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/fail.h>
52 #include <sys/limits.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mount.h>
56 #include <sys/mutex.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
59 #include <sys/proc.h>
60 #include <sys/resourcevar.h>
61 #include <sys/sysctl.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
64 #include <geom/geom.h>
65 #include <vm/vm.h>
66 #include <vm/vm_param.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_pageout.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_extern.h>
72 #include <vm/vm_map.h>
73 #include "opt_compat.h"
74 #include "opt_directio.h"
75 #include "opt_swap.h"
76
77 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
78
79 struct bio_ops bioops; /* I/O operation notification */
80
81 struct buf_ops buf_ops_bio = {
82 .bop_name = "buf_ops_bio",
83 .bop_write = bufwrite,
84 .bop_strategy = bufstrategy,
85 .bop_sync = bufsync,
86 .bop_bdflush = bufbdflush,
87 };
88
89 /*
90 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
91 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
92 */
93 struct buf *buf; /* buffer header pool */
94
95 static struct proc *bufdaemonproc;
96
97 static int inmem(struct vnode *vp, daddr_t blkno);
98 static void vm_hold_free_pages(struct buf *bp, int newbsize);
99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
100 vm_offset_t to);
101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
103 vm_page_t m);
104 static void vfs_drain_busy_pages(struct buf *bp);
105 static void vfs_clean_pages_dirty_buf(struct buf *bp);
106 static void vfs_setdirty_locked_object(struct buf *bp);
107 static void vfs_vmio_release(struct buf *bp);
108 static int vfs_bio_clcheck(struct vnode *vp, int size,
109 daddr_t lblkno, daddr_t blkno);
110 static int buf_do_flush(struct vnode *vp);
111 static int flushbufqueues(struct vnode *, int, int);
112 static void buf_daemon(void);
113 static void bremfreel(struct buf *bp);
114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
115 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
117 #endif
118
119 int vmiodirenable = TRUE;
120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
121 "Use the VM system for directory writes");
122 long runningbufspace;
123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
124 "Amount of presently outstanding async buffer io");
125 static long bufspace;
126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
127 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
129 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
130 #else
131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
132 "Virtual memory used for buffers");
133 #endif
134 static long maxbufspace;
135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
136 "Maximum allowed value of bufspace (including buf_daemon)");
137 static long bufmallocspace;
138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
139 "Amount of malloced memory for buffers");
140 static long maxbufmallocspace;
141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
142 "Maximum amount of malloced memory for buffers");
143 static long lobufspace;
144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
145 "Minimum amount of buffers we want to have");
146 long hibufspace;
147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
148 "Maximum allowed value of bufspace (excluding buf_daemon)");
149 static int bufreusecnt;
150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
151 "Number of times we have reused a buffer");
152 static int buffreekvacnt;
153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
154 "Number of times we have freed the KVA space from some buffer");
155 static int bufdefragcnt;
156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
157 "Number of times we have had to repeat buffer allocation to defragment");
158 static long lorunningspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
160 "Minimum preferred space used for in-progress I/O");
161 static long hirunningspace;
162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
163 "Maximum amount of space to use for in-progress I/O");
164 int dirtybufferflushes;
165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
166 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
167 int bdwriteskip;
168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
169 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
170 int altbufferflushes;
171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
172 0, "Number of fsync flushes to limit dirty buffers");
173 static int recursiveflushes;
174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
175 0, "Number of flushes skipped due to being recursive");
176 static int numdirtybuffers;
177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
178 "Number of buffers that are dirty (has unwritten changes) at the moment");
179 static int lodirtybuffers;
180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
181 "How many buffers we want to have free before bufdaemon can sleep");
182 static int hidirtybuffers;
183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
184 "When the number of dirty buffers is considered severe");
185 int dirtybufthresh;
186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
187 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
188 static int numfreebuffers;
189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
190 "Number of free buffers");
191 static int lofreebuffers;
192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
193 "XXX Unused");
194 static int hifreebuffers;
195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
196 "XXX Complicatedly unused");
197 static int getnewbufcalls;
198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
199 "Number of calls to getnewbuf");
200 static int getnewbufrestarts;
201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
202 "Number of times getnewbuf has had to restart a buffer aquisition");
203 static int flushbufqtarget = 100;
204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
205 "Amount of work to do in flushbufqueues when helping bufdaemon");
206 static long notbufdflashes;
207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0,
208 "Number of dirty buffer flushes done by the bufdaemon helpers");
209 static long barrierwrites;
210 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
211 "Number of barrier writes");
212
213 /*
214 * Wakeup point for bufdaemon, as well as indicator of whether it is already
215 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
216 * is idling.
217 */
218 static int bd_request;
219
220 /*
221 * This lock synchronizes access to bd_request.
222 */
223 static struct mtx bdlock;
224
225 /*
226 * bogus page -- for I/O to/from partially complete buffers
227 * this is a temporary solution to the problem, but it is not
228 * really that bad. it would be better to split the buffer
229 * for input in the case of buffers partially already in memory,
230 * but the code is intricate enough already.
231 */
232 vm_page_t bogus_page;
233
234 /*
235 * Synchronization (sleep/wakeup) variable for active buffer space requests.
236 * Set when wait starts, cleared prior to wakeup().
237 * Used in runningbufwakeup() and waitrunningbufspace().
238 */
239 static int runningbufreq;
240
241 /*
242 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
243 * waitrunningbufspace().
244 */
245 static struct mtx rbreqlock;
246
247 /*
248 * Synchronization (sleep/wakeup) variable for buffer requests.
249 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
250 * by and/or.
251 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
252 * getnewbuf(), and getblk().
253 */
254 static int needsbuffer;
255
256 /*
257 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
258 */
259 static struct mtx nblock;
260
261 /*
262 * Definitions for the buffer free lists.
263 */
264 #define BUFFER_QUEUES 6 /* number of free buffer queues */
265
266 #define QUEUE_NONE 0 /* on no queue */
267 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
268 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
269 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */
270 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */
271 #define QUEUE_EMPTY 5 /* empty buffer headers */
272 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */
273
274 /* Queues for free buffers with various properties */
275 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
276
277 /* Lock for the bufqueues */
278 static struct mtx bqlock;
279
280 /*
281 * Single global constant for BUF_WMESG, to avoid getting multiple references.
282 * buf_wmesg is referred from macros.
283 */
284 const char *buf_wmesg = BUF_WMESG;
285
286 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
287 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
288 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
289 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
290
291 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
292 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
293 static int
294 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
295 {
296 long lvalue;
297 int ivalue;
298
299 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
300 return (sysctl_handle_long(oidp, arg1, arg2, req));
301 lvalue = *(long *)arg1;
302 if (lvalue > INT_MAX)
303 /* On overflow, still write out a long to trigger ENOMEM. */
304 return (sysctl_handle_long(oidp, &lvalue, 0, req));
305 ivalue = lvalue;
306 return (sysctl_handle_int(oidp, &ivalue, 0, req));
307 }
308 #endif
309
310 #ifdef DIRECTIO
311 extern void ffs_rawread_setup(void);
312 #endif /* DIRECTIO */
313 /*
314 * numdirtywakeup:
315 *
316 * If someone is blocked due to there being too many dirty buffers,
317 * and numdirtybuffers is now reasonable, wake them up.
318 */
319
320 static __inline void
321 numdirtywakeup(int level)
322 {
323
324 if (numdirtybuffers <= level) {
325 mtx_lock(&nblock);
326 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
327 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
328 wakeup(&needsbuffer);
329 }
330 mtx_unlock(&nblock);
331 }
332 }
333
334 /*
335 * bufspacewakeup:
336 *
337 * Called when buffer space is potentially available for recovery.
338 * getnewbuf() will block on this flag when it is unable to free
339 * sufficient buffer space. Buffer space becomes recoverable when
340 * bp's get placed back in the queues.
341 */
342
343 static __inline void
344 bufspacewakeup(void)
345 {
346
347 /*
348 * If someone is waiting for BUF space, wake them up. Even
349 * though we haven't freed the kva space yet, the waiting
350 * process will be able to now.
351 */
352 mtx_lock(&nblock);
353 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
354 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
355 wakeup(&needsbuffer);
356 }
357 mtx_unlock(&nblock);
358 }
359
360 /*
361 * runningbufwakeup() - in-progress I/O accounting.
362 *
363 */
364 void
365 runningbufwakeup(struct buf *bp)
366 {
367
368 if (bp->b_runningbufspace) {
369 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
370 bp->b_runningbufspace = 0;
371 mtx_lock(&rbreqlock);
372 if (runningbufreq && runningbufspace <= lorunningspace) {
373 runningbufreq = 0;
374 wakeup(&runningbufreq);
375 }
376 mtx_unlock(&rbreqlock);
377 }
378 }
379
380 /*
381 * bufcountwakeup:
382 *
383 * Called when a buffer has been added to one of the free queues to
384 * account for the buffer and to wakeup anyone waiting for free buffers.
385 * This typically occurs when large amounts of metadata are being handled
386 * by the buffer cache ( else buffer space runs out first, usually ).
387 */
388
389 static __inline void
390 bufcountwakeup(struct buf *bp)
391 {
392 int old;
393
394 KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
395 ("buf %p already counted as free", bp));
396 if (bp->b_bufobj != NULL)
397 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
398 bp->b_vflags |= BV_INFREECNT;
399 old = atomic_fetchadd_int(&numfreebuffers, 1);
400 KASSERT(old >= 0 && old < nbuf,
401 ("numfreebuffers climbed to %d", old + 1));
402 mtx_lock(&nblock);
403 if (needsbuffer) {
404 needsbuffer &= ~VFS_BIO_NEED_ANY;
405 if (numfreebuffers >= hifreebuffers)
406 needsbuffer &= ~VFS_BIO_NEED_FREE;
407 wakeup(&needsbuffer);
408 }
409 mtx_unlock(&nblock);
410 }
411
412 /*
413 * waitrunningbufspace()
414 *
415 * runningbufspace is a measure of the amount of I/O currently
416 * running. This routine is used in async-write situations to
417 * prevent creating huge backups of pending writes to a device.
418 * Only asynchronous writes are governed by this function.
419 *
420 * Reads will adjust runningbufspace, but will not block based on it.
421 * The read load has a side effect of reducing the allowed write load.
422 *
423 * This does NOT turn an async write into a sync write. It waits
424 * for earlier writes to complete and generally returns before the
425 * caller's write has reached the device.
426 */
427 void
428 waitrunningbufspace(void)
429 {
430
431 mtx_lock(&rbreqlock);
432 while (runningbufspace > hirunningspace) {
433 ++runningbufreq;
434 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
435 }
436 mtx_unlock(&rbreqlock);
437 }
438
439
440 /*
441 * vfs_buf_test_cache:
442 *
443 * Called when a buffer is extended. This function clears the B_CACHE
444 * bit if the newly extended portion of the buffer does not contain
445 * valid data.
446 */
447 static __inline
448 void
449 vfs_buf_test_cache(struct buf *bp,
450 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
451 vm_page_t m)
452 {
453
454 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
455 if (bp->b_flags & B_CACHE) {
456 int base = (foff + off) & PAGE_MASK;
457 if (vm_page_is_valid(m, base, size) == 0)
458 bp->b_flags &= ~B_CACHE;
459 }
460 }
461
462 /* Wake up the buffer daemon if necessary */
463 static __inline
464 void
465 bd_wakeup(int dirtybuflevel)
466 {
467
468 mtx_lock(&bdlock);
469 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
470 bd_request = 1;
471 wakeup(&bd_request);
472 }
473 mtx_unlock(&bdlock);
474 }
475
476 /*
477 * bd_speedup - speedup the buffer cache flushing code
478 */
479
480 static __inline
481 void
482 bd_speedup(void)
483 {
484
485 bd_wakeup(1);
486 }
487
488 /*
489 * Calculating buffer cache scaling values and reserve space for buffer
490 * headers. This is called during low level kernel initialization and
491 * may be called more then once. We CANNOT write to the memory area
492 * being reserved at this time.
493 */
494 caddr_t
495 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
496 {
497 int tuned_nbuf;
498 long maxbuf;
499
500 /*
501 * physmem_est is in pages. Convert it to kilobytes (assumes
502 * PAGE_SIZE is >= 1K)
503 */
504 physmem_est = physmem_est * (PAGE_SIZE / 1024);
505
506 /*
507 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
508 * For the first 64MB of ram nominally allocate sufficient buffers to
509 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
510 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
511 * the buffer cache we limit the eventual kva reservation to
512 * maxbcache bytes.
513 *
514 * factor represents the 1/4 x ram conversion.
515 */
516 if (nbuf == 0) {
517 int factor = 4 * BKVASIZE / 1024;
518
519 nbuf = 50;
520 if (physmem_est > 4096)
521 nbuf += min((physmem_est - 4096) / factor,
522 65536 / factor);
523 if (physmem_est > 65536)
524 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
525
526 if (maxbcache && nbuf > maxbcache / BKVASIZE)
527 nbuf = maxbcache / BKVASIZE;
528 tuned_nbuf = 1;
529 } else
530 tuned_nbuf = 0;
531
532 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
533 maxbuf = (LONG_MAX / 3) / BKVASIZE;
534 if (nbuf > maxbuf) {
535 if (!tuned_nbuf)
536 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
537 maxbuf);
538 nbuf = maxbuf;
539 }
540
541 /*
542 * swbufs are used as temporary holders for I/O, such as paging I/O.
543 * We have no less then 16 and no more then 256.
544 */
545 nswbuf = max(min(nbuf/4, 256), 16);
546 #ifdef NSWBUF_MIN
547 if (nswbuf < NSWBUF_MIN)
548 nswbuf = NSWBUF_MIN;
549 #endif
550 #ifdef DIRECTIO
551 ffs_rawread_setup();
552 #endif
553
554 /*
555 * Reserve space for the buffer cache buffers
556 */
557 swbuf = (void *)v;
558 v = (caddr_t)(swbuf + nswbuf);
559 buf = (void *)v;
560 v = (caddr_t)(buf + nbuf);
561
562 return(v);
563 }
564
565 /* Initialize the buffer subsystem. Called before use of any buffers. */
566 void
567 bufinit(void)
568 {
569 struct buf *bp;
570 int i;
571
572 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
573 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
574 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
575 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
576
577 /* next, make a null set of free lists */
578 for (i = 0; i < BUFFER_QUEUES; i++)
579 TAILQ_INIT(&bufqueues[i]);
580
581 /* finally, initialize each buffer header and stick on empty q */
582 for (i = 0; i < nbuf; i++) {
583 bp = &buf[i];
584 bzero(bp, sizeof *bp);
585 bp->b_flags = B_INVAL; /* we're just an empty header */
586 bp->b_rcred = NOCRED;
587 bp->b_wcred = NOCRED;
588 bp->b_qindex = QUEUE_EMPTY;
589 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */
590 bp->b_xflags = 0;
591 LIST_INIT(&bp->b_dep);
592 BUF_LOCKINIT(bp);
593 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
594 }
595
596 /*
597 * maxbufspace is the absolute maximum amount of buffer space we are
598 * allowed to reserve in KVM and in real terms. The absolute maximum
599 * is nominally used by buf_daemon. hibufspace is the nominal maximum
600 * used by most other processes. The differential is required to
601 * ensure that buf_daemon is able to run when other processes might
602 * be blocked waiting for buffer space.
603 *
604 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
605 * this may result in KVM fragmentation which is not handled optimally
606 * by the system.
607 */
608 maxbufspace = (long)nbuf * BKVASIZE;
609 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
610 lobufspace = hibufspace - MAXBSIZE;
611
612 lorunningspace = 512 * 1024;
613 hirunningspace = 1024 * 1024;
614
615 /*
616 * Limit the amount of malloc memory since it is wired permanently into
617 * the kernel space. Even though this is accounted for in the buffer
618 * allocation, we don't want the malloced region to grow uncontrolled.
619 * The malloc scheme improves memory utilization significantly on average
620 * (small) directories.
621 */
622 maxbufmallocspace = hibufspace / 20;
623
624 /*
625 * Reduce the chance of a deadlock occuring by limiting the number
626 * of delayed-write dirty buffers we allow to stack up.
627 */
628 hidirtybuffers = nbuf / 4 + 20;
629 dirtybufthresh = hidirtybuffers * 9 / 10;
630 numdirtybuffers = 0;
631 /*
632 * To support extreme low-memory systems, make sure hidirtybuffers cannot
633 * eat up all available buffer space. This occurs when our minimum cannot
634 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
635 * BKVASIZE'd (8K) buffers.
636 */
637 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
638 hidirtybuffers >>= 1;
639 }
640 lodirtybuffers = hidirtybuffers / 2;
641
642 /*
643 * Try to keep the number of free buffers in the specified range,
644 * and give special processes (e.g. like buf_daemon) access to an
645 * emergency reserve.
646 */
647 lofreebuffers = nbuf / 18 + 5;
648 hifreebuffers = 2 * lofreebuffers;
649 numfreebuffers = nbuf;
650
651 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
652 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
653 }
654
655 /*
656 * bfreekva() - free the kva allocation for a buffer.
657 *
658 * Since this call frees up buffer space, we call bufspacewakeup().
659 */
660 static void
661 bfreekva(struct buf *bp)
662 {
663
664 if (bp->b_kvasize) {
665 atomic_add_int(&buffreekvacnt, 1);
666 atomic_subtract_long(&bufspace, bp->b_kvasize);
667 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
668 (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
669 bp->b_kvasize = 0;
670 bufspacewakeup();
671 }
672 }
673
674 /*
675 * bremfree:
676 *
677 * Mark the buffer for removal from the appropriate free list in brelse.
678 *
679 */
680 void
681 bremfree(struct buf *bp)
682 {
683 int old;
684
685 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
686 KASSERT((bp->b_flags & B_REMFREE) == 0,
687 ("bremfree: buffer %p already marked for delayed removal.", bp));
688 KASSERT(bp->b_qindex != QUEUE_NONE,
689 ("bremfree: buffer %p not on a queue.", bp));
690 BUF_ASSERT_HELD(bp);
691
692 bp->b_flags |= B_REMFREE;
693 /* Fixup numfreebuffers count. */
694 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
695 KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
696 ("buf %p not counted in numfreebuffers", bp));
697 if (bp->b_bufobj != NULL)
698 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
699 bp->b_vflags &= ~BV_INFREECNT;
700 old = atomic_fetchadd_int(&numfreebuffers, -1);
701 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
702 }
703 }
704
705 /*
706 * bremfreef:
707 *
708 * Force an immediate removal from a free list. Used only in nfs when
709 * it abuses the b_freelist pointer.
710 */
711 void
712 bremfreef(struct buf *bp)
713 {
714 mtx_lock(&bqlock);
715 bremfreel(bp);
716 mtx_unlock(&bqlock);
717 }
718
719 /*
720 * bremfreel:
721 *
722 * Removes a buffer from the free list, must be called with the
723 * bqlock held.
724 */
725 static void
726 bremfreel(struct buf *bp)
727 {
728 int old;
729
730 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
731 bp, bp->b_vp, bp->b_flags);
732 KASSERT(bp->b_qindex != QUEUE_NONE,
733 ("bremfreel: buffer %p not on a queue.", bp));
734 BUF_ASSERT_HELD(bp);
735 mtx_assert(&bqlock, MA_OWNED);
736
737 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
738 bp->b_qindex = QUEUE_NONE;
739 /*
740 * If this was a delayed bremfree() we only need to remove the buffer
741 * from the queue and return the stats are already done.
742 */
743 if (bp->b_flags & B_REMFREE) {
744 bp->b_flags &= ~B_REMFREE;
745 return;
746 }
747 /*
748 * Fixup numfreebuffers count. If the buffer is invalid or not
749 * delayed-write, the buffer was free and we must decrement
750 * numfreebuffers.
751 */
752 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
753 KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
754 ("buf %p not counted in numfreebuffers", bp));
755 if (bp->b_bufobj != NULL)
756 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
757 bp->b_vflags &= ~BV_INFREECNT;
758 old = atomic_fetchadd_int(&numfreebuffers, -1);
759 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
760 }
761 }
762
763
764 /*
765 * Get a buffer with the specified data. Look in the cache first. We
766 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
767 * is set, the buffer is valid and we do not have to do anything ( see
768 * getblk() ). This is really just a special case of breadn().
769 */
770 int
771 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
772 struct buf **bpp)
773 {
774
775 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
776 }
777
778 /*
779 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
780 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
781 * the buffer is valid and we do not have to do anything.
782 */
783 void
784 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
785 int cnt, struct ucred * cred)
786 {
787 struct buf *rabp;
788 int i;
789
790 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
791 if (inmem(vp, *rablkno))
792 continue;
793 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
794
795 if ((rabp->b_flags & B_CACHE) == 0) {
796 if (!TD_IS_IDLETHREAD(curthread))
797 curthread->td_ru.ru_inblock++;
798 rabp->b_flags |= B_ASYNC;
799 rabp->b_flags &= ~B_INVAL;
800 rabp->b_ioflags &= ~BIO_ERROR;
801 rabp->b_iocmd = BIO_READ;
802 if (rabp->b_rcred == NOCRED && cred != NOCRED)
803 rabp->b_rcred = crhold(cred);
804 vfs_busy_pages(rabp, 0);
805 BUF_KERNPROC(rabp);
806 rabp->b_iooffset = dbtob(rabp->b_blkno);
807 bstrategy(rabp);
808 } else {
809 brelse(rabp);
810 }
811 }
812 }
813
814 /*
815 * Operates like bread, but also starts asynchronous I/O on
816 * read-ahead blocks.
817 */
818 int
819 breadn(struct vnode * vp, daddr_t blkno, int size,
820 daddr_t * rablkno, int *rabsize,
821 int cnt, struct ucred * cred, struct buf **bpp)
822 {
823 struct buf *bp;
824 int rv = 0, readwait = 0;
825
826 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
827 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
828
829 /* if not found in cache, do some I/O */
830 if ((bp->b_flags & B_CACHE) == 0) {
831 if (!TD_IS_IDLETHREAD(curthread))
832 curthread->td_ru.ru_inblock++;
833 bp->b_iocmd = BIO_READ;
834 bp->b_flags &= ~B_INVAL;
835 bp->b_ioflags &= ~BIO_ERROR;
836 if (bp->b_rcred == NOCRED && cred != NOCRED)
837 bp->b_rcred = crhold(cred);
838 vfs_busy_pages(bp, 0);
839 bp->b_iooffset = dbtob(bp->b_blkno);
840 bstrategy(bp);
841 ++readwait;
842 }
843
844 breada(vp, rablkno, rabsize, cnt, cred);
845
846 if (readwait) {
847 rv = bufwait(bp);
848 }
849 return (rv);
850 }
851
852 /*
853 * Write, release buffer on completion. (Done by iodone
854 * if async). Do not bother writing anything if the buffer
855 * is invalid.
856 *
857 * Note that we set B_CACHE here, indicating that buffer is
858 * fully valid and thus cacheable. This is true even of NFS
859 * now so we set it generally. This could be set either here
860 * or in biodone() since the I/O is synchronous. We put it
861 * here.
862 */
863 int
864 bufwrite(struct buf *bp)
865 {
866 int oldflags;
867 struct vnode *vp;
868 int vp_md;
869
870 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
871 if (bp->b_flags & B_INVAL) {
872 brelse(bp);
873 return (0);
874 }
875
876 if (bp->b_flags & B_BARRIER)
877 barrierwrites++;
878
879 oldflags = bp->b_flags;
880
881 BUF_ASSERT_HELD(bp);
882
883 if (bp->b_pin_count > 0)
884 bunpin_wait(bp);
885
886 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
887 ("FFS background buffer should not get here %p", bp));
888
889 vp = bp->b_vp;
890 if (vp)
891 vp_md = vp->v_vflag & VV_MD;
892 else
893 vp_md = 0;
894
895 /*
896 * Mark the buffer clean. Increment the bufobj write count
897 * before bundirty() call, to prevent other thread from seeing
898 * empty dirty list and zero counter for writes in progress,
899 * falsely indicating that the bufobj is clean.
900 */
901 bufobj_wref(bp->b_bufobj);
902 bundirty(bp);
903
904 bp->b_flags &= ~B_DONE;
905 bp->b_ioflags &= ~BIO_ERROR;
906 bp->b_flags |= B_CACHE;
907 bp->b_iocmd = BIO_WRITE;
908
909 vfs_busy_pages(bp, 1);
910
911 /*
912 * Normal bwrites pipeline writes
913 */
914 bp->b_runningbufspace = bp->b_bufsize;
915 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
916
917 if (!TD_IS_IDLETHREAD(curthread))
918 curthread->td_ru.ru_oublock++;
919 if (oldflags & B_ASYNC)
920 BUF_KERNPROC(bp);
921 bp->b_iooffset = dbtob(bp->b_blkno);
922 bstrategy(bp);
923
924 if ((oldflags & B_ASYNC) == 0) {
925 int rtval = bufwait(bp);
926 brelse(bp);
927 return (rtval);
928 } else {
929 /*
930 * don't allow the async write to saturate the I/O
931 * system. We will not deadlock here because
932 * we are blocking waiting for I/O that is already in-progress
933 * to complete. We do not block here if it is the update
934 * or syncer daemon trying to clean up as that can lead
935 * to deadlock.
936 */
937 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
938 waitrunningbufspace();
939 }
940
941 return (0);
942 }
943
944 void
945 bufbdflush(struct bufobj *bo, struct buf *bp)
946 {
947 struct buf *nbp;
948
949 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
950 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
951 altbufferflushes++;
952 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
953 BO_LOCK(bo);
954 /*
955 * Try to find a buffer to flush.
956 */
957 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
958 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
959 BUF_LOCK(nbp,
960 LK_EXCLUSIVE | LK_NOWAIT, NULL))
961 continue;
962 if (bp == nbp)
963 panic("bdwrite: found ourselves");
964 BO_UNLOCK(bo);
965 /* Don't countdeps with the bo lock held. */
966 if (buf_countdeps(nbp, 0)) {
967 BO_LOCK(bo);
968 BUF_UNLOCK(nbp);
969 continue;
970 }
971 if (nbp->b_flags & B_CLUSTEROK) {
972 vfs_bio_awrite(nbp);
973 } else {
974 bremfree(nbp);
975 bawrite(nbp);
976 }
977 dirtybufferflushes++;
978 break;
979 }
980 if (nbp == NULL)
981 BO_UNLOCK(bo);
982 }
983 }
984
985 /*
986 * Delayed write. (Buffer is marked dirty). Do not bother writing
987 * anything if the buffer is marked invalid.
988 *
989 * Note that since the buffer must be completely valid, we can safely
990 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
991 * biodone() in order to prevent getblk from writing the buffer
992 * out synchronously.
993 */
994 void
995 bdwrite(struct buf *bp)
996 {
997 struct thread *td = curthread;
998 struct vnode *vp;
999 struct bufobj *bo;
1000
1001 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1002 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1003 KASSERT((bp->b_flags & B_BARRIER) == 0,
1004 ("Barrier request in delayed write %p", bp));
1005 BUF_ASSERT_HELD(bp);
1006
1007 if (bp->b_flags & B_INVAL) {
1008 brelse(bp);
1009 return;
1010 }
1011
1012 /*
1013 * If we have too many dirty buffers, don't create any more.
1014 * If we are wildly over our limit, then force a complete
1015 * cleanup. Otherwise, just keep the situation from getting
1016 * out of control. Note that we have to avoid a recursive
1017 * disaster and not try to clean up after our own cleanup!
1018 */
1019 vp = bp->b_vp;
1020 bo = bp->b_bufobj;
1021 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1022 td->td_pflags |= TDP_INBDFLUSH;
1023 BO_BDFLUSH(bo, bp);
1024 td->td_pflags &= ~TDP_INBDFLUSH;
1025 } else
1026 recursiveflushes++;
1027
1028 bdirty(bp);
1029 /*
1030 * Set B_CACHE, indicating that the buffer is fully valid. This is
1031 * true even of NFS now.
1032 */
1033 bp->b_flags |= B_CACHE;
1034
1035 /*
1036 * This bmap keeps the system from needing to do the bmap later,
1037 * perhaps when the system is attempting to do a sync. Since it
1038 * is likely that the indirect block -- or whatever other datastructure
1039 * that the filesystem needs is still in memory now, it is a good
1040 * thing to do this. Note also, that if the pageout daemon is
1041 * requesting a sync -- there might not be enough memory to do
1042 * the bmap then... So, this is important to do.
1043 */
1044 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1045 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1046 }
1047
1048 /*
1049 * Set the *dirty* buffer range based upon the VM system dirty
1050 * pages.
1051 *
1052 * Mark the buffer pages as clean. We need to do this here to
1053 * satisfy the vnode_pager and the pageout daemon, so that it
1054 * thinks that the pages have been "cleaned". Note that since
1055 * the pages are in a delayed write buffer -- the VFS layer
1056 * "will" see that the pages get written out on the next sync,
1057 * or perhaps the cluster will be completed.
1058 */
1059 vfs_clean_pages_dirty_buf(bp);
1060 bqrelse(bp);
1061
1062 /*
1063 * Wakeup the buffer flushing daemon if we have a lot of dirty
1064 * buffers (midpoint between our recovery point and our stall
1065 * point).
1066 */
1067 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1068
1069 /*
1070 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1071 * due to the softdep code.
1072 */
1073 }
1074
1075 /*
1076 * bdirty:
1077 *
1078 * Turn buffer into delayed write request. We must clear BIO_READ and
1079 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1080 * itself to properly update it in the dirty/clean lists. We mark it
1081 * B_DONE to ensure that any asynchronization of the buffer properly
1082 * clears B_DONE ( else a panic will occur later ).
1083 *
1084 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1085 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1086 * should only be called if the buffer is known-good.
1087 *
1088 * Since the buffer is not on a queue, we do not update the numfreebuffers
1089 * count.
1090 *
1091 * The buffer must be on QUEUE_NONE.
1092 */
1093 void
1094 bdirty(struct buf *bp)
1095 {
1096
1097 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1098 bp, bp->b_vp, bp->b_flags);
1099 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1100 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1101 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1102 BUF_ASSERT_HELD(bp);
1103 bp->b_flags &= ~(B_RELBUF);
1104 bp->b_iocmd = BIO_WRITE;
1105
1106 if ((bp->b_flags & B_DELWRI) == 0) {
1107 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1108 reassignbuf(bp);
1109 atomic_add_int(&numdirtybuffers, 1);
1110 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1111 }
1112 }
1113
1114 /*
1115 * bundirty:
1116 *
1117 * Clear B_DELWRI for buffer.
1118 *
1119 * Since the buffer is not on a queue, we do not update the numfreebuffers
1120 * count.
1121 *
1122 * The buffer must be on QUEUE_NONE.
1123 */
1124
1125 void
1126 bundirty(struct buf *bp)
1127 {
1128
1129 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1130 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1131 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1132 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1133 BUF_ASSERT_HELD(bp);
1134
1135 if (bp->b_flags & B_DELWRI) {
1136 bp->b_flags &= ~B_DELWRI;
1137 reassignbuf(bp);
1138 atomic_subtract_int(&numdirtybuffers, 1);
1139 numdirtywakeup(lodirtybuffers);
1140 }
1141 /*
1142 * Since it is now being written, we can clear its deferred write flag.
1143 */
1144 bp->b_flags &= ~B_DEFERRED;
1145 }
1146
1147 /*
1148 * bawrite:
1149 *
1150 * Asynchronous write. Start output on a buffer, but do not wait for
1151 * it to complete. The buffer is released when the output completes.
1152 *
1153 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1154 * B_INVAL buffers. Not us.
1155 */
1156 void
1157 bawrite(struct buf *bp)
1158 {
1159
1160 bp->b_flags |= B_ASYNC;
1161 (void) bwrite(bp);
1162 }
1163
1164 /*
1165 * babarrierwrite:
1166 *
1167 * Asynchronous barrier write. Start output on a buffer, but do not
1168 * wait for it to complete. Place a write barrier after this write so
1169 * that this buffer and all buffers written before it are committed to
1170 * the disk before any buffers written after this write are committed
1171 * to the disk. The buffer is released when the output completes.
1172 */
1173 void
1174 babarrierwrite(struct buf *bp)
1175 {
1176
1177 bp->b_flags |= B_ASYNC | B_BARRIER;
1178 (void) bwrite(bp);
1179 }
1180
1181 /*
1182 * bbarrierwrite:
1183 *
1184 * Synchronous barrier write. Start output on a buffer and wait for
1185 * it to complete. Place a write barrier after this write so that
1186 * this buffer and all buffers written before it are committed to
1187 * the disk before any buffers written after this write are committed
1188 * to the disk. The buffer is released when the output completes.
1189 */
1190 int
1191 bbarrierwrite(struct buf *bp)
1192 {
1193
1194 bp->b_flags |= B_BARRIER;
1195 return (bwrite(bp));
1196 }
1197
1198 /*
1199 * bwillwrite:
1200 *
1201 * Called prior to the locking of any vnodes when we are expecting to
1202 * write. We do not want to starve the buffer cache with too many
1203 * dirty buffers so we block here. By blocking prior to the locking
1204 * of any vnodes we attempt to avoid the situation where a locked vnode
1205 * prevents the various system daemons from flushing related buffers.
1206 */
1207
1208 void
1209 bwillwrite(void)
1210 {
1211
1212 if (numdirtybuffers >= hidirtybuffers) {
1213 mtx_lock(&nblock);
1214 while (numdirtybuffers >= hidirtybuffers) {
1215 bd_wakeup(1);
1216 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1217 msleep(&needsbuffer, &nblock,
1218 (PRIBIO + 4), "flswai", 0);
1219 }
1220 mtx_unlock(&nblock);
1221 }
1222 }
1223
1224 /*
1225 * Return true if we have too many dirty buffers.
1226 */
1227 int
1228 buf_dirty_count_severe(void)
1229 {
1230
1231 return(numdirtybuffers >= hidirtybuffers);
1232 }
1233
1234 static __noinline int
1235 buf_vm_page_count_severe(void)
1236 {
1237
1238 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1239
1240 return vm_page_count_severe();
1241 }
1242
1243 /*
1244 * brelse:
1245 *
1246 * Release a busy buffer and, if requested, free its resources. The
1247 * buffer will be stashed in the appropriate bufqueue[] allowing it
1248 * to be accessed later as a cache entity or reused for other purposes.
1249 */
1250 void
1251 brelse(struct buf *bp)
1252 {
1253 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1254 bp, bp->b_vp, bp->b_flags);
1255 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1256 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1257
1258 if (bp->b_flags & B_MANAGED) {
1259 bqrelse(bp);
1260 return;
1261 }
1262
1263 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1264 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1265 /*
1266 * Failed write, redirty. Must clear BIO_ERROR to prevent
1267 * pages from being scrapped. If the error is anything
1268 * other than an I/O error (EIO), assume that retrying
1269 * is futile.
1270 */
1271 bp->b_ioflags &= ~BIO_ERROR;
1272 bdirty(bp);
1273 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1274 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1275 /*
1276 * Either a failed I/O or we were asked to free or not
1277 * cache the buffer.
1278 */
1279 bp->b_flags |= B_INVAL;
1280 if (!LIST_EMPTY(&bp->b_dep))
1281 buf_deallocate(bp);
1282 if (bp->b_flags & B_DELWRI) {
1283 atomic_subtract_int(&numdirtybuffers, 1);
1284 numdirtywakeup(lodirtybuffers);
1285 }
1286 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1287 if ((bp->b_flags & B_VMIO) == 0) {
1288 if (bp->b_bufsize)
1289 allocbuf(bp, 0);
1290 if (bp->b_vp)
1291 brelvp(bp);
1292 }
1293 }
1294
1295 /*
1296 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1297 * is called with B_DELWRI set, the underlying pages may wind up
1298 * getting freed causing a previous write (bdwrite()) to get 'lost'
1299 * because pages associated with a B_DELWRI bp are marked clean.
1300 *
1301 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1302 * if B_DELWRI is set.
1303 *
1304 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1305 * on pages to return pages to the VM page queues.
1306 */
1307 if (bp->b_flags & B_DELWRI)
1308 bp->b_flags &= ~B_RELBUF;
1309 else if (buf_vm_page_count_severe()) {
1310 /*
1311 * The locking of the BO_LOCK is not necessary since
1312 * BKGRDINPROG cannot be set while we hold the buf
1313 * lock, it can only be cleared if it is already
1314 * pending.
1315 */
1316 if (bp->b_vp) {
1317 if (!(bp->b_vflags & BV_BKGRDINPROG))
1318 bp->b_flags |= B_RELBUF;
1319 } else
1320 bp->b_flags |= B_RELBUF;
1321 }
1322
1323 /*
1324 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1325 * constituted, not even NFS buffers now. Two flags effect this. If
1326 * B_INVAL, the struct buf is invalidated but the VM object is kept
1327 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1328 *
1329 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1330 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1331 * buffer is also B_INVAL because it hits the re-dirtying code above.
1332 *
1333 * Normally we can do this whether a buffer is B_DELWRI or not. If
1334 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1335 * the commit state and we cannot afford to lose the buffer. If the
1336 * buffer has a background write in progress, we need to keep it
1337 * around to prevent it from being reconstituted and starting a second
1338 * background write.
1339 */
1340 if ((bp->b_flags & B_VMIO)
1341 && !(bp->b_vp->v_mount != NULL &&
1342 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1343 !vn_isdisk(bp->b_vp, NULL) &&
1344 (bp->b_flags & B_DELWRI))
1345 ) {
1346
1347 int i, j, resid;
1348 vm_page_t m;
1349 off_t foff;
1350 vm_pindex_t poff;
1351 vm_object_t obj;
1352
1353 obj = bp->b_bufobj->bo_object;
1354
1355 /*
1356 * Get the base offset and length of the buffer. Note that
1357 * in the VMIO case if the buffer block size is not
1358 * page-aligned then b_data pointer may not be page-aligned.
1359 * But our b_pages[] array *IS* page aligned.
1360 *
1361 * block sizes less then DEV_BSIZE (usually 512) are not
1362 * supported due to the page granularity bits (m->valid,
1363 * m->dirty, etc...).
1364 *
1365 * See man buf(9) for more information
1366 */
1367 resid = bp->b_bufsize;
1368 foff = bp->b_offset;
1369 VM_OBJECT_LOCK(obj);
1370 for (i = 0; i < bp->b_npages; i++) {
1371 int had_bogus = 0;
1372
1373 m = bp->b_pages[i];
1374
1375 /*
1376 * If we hit a bogus page, fixup *all* the bogus pages
1377 * now.
1378 */
1379 if (m == bogus_page) {
1380 poff = OFF_TO_IDX(bp->b_offset);
1381 had_bogus = 1;
1382
1383 for (j = i; j < bp->b_npages; j++) {
1384 vm_page_t mtmp;
1385 mtmp = bp->b_pages[j];
1386 if (mtmp == bogus_page) {
1387 mtmp = vm_page_lookup(obj, poff + j);
1388 if (!mtmp) {
1389 panic("brelse: page missing\n");
1390 }
1391 bp->b_pages[j] = mtmp;
1392 }
1393 }
1394
1395 if ((bp->b_flags & B_INVAL) == 0) {
1396 pmap_qenter(
1397 trunc_page((vm_offset_t)bp->b_data),
1398 bp->b_pages, bp->b_npages);
1399 }
1400 m = bp->b_pages[i];
1401 }
1402 if ((bp->b_flags & B_NOCACHE) ||
1403 (bp->b_ioflags & BIO_ERROR &&
1404 bp->b_iocmd == BIO_READ)) {
1405 int poffset = foff & PAGE_MASK;
1406 int presid = resid > (PAGE_SIZE - poffset) ?
1407 (PAGE_SIZE - poffset) : resid;
1408
1409 KASSERT(presid >= 0, ("brelse: extra page"));
1410 vm_page_lock_queues();
1411 vm_page_set_invalid(m, poffset, presid);
1412 vm_page_unlock_queues();
1413 if (had_bogus)
1414 printf("avoided corruption bug in bogus_page/brelse code\n");
1415 }
1416 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1417 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1418 }
1419 VM_OBJECT_UNLOCK(obj);
1420 if (bp->b_flags & (B_INVAL | B_RELBUF))
1421 vfs_vmio_release(bp);
1422
1423 } else if (bp->b_flags & B_VMIO) {
1424
1425 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1426 vfs_vmio_release(bp);
1427 }
1428
1429 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1430 if (bp->b_bufsize != 0)
1431 allocbuf(bp, 0);
1432 if (bp->b_vp != NULL)
1433 brelvp(bp);
1434 }
1435
1436 if (BUF_LOCKRECURSED(bp)) {
1437 /* do not release to free list */
1438 BUF_UNLOCK(bp);
1439 return;
1440 }
1441
1442 /* enqueue */
1443 mtx_lock(&bqlock);
1444 /* Handle delayed bremfree() processing. */
1445 if (bp->b_flags & B_REMFREE) {
1446 struct bufobj *bo;
1447
1448 bo = bp->b_bufobj;
1449 if (bo != NULL)
1450 BO_LOCK(bo);
1451 bremfreel(bp);
1452 if (bo != NULL)
1453 BO_UNLOCK(bo);
1454 }
1455 if (bp->b_qindex != QUEUE_NONE)
1456 panic("brelse: free buffer onto another queue???");
1457
1458 /*
1459 * If the buffer has junk contents signal it and eventually
1460 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1461 * doesn't find it.
1462 */
1463 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1464 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1465 bp->b_flags |= B_INVAL;
1466 if (bp->b_flags & B_INVAL) {
1467 if (bp->b_flags & B_DELWRI)
1468 bundirty(bp);
1469 if (bp->b_vp)
1470 brelvp(bp);
1471 }
1472
1473 /* buffers with no memory */
1474 if (bp->b_bufsize == 0) {
1475 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1476 if (bp->b_vflags & BV_BKGRDINPROG)
1477 panic("losing buffer 1");
1478 if (bp->b_kvasize) {
1479 bp->b_qindex = QUEUE_EMPTYKVA;
1480 } else {
1481 bp->b_qindex = QUEUE_EMPTY;
1482 }
1483 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1484 /* buffers with junk contents */
1485 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1486 (bp->b_ioflags & BIO_ERROR)) {
1487 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1488 if (bp->b_vflags & BV_BKGRDINPROG)
1489 panic("losing buffer 2");
1490 bp->b_qindex = QUEUE_CLEAN;
1491 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1492 /* remaining buffers */
1493 } else {
1494 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1495 (B_DELWRI|B_NEEDSGIANT))
1496 bp->b_qindex = QUEUE_DIRTY_GIANT;
1497 else if (bp->b_flags & B_DELWRI)
1498 bp->b_qindex = QUEUE_DIRTY;
1499 else
1500 bp->b_qindex = QUEUE_CLEAN;
1501 if (bp->b_flags & B_AGE)
1502 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1503 else
1504 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1505 }
1506 mtx_unlock(&bqlock);
1507
1508 /*
1509 * Fixup numfreebuffers count. The bp is on an appropriate queue
1510 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1511 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1512 * if B_INVAL is set ).
1513 */
1514
1515 if (!(bp->b_flags & B_DELWRI)) {
1516 struct bufobj *bo;
1517
1518 bo = bp->b_bufobj;
1519 if (bo != NULL)
1520 BO_LOCK(bo);
1521 bufcountwakeup(bp);
1522 if (bo != NULL)
1523 BO_UNLOCK(bo);
1524 }
1525
1526 /*
1527 * Something we can maybe free or reuse
1528 */
1529 if (bp->b_bufsize || bp->b_kvasize)
1530 bufspacewakeup();
1531
1532 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1533 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1534 panic("brelse: not dirty");
1535 /* unlock */
1536 BUF_UNLOCK(bp);
1537 }
1538
1539 /*
1540 * Release a buffer back to the appropriate queue but do not try to free
1541 * it. The buffer is expected to be used again soon.
1542 *
1543 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1544 * biodone() to requeue an async I/O on completion. It is also used when
1545 * known good buffers need to be requeued but we think we may need the data
1546 * again soon.
1547 *
1548 * XXX we should be able to leave the B_RELBUF hint set on completion.
1549 */
1550 void
1551 bqrelse(struct buf *bp)
1552 {
1553 struct bufobj *bo;
1554
1555 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1556 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1557 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1558
1559 if (BUF_LOCKRECURSED(bp)) {
1560 /* do not release to free list */
1561 BUF_UNLOCK(bp);
1562 return;
1563 }
1564
1565 bo = bp->b_bufobj;
1566 if (bp->b_flags & B_MANAGED) {
1567 if (bp->b_flags & B_REMFREE) {
1568 mtx_lock(&bqlock);
1569 if (bo != NULL)
1570 BO_LOCK(bo);
1571 bremfreel(bp);
1572 if (bo != NULL)
1573 BO_UNLOCK(bo);
1574 mtx_unlock(&bqlock);
1575 }
1576 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1577 BUF_UNLOCK(bp);
1578 return;
1579 }
1580
1581 mtx_lock(&bqlock);
1582 /* Handle delayed bremfree() processing. */
1583 if (bp->b_flags & B_REMFREE) {
1584 if (bo != NULL)
1585 BO_LOCK(bo);
1586 bremfreel(bp);
1587 if (bo != NULL)
1588 BO_UNLOCK(bo);
1589 }
1590 if (bp->b_qindex != QUEUE_NONE)
1591 panic("bqrelse: free buffer onto another queue???");
1592 /* buffers with stale but valid contents */
1593 if (bp->b_flags & B_DELWRI) {
1594 if (bp->b_flags & B_NEEDSGIANT)
1595 bp->b_qindex = QUEUE_DIRTY_GIANT;
1596 else
1597 bp->b_qindex = QUEUE_DIRTY;
1598 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1599 } else {
1600 /*
1601 * The locking of the BO_LOCK for checking of the
1602 * BV_BKGRDINPROG is not necessary since the
1603 * BV_BKGRDINPROG cannot be set while we hold the buf
1604 * lock, it can only be cleared if it is already
1605 * pending.
1606 */
1607 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1608 bp->b_qindex = QUEUE_CLEAN;
1609 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1610 b_freelist);
1611 } else {
1612 /*
1613 * We are too low on memory, we have to try to free
1614 * the buffer (most importantly: the wired pages
1615 * making up its backing store) *now*.
1616 */
1617 mtx_unlock(&bqlock);
1618 brelse(bp);
1619 return;
1620 }
1621 }
1622 mtx_unlock(&bqlock);
1623
1624 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) {
1625 if (bo != NULL)
1626 BO_LOCK(bo);
1627 bufcountwakeup(bp);
1628 if (bo != NULL)
1629 BO_UNLOCK(bo);
1630 }
1631
1632 /*
1633 * Something we can maybe free or reuse.
1634 */
1635 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1636 bufspacewakeup();
1637
1638 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1639 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1640 panic("bqrelse: not dirty");
1641 /* unlock */
1642 BUF_UNLOCK(bp);
1643 }
1644
1645 /* Give pages used by the bp back to the VM system (where possible) */
1646 static void
1647 vfs_vmio_release(struct buf *bp)
1648 {
1649 int i;
1650 vm_page_t m;
1651
1652 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1653 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1654 vm_page_lock_queues();
1655 for (i = 0; i < bp->b_npages; i++) {
1656 m = bp->b_pages[i];
1657 bp->b_pages[i] = NULL;
1658 /*
1659 * In order to keep page LRU ordering consistent, put
1660 * everything on the inactive queue.
1661 */
1662 vm_page_unwire(m, 0);
1663 /*
1664 * We don't mess with busy pages, it is
1665 * the responsibility of the process that
1666 * busied the pages to deal with them.
1667 */
1668 if ((m->oflags & VPO_BUSY) || (m->busy != 0))
1669 continue;
1670
1671 if (m->wire_count == 0) {
1672 /*
1673 * Might as well free the page if we can and it has
1674 * no valid data. We also free the page if the
1675 * buffer was used for direct I/O
1676 */
1677 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1678 m->hold_count == 0) {
1679 vm_page_free(m);
1680 } else if (bp->b_flags & B_DIRECT) {
1681 vm_page_try_to_free(m);
1682 } else if (buf_vm_page_count_severe()) {
1683 vm_page_try_to_cache(m);
1684 }
1685 }
1686 }
1687 vm_page_unlock_queues();
1688 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1689
1690 if (bp->b_bufsize) {
1691 bufspacewakeup();
1692 bp->b_bufsize = 0;
1693 }
1694 bp->b_npages = 0;
1695 bp->b_flags &= ~B_VMIO;
1696 if (bp->b_vp)
1697 brelvp(bp);
1698 }
1699
1700 /*
1701 * Check to see if a block at a particular lbn is available for a clustered
1702 * write.
1703 */
1704 static int
1705 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1706 {
1707 struct buf *bpa;
1708 int match;
1709
1710 match = 0;
1711
1712 /* If the buf isn't in core skip it */
1713 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1714 return (0);
1715
1716 /* If the buf is busy we don't want to wait for it */
1717 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1718 return (0);
1719
1720 /* Only cluster with valid clusterable delayed write buffers */
1721 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1722 (B_DELWRI | B_CLUSTEROK))
1723 goto done;
1724
1725 if (bpa->b_bufsize != size)
1726 goto done;
1727
1728 /*
1729 * Check to see if it is in the expected place on disk and that the
1730 * block has been mapped.
1731 */
1732 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1733 match = 1;
1734 done:
1735 BUF_UNLOCK(bpa);
1736 return (match);
1737 }
1738
1739 /*
1740 * vfs_bio_awrite:
1741 *
1742 * Implement clustered async writes for clearing out B_DELWRI buffers.
1743 * This is much better then the old way of writing only one buffer at
1744 * a time. Note that we may not be presented with the buffers in the
1745 * correct order, so we search for the cluster in both directions.
1746 */
1747 int
1748 vfs_bio_awrite(struct buf *bp)
1749 {
1750 struct bufobj *bo;
1751 int i;
1752 int j;
1753 daddr_t lblkno = bp->b_lblkno;
1754 struct vnode *vp = bp->b_vp;
1755 int ncl;
1756 int nwritten;
1757 int size;
1758 int maxcl;
1759
1760 bo = &vp->v_bufobj;
1761 /*
1762 * right now we support clustered writing only to regular files. If
1763 * we find a clusterable block we could be in the middle of a cluster
1764 * rather then at the beginning.
1765 */
1766 if ((vp->v_type == VREG) &&
1767 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1768 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1769
1770 size = vp->v_mount->mnt_stat.f_iosize;
1771 maxcl = MAXPHYS / size;
1772
1773 BO_LOCK(bo);
1774 for (i = 1; i < maxcl; i++)
1775 if (vfs_bio_clcheck(vp, size, lblkno + i,
1776 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1777 break;
1778
1779 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1780 if (vfs_bio_clcheck(vp, size, lblkno - j,
1781 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1782 break;
1783 BO_UNLOCK(bo);
1784 --j;
1785 ncl = i + j;
1786 /*
1787 * this is a possible cluster write
1788 */
1789 if (ncl != 1) {
1790 BUF_UNLOCK(bp);
1791 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1792 return nwritten;
1793 }
1794 }
1795 bremfree(bp);
1796 bp->b_flags |= B_ASYNC;
1797 /*
1798 * default (old) behavior, writing out only one block
1799 *
1800 * XXX returns b_bufsize instead of b_bcount for nwritten?
1801 */
1802 nwritten = bp->b_bufsize;
1803 (void) bwrite(bp);
1804
1805 return nwritten;
1806 }
1807
1808 /*
1809 * getnewbuf:
1810 *
1811 * Find and initialize a new buffer header, freeing up existing buffers
1812 * in the bufqueues as necessary. The new buffer is returned locked.
1813 *
1814 * Important: B_INVAL is not set. If the caller wishes to throw the
1815 * buffer away, the caller must set B_INVAL prior to calling brelse().
1816 *
1817 * We block if:
1818 * We have insufficient buffer headers
1819 * We have insufficient buffer space
1820 * buffer_map is too fragmented ( space reservation fails )
1821 * If we have to flush dirty buffers ( but we try to avoid this )
1822 *
1823 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1824 * Instead we ask the buf daemon to do it for us. We attempt to
1825 * avoid piecemeal wakeups of the pageout daemon.
1826 */
1827
1828 static struct buf *
1829 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1830 int gbflags)
1831 {
1832 struct thread *td;
1833 struct buf *bp;
1834 struct buf *nbp;
1835 int defrag = 0;
1836 int nqindex;
1837 static int flushingbufs;
1838
1839 td = curthread;
1840 /*
1841 * We can't afford to block since we might be holding a vnode lock,
1842 * which may prevent system daemons from running. We deal with
1843 * low-memory situations by proactively returning memory and running
1844 * async I/O rather then sync I/O.
1845 */
1846 atomic_add_int(&getnewbufcalls, 1);
1847 atomic_subtract_int(&getnewbufrestarts, 1);
1848 restart:
1849 atomic_add_int(&getnewbufrestarts, 1);
1850
1851 /*
1852 * Setup for scan. If we do not have enough free buffers,
1853 * we setup a degenerate case that immediately fails. Note
1854 * that if we are specially marked process, we are allowed to
1855 * dip into our reserves.
1856 *
1857 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1858 *
1859 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1860 * However, there are a number of cases (defragging, reusing, ...)
1861 * where we cannot backup.
1862 */
1863 mtx_lock(&bqlock);
1864 nqindex = QUEUE_EMPTYKVA;
1865 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1866
1867 if (nbp == NULL) {
1868 /*
1869 * If no EMPTYKVA buffers and we are either
1870 * defragging or reusing, locate a CLEAN buffer
1871 * to free or reuse. If bufspace useage is low
1872 * skip this step so we can allocate a new buffer.
1873 */
1874 if (defrag || bufspace >= lobufspace) {
1875 nqindex = QUEUE_CLEAN;
1876 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1877 }
1878
1879 /*
1880 * If we could not find or were not allowed to reuse a
1881 * CLEAN buffer, check to see if it is ok to use an EMPTY
1882 * buffer. We can only use an EMPTY buffer if allocating
1883 * its KVA would not otherwise run us out of buffer space.
1884 */
1885 if (nbp == NULL && defrag == 0 &&
1886 bufspace + maxsize < hibufspace) {
1887 nqindex = QUEUE_EMPTY;
1888 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1889 }
1890 }
1891
1892 /*
1893 * Run scan, possibly freeing data and/or kva mappings on the fly
1894 * depending.
1895 */
1896
1897 while ((bp = nbp) != NULL) {
1898 int qindex = nqindex;
1899
1900 /*
1901 * Calculate next bp ( we can only use it if we do not block
1902 * or do other fancy things ).
1903 */
1904 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1905 switch(qindex) {
1906 case QUEUE_EMPTY:
1907 nqindex = QUEUE_EMPTYKVA;
1908 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1909 break;
1910 /* FALLTHROUGH */
1911 case QUEUE_EMPTYKVA:
1912 nqindex = QUEUE_CLEAN;
1913 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1914 break;
1915 /* FALLTHROUGH */
1916 case QUEUE_CLEAN:
1917 /*
1918 * nbp is NULL.
1919 */
1920 break;
1921 }
1922 }
1923 /*
1924 * If we are defragging then we need a buffer with
1925 * b_kvasize != 0. XXX this situation should no longer
1926 * occur, if defrag is non-zero the buffer's b_kvasize
1927 * should also be non-zero at this point. XXX
1928 */
1929 if (defrag && bp->b_kvasize == 0) {
1930 printf("Warning: defrag empty buffer %p\n", bp);
1931 continue;
1932 }
1933
1934 /*
1935 * Start freeing the bp. This is somewhat involved. nbp
1936 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1937 */
1938 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1939 continue;
1940 if (bp->b_vp) {
1941 BO_LOCK(bp->b_bufobj);
1942 if (bp->b_vflags & BV_BKGRDINPROG) {
1943 BO_UNLOCK(bp->b_bufobj);
1944 BUF_UNLOCK(bp);
1945 continue;
1946 }
1947 BO_UNLOCK(bp->b_bufobj);
1948 }
1949 CTR6(KTR_BUF,
1950 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1951 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1952 bp->b_kvasize, bp->b_bufsize, qindex);
1953
1954 /*
1955 * Sanity Checks
1956 */
1957 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1958
1959 /*
1960 * Note: we no longer distinguish between VMIO and non-VMIO
1961 * buffers.
1962 */
1963
1964 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1965
1966 if (bp->b_bufobj != NULL)
1967 BO_LOCK(bp->b_bufobj);
1968 bremfreel(bp);
1969 if (bp->b_bufobj != NULL)
1970 BO_UNLOCK(bp->b_bufobj);
1971 mtx_unlock(&bqlock);
1972
1973 if (qindex == QUEUE_CLEAN) {
1974 if (bp->b_flags & B_VMIO) {
1975 bp->b_flags &= ~B_ASYNC;
1976 vfs_vmio_release(bp);
1977 }
1978 if (bp->b_vp)
1979 brelvp(bp);
1980 }
1981
1982 /*
1983 * NOTE: nbp is now entirely invalid. We can only restart
1984 * the scan from this point on.
1985 *
1986 * Get the rest of the buffer freed up. b_kva* is still
1987 * valid after this operation.
1988 */
1989
1990 if (bp->b_rcred != NOCRED) {
1991 crfree(bp->b_rcred);
1992 bp->b_rcred = NOCRED;
1993 }
1994 if (bp->b_wcred != NOCRED) {
1995 crfree(bp->b_wcred);
1996 bp->b_wcred = NOCRED;
1997 }
1998 if (!LIST_EMPTY(&bp->b_dep))
1999 buf_deallocate(bp);
2000 if (bp->b_vflags & BV_BKGRDINPROG)
2001 panic("losing buffer 3");
2002 KASSERT(bp->b_vp == NULL,
2003 ("bp: %p still has vnode %p. qindex: %d",
2004 bp, bp->b_vp, qindex));
2005 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2006 ("bp: %p still on a buffer list. xflags %X",
2007 bp, bp->b_xflags));
2008
2009 if (bp->b_bufsize)
2010 allocbuf(bp, 0);
2011
2012 bp->b_flags = 0;
2013 bp->b_ioflags = 0;
2014 bp->b_xflags = 0;
2015 KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
2016 ("buf %p still counted as free?", bp));
2017 bp->b_vflags = 0;
2018 bp->b_vp = NULL;
2019 bp->b_blkno = bp->b_lblkno = 0;
2020 bp->b_offset = NOOFFSET;
2021 bp->b_iodone = 0;
2022 bp->b_error = 0;
2023 bp->b_resid = 0;
2024 bp->b_bcount = 0;
2025 bp->b_npages = 0;
2026 bp->b_dirtyoff = bp->b_dirtyend = 0;
2027 bp->b_bufobj = NULL;
2028 bp->b_pin_count = 0;
2029 bp->b_fsprivate1 = NULL;
2030 bp->b_fsprivate2 = NULL;
2031 bp->b_fsprivate3 = NULL;
2032
2033 LIST_INIT(&bp->b_dep);
2034
2035 /*
2036 * If we are defragging then free the buffer.
2037 */
2038 if (defrag) {
2039 bp->b_flags |= B_INVAL;
2040 bfreekva(bp);
2041 brelse(bp);
2042 defrag = 0;
2043 goto restart;
2044 }
2045
2046 /*
2047 * Notify any waiters for the buffer lock about
2048 * identity change by freeing the buffer.
2049 */
2050 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2051 bp->b_flags |= B_INVAL;
2052 bfreekva(bp);
2053 brelse(bp);
2054 goto restart;
2055 }
2056
2057 /*
2058 * If we are overcomitted then recover the buffer and its
2059 * KVM space. This occurs in rare situations when multiple
2060 * processes are blocked in getnewbuf() or allocbuf().
2061 */
2062 if (bufspace >= hibufspace)
2063 flushingbufs = 1;
2064 if (flushingbufs && bp->b_kvasize != 0) {
2065 bp->b_flags |= B_INVAL;
2066 bfreekva(bp);
2067 brelse(bp);
2068 goto restart;
2069 }
2070 if (bufspace < lobufspace)
2071 flushingbufs = 0;
2072 break;
2073 }
2074
2075 /*
2076 * If we exhausted our list, sleep as appropriate. We may have to
2077 * wakeup various daemons and write out some dirty buffers.
2078 *
2079 * Generally we are sleeping due to insufficient buffer space.
2080 */
2081
2082 if (bp == NULL) {
2083 int flags, norunbuf;
2084 char *waitmsg;
2085 int fl;
2086
2087 if (defrag) {
2088 flags = VFS_BIO_NEED_BUFSPACE;
2089 waitmsg = "nbufkv";
2090 } else if (bufspace >= hibufspace) {
2091 waitmsg = "nbufbs";
2092 flags = VFS_BIO_NEED_BUFSPACE;
2093 } else {
2094 waitmsg = "newbuf";
2095 flags = VFS_BIO_NEED_ANY;
2096 }
2097 mtx_lock(&nblock);
2098 needsbuffer |= flags;
2099 mtx_unlock(&nblock);
2100 mtx_unlock(&bqlock);
2101
2102 bd_speedup(); /* heeeelp */
2103 if (gbflags & GB_NOWAIT_BD)
2104 return (NULL);
2105
2106 mtx_lock(&nblock);
2107 while (needsbuffer & flags) {
2108 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2109 mtx_unlock(&nblock);
2110 /*
2111 * getblk() is called with a vnode
2112 * locked, and some majority of the
2113 * dirty buffers may as well belong to
2114 * the vnode. Flushing the buffers
2115 * there would make a progress that
2116 * cannot be achieved by the
2117 * buf_daemon, that cannot lock the
2118 * vnode.
2119 */
2120 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2121 (td->td_pflags & TDP_NORUNNINGBUF);
2122 /* play bufdaemon */
2123 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2124 fl = buf_do_flush(vp);
2125 td->td_pflags &= norunbuf;
2126 mtx_lock(&nblock);
2127 if (fl != 0)
2128 continue;
2129 if ((needsbuffer & flags) == 0)
2130 break;
2131 }
2132 if (msleep(&needsbuffer, &nblock,
2133 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2134 mtx_unlock(&nblock);
2135 return (NULL);
2136 }
2137 }
2138 mtx_unlock(&nblock);
2139 } else {
2140 /*
2141 * We finally have a valid bp. We aren't quite out of the
2142 * woods, we still have to reserve kva space. In order
2143 * to keep fragmentation sane we only allocate kva in
2144 * BKVASIZE chunks.
2145 */
2146 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2147
2148 if (maxsize != bp->b_kvasize) {
2149 vm_offset_t addr = 0;
2150
2151 bfreekva(bp);
2152
2153 vm_map_lock(buffer_map);
2154 if (vm_map_findspace(buffer_map,
2155 vm_map_min(buffer_map), maxsize, &addr)) {
2156 /*
2157 * Uh oh. Buffer map is to fragmented. We
2158 * must defragment the map.
2159 */
2160 atomic_add_int(&bufdefragcnt, 1);
2161 vm_map_unlock(buffer_map);
2162 defrag = 1;
2163 bp->b_flags |= B_INVAL;
2164 brelse(bp);
2165 goto restart;
2166 }
2167 if (addr) {
2168 vm_map_insert(buffer_map, NULL, 0,
2169 addr, addr + maxsize,
2170 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2171
2172 bp->b_kvabase = (caddr_t) addr;
2173 bp->b_kvasize = maxsize;
2174 atomic_add_long(&bufspace, bp->b_kvasize);
2175 atomic_add_int(&bufreusecnt, 1);
2176 }
2177 vm_map_unlock(buffer_map);
2178 }
2179 bp->b_saveaddr = bp->b_kvabase;
2180 bp->b_data = bp->b_saveaddr;
2181 }
2182 return(bp);
2183 }
2184
2185 /*
2186 * buf_daemon:
2187 *
2188 * buffer flushing daemon. Buffers are normally flushed by the
2189 * update daemon but if it cannot keep up this process starts to
2190 * take the load in an attempt to prevent getnewbuf() from blocking.
2191 */
2192
2193 static struct kproc_desc buf_kp = {
2194 "bufdaemon",
2195 buf_daemon,
2196 &bufdaemonproc
2197 };
2198 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2199
2200 static int
2201 buf_do_flush(struct vnode *vp)
2202 {
2203 int flushed;
2204
2205 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2206 /* The list empty check here is slightly racy */
2207 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2208 mtx_lock(&Giant);
2209 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0);
2210 mtx_unlock(&Giant);
2211 }
2212 if (flushed == 0) {
2213 /*
2214 * Could not find any buffers without rollback
2215 * dependencies, so just write the first one
2216 * in the hopes of eventually making progress.
2217 */
2218 flushbufqueues(vp, QUEUE_DIRTY, 1);
2219 if (!TAILQ_EMPTY(
2220 &bufqueues[QUEUE_DIRTY_GIANT])) {
2221 mtx_lock(&Giant);
2222 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1);
2223 mtx_unlock(&Giant);
2224 }
2225 }
2226 return (flushed);
2227 }
2228
2229 static void
2230 buf_daemon()
2231 {
2232
2233 /*
2234 * This process needs to be suspended prior to shutdown sync.
2235 */
2236 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2237 SHUTDOWN_PRI_LAST);
2238
2239 /*
2240 * This process is allowed to take the buffer cache to the limit
2241 */
2242 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2243 mtx_lock(&bdlock);
2244 for (;;) {
2245 bd_request = 0;
2246 mtx_unlock(&bdlock);
2247
2248 kproc_suspend_check(bufdaemonproc);
2249
2250 /*
2251 * Do the flush. Limit the amount of in-transit I/O we
2252 * allow to build up, otherwise we would completely saturate
2253 * the I/O system. Wakeup any waiting processes before we
2254 * normally would so they can run in parallel with our drain.
2255 */
2256 while (numdirtybuffers > lodirtybuffers) {
2257 if (buf_do_flush(NULL) == 0)
2258 break;
2259 uio_yield();
2260 }
2261
2262 /*
2263 * Only clear bd_request if we have reached our low water
2264 * mark. The buf_daemon normally waits 1 second and
2265 * then incrementally flushes any dirty buffers that have
2266 * built up, within reason.
2267 *
2268 * If we were unable to hit our low water mark and couldn't
2269 * find any flushable buffers, we sleep half a second.
2270 * Otherwise we loop immediately.
2271 */
2272 mtx_lock(&bdlock);
2273 if (numdirtybuffers <= lodirtybuffers) {
2274 /*
2275 * We reached our low water mark, reset the
2276 * request and sleep until we are needed again.
2277 * The sleep is just so the suspend code works.
2278 */
2279 bd_request = 0;
2280 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2281 } else {
2282 /*
2283 * We couldn't find any flushable dirty buffers but
2284 * still have too many dirty buffers, we
2285 * have to sleep and try again. (rare)
2286 */
2287 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2288 }
2289 }
2290 }
2291
2292 /*
2293 * flushbufqueues:
2294 *
2295 * Try to flush a buffer in the dirty queue. We must be careful to
2296 * free up B_INVAL buffers instead of write them, which NFS is
2297 * particularly sensitive to.
2298 */
2299 static int flushwithdeps = 0;
2300 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2301 0, "Number of buffers flushed with dependecies that require rollbacks");
2302
2303 static int
2304 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2305 {
2306 struct buf *sentinel;
2307 struct vnode *vp;
2308 struct mount *mp;
2309 struct buf *bp;
2310 int hasdeps;
2311 int flushed;
2312 int target;
2313
2314 if (lvp == NULL) {
2315 target = numdirtybuffers - lodirtybuffers;
2316 if (flushdeps && target > 2)
2317 target /= 2;
2318 } else
2319 target = flushbufqtarget;
2320 flushed = 0;
2321 bp = NULL;
2322 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2323 sentinel->b_qindex = QUEUE_SENTINEL;
2324 mtx_lock(&bqlock);
2325 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2326 while (flushed != target) {
2327 bp = TAILQ_NEXT(sentinel, b_freelist);
2328 if (bp != NULL) {
2329 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2330 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2331 b_freelist);
2332 } else
2333 break;
2334 /*
2335 * Skip sentinels inserted by other invocations of the
2336 * flushbufqueues(), taking care to not reorder them.
2337 */
2338 if (bp->b_qindex == QUEUE_SENTINEL)
2339 continue;
2340 /*
2341 * Only flush the buffers that belong to the
2342 * vnode locked by the curthread.
2343 */
2344 if (lvp != NULL && bp->b_vp != lvp)
2345 continue;
2346 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2347 continue;
2348 if (bp->b_pin_count > 0) {
2349 BUF_UNLOCK(bp);
2350 continue;
2351 }
2352 BO_LOCK(bp->b_bufobj);
2353 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2354 (bp->b_flags & B_DELWRI) == 0) {
2355 BO_UNLOCK(bp->b_bufobj);
2356 BUF_UNLOCK(bp);
2357 continue;
2358 }
2359 BO_UNLOCK(bp->b_bufobj);
2360 if (bp->b_flags & B_INVAL) {
2361 bremfreel(bp);
2362 mtx_unlock(&bqlock);
2363 brelse(bp);
2364 flushed++;
2365 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2366 mtx_lock(&bqlock);
2367 continue;
2368 }
2369
2370 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2371 if (flushdeps == 0) {
2372 BUF_UNLOCK(bp);
2373 continue;
2374 }
2375 hasdeps = 1;
2376 } else
2377 hasdeps = 0;
2378 /*
2379 * We must hold the lock on a vnode before writing
2380 * one of its buffers. Otherwise we may confuse, or
2381 * in the case of a snapshot vnode, deadlock the
2382 * system.
2383 *
2384 * The lock order here is the reverse of the normal
2385 * of vnode followed by buf lock. This is ok because
2386 * the NOWAIT will prevent deadlock.
2387 */
2388 vp = bp->b_vp;
2389 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2390 BUF_UNLOCK(bp);
2391 continue;
2392 }
2393 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2394 mtx_unlock(&bqlock);
2395 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2396 bp, bp->b_vp, bp->b_flags);
2397 if (curproc == bufdaemonproc)
2398 vfs_bio_awrite(bp);
2399 else {
2400 bremfree(bp);
2401 bwrite(bp);
2402 notbufdflashes++;
2403 }
2404 vn_finished_write(mp);
2405 VOP_UNLOCK(vp, 0);
2406 flushwithdeps += hasdeps;
2407 flushed++;
2408
2409 /*
2410 * Sleeping on runningbufspace while holding
2411 * vnode lock leads to deadlock.
2412 */
2413 if (curproc == bufdaemonproc)
2414 waitrunningbufspace();
2415 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2416 mtx_lock(&bqlock);
2417 continue;
2418 }
2419 vn_finished_write(mp);
2420 BUF_UNLOCK(bp);
2421 }
2422 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2423 mtx_unlock(&bqlock);
2424 free(sentinel, M_TEMP);
2425 return (flushed);
2426 }
2427
2428 /*
2429 * Check to see if a block is currently memory resident.
2430 */
2431 struct buf *
2432 incore(struct bufobj *bo, daddr_t blkno)
2433 {
2434 struct buf *bp;
2435
2436 BO_LOCK(bo);
2437 bp = gbincore(bo, blkno);
2438 BO_UNLOCK(bo);
2439 return (bp);
2440 }
2441
2442 /*
2443 * Returns true if no I/O is needed to access the
2444 * associated VM object. This is like incore except
2445 * it also hunts around in the VM system for the data.
2446 */
2447
2448 static int
2449 inmem(struct vnode * vp, daddr_t blkno)
2450 {
2451 vm_object_t obj;
2452 vm_offset_t toff, tinc, size;
2453 vm_page_t m;
2454 vm_ooffset_t off;
2455
2456 ASSERT_VOP_LOCKED(vp, "inmem");
2457
2458 if (incore(&vp->v_bufobj, blkno))
2459 return 1;
2460 if (vp->v_mount == NULL)
2461 return 0;
2462 obj = vp->v_object;
2463 if (obj == NULL)
2464 return (0);
2465
2466 size = PAGE_SIZE;
2467 if (size > vp->v_mount->mnt_stat.f_iosize)
2468 size = vp->v_mount->mnt_stat.f_iosize;
2469 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2470
2471 VM_OBJECT_LOCK(obj);
2472 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2473 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2474 if (!m)
2475 goto notinmem;
2476 tinc = size;
2477 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2478 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2479 if (vm_page_is_valid(m,
2480 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2481 goto notinmem;
2482 }
2483 VM_OBJECT_UNLOCK(obj);
2484 return 1;
2485
2486 notinmem:
2487 VM_OBJECT_UNLOCK(obj);
2488 return (0);
2489 }
2490
2491 /*
2492 * Set the dirty range for a buffer based on the status of the dirty
2493 * bits in the pages comprising the buffer. The range is limited
2494 * to the size of the buffer.
2495 *
2496 * Tell the VM system that the pages associated with this buffer
2497 * are clean. This is used for delayed writes where the data is
2498 * going to go to disk eventually without additional VM intevention.
2499 *
2500 * Note that while we only really need to clean through to b_bcount, we
2501 * just go ahead and clean through to b_bufsize.
2502 */
2503 static void
2504 vfs_clean_pages_dirty_buf(struct buf *bp)
2505 {
2506 vm_ooffset_t foff, noff, eoff;
2507 vm_page_t m;
2508 int i;
2509
2510 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2511 return;
2512
2513 foff = bp->b_offset;
2514 KASSERT(bp->b_offset != NOOFFSET,
2515 ("vfs_clean_pages_dirty_buf: no buffer offset"));
2516
2517 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2518 vfs_drain_busy_pages(bp);
2519 vfs_setdirty_locked_object(bp);
2520 vm_page_lock_queues();
2521 for (i = 0; i < bp->b_npages; i++) {
2522 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2523 eoff = noff;
2524 if (eoff > bp->b_offset + bp->b_bufsize)
2525 eoff = bp->b_offset + bp->b_bufsize;
2526 m = bp->b_pages[i];
2527 vfs_page_set_validclean(bp, foff, m);
2528 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2529 foff = noff;
2530 }
2531 vm_page_unlock_queues();
2532 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2533 }
2534
2535 static void
2536 vfs_setdirty_locked_object(struct buf *bp)
2537 {
2538 vm_object_t object;
2539 int i;
2540
2541 object = bp->b_bufobj->bo_object;
2542 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2543
2544 /*
2545 * We qualify the scan for modified pages on whether the
2546 * object has been flushed yet.
2547 */
2548 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2549 vm_offset_t boffset;
2550 vm_offset_t eoffset;
2551
2552 vm_page_lock_queues();
2553 /*
2554 * test the pages to see if they have been modified directly
2555 * by users through the VM system.
2556 */
2557 for (i = 0; i < bp->b_npages; i++)
2558 vm_page_test_dirty(bp->b_pages[i]);
2559
2560 /*
2561 * Calculate the encompassing dirty range, boffset and eoffset,
2562 * (eoffset - boffset) bytes.
2563 */
2564
2565 for (i = 0; i < bp->b_npages; i++) {
2566 if (bp->b_pages[i]->dirty)
2567 break;
2568 }
2569 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2570
2571 for (i = bp->b_npages - 1; i >= 0; --i) {
2572 if (bp->b_pages[i]->dirty) {
2573 break;
2574 }
2575 }
2576 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2577
2578 vm_page_unlock_queues();
2579 /*
2580 * Fit it to the buffer.
2581 */
2582
2583 if (eoffset > bp->b_bcount)
2584 eoffset = bp->b_bcount;
2585
2586 /*
2587 * If we have a good dirty range, merge with the existing
2588 * dirty range.
2589 */
2590
2591 if (boffset < eoffset) {
2592 if (bp->b_dirtyoff > boffset)
2593 bp->b_dirtyoff = boffset;
2594 if (bp->b_dirtyend < eoffset)
2595 bp->b_dirtyend = eoffset;
2596 }
2597 }
2598 }
2599
2600 /*
2601 * getblk:
2602 *
2603 * Get a block given a specified block and offset into a file/device.
2604 * The buffers B_DONE bit will be cleared on return, making it almost
2605 * ready for an I/O initiation. B_INVAL may or may not be set on
2606 * return. The caller should clear B_INVAL prior to initiating a
2607 * READ.
2608 *
2609 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2610 * an existing buffer.
2611 *
2612 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2613 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2614 * and then cleared based on the backing VM. If the previous buffer is
2615 * non-0-sized but invalid, B_CACHE will be cleared.
2616 *
2617 * If getblk() must create a new buffer, the new buffer is returned with
2618 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2619 * case it is returned with B_INVAL clear and B_CACHE set based on the
2620 * backing VM.
2621 *
2622 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2623 * B_CACHE bit is clear.
2624 *
2625 * What this means, basically, is that the caller should use B_CACHE to
2626 * determine whether the buffer is fully valid or not and should clear
2627 * B_INVAL prior to issuing a read. If the caller intends to validate
2628 * the buffer by loading its data area with something, the caller needs
2629 * to clear B_INVAL. If the caller does this without issuing an I/O,
2630 * the caller should set B_CACHE ( as an optimization ), else the caller
2631 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2632 * a write attempt or if it was a successfull read. If the caller
2633 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2634 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2635 */
2636 struct buf *
2637 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2638 int flags)
2639 {
2640 struct buf *bp;
2641 struct bufobj *bo;
2642 int error;
2643
2644 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2645 ASSERT_VOP_LOCKED(vp, "getblk");
2646 if (size > MAXBSIZE)
2647 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2648
2649 bo = &vp->v_bufobj;
2650 loop:
2651 /*
2652 * Block if we are low on buffers. Certain processes are allowed
2653 * to completely exhaust the buffer cache.
2654 *
2655 * If this check ever becomes a bottleneck it may be better to
2656 * move it into the else, when gbincore() fails. At the moment
2657 * it isn't a problem.
2658 *
2659 * XXX remove if 0 sections (clean this up after its proven)
2660 */
2661 if (numfreebuffers == 0) {
2662 if (TD_IS_IDLETHREAD(curthread))
2663 return NULL;
2664 mtx_lock(&nblock);
2665 needsbuffer |= VFS_BIO_NEED_ANY;
2666 mtx_unlock(&nblock);
2667 }
2668
2669 BO_LOCK(bo);
2670 bp = gbincore(bo, blkno);
2671 if (bp != NULL) {
2672 int lockflags;
2673 /*
2674 * Buffer is in-core. If the buffer is not busy, it must
2675 * be on a queue.
2676 */
2677 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2678
2679 if (flags & GB_LOCK_NOWAIT)
2680 lockflags |= LK_NOWAIT;
2681
2682 error = BUF_TIMELOCK(bp, lockflags,
2683 BO_MTX(bo), "getblk", slpflag, slptimeo);
2684
2685 /*
2686 * If we slept and got the lock we have to restart in case
2687 * the buffer changed identities.
2688 */
2689 if (error == ENOLCK)
2690 goto loop;
2691 /* We timed out or were interrupted. */
2692 else if (error)
2693 return (NULL);
2694
2695 /*
2696 * The buffer is locked. B_CACHE is cleared if the buffer is
2697 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2698 * and for a VMIO buffer B_CACHE is adjusted according to the
2699 * backing VM cache.
2700 */
2701 if (bp->b_flags & B_INVAL)
2702 bp->b_flags &= ~B_CACHE;
2703 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2704 bp->b_flags |= B_CACHE;
2705 BO_LOCK(bo);
2706 bremfree(bp);
2707 BO_UNLOCK(bo);
2708
2709 /*
2710 * check for size inconsistancies for non-VMIO case.
2711 */
2712
2713 if (bp->b_bcount != size) {
2714 if ((bp->b_flags & B_VMIO) == 0 ||
2715 (size > bp->b_kvasize)) {
2716 if (bp->b_flags & B_DELWRI) {
2717 /*
2718 * If buffer is pinned and caller does
2719 * not want sleep waiting for it to be
2720 * unpinned, bail out
2721 * */
2722 if (bp->b_pin_count > 0) {
2723 if (flags & GB_LOCK_NOWAIT) {
2724 bqrelse(bp);
2725 return (NULL);
2726 } else {
2727 bunpin_wait(bp);
2728 }
2729 }
2730 bp->b_flags |= B_NOCACHE;
2731 bwrite(bp);
2732 } else {
2733 if (LIST_EMPTY(&bp->b_dep)) {
2734 bp->b_flags |= B_RELBUF;
2735 brelse(bp);
2736 } else {
2737 bp->b_flags |= B_NOCACHE;
2738 bwrite(bp);
2739 }
2740 }
2741 goto loop;
2742 }
2743 }
2744
2745 /*
2746 * If the size is inconsistant in the VMIO case, we can resize
2747 * the buffer. This might lead to B_CACHE getting set or
2748 * cleared. If the size has not changed, B_CACHE remains
2749 * unchanged from its previous state.
2750 */
2751
2752 if (bp->b_bcount != size)
2753 allocbuf(bp, size);
2754
2755 KASSERT(bp->b_offset != NOOFFSET,
2756 ("getblk: no buffer offset"));
2757
2758 /*
2759 * A buffer with B_DELWRI set and B_CACHE clear must
2760 * be committed before we can return the buffer in
2761 * order to prevent the caller from issuing a read
2762 * ( due to B_CACHE not being set ) and overwriting
2763 * it.
2764 *
2765 * Most callers, including NFS and FFS, need this to
2766 * operate properly either because they assume they
2767 * can issue a read if B_CACHE is not set, or because
2768 * ( for example ) an uncached B_DELWRI might loop due
2769 * to softupdates re-dirtying the buffer. In the latter
2770 * case, B_CACHE is set after the first write completes,
2771 * preventing further loops.
2772 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2773 * above while extending the buffer, we cannot allow the
2774 * buffer to remain with B_CACHE set after the write
2775 * completes or it will represent a corrupt state. To
2776 * deal with this we set B_NOCACHE to scrap the buffer
2777 * after the write.
2778 *
2779 * We might be able to do something fancy, like setting
2780 * B_CACHE in bwrite() except if B_DELWRI is already set,
2781 * so the below call doesn't set B_CACHE, but that gets real
2782 * confusing. This is much easier.
2783 */
2784
2785 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2786 bp->b_flags |= B_NOCACHE;
2787 bwrite(bp);
2788 goto loop;
2789 }
2790 bp->b_flags &= ~B_DONE;
2791 } else {
2792 int bsize, maxsize, vmio;
2793 off_t offset;
2794
2795 /*
2796 * Buffer is not in-core, create new buffer. The buffer
2797 * returned by getnewbuf() is locked. Note that the returned
2798 * buffer is also considered valid (not marked B_INVAL).
2799 */
2800 BO_UNLOCK(bo);
2801 /*
2802 * If the user does not want us to create the buffer, bail out
2803 * here.
2804 */
2805 if (flags & GB_NOCREAT)
2806 return NULL;
2807 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
2808 offset = blkno * bsize;
2809 vmio = vp->v_object != NULL;
2810 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2811 maxsize = imax(maxsize, bsize);
2812
2813 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2814 if (bp == NULL) {
2815 if (slpflag || slptimeo)
2816 return NULL;
2817 goto loop;
2818 }
2819
2820 /*
2821 * This code is used to make sure that a buffer is not
2822 * created while the getnewbuf routine is blocked.
2823 * This can be a problem whether the vnode is locked or not.
2824 * If the buffer is created out from under us, we have to
2825 * throw away the one we just created.
2826 *
2827 * Note: this must occur before we associate the buffer
2828 * with the vp especially considering limitations in
2829 * the splay tree implementation when dealing with duplicate
2830 * lblkno's.
2831 */
2832 BO_LOCK(bo);
2833 if (gbincore(bo, blkno)) {
2834 BO_UNLOCK(bo);
2835 bp->b_flags |= B_INVAL;
2836 brelse(bp);
2837 goto loop;
2838 }
2839
2840 /*
2841 * Insert the buffer into the hash, so that it can
2842 * be found by incore.
2843 */
2844 bp->b_blkno = bp->b_lblkno = blkno;
2845 bp->b_offset = offset;
2846 bgetvp(vp, bp);
2847 BO_UNLOCK(bo);
2848
2849 /*
2850 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2851 * buffer size starts out as 0, B_CACHE will be set by
2852 * allocbuf() for the VMIO case prior to it testing the
2853 * backing store for validity.
2854 */
2855
2856 if (vmio) {
2857 bp->b_flags |= B_VMIO;
2858 #if defined(VFS_BIO_DEBUG)
2859 if (vn_canvmio(vp) != TRUE)
2860 printf("getblk: VMIO on vnode type %d\n",
2861 vp->v_type);
2862 #endif
2863 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2864 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2865 bp, vp->v_object, bp->b_bufobj->bo_object));
2866 } else {
2867 bp->b_flags &= ~B_VMIO;
2868 KASSERT(bp->b_bufobj->bo_object == NULL,
2869 ("ARGH! has b_bufobj->bo_object %p %p\n",
2870 bp, bp->b_bufobj->bo_object));
2871 }
2872
2873 allocbuf(bp, size);
2874 bp->b_flags &= ~B_DONE;
2875 }
2876 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2877 BUF_ASSERT_HELD(bp);
2878 KASSERT(bp->b_bufobj == bo,
2879 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2880 return (bp);
2881 }
2882
2883 /*
2884 * Get an empty, disassociated buffer of given size. The buffer is initially
2885 * set to B_INVAL.
2886 */
2887 struct buf *
2888 geteblk(int size, int flags)
2889 {
2890 struct buf *bp;
2891 int maxsize;
2892
2893 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2894 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2895 if ((flags & GB_NOWAIT_BD) &&
2896 (curthread->td_pflags & TDP_BUFNEED) != 0)
2897 return (NULL);
2898 }
2899 allocbuf(bp, size);
2900 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2901 BUF_ASSERT_HELD(bp);
2902 return (bp);
2903 }
2904
2905
2906 /*
2907 * This code constitutes the buffer memory from either anonymous system
2908 * memory (in the case of non-VMIO operations) or from an associated
2909 * VM object (in the case of VMIO operations). This code is able to
2910 * resize a buffer up or down.
2911 *
2912 * Note that this code is tricky, and has many complications to resolve
2913 * deadlock or inconsistant data situations. Tread lightly!!!
2914 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2915 * the caller. Calling this code willy nilly can result in the loss of data.
2916 *
2917 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2918 * B_CACHE for the non-VMIO case.
2919 */
2920
2921 int
2922 allocbuf(struct buf *bp, int size)
2923 {
2924 int newbsize, mbsize;
2925 int i;
2926
2927 BUF_ASSERT_HELD(bp);
2928
2929 if (bp->b_kvasize < size)
2930 panic("allocbuf: buffer too small");
2931
2932 if ((bp->b_flags & B_VMIO) == 0) {
2933 caddr_t origbuf;
2934 int origbufsize;
2935 /*
2936 * Just get anonymous memory from the kernel. Don't
2937 * mess with B_CACHE.
2938 */
2939 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2940 if (bp->b_flags & B_MALLOC)
2941 newbsize = mbsize;
2942 else
2943 newbsize = round_page(size);
2944
2945 if (newbsize < bp->b_bufsize) {
2946 /*
2947 * malloced buffers are not shrunk
2948 */
2949 if (bp->b_flags & B_MALLOC) {
2950 if (newbsize) {
2951 bp->b_bcount = size;
2952 } else {
2953 free(bp->b_data, M_BIOBUF);
2954 if (bp->b_bufsize) {
2955 atomic_subtract_long(
2956 &bufmallocspace,
2957 bp->b_bufsize);
2958 bufspacewakeup();
2959 bp->b_bufsize = 0;
2960 }
2961 bp->b_saveaddr = bp->b_kvabase;
2962 bp->b_data = bp->b_saveaddr;
2963 bp->b_bcount = 0;
2964 bp->b_flags &= ~B_MALLOC;
2965 }
2966 return 1;
2967 }
2968 vm_hold_free_pages(bp, newbsize);
2969 } else if (newbsize > bp->b_bufsize) {
2970 /*
2971 * We only use malloced memory on the first allocation.
2972 * and revert to page-allocated memory when the buffer
2973 * grows.
2974 */
2975 /*
2976 * There is a potential smp race here that could lead
2977 * to bufmallocspace slightly passing the max. It
2978 * is probably extremely rare and not worth worrying
2979 * over.
2980 */
2981 if ( (bufmallocspace < maxbufmallocspace) &&
2982 (bp->b_bufsize == 0) &&
2983 (mbsize <= PAGE_SIZE/2)) {
2984
2985 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2986 bp->b_bufsize = mbsize;
2987 bp->b_bcount = size;
2988 bp->b_flags |= B_MALLOC;
2989 atomic_add_long(&bufmallocspace, mbsize);
2990 return 1;
2991 }
2992 origbuf = NULL;
2993 origbufsize = 0;
2994 /*
2995 * If the buffer is growing on its other-than-first allocation,
2996 * then we revert to the page-allocation scheme.
2997 */
2998 if (bp->b_flags & B_MALLOC) {
2999 origbuf = bp->b_data;
3000 origbufsize = bp->b_bufsize;
3001 bp->b_data = bp->b_kvabase;
3002 if (bp->b_bufsize) {
3003 atomic_subtract_long(&bufmallocspace,
3004 bp->b_bufsize);
3005 bufspacewakeup();
3006 bp->b_bufsize = 0;
3007 }
3008 bp->b_flags &= ~B_MALLOC;
3009 newbsize = round_page(newbsize);
3010 }
3011 vm_hold_load_pages(
3012 bp,
3013 (vm_offset_t) bp->b_data + bp->b_bufsize,
3014 (vm_offset_t) bp->b_data + newbsize);
3015 if (origbuf) {
3016 bcopy(origbuf, bp->b_data, origbufsize);
3017 free(origbuf, M_BIOBUF);
3018 }
3019 }
3020 } else {
3021 int desiredpages;
3022
3023 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3024 desiredpages = (size == 0) ? 0 :
3025 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3026
3027 if (bp->b_flags & B_MALLOC)
3028 panic("allocbuf: VMIO buffer can't be malloced");
3029 /*
3030 * Set B_CACHE initially if buffer is 0 length or will become
3031 * 0-length.
3032 */
3033 if (size == 0 || bp->b_bufsize == 0)
3034 bp->b_flags |= B_CACHE;
3035
3036 if (newbsize < bp->b_bufsize) {
3037 /*
3038 * DEV_BSIZE aligned new buffer size is less then the
3039 * DEV_BSIZE aligned existing buffer size. Figure out
3040 * if we have to remove any pages.
3041 */
3042 if (desiredpages < bp->b_npages) {
3043 vm_page_t m;
3044
3045 pmap_qremove((vm_offset_t)trunc_page(
3046 (vm_offset_t)bp->b_data) +
3047 (desiredpages << PAGE_SHIFT),
3048 (bp->b_npages - desiredpages));
3049 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3050 vm_page_lock_queues();
3051 for (i = desiredpages; i < bp->b_npages; i++) {
3052 /*
3053 * the page is not freed here -- it
3054 * is the responsibility of
3055 * vnode_pager_setsize
3056 */
3057 m = bp->b_pages[i];
3058 KASSERT(m != bogus_page,
3059 ("allocbuf: bogus page found"));
3060 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
3061 vm_page_lock_queues();
3062
3063 bp->b_pages[i] = NULL;
3064 vm_page_unwire(m, 0);
3065 }
3066 vm_page_unlock_queues();
3067 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3068 bp->b_npages = desiredpages;
3069 }
3070 } else if (size > bp->b_bcount) {
3071 /*
3072 * We are growing the buffer, possibly in a
3073 * byte-granular fashion.
3074 */
3075 vm_object_t obj;
3076 vm_offset_t toff;
3077 vm_offset_t tinc;
3078
3079 /*
3080 * Step 1, bring in the VM pages from the object,
3081 * allocating them if necessary. We must clear
3082 * B_CACHE if these pages are not valid for the
3083 * range covered by the buffer.
3084 */
3085
3086 obj = bp->b_bufobj->bo_object;
3087
3088 VM_OBJECT_LOCK(obj);
3089 while (bp->b_npages < desiredpages) {
3090 vm_page_t m;
3091 vm_pindex_t pi;
3092
3093 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
3094 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3095 /*
3096 * note: must allocate system pages
3097 * since blocking here could intefere
3098 * with paging I/O, no matter which
3099 * process we are.
3100 */
3101 m = vm_page_alloc(obj, pi,
3102 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
3103 VM_ALLOC_WIRED);
3104 if (m == NULL) {
3105 atomic_add_int(&vm_pageout_deficit,
3106 desiredpages - bp->b_npages);
3107 VM_OBJECT_UNLOCK(obj);
3108 VM_WAIT;
3109 VM_OBJECT_LOCK(obj);
3110 } else {
3111 if (m->valid == 0)
3112 bp->b_flags &= ~B_CACHE;
3113 bp->b_pages[bp->b_npages] = m;
3114 ++bp->b_npages;
3115 }
3116 continue;
3117 }
3118
3119 /*
3120 * We found a page. If we have to sleep on it,
3121 * retry because it might have gotten freed out
3122 * from under us.
3123 *
3124 * We can only test VPO_BUSY here. Blocking on
3125 * m->busy might lead to a deadlock:
3126 *
3127 * vm_fault->getpages->cluster_read->allocbuf
3128 *
3129 */
3130 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
3131 continue;
3132
3133 /*
3134 * We have a good page.
3135 */
3136 vm_page_lock_queues();
3137 vm_page_wire(m);
3138 vm_page_unlock_queues();
3139 bp->b_pages[bp->b_npages] = m;
3140 ++bp->b_npages;
3141 }
3142
3143 /*
3144 * Step 2. We've loaded the pages into the buffer,
3145 * we have to figure out if we can still have B_CACHE
3146 * set. Note that B_CACHE is set according to the
3147 * byte-granular range ( bcount and size ), new the
3148 * aligned range ( newbsize ).
3149 *
3150 * The VM test is against m->valid, which is DEV_BSIZE
3151 * aligned. Needless to say, the validity of the data
3152 * needs to also be DEV_BSIZE aligned. Note that this
3153 * fails with NFS if the server or some other client
3154 * extends the file's EOF. If our buffer is resized,
3155 * B_CACHE may remain set! XXX
3156 */
3157
3158 toff = bp->b_bcount;
3159 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3160
3161 while ((bp->b_flags & B_CACHE) && toff < size) {
3162 vm_pindex_t pi;
3163
3164 if (tinc > (size - toff))
3165 tinc = size - toff;
3166
3167 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3168 PAGE_SHIFT;
3169
3170 vfs_buf_test_cache(
3171 bp,
3172 bp->b_offset,
3173 toff,
3174 tinc,
3175 bp->b_pages[pi]
3176 );
3177 toff += tinc;
3178 tinc = PAGE_SIZE;
3179 }
3180 VM_OBJECT_UNLOCK(obj);
3181
3182 /*
3183 * Step 3, fixup the KVM pmap. Remember that
3184 * bp->b_data is relative to bp->b_offset, but
3185 * bp->b_offset may be offset into the first page.
3186 */
3187
3188 bp->b_data = (caddr_t)
3189 trunc_page((vm_offset_t)bp->b_data);
3190 pmap_qenter(
3191 (vm_offset_t)bp->b_data,
3192 bp->b_pages,
3193 bp->b_npages
3194 );
3195
3196 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3197 (vm_offset_t)(bp->b_offset & PAGE_MASK));
3198 }
3199 }
3200 if (newbsize < bp->b_bufsize)
3201 bufspacewakeup();
3202 bp->b_bufsize = newbsize; /* actual buffer allocation */
3203 bp->b_bcount = size; /* requested buffer size */
3204 return 1;
3205 }
3206
3207 void
3208 biodone(struct bio *bp)
3209 {
3210 struct mtx *mtxp;
3211 void (*done)(struct bio *);
3212
3213 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3214 mtx_lock(mtxp);
3215 bp->bio_flags |= BIO_DONE;
3216 done = bp->bio_done;
3217 if (done == NULL)
3218 wakeup(bp);
3219 mtx_unlock(mtxp);
3220 if (done != NULL)
3221 done(bp);
3222 }
3223
3224 /*
3225 * Wait for a BIO to finish.
3226 *
3227 * XXX: resort to a timeout for now. The optimal locking (if any) for this
3228 * case is not yet clear.
3229 */
3230 int
3231 biowait(struct bio *bp, const char *wchan)
3232 {
3233 struct mtx *mtxp;
3234
3235 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3236 mtx_lock(mtxp);
3237 while ((bp->bio_flags & BIO_DONE) == 0)
3238 msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3239 mtx_unlock(mtxp);
3240 if (bp->bio_error != 0)
3241 return (bp->bio_error);
3242 if (!(bp->bio_flags & BIO_ERROR))
3243 return (0);
3244 return (EIO);
3245 }
3246
3247 void
3248 biofinish(struct bio *bp, struct devstat *stat, int error)
3249 {
3250
3251 if (error) {
3252 bp->bio_error = error;
3253 bp->bio_flags |= BIO_ERROR;
3254 }
3255 if (stat != NULL)
3256 devstat_end_transaction_bio(stat, bp);
3257 biodone(bp);
3258 }
3259
3260 /*
3261 * bufwait:
3262 *
3263 * Wait for buffer I/O completion, returning error status. The buffer
3264 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3265 * error and cleared.
3266 */
3267 int
3268 bufwait(struct buf *bp)
3269 {
3270 if (bp->b_iocmd == BIO_READ)
3271 bwait(bp, PRIBIO, "biord");
3272 else
3273 bwait(bp, PRIBIO, "biowr");
3274 if (bp->b_flags & B_EINTR) {
3275 bp->b_flags &= ~B_EINTR;
3276 return (EINTR);
3277 }
3278 if (bp->b_ioflags & BIO_ERROR) {
3279 return (bp->b_error ? bp->b_error : EIO);
3280 } else {
3281 return (0);
3282 }
3283 }
3284
3285 /*
3286 * Call back function from struct bio back up to struct buf.
3287 */
3288 static void
3289 bufdonebio(struct bio *bip)
3290 {
3291 struct buf *bp;
3292
3293 bp = bip->bio_caller2;
3294 bp->b_resid = bp->b_bcount - bip->bio_completed;
3295 bp->b_resid = bip->bio_resid; /* XXX: remove */
3296 bp->b_ioflags = bip->bio_flags;
3297 bp->b_error = bip->bio_error;
3298 if (bp->b_error)
3299 bp->b_ioflags |= BIO_ERROR;
3300 bufdone(bp);
3301 g_destroy_bio(bip);
3302 }
3303
3304 void
3305 dev_strategy(struct cdev *dev, struct buf *bp)
3306 {
3307 struct cdevsw *csw;
3308 struct bio *bip;
3309 int ref;
3310
3311 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3312 panic("b_iocmd botch");
3313 for (;;) {
3314 bip = g_new_bio();
3315 if (bip != NULL)
3316 break;
3317 /* Try again later */
3318 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3319 }
3320 bip->bio_cmd = bp->b_iocmd;
3321 bip->bio_offset = bp->b_iooffset;
3322 bip->bio_length = bp->b_bcount;
3323 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3324 bip->bio_data = bp->b_data;
3325 bip->bio_done = bufdonebio;
3326 bip->bio_caller2 = bp;
3327 bip->bio_dev = dev;
3328 KASSERT(dev->si_refcount > 0,
3329 ("dev_strategy on un-referenced struct cdev *(%s)",
3330 devtoname(dev)));
3331 csw = dev_refthread(dev, &ref);
3332 if (csw == NULL) {
3333 g_destroy_bio(bip);
3334 bp->b_error = ENXIO;
3335 bp->b_ioflags = BIO_ERROR;
3336 bufdone(bp);
3337 return;
3338 }
3339 (*csw->d_strategy)(bip);
3340 dev_relthread(dev, ref);
3341 }
3342
3343 /*
3344 * bufdone:
3345 *
3346 * Finish I/O on a buffer, optionally calling a completion function.
3347 * This is usually called from an interrupt so process blocking is
3348 * not allowed.
3349 *
3350 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3351 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3352 * assuming B_INVAL is clear.
3353 *
3354 * For the VMIO case, we set B_CACHE if the op was a read and no
3355 * read error occured, or if the op was a write. B_CACHE is never
3356 * set if the buffer is invalid or otherwise uncacheable.
3357 *
3358 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3359 * initiator to leave B_INVAL set to brelse the buffer out of existance
3360 * in the biodone routine.
3361 */
3362 void
3363 bufdone(struct buf *bp)
3364 {
3365 struct bufobj *dropobj;
3366 void (*biodone)(struct buf *);
3367
3368 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3369 dropobj = NULL;
3370
3371 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3372 BUF_ASSERT_HELD(bp);
3373
3374 runningbufwakeup(bp);
3375 if (bp->b_iocmd == BIO_WRITE)
3376 dropobj = bp->b_bufobj;
3377 /* call optional completion function if requested */
3378 if (bp->b_iodone != NULL) {
3379 biodone = bp->b_iodone;
3380 bp->b_iodone = NULL;
3381 (*biodone) (bp);
3382 if (dropobj)
3383 bufobj_wdrop(dropobj);
3384 return;
3385 }
3386
3387 bufdone_finish(bp);
3388
3389 if (dropobj)
3390 bufobj_wdrop(dropobj);
3391 }
3392
3393 void
3394 bufdone_finish(struct buf *bp)
3395 {
3396 BUF_ASSERT_HELD(bp);
3397
3398 if (!LIST_EMPTY(&bp->b_dep))
3399 buf_complete(bp);
3400
3401 if (bp->b_flags & B_VMIO) {
3402 int i;
3403 vm_ooffset_t foff;
3404 vm_page_t m;
3405 vm_object_t obj;
3406 int bogus, iosize;
3407 struct vnode *vp = bp->b_vp;
3408
3409 obj = bp->b_bufobj->bo_object;
3410
3411 #if defined(VFS_BIO_DEBUG)
3412 mp_fixme("usecount and vflag accessed without locks.");
3413 if (vp->v_usecount == 0) {
3414 panic("biodone: zero vnode ref count");
3415 }
3416
3417 KASSERT(vp->v_object != NULL,
3418 ("biodone: vnode %p has no vm_object", vp));
3419 #endif
3420
3421 foff = bp->b_offset;
3422 KASSERT(bp->b_offset != NOOFFSET,
3423 ("biodone: no buffer offset"));
3424
3425 VM_OBJECT_LOCK(obj);
3426 #if defined(VFS_BIO_DEBUG)
3427 if (obj->paging_in_progress < bp->b_npages) {
3428 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3429 obj->paging_in_progress, bp->b_npages);
3430 }
3431 #endif
3432
3433 /*
3434 * Set B_CACHE if the op was a normal read and no error
3435 * occured. B_CACHE is set for writes in the b*write()
3436 * routines.
3437 */
3438 iosize = bp->b_bcount - bp->b_resid;
3439 if (bp->b_iocmd == BIO_READ &&
3440 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3441 !(bp->b_ioflags & BIO_ERROR)) {
3442 bp->b_flags |= B_CACHE;
3443 }
3444 bogus = 0;
3445 for (i = 0; i < bp->b_npages; i++) {
3446 int bogusflag = 0;
3447 int resid;
3448
3449 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3450 if (resid > iosize)
3451 resid = iosize;
3452
3453 /*
3454 * cleanup bogus pages, restoring the originals
3455 */
3456 m = bp->b_pages[i];
3457 if (m == bogus_page) {
3458 bogus = bogusflag = 1;
3459 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3460 if (m == NULL)
3461 panic("biodone: page disappeared!");
3462 bp->b_pages[i] = m;
3463 }
3464 #if defined(VFS_BIO_DEBUG)
3465 if (OFF_TO_IDX(foff) != m->pindex) {
3466 printf(
3467 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3468 (intmax_t)foff, (uintmax_t)m->pindex);
3469 }
3470 #endif
3471
3472 /*
3473 * In the write case, the valid and clean bits are
3474 * already changed correctly ( see bdwrite() ), so we
3475 * only need to do this here in the read case.
3476 */
3477 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3478 KASSERT((m->dirty & vm_page_bits(foff &
3479 PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3480 " page %p has unexpected dirty bits", m));
3481 vfs_page_set_valid(bp, foff, m);
3482 }
3483
3484 /*
3485 * when debugging new filesystems or buffer I/O methods, this
3486 * is the most common error that pops up. if you see this, you
3487 * have not set the page busy flag correctly!!!
3488 */
3489 if (m->busy == 0) {
3490 printf("biodone: page busy < 0, "
3491 "pindex: %d, foff: 0x(%x,%x), "
3492 "resid: %d, index: %d\n",
3493 (int) m->pindex, (int)(foff >> 32),
3494 (int) foff & 0xffffffff, resid, i);
3495 if (!vn_isdisk(vp, NULL))
3496 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3497 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3498 (intmax_t) bp->b_lblkno,
3499 bp->b_flags, bp->b_npages);
3500 else
3501 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3502 (intmax_t) bp->b_lblkno,
3503 bp->b_flags, bp->b_npages);
3504 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3505 (u_long)m->valid, (u_long)m->dirty,
3506 m->wire_count);
3507 panic("biodone: page busy < 0\n");
3508 }
3509 vm_page_io_finish(m);
3510 vm_object_pip_subtract(obj, 1);
3511 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3512 iosize -= resid;
3513 }
3514 vm_object_pip_wakeupn(obj, 0);
3515 VM_OBJECT_UNLOCK(obj);
3516 if (bogus)
3517 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3518 bp->b_pages, bp->b_npages);
3519 }
3520
3521 /*
3522 * For asynchronous completions, release the buffer now. The brelse
3523 * will do a wakeup there if necessary - so no need to do a wakeup
3524 * here in the async case. The sync case always needs to do a wakeup.
3525 */
3526
3527 if (bp->b_flags & B_ASYNC) {
3528 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3529 brelse(bp);
3530 else
3531 bqrelse(bp);
3532 } else
3533 bdone(bp);
3534 }
3535
3536 /*
3537 * This routine is called in lieu of iodone in the case of
3538 * incomplete I/O. This keeps the busy status for pages
3539 * consistant.
3540 */
3541 void
3542 vfs_unbusy_pages(struct buf *bp)
3543 {
3544 int i;
3545 vm_object_t obj;
3546 vm_page_t m;
3547
3548 runningbufwakeup(bp);
3549 if (!(bp->b_flags & B_VMIO))
3550 return;
3551
3552 obj = bp->b_bufobj->bo_object;
3553 VM_OBJECT_LOCK(obj);
3554 for (i = 0; i < bp->b_npages; i++) {
3555 m = bp->b_pages[i];
3556 if (m == bogus_page) {
3557 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3558 if (!m)
3559 panic("vfs_unbusy_pages: page missing\n");
3560 bp->b_pages[i] = m;
3561 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3562 bp->b_pages, bp->b_npages);
3563 }
3564 vm_object_pip_subtract(obj, 1);
3565 vm_page_io_finish(m);
3566 }
3567 vm_object_pip_wakeupn(obj, 0);
3568 VM_OBJECT_UNLOCK(obj);
3569 }
3570
3571 /*
3572 * vfs_page_set_valid:
3573 *
3574 * Set the valid bits in a page based on the supplied offset. The
3575 * range is restricted to the buffer's size.
3576 *
3577 * This routine is typically called after a read completes.
3578 */
3579 static void
3580 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3581 {
3582 vm_ooffset_t eoff;
3583
3584 /*
3585 * Compute the end offset, eoff, such that [off, eoff) does not span a
3586 * page boundary and eoff is not greater than the end of the buffer.
3587 * The end of the buffer, in this case, is our file EOF, not the
3588 * allocation size of the buffer.
3589 */
3590 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3591 if (eoff > bp->b_offset + bp->b_bcount)
3592 eoff = bp->b_offset + bp->b_bcount;
3593
3594 /*
3595 * Set valid range. This is typically the entire buffer and thus the
3596 * entire page.
3597 */
3598 if (eoff > off)
3599 vm_page_set_valid(m, off & PAGE_MASK, eoff - off);
3600 }
3601
3602 /*
3603 * vfs_page_set_validclean:
3604 *
3605 * Set the valid bits and clear the dirty bits in a page based on the
3606 * supplied offset. The range is restricted to the buffer's size.
3607 */
3608 static void
3609 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3610 {
3611 vm_ooffset_t soff, eoff;
3612
3613 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3614 /*
3615 * Start and end offsets in buffer. eoff - soff may not cross a
3616 * page boundry or cross the end of the buffer. The end of the
3617 * buffer, in this case, is our file EOF, not the allocation size
3618 * of the buffer.
3619 */
3620 soff = off;
3621 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3622 if (eoff > bp->b_offset + bp->b_bcount)
3623 eoff = bp->b_offset + bp->b_bcount;
3624
3625 /*
3626 * Set valid range. This is typically the entire buffer and thus the
3627 * entire page.
3628 */
3629 if (eoff > soff) {
3630 vm_page_set_validclean(
3631 m,
3632 (vm_offset_t) (soff & PAGE_MASK),
3633 (vm_offset_t) (eoff - soff)
3634 );
3635 }
3636 }
3637
3638 /*
3639 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3640 * any page is busy, drain the flag.
3641 */
3642 static void
3643 vfs_drain_busy_pages(struct buf *bp)
3644 {
3645 vm_page_t m;
3646 int i, last_busied;
3647
3648 VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED);
3649 last_busied = 0;
3650 for (i = 0; i < bp->b_npages; i++) {
3651 m = bp->b_pages[i];
3652 if ((m->oflags & VPO_BUSY) != 0) {
3653 for (; last_busied < i; last_busied++)
3654 vm_page_busy(bp->b_pages[last_busied]);
3655 while ((m->oflags & VPO_BUSY) != 0)
3656 vm_page_sleep(m, "vbpage");
3657 }
3658 }
3659 for (i = 0; i < last_busied; i++)
3660 vm_page_wakeup(bp->b_pages[i]);
3661 }
3662
3663 /*
3664 * This routine is called before a device strategy routine.
3665 * It is used to tell the VM system that paging I/O is in
3666 * progress, and treat the pages associated with the buffer
3667 * almost as being VPO_BUSY. Also the object paging_in_progress
3668 * flag is handled to make sure that the object doesn't become
3669 * inconsistant.
3670 *
3671 * Since I/O has not been initiated yet, certain buffer flags
3672 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3673 * and should be ignored.
3674 */
3675 void
3676 vfs_busy_pages(struct buf *bp, int clear_modify)
3677 {
3678 int i, bogus;
3679 vm_object_t obj;
3680 vm_ooffset_t foff;
3681 vm_page_t m;
3682
3683 if (!(bp->b_flags & B_VMIO))
3684 return;
3685
3686 obj = bp->b_bufobj->bo_object;
3687 foff = bp->b_offset;
3688 KASSERT(bp->b_offset != NOOFFSET,
3689 ("vfs_busy_pages: no buffer offset"));
3690 VM_OBJECT_LOCK(obj);
3691 vfs_drain_busy_pages(bp);
3692 if (bp->b_bufsize != 0)
3693 vfs_setdirty_locked_object(bp);
3694 bogus = 0;
3695 if (clear_modify)
3696 vm_page_lock_queues();
3697 for (i = 0; i < bp->b_npages; i++) {
3698 m = bp->b_pages[i];
3699
3700 if ((bp->b_flags & B_CLUSTER) == 0) {
3701 vm_object_pip_add(obj, 1);
3702 vm_page_io_start(m);
3703 }
3704 /*
3705 * When readying a buffer for a read ( i.e
3706 * clear_modify == 0 ), it is important to do
3707 * bogus_page replacement for valid pages in
3708 * partially instantiated buffers. Partially
3709 * instantiated buffers can, in turn, occur when
3710 * reconstituting a buffer from its VM backing store
3711 * base. We only have to do this if B_CACHE is
3712 * clear ( which causes the I/O to occur in the
3713 * first place ). The replacement prevents the read
3714 * I/O from overwriting potentially dirty VM-backed
3715 * pages. XXX bogus page replacement is, uh, bogus.
3716 * It may not work properly with small-block devices.
3717 * We need to find a better way.
3718 */
3719 if (clear_modify) {
3720 pmap_remove_write(m);
3721 vfs_page_set_validclean(bp, foff, m);
3722 } else if (m->valid == VM_PAGE_BITS_ALL &&
3723 (bp->b_flags & B_CACHE) == 0) {
3724 bp->b_pages[i] = bogus_page;
3725 bogus++;
3726 }
3727 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3728 }
3729 if (clear_modify)
3730 vm_page_unlock_queues();
3731 VM_OBJECT_UNLOCK(obj);
3732 if (bogus)
3733 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3734 bp->b_pages, bp->b_npages);
3735 }
3736
3737 /*
3738 * vfs_bio_set_valid:
3739 *
3740 * Set the range within the buffer to valid. The range is
3741 * relative to the beginning of the buffer, b_offset. Note that
3742 * b_offset itself may be offset from the beginning of the first
3743 * page.
3744 */
3745 void
3746 vfs_bio_set_valid(struct buf *bp, int base, int size)
3747 {
3748 int i, n;
3749 vm_page_t m;
3750
3751 if (!(bp->b_flags & B_VMIO))
3752 return;
3753
3754 /*
3755 * Fixup base to be relative to beginning of first page.
3756 * Set initial n to be the maximum number of bytes in the
3757 * first page that can be validated.
3758 */
3759 base += (bp->b_offset & PAGE_MASK);
3760 n = PAGE_SIZE - (base & PAGE_MASK);
3761
3762 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3763 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3764 m = bp->b_pages[i];
3765 if (n > size)
3766 n = size;
3767 vm_page_set_valid(m, base & PAGE_MASK, n);
3768 base += n;
3769 size -= n;
3770 n = PAGE_SIZE;
3771 }
3772 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3773 }
3774
3775 /*
3776 * vfs_bio_clrbuf:
3777 *
3778 * If the specified buffer is a non-VMIO buffer, clear the entire
3779 * buffer. If the specified buffer is a VMIO buffer, clear and
3780 * validate only the previously invalid portions of the buffer.
3781 * This routine essentially fakes an I/O, so we need to clear
3782 * BIO_ERROR and B_INVAL.
3783 *
3784 * Note that while we only theoretically need to clear through b_bcount,
3785 * we go ahead and clear through b_bufsize.
3786 */
3787 void
3788 vfs_bio_clrbuf(struct buf *bp)
3789 {
3790 int i, j, mask;
3791 caddr_t sa, ea;
3792
3793 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3794 clrbuf(bp);
3795 return;
3796 }
3797 bp->b_flags &= ~B_INVAL;
3798 bp->b_ioflags &= ~BIO_ERROR;
3799 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3800 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3801 (bp->b_offset & PAGE_MASK) == 0) {
3802 if (bp->b_pages[0] == bogus_page)
3803 goto unlock;
3804 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3805 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3806 if ((bp->b_pages[0]->valid & mask) == mask)
3807 goto unlock;
3808 if ((bp->b_pages[0]->valid & mask) == 0) {
3809 bzero(bp->b_data, bp->b_bufsize);
3810 bp->b_pages[0]->valid |= mask;
3811 goto unlock;
3812 }
3813 }
3814 ea = sa = bp->b_data;
3815 for(i = 0; i < bp->b_npages; i++, sa = ea) {
3816 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3817 ea = (caddr_t)(vm_offset_t)ulmin(
3818 (u_long)(vm_offset_t)ea,
3819 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3820 if (bp->b_pages[i] == bogus_page)
3821 continue;
3822 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3823 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3824 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3825 if ((bp->b_pages[i]->valid & mask) == mask)
3826 continue;
3827 if ((bp->b_pages[i]->valid & mask) == 0)
3828 bzero(sa, ea - sa);
3829 else {
3830 for (; sa < ea; sa += DEV_BSIZE, j++) {
3831 if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3832 bzero(sa, DEV_BSIZE);
3833 }
3834 }
3835 bp->b_pages[i]->valid |= mask;
3836 }
3837 unlock:
3838 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3839 bp->b_resid = 0;
3840 }
3841
3842 /*
3843 * vm_hold_load_pages and vm_hold_free_pages get pages into
3844 * a buffers address space. The pages are anonymous and are
3845 * not associated with a file object.
3846 */
3847 static void
3848 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3849 {
3850 vm_offset_t pg;
3851 vm_page_t p;
3852 int index;
3853
3854 to = round_page(to);
3855 from = round_page(from);
3856 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3857
3858 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3859 tryagain:
3860 /*
3861 * note: must allocate system pages since blocking here
3862 * could interfere with paging I/O, no matter which
3863 * process we are.
3864 */
3865 p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ |
3866 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3867 if (!p) {
3868 atomic_add_int(&vm_pageout_deficit,
3869 (to - pg) >> PAGE_SHIFT);
3870 VM_WAIT;
3871 goto tryagain;
3872 }
3873 pmap_qenter(pg, &p, 1);
3874 bp->b_pages[index] = p;
3875 }
3876 bp->b_npages = index;
3877 }
3878
3879 /* Return pages associated with this buf to the vm system */
3880 static void
3881 vm_hold_free_pages(struct buf *bp, int newbsize)
3882 {
3883 vm_offset_t from;
3884 vm_page_t p;
3885 int index, newnpages;
3886
3887 from = round_page((vm_offset_t)bp->b_data + newbsize);
3888 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3889 if (bp->b_npages > newnpages)
3890 pmap_qremove(from, bp->b_npages - newnpages);
3891 for (index = newnpages; index < bp->b_npages; index++) {
3892 p = bp->b_pages[index];
3893 bp->b_pages[index] = NULL;
3894 if (p->busy != 0)
3895 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3896 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
3897 p->wire_count--;
3898 vm_page_free(p);
3899 atomic_subtract_int(&cnt.v_wire_count, 1);
3900 }
3901 bp->b_npages = newnpages;
3902 }
3903
3904 /*
3905 * Map an IO request into kernel virtual address space.
3906 *
3907 * All requests are (re)mapped into kernel VA space.
3908 * Notice that we use b_bufsize for the size of the buffer
3909 * to be mapped. b_bcount might be modified by the driver.
3910 *
3911 * Note that even if the caller determines that the address space should
3912 * be valid, a race or a smaller-file mapped into a larger space may
3913 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3914 * check the return value.
3915 */
3916 int
3917 vmapbuf(struct buf *bp)
3918 {
3919 caddr_t addr, kva;
3920 vm_prot_t prot;
3921 int pidx, i;
3922 struct vm_page *m;
3923 struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3924
3925 if (bp->b_bufsize < 0)
3926 return (-1);
3927 prot = VM_PROT_READ;
3928 if (bp->b_iocmd == BIO_READ)
3929 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
3930 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3931 addr < bp->b_data + bp->b_bufsize;
3932 addr += PAGE_SIZE, pidx++) {
3933 /*
3934 * Do the vm_fault if needed; do the copy-on-write thing
3935 * when reading stuff off device into memory.
3936 *
3937 * NOTE! Must use pmap_extract() because addr may be in
3938 * the userland address space, and kextract is only guarenteed
3939 * to work for the kernland address space (see: sparc64 port).
3940 */
3941 retry:
3942 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3943 prot) < 0) {
3944 vm_page_lock_queues();
3945 for (i = 0; i < pidx; ++i) {
3946 vm_page_unhold(bp->b_pages[i]);
3947 bp->b_pages[i] = NULL;
3948 }
3949 vm_page_unlock_queues();
3950 return(-1);
3951 }
3952 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3953 if (m == NULL)
3954 goto retry;
3955 bp->b_pages[pidx] = m;
3956 }
3957 if (pidx > btoc(MAXPHYS))
3958 panic("vmapbuf: mapped more than MAXPHYS");
3959 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3960
3961 kva = bp->b_saveaddr;
3962 bp->b_npages = pidx;
3963 bp->b_saveaddr = bp->b_data;
3964 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3965 return(0);
3966 }
3967
3968 /*
3969 * Free the io map PTEs associated with this IO operation.
3970 * We also invalidate the TLB entries and restore the original b_addr.
3971 */
3972 void
3973 vunmapbuf(struct buf *bp)
3974 {
3975 int pidx;
3976 int npages;
3977
3978 npages = bp->b_npages;
3979 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3980 vm_page_lock_queues();
3981 for (pidx = 0; pidx < npages; pidx++)
3982 vm_page_unhold(bp->b_pages[pidx]);
3983 vm_page_unlock_queues();
3984
3985 bp->b_data = bp->b_saveaddr;
3986 }
3987
3988 void
3989 bdone(struct buf *bp)
3990 {
3991 struct mtx *mtxp;
3992
3993 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3994 mtx_lock(mtxp);
3995 bp->b_flags |= B_DONE;
3996 wakeup(bp);
3997 mtx_unlock(mtxp);
3998 }
3999
4000 void
4001 bwait(struct buf *bp, u_char pri, const char *wchan)
4002 {
4003 struct mtx *mtxp;
4004
4005 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4006 mtx_lock(mtxp);
4007 while ((bp->b_flags & B_DONE) == 0)
4008 msleep(bp, mtxp, pri, wchan, 0);
4009 mtx_unlock(mtxp);
4010 }
4011
4012 int
4013 bufsync(struct bufobj *bo, int waitfor)
4014 {
4015
4016 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4017 }
4018
4019 void
4020 bufstrategy(struct bufobj *bo, struct buf *bp)
4021 {
4022 int i = 0;
4023 struct vnode *vp;
4024
4025 vp = bp->b_vp;
4026 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4027 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4028 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4029 i = VOP_STRATEGY(vp, bp);
4030 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4031 }
4032
4033 void
4034 bufobj_wrefl(struct bufobj *bo)
4035 {
4036
4037 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4038 ASSERT_BO_LOCKED(bo);
4039 bo->bo_numoutput++;
4040 }
4041
4042 void
4043 bufobj_wref(struct bufobj *bo)
4044 {
4045
4046 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4047 BO_LOCK(bo);
4048 bo->bo_numoutput++;
4049 BO_UNLOCK(bo);
4050 }
4051
4052 void
4053 bufobj_wdrop(struct bufobj *bo)
4054 {
4055
4056 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4057 BO_LOCK(bo);
4058 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4059 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4060 bo->bo_flag &= ~BO_WWAIT;
4061 wakeup(&bo->bo_numoutput);
4062 }
4063 BO_UNLOCK(bo);
4064 }
4065
4066 int
4067 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4068 {
4069 int error;
4070
4071 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4072 ASSERT_BO_LOCKED(bo);
4073 error = 0;
4074 while (bo->bo_numoutput) {
4075 bo->bo_flag |= BO_WWAIT;
4076 error = msleep(&bo->bo_numoutput, BO_MTX(bo),
4077 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4078 if (error)
4079 break;
4080 }
4081 return (error);
4082 }
4083
4084 void
4085 bpin(struct buf *bp)
4086 {
4087 struct mtx *mtxp;
4088
4089 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4090 mtx_lock(mtxp);
4091 bp->b_pin_count++;
4092 mtx_unlock(mtxp);
4093 }
4094
4095 void
4096 bunpin(struct buf *bp)
4097 {
4098 struct mtx *mtxp;
4099
4100 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4101 mtx_lock(mtxp);
4102 if (--bp->b_pin_count == 0)
4103 wakeup(bp);
4104 mtx_unlock(mtxp);
4105 }
4106
4107 void
4108 bunpin_wait(struct buf *bp)
4109 {
4110 struct mtx *mtxp;
4111
4112 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4113 mtx_lock(mtxp);
4114 while (bp->b_pin_count > 0)
4115 msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4116 mtx_unlock(mtxp);
4117 }
4118
4119 #include "opt_ddb.h"
4120 #ifdef DDB
4121 #include <ddb/ddb.h>
4122
4123 /* DDB command to show buffer data */
4124 DB_SHOW_COMMAND(buffer, db_show_buffer)
4125 {
4126 /* get args */
4127 struct buf *bp = (struct buf *)addr;
4128
4129 if (!have_addr) {
4130 db_printf("usage: show buffer <addr>\n");
4131 return;
4132 }
4133
4134 db_printf("buf at %p\n", bp);
4135 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4136 db_printf(
4137 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4138 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n",
4139 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4140 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4141 bp->b_dep.lh_first);
4142 if (bp->b_npages) {
4143 int i;
4144 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4145 for (i = 0; i < bp->b_npages; i++) {
4146 vm_page_t m;
4147 m = bp->b_pages[i];
4148 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4149 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4150 if ((i + 1) < bp->b_npages)
4151 db_printf(",");
4152 }
4153 db_printf("\n");
4154 }
4155 db_printf(" ");
4156 BUF_LOCKPRINTINFO(bp);
4157 }
4158
4159 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4160 {
4161 struct buf *bp;
4162 int i;
4163
4164 for (i = 0; i < nbuf; i++) {
4165 bp = &buf[i];
4166 if (BUF_ISLOCKED(bp)) {
4167 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4168 db_printf("\n");
4169 }
4170 }
4171 }
4172
4173 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4174 {
4175 struct vnode *vp;
4176 struct buf *bp;
4177
4178 if (!have_addr) {
4179 db_printf("usage: show vnodebufs <addr>\n");
4180 return;
4181 }
4182 vp = (struct vnode *)addr;
4183 db_printf("Clean buffers:\n");
4184 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4185 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4186 db_printf("\n");
4187 }
4188 db_printf("Dirty buffers:\n");
4189 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4190 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4191 db_printf("\n");
4192 }
4193 }
4194
4195 DB_COMMAND(countfreebufs, db_coundfreebufs)
4196 {
4197 struct buf *bp;
4198 int i, used = 0, nfree = 0;
4199
4200 if (have_addr) {
4201 db_printf("usage: countfreebufs\n");
4202 return;
4203 }
4204
4205 for (i = 0; i < nbuf; i++) {
4206 bp = &buf[i];
4207 if ((bp->b_vflags & BV_INFREECNT) != 0)
4208 nfree++;
4209 else
4210 used++;
4211 }
4212
4213 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4214 nfree + used);
4215 db_printf("numfreebuffers is %d\n", numfreebuffers);
4216 }
4217 #endif /* DDB */
Cache object: d4163623c53f8c757c6bf799f3465981
|