object,
3650 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3651 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3652 if (!p) {
3653 atomic_add_int(&vm_pageout_deficit,
3654 (to - pg) >> PAGE_SHIFT);
3655 VM_OBJECT_UNLOCK(kernel_object);
3656 VM_WAIT;
3657 VM_OBJECT_LOCK(kernel_object);
3658 goto tryagain;
3659 }
3660 p->valid = VM_PAGE_BITS_ALL;
3661 pmap_qenter(pg, &p, 1);
3662 bp->b_pages[index] = p;
3663 }
3664 VM_OBJECT_UNLOCK(kernel_object);
3665 bp->b_npages = index;
3666 }
3667
3668 /* Return pages associated with this buf to the vm system */
3669 static void
3670 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3671 {
3672 vm_offset_t pg;
3673 vm_page_t p;
3674 int index, newnpages;
3675
3676 from = round_page(from);
3677 to = round_page(to);
3678 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3679
3680 VM_OBJECT_LOCK(kernel_object);
3681 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3682 p = bp->b_pages[index];
3683 if (p && (index < bp->b_npages)) {
3684 if (p->busy) {
3685 printf(
3686 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3687 (intmax_t)bp->b_blkno,
3688 (intmax_t)bp->b_lblkno);
3689 }
3690 bp->b_pages[index] = NULL;
3691 pmap_qremove(pg, 1);
3692 vm_page_lock_queues();
3693 vm_page_unwire(p, 0);
3694 vm_page_free(p);
3695 vm_page_unlock_queues();
3696 }
3697 }
3698 VM_OBJECT_UNLOCK(kernel_object);
3699 bp->b_npages = newnpages;
3700 }
3701
3702 /*
3703 * Map an IO request into kernel virtual address space.
3704 *
3705 * All requests are (re)mapped into kernel VA space.
3706 * Notice that we use b_bufsize for the size of the buffer
3707 * to be mapped. b_bcount might be modified by the driver.
3708 *
3709 * Note that even if the caller determines that the address space should
3710 * be valid, a race or a smaller-file mapped into a larger space may
3711 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3712 * check the return value.
3713 */
3714 int
3715 vmapbuf(struct buf *bp)
3716 {
3717 caddr_t addr, kva;
3718 vm_prot_t prot;
3719 int pidx, i;
3720 struct vm_page *m;
3721 struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3722
3723 if (bp->b_bufsize < 0)
3724 return (-1);
3725 prot = VM_PROT_READ;
3726 if (bp->b_iocmd == BIO_READ)
3727 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
3728 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3729 addr < bp->b_data + bp->b_bufsize;
3730 addr += PAGE_SIZE, pidx++) {
3731 /*
3732 * Do the vm_fault if needed; do the copy-on-write thing
3733 * when reading stuff off device into memory.
3734 *
3735 * NOTE! Must use pmap_extract() because addr may be in
3736 * the userland address space, and kextract is only guarenteed
3737 * to work for the kernland address space (see: sparc64 port).
3738 */
3739 retry:
3740 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3741 prot) < 0) {
3742 vm_page_lock_queues();
3743 for (i = 0; i < pidx; ++i) {
3744 vm_page_unhold(bp->b_pages[i]);
3745 bp->b_pages[i] = NULL;
3746 }
3747 vm_page_unlock_queues();
3748 return(-1);
3749 }
3750 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3751 if (m == NULL)
3752 goto retry;
3753 bp->b_pages[pidx] = m;
3754 }
3755 if (pidx > btoc(MAXPHYS))
3756 panic("vmapbuf: mapped more than MAXPHYS");
3757 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3758
3759 kva = bp->b_saveaddr;
3760 bp->b_npages = pidx;
3761 bp->b_saveaddr = bp->b_data;
3762 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3763 return(0);
3764 }
3765
3766 /*
3767 * Free the io map PTEs associated with this IO operation.
3768 * We also invalidate the TLB entries and restore the original b_addr.
3769 */
3770 void
3771 vunmapbuf(struct buf *bp)
3772 {
3773 int pidx;
3774 int npages;
3775
3776 npages = bp->b_npages;
3777 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3778 vm_page_lock_queues();
3779 for (pidx = 0; pidx < npages; pidx++)
3780 vm_page_unhold(bp->b_pages[pidx]);
3781 vm_page_unlock_queues();
3782
3783 bp->b_data = bp->b_saveaddr;
3784 }
3785
3786 void
3787 bdone(struct buf *bp)
3788 {
3789
3790 mtx_lock(&bdonelock);
3791 bp->b_flags |= B_DONE;
3792 wakeup(bp);
3793 mtx_unlock(&bdonelock);
3794 }
3795
3796 void
3797 bwait(struct buf *bp, u_char pri, const char *wchan)
3798 {
3799
3800 mtx_lock(&bdonelock);
3801 while ((bp->b_flags & B_DONE) == 0)
3802 msleep(bp, &bdonelock, pri, wchan, 0);
3803 mtx_unlock(&bdonelock);
3804 }
3805
3806 int
3807 bufsync(struct bufobj *bo, int waitfor, struct thread *td)
3808 {
3809
3810 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td));
3811 }
3812
3813 void
3814 bufstrategy(struct bufobj *bo, struct buf *bp)
3815 {
3816 int i = 0;
3817 struct vnode *vp;
3818
3819 vp = bp->b_vp;
3820 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3821 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3822 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3823 i = VOP_STRATEGY(vp, bp);
3824 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3825 }
3826
3827 void
3828 bufobj_wrefl(struct bufobj *bo)
3829 {
3830
3831 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3832 ASSERT_BO_LOCKED(bo);
3833 bo->bo_numoutput++;
3834 }
3835
3836 void
3837 bufobj_wref(struct bufobj *bo)
3838 {
3839
3840 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3841 BO_LOCK(bo);
3842 bo->bo_numoutput++;
3843 BO_UNLOCK(bo);
3844 }
3845
3846 void
3847 bufobj_wdrop(struct bufobj *bo)
3848 {
3849
3850 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3851 BO_LOCK(bo);
3852 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3853 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3854 bo->bo_flag &= ~BO_WWAIT;
3855 wakeup(&bo->bo_numoutput);
3856 }
3857 BO_UNLOCK(bo);
3858 }
3859
3860 int
3861 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3862 {
3863 int error;
3864
3865 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3866 ASSERT_BO_LOCKED(bo);
3867 error = 0;
3868 while (bo->bo_numoutput) {
3869 bo->bo_flag |= BO_WWAIT;
3870 error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3871 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3872 if (error)
3873 break;
3874 }
3875 return (error);
3876 }
3877
3878 void
3879 bpin(struct buf *bp)
3880 {
3881 mtx_lock(&bpinlock);
3882 bp->b_pin_count++;
3883 mtx_unlock(&bpinlock);
3884 }
3885
3886 void
3887 bunpin(struct buf *bp)
3888 {
3889 mtx_lock(&bpinlock);
3890 if (--bp->b_pin_count == 0)
3891 wakeup(bp);
3892 mtx_unlock(&bpinlock);
3893 }
3894
3895 void
3896 bunpin_wait(struct buf *bp)
3897 {
3898 mtx_lock(&bpinlock);
3899 while (bp->b_pin_count > 0)
3900 msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0);
3901 mtx_unlock(&bpinlock);
3902 }
3903
3904 #include "opt_ddb.h"
3905 #ifdef DDB
3906 #include <ddb/ddb.h>
3907
3908 /* DDB command to show buffer data */
3909 DB_SHOW_COMMAND(buffer, db_show_buffer)
3910 {
3911 /* get args */
3912 struct buf *bp = (struct buf *)addr;
3913
3914 if (!have_addr) {
3915 db_printf("usage: show buffer <addr>\n");
3916 return;
3917 }
3918
3919 db_printf("buf at %p\n", bp);
3920 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3921 db_printf(
3922 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3923 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n",
3924 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3925 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno);
3926 if (bp->b_npages) {
3927 int i;
3928 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3929 for (i = 0; i < bp->b_npages; i++) {
3930 vm_page_t m;
3931 m = bp->b_pages[i];
3932 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3933 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3934 if ((i + 1) < bp->b_npages)
3935 db_printf(",");
3936 }
3937 db_printf("\n");
3938 }
3939 lockmgr_printinfo(&bp->b_lock);
3940 }
3941
3942 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
3943 {
3944 struct buf *bp;
3945 int i;
3946
3947 for (i = 0; i < nbuf; i++) {
3948 bp = &buf[i];
3949 if (lockcount(&bp->b_lock)) {
3950 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
3951 db_printf("\n");
3952 }
3953 }
3954 }
3955 #endif /* DDB */
Cache object: 251bd70f53957153d9540a0dda387b01
|