FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_bio.c
1 /*
2 * Copyright (c) 1994,1997 John S. Dyson
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
12 * John S. Dyson.
13 *
14 * $FreeBSD: releng/5.1/sys/kern/vfs_bio.c 114147 2003-04-28 05:40:45Z alc $
15 */
16
17 /*
18 * this file contains a new buffer I/O scheme implementing a coherent
19 * VM object and buffer cache scheme. Pains have been taken to make
20 * sure that the performance degradation associated with schemes such
21 * as this is not realized.
22 *
23 * Author: John S. Dyson
24 * Significant help during the development and debugging phases
25 * had been provided by David Greenman, also of the FreeBSD core team.
26 *
27 * see man buf(9) for more info.
28 */
29
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/bio.h>
33 #include <sys/buf.h>
34 #include <sys/devicestat.h>
35 #include <sys/eventhandler.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/mutex.h>
40 #include <sys/kernel.h>
41 #include <sys/kthread.h>
42 #include <sys/proc.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 #include <vm/vm.h>
48 #include <vm/vm_param.h>
49 #include <vm/vm_kern.h>
50 #include <vm/vm_pageout.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_map.h>
55 #include "opt_directio.h"
56 #include "opt_swap.h"
57
58 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
59
60 struct bio_ops bioops; /* I/O operation notification */
61
62 struct buf_ops buf_ops_bio = {
63 "buf_ops_bio",
64 bwrite
65 };
66
67 /*
68 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
69 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
70 */
71 struct buf *buf; /* buffer header pool */
72
73 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
74 vm_offset_t to);
75 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
76 vm_offset_t to);
77 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
78 int pageno, vm_page_t m);
79 static void vfs_clean_pages(struct buf * bp);
80 static void vfs_setdirty(struct buf *bp);
81 static void vfs_vmio_release(struct buf *bp);
82 static void vfs_backgroundwritedone(struct buf *bp);
83 static int vfs_bio_clcheck(struct vnode *vp, int size,
84 daddr_t lblkno, daddr_t blkno);
85 static int flushbufqueues(int flushdeps);
86 static void buf_daemon(void);
87 void bremfreel(struct buf * bp);
88
89 int vmiodirenable = TRUE;
90 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
91 "Use the VM system for directory writes");
92 int runningbufspace;
93 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
94 "Amount of presently outstanding async buffer io");
95 static int bufspace;
96 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
97 "KVA memory used for bufs");
98 static int maxbufspace;
99 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
100 "Maximum allowed value of bufspace (including buf_daemon)");
101 static int bufmallocspace;
102 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
103 "Amount of malloced memory for buffers");
104 static int maxbufmallocspace;
105 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
106 "Maximum amount of malloced memory for buffers");
107 static int lobufspace;
108 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
109 "Minimum amount of buffers we want to have");
110 static int hibufspace;
111 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
112 "Maximum allowed value of bufspace (excluding buf_daemon)");
113 static int bufreusecnt;
114 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
115 "Number of times we have reused a buffer");
116 static int buffreekvacnt;
117 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
118 "Number of times we have freed the KVA space from some buffer");
119 static int bufdefragcnt;
120 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
121 "Number of times we have had to repeat buffer allocation to defragment");
122 static int lorunningspace;
123 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
124 "Minimum preferred space used for in-progress I/O");
125 static int hirunningspace;
126 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
127 "Maximum amount of space to use for in-progress I/O");
128 static int dirtybufferflushes;
129 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
130 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
131 static int altbufferflushes;
132 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
133 0, "Number of fsync flushes to limit dirty buffers");
134 static int recursiveflushes;
135 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
136 0, "Number of flushes skipped due to being recursive");
137 static int numdirtybuffers;
138 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
139 "Number of buffers that are dirty (has unwritten changes) at the moment");
140 static int lodirtybuffers;
141 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
142 "How many buffers we want to have free before bufdaemon can sleep");
143 static int hidirtybuffers;
144 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
145 "When the number of dirty buffers is considered severe");
146 static int dirtybufthresh;
147 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
148 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
149 static int numfreebuffers;
150 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
151 "Number of free buffers");
152 static int lofreebuffers;
153 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
154 "XXX Unused");
155 static int hifreebuffers;
156 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
157 "XXX Complicatedly unused");
158 static int getnewbufcalls;
159 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
160 "Number of calls to getnewbuf");
161 static int getnewbufrestarts;
162 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
163 "Number of times getnewbuf has had to restart a buffer aquisition");
164 static int dobkgrdwrite = 1;
165 SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0,
166 "Do background writes (honoring the BX_BKGRDWRITE flag)?");
167
168 /*
169 * Wakeup point for bufdaemon, as well as indicator of whether it is already
170 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
171 * is idling.
172 */
173 static int bd_request;
174
175 /*
176 * This lock synchronizes access to bd_request.
177 */
178 static struct mtx bdlock;
179
180 /*
181 * bogus page -- for I/O to/from partially complete buffers
182 * this is a temporary solution to the problem, but it is not
183 * really that bad. it would be better to split the buffer
184 * for input in the case of buffers partially already in memory,
185 * but the code is intricate enough already.
186 */
187 vm_page_t bogus_page;
188
189 /*
190 * Synchronization (sleep/wakeup) variable for active buffer space requests.
191 * Set when wait starts, cleared prior to wakeup().
192 * Used in runningbufwakeup() and waitrunningbufspace().
193 */
194 static int runningbufreq;
195
196 /*
197 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
198 * waitrunningbufspace().
199 */
200 static struct mtx rbreqlock;
201
202 /*
203 * Synchronization (sleep/wakeup) variable for buffer requests.
204 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
205 * by and/or.
206 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
207 * getnewbuf(), and getblk().
208 */
209 static int needsbuffer;
210
211 /*
212 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
213 */
214 static struct mtx nblock;
215
216 /*
217 * Lock that protects against bwait()/bdone()/B_DONE races.
218 */
219
220 static struct mtx bdonelock;
221
222 /*
223 * Definitions for the buffer free lists.
224 */
225 #define BUFFER_QUEUES 6 /* number of free buffer queues */
226
227 #define QUEUE_NONE 0 /* on no queue */
228 #define QUEUE_LOCKED 1 /* locked buffers */
229 #define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */
230 #define QUEUE_DIRTY 3 /* B_DELWRI buffers */
231 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */
232 #define QUEUE_EMPTY 5 /* empty buffer headers */
233
234 /* Queues for free buffers with various properties */
235 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
236
237 /* Lock for the bufqueues */
238 static struct mtx bqlock;
239
240 /*
241 * Single global constant for BUF_WMESG, to avoid getting multiple references.
242 * buf_wmesg is referred from macros.
243 */
244 const char *buf_wmesg = BUF_WMESG;
245
246 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
247 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
248 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
249 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
250
251 #ifdef DIRECTIO
252 extern void ffs_rawread_setup(void);
253 #endif /* DIRECTIO */
254 /*
255 * numdirtywakeup:
256 *
257 * If someone is blocked due to there being too many dirty buffers,
258 * and numdirtybuffers is now reasonable, wake them up.
259 */
260
261 static __inline void
262 numdirtywakeup(int level)
263 {
264 if (numdirtybuffers <= level) {
265 mtx_lock(&nblock);
266 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
267 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
268 wakeup(&needsbuffer);
269 }
270 mtx_unlock(&nblock);
271 }
272 }
273
274 /*
275 * bufspacewakeup:
276 *
277 * Called when buffer space is potentially available for recovery.
278 * getnewbuf() will block on this flag when it is unable to free
279 * sufficient buffer space. Buffer space becomes recoverable when
280 * bp's get placed back in the queues.
281 */
282
283 static __inline void
284 bufspacewakeup(void)
285 {
286 /*
287 * If someone is waiting for BUF space, wake them up. Even
288 * though we haven't freed the kva space yet, the waiting
289 * process will be able to now.
290 */
291 mtx_lock(&nblock);
292 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
293 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
294 wakeup(&needsbuffer);
295 }
296 mtx_unlock(&nblock);
297 }
298
299 /*
300 * runningbufwakeup() - in-progress I/O accounting.
301 *
302 */
303 static __inline void
304 runningbufwakeup(struct buf *bp)
305 {
306 if (bp->b_runningbufspace) {
307 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
308 bp->b_runningbufspace = 0;
309 mtx_lock(&rbreqlock);
310 if (runningbufreq && runningbufspace <= lorunningspace) {
311 runningbufreq = 0;
312 wakeup(&runningbufreq);
313 }
314 mtx_unlock(&rbreqlock);
315 }
316 }
317
318 /*
319 * bufcountwakeup:
320 *
321 * Called when a buffer has been added to one of the free queues to
322 * account for the buffer and to wakeup anyone waiting for free buffers.
323 * This typically occurs when large amounts of metadata are being handled
324 * by the buffer cache ( else buffer space runs out first, usually ).
325 */
326
327 static __inline void
328 bufcountwakeup(void)
329 {
330 atomic_add_int(&numfreebuffers, 1);
331 mtx_lock(&nblock);
332 if (needsbuffer) {
333 needsbuffer &= ~VFS_BIO_NEED_ANY;
334 if (numfreebuffers >= hifreebuffers)
335 needsbuffer &= ~VFS_BIO_NEED_FREE;
336 wakeup(&needsbuffer);
337 }
338 mtx_unlock(&nblock);
339 }
340
341 /*
342 * waitrunningbufspace()
343 *
344 * runningbufspace is a measure of the amount of I/O currently
345 * running. This routine is used in async-write situations to
346 * prevent creating huge backups of pending writes to a device.
347 * Only asynchronous writes are governed by this function.
348 *
349 * Reads will adjust runningbufspace, but will not block based on it.
350 * The read load has a side effect of reducing the allowed write load.
351 *
352 * This does NOT turn an async write into a sync write. It waits
353 * for earlier writes to complete and generally returns before the
354 * caller's write has reached the device.
355 */
356 static __inline void
357 waitrunningbufspace(void)
358 {
359 mtx_lock(&rbreqlock);
360 while (runningbufspace > hirunningspace) {
361 ++runningbufreq;
362 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
363 }
364 mtx_unlock(&rbreqlock);
365 }
366
367
368 /*
369 * vfs_buf_test_cache:
370 *
371 * Called when a buffer is extended. This function clears the B_CACHE
372 * bit if the newly extended portion of the buffer does not contain
373 * valid data.
374 */
375 static __inline__
376 void
377 vfs_buf_test_cache(struct buf *bp,
378 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
379 vm_page_t m)
380 {
381 GIANT_REQUIRED;
382
383 if (bp->b_flags & B_CACHE) {
384 int base = (foff + off) & PAGE_MASK;
385 if (vm_page_is_valid(m, base, size) == 0)
386 bp->b_flags &= ~B_CACHE;
387 }
388 }
389
390 /* Wake up the buffer deamon if necessary */
391 static __inline__
392 void
393 bd_wakeup(int dirtybuflevel)
394 {
395 mtx_lock(&bdlock);
396 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
397 bd_request = 1;
398 wakeup(&bd_request);
399 }
400 mtx_unlock(&bdlock);
401 }
402
403 /*
404 * bd_speedup - speedup the buffer cache flushing code
405 */
406
407 static __inline__
408 void
409 bd_speedup(void)
410 {
411 bd_wakeup(1);
412 }
413
414 /*
415 * Calculating buffer cache scaling values and reserve space for buffer
416 * headers. This is called during low level kernel initialization and
417 * may be called more then once. We CANNOT write to the memory area
418 * being reserved at this time.
419 */
420 caddr_t
421 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
422 {
423 /*
424 * physmem_est is in pages. Convert it to kilobytes (assumes
425 * PAGE_SIZE is >= 1K)
426 */
427 physmem_est = physmem_est * (PAGE_SIZE / 1024);
428
429 /*
430 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
431 * For the first 64MB of ram nominally allocate sufficient buffers to
432 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
433 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing
434 * the buffer cache we limit the eventual kva reservation to
435 * maxbcache bytes.
436 *
437 * factor represents the 1/4 x ram conversion.
438 */
439 if (nbuf == 0) {
440 int factor = 4 * BKVASIZE / 1024;
441
442 nbuf = 50;
443 if (physmem_est > 4096)
444 nbuf += min((physmem_est - 4096) / factor,
445 65536 / factor);
446 if (physmem_est > 65536)
447 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
448
449 if (maxbcache && nbuf > maxbcache / BKVASIZE)
450 nbuf = maxbcache / BKVASIZE;
451 }
452
453 #if 0
454 /*
455 * Do not allow the buffer_map to be more then 1/2 the size of the
456 * kernel_map.
457 */
458 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
459 (BKVASIZE * 2)) {
460 nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
461 (BKVASIZE * 2);
462 printf("Warning: nbufs capped at %d\n", nbuf);
463 }
464 #endif
465
466 /*
467 * swbufs are used as temporary holders for I/O, such as paging I/O.
468 * We have no less then 16 and no more then 256.
469 */
470 nswbuf = max(min(nbuf/4, 256), 16);
471 #ifdef NSWBUF_MIN
472 if (nswbuf < NSWBUF_MIN)
473 nswbuf = NSWBUF_MIN;
474 #endif
475 #ifdef DIRECTIO
476 ffs_rawread_setup();
477 #endif
478
479 /*
480 * Reserve space for the buffer cache buffers
481 */
482 swbuf = (void *)v;
483 v = (caddr_t)(swbuf + nswbuf);
484 buf = (void *)v;
485 v = (caddr_t)(buf + nbuf);
486
487 return(v);
488 }
489
490 /* Initialize the buffer subsystem. Called before use of any buffers. */
491 void
492 bufinit(void)
493 {
494 struct buf *bp;
495 vm_offset_t bogus_offset;
496 int i;
497
498 GIANT_REQUIRED;
499
500 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
501 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
502 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
503 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
504 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
505
506 /* next, make a null set of free lists */
507 for (i = 0; i < BUFFER_QUEUES; i++)
508 TAILQ_INIT(&bufqueues[i]);
509
510 /* finally, initialize each buffer header and stick on empty q */
511 for (i = 0; i < nbuf; i++) {
512 bp = &buf[i];
513 bzero(bp, sizeof *bp);
514 bp->b_flags = B_INVAL; /* we're just an empty header */
515 bp->b_dev = NODEV;
516 bp->b_rcred = NOCRED;
517 bp->b_wcred = NOCRED;
518 bp->b_qindex = QUEUE_EMPTY;
519 bp->b_vflags = 0;
520 bp->b_xflags = 0;
521 LIST_INIT(&bp->b_dep);
522 BUF_LOCKINIT(bp);
523 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
524 }
525
526 /*
527 * maxbufspace is the absolute maximum amount of buffer space we are
528 * allowed to reserve in KVM and in real terms. The absolute maximum
529 * is nominally used by buf_daemon. hibufspace is the nominal maximum
530 * used by most other processes. The differential is required to
531 * ensure that buf_daemon is able to run when other processes might
532 * be blocked waiting for buffer space.
533 *
534 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
535 * this may result in KVM fragmentation which is not handled optimally
536 * by the system.
537 */
538 maxbufspace = nbuf * BKVASIZE;
539 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
540 lobufspace = hibufspace - MAXBSIZE;
541
542 lorunningspace = 512 * 1024;
543 hirunningspace = 1024 * 1024;
544
545 /*
546 * Limit the amount of malloc memory since it is wired permanently into
547 * the kernel space. Even though this is accounted for in the buffer
548 * allocation, we don't want the malloced region to grow uncontrolled.
549 * The malloc scheme improves memory utilization significantly on average
550 * (small) directories.
551 */
552 maxbufmallocspace = hibufspace / 20;
553
554 /*
555 * Reduce the chance of a deadlock occuring by limiting the number
556 * of delayed-write dirty buffers we allow to stack up.
557 */
558 hidirtybuffers = nbuf / 4 + 20;
559 dirtybufthresh = hidirtybuffers * 9 / 10;
560 numdirtybuffers = 0;
561 /*
562 * To support extreme low-memory systems, make sure hidirtybuffers cannot
563 * eat up all available buffer space. This occurs when our minimum cannot
564 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
565 * BKVASIZE'd (8K) buffers.
566 */
567 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
568 hidirtybuffers >>= 1;
569 }
570 lodirtybuffers = hidirtybuffers / 2;
571
572 /*
573 * Try to keep the number of free buffers in the specified range,
574 * and give special processes (e.g. like buf_daemon) access to an
575 * emergency reserve.
576 */
577 lofreebuffers = nbuf / 18 + 5;
578 hifreebuffers = 2 * lofreebuffers;
579 numfreebuffers = nbuf;
580
581 /*
582 * Maximum number of async ops initiated per buf_daemon loop. This is
583 * somewhat of a hack at the moment, we really need to limit ourselves
584 * based on the number of bytes of I/O in-transit that were initiated
585 * from buf_daemon.
586 */
587
588 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
589 VM_OBJECT_LOCK(kernel_object);
590 bogus_page = vm_page_alloc(kernel_object,
591 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
592 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
593 VM_OBJECT_UNLOCK(kernel_object);
594 }
595
596 /*
597 * bfreekva() - free the kva allocation for a buffer.
598 *
599 * Must be called at splbio() or higher as this is the only locking for
600 * buffer_map.
601 *
602 * Since this call frees up buffer space, we call bufspacewakeup().
603 */
604 static void
605 bfreekva(struct buf * bp)
606 {
607 GIANT_REQUIRED;
608
609 if (bp->b_kvasize) {
610 atomic_add_int(&buffreekvacnt, 1);
611 atomic_subtract_int(&bufspace, bp->b_kvasize);
612 vm_map_delete(buffer_map,
613 (vm_offset_t) bp->b_kvabase,
614 (vm_offset_t) bp->b_kvabase + bp->b_kvasize
615 );
616 bp->b_kvasize = 0;
617 bufspacewakeup();
618 }
619 }
620
621 /*
622 * bremfree:
623 *
624 * Remove the buffer from the appropriate free list.
625 */
626 void
627 bremfree(struct buf * bp)
628 {
629 mtx_lock(&bqlock);
630 bremfreel(bp);
631 mtx_unlock(&bqlock);
632 }
633
634 void
635 bremfreel(struct buf * bp)
636 {
637 int s = splbio();
638 int old_qindex = bp->b_qindex;
639
640 GIANT_REQUIRED;
641
642 if (bp->b_qindex != QUEUE_NONE) {
643 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
644 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
645 bp->b_qindex = QUEUE_NONE;
646 } else {
647 if (BUF_REFCNT(bp) <= 1)
648 panic("bremfree: removing a buffer not on a queue");
649 }
650
651 /*
652 * Fixup numfreebuffers count. If the buffer is invalid or not
653 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
654 * the buffer was free and we must decrement numfreebuffers.
655 */
656 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
657 switch(old_qindex) {
658 case QUEUE_DIRTY:
659 case QUEUE_CLEAN:
660 case QUEUE_EMPTY:
661 case QUEUE_EMPTYKVA:
662 atomic_subtract_int(&numfreebuffers, 1);
663 break;
664 default:
665 break;
666 }
667 }
668 splx(s);
669 }
670
671
672 /*
673 * Get a buffer with the specified data. Look in the cache first. We
674 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
675 * is set, the buffer is valid and we do not have to do anything ( see
676 * getblk() ). This is really just a special case of breadn().
677 */
678 int
679 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
680 struct buf ** bpp)
681 {
682
683 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
684 }
685
686 /*
687 * Operates like bread, but also starts asynchronous I/O on
688 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior
689 * to initiating I/O . If B_CACHE is set, the buffer is valid
690 * and we do not have to do anything.
691 */
692 int
693 breadn(struct vnode * vp, daddr_t blkno, int size,
694 daddr_t * rablkno, int *rabsize,
695 int cnt, struct ucred * cred, struct buf ** bpp)
696 {
697 struct buf *bp, *rabp;
698 int i;
699 int rv = 0, readwait = 0;
700
701 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
702
703 /* if not found in cache, do some I/O */
704 if ((bp->b_flags & B_CACHE) == 0) {
705 if (curthread != PCPU_GET(idlethread))
706 curthread->td_proc->p_stats->p_ru.ru_inblock++;
707 bp->b_iocmd = BIO_READ;
708 bp->b_flags &= ~B_INVAL;
709 bp->b_ioflags &= ~BIO_ERROR;
710 if (bp->b_rcred == NOCRED && cred != NOCRED)
711 bp->b_rcred = crhold(cred);
712 vfs_busy_pages(bp, 0);
713 if (vp->v_type == VCHR)
714 VOP_SPECSTRATEGY(vp, bp);
715 else
716 VOP_STRATEGY(vp, bp);
717 ++readwait;
718 }
719
720 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
721 if (inmem(vp, *rablkno))
722 continue;
723 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
724
725 if ((rabp->b_flags & B_CACHE) == 0) {
726 if (curthread != PCPU_GET(idlethread))
727 curthread->td_proc->p_stats->p_ru.ru_inblock++;
728 rabp->b_flags |= B_ASYNC;
729 rabp->b_flags &= ~B_INVAL;
730 rabp->b_ioflags &= ~BIO_ERROR;
731 rabp->b_iocmd = BIO_READ;
732 if (rabp->b_rcred == NOCRED && cred != NOCRED)
733 rabp->b_rcred = crhold(cred);
734 vfs_busy_pages(rabp, 0);
735 BUF_KERNPROC(rabp);
736 if (vp->v_type == VCHR)
737 VOP_SPECSTRATEGY(vp, rabp);
738 else
739 VOP_STRATEGY(vp, rabp);
740 } else {
741 brelse(rabp);
742 }
743 }
744
745 if (readwait) {
746 rv = bufwait(bp);
747 }
748 return (rv);
749 }
750
751 /*
752 * Write, release buffer on completion. (Done by iodone
753 * if async). Do not bother writing anything if the buffer
754 * is invalid.
755 *
756 * Note that we set B_CACHE here, indicating that buffer is
757 * fully valid and thus cacheable. This is true even of NFS
758 * now so we set it generally. This could be set either here
759 * or in biodone() since the I/O is synchronous. We put it
760 * here.
761 */
762
763 int
764 bwrite(struct buf * bp)
765 {
766 int oldflags, s;
767 struct buf *newbp;
768
769 if (bp->b_flags & B_INVAL) {
770 brelse(bp);
771 return (0);
772 }
773
774 oldflags = bp->b_flags;
775
776 if (BUF_REFCNT(bp) == 0)
777 panic("bwrite: buffer is not busy???");
778 s = splbio();
779 /*
780 * If a background write is already in progress, delay
781 * writing this block if it is asynchronous. Otherwise
782 * wait for the background write to complete.
783 */
784 if (bp->b_xflags & BX_BKGRDINPROG) {
785 if (bp->b_flags & B_ASYNC) {
786 splx(s);
787 bdwrite(bp);
788 return (0);
789 }
790 bp->b_xflags |= BX_BKGRDWAIT;
791 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0);
792 if (bp->b_xflags & BX_BKGRDINPROG)
793 panic("bwrite: still writing");
794 }
795
796 /* Mark the buffer clean */
797 bundirty(bp);
798
799 /*
800 * If this buffer is marked for background writing and we
801 * do not have to wait for it, make a copy and write the
802 * copy so as to leave this buffer ready for further use.
803 *
804 * This optimization eats a lot of memory. If we have a page
805 * or buffer shortfall we can't do it.
806 */
807 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) &&
808 (bp->b_flags & B_ASYNC) &&
809 !vm_page_count_severe() &&
810 !buf_dirty_count_severe()) {
811 if (bp->b_iodone != NULL) {
812 printf("bp->b_iodone = %p\n", bp->b_iodone);
813 panic("bwrite: need chained iodone");
814 }
815
816 /* get a new block */
817 newbp = geteblk(bp->b_bufsize);
818
819 /*
820 * set it to be identical to the old block. We have to
821 * set b_lblkno and BKGRDMARKER before calling bgetvp()
822 * to avoid confusing the splay tree and gbincore().
823 */
824 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
825 newbp->b_lblkno = bp->b_lblkno;
826 newbp->b_xflags |= BX_BKGRDMARKER;
827 /* XXX The BX_ flags need to be protected as well */
828 VI_LOCK(bp->b_vp);
829 bgetvp(bp->b_vp, newbp);
830 VI_UNLOCK(bp->b_vp);
831 newbp->b_blkno = bp->b_blkno;
832 newbp->b_offset = bp->b_offset;
833 newbp->b_iodone = vfs_backgroundwritedone;
834 newbp->b_flags |= B_ASYNC;
835 newbp->b_flags &= ~B_INVAL;
836
837 /* move over the dependencies */
838 if (LIST_FIRST(&bp->b_dep) != NULL)
839 buf_movedeps(bp, newbp);
840
841 /*
842 * Initiate write on the copy, release the original to
843 * the B_LOCKED queue so that it cannot go away until
844 * the background write completes. If not locked it could go
845 * away and then be reconstituted while it was being written.
846 * If the reconstituted buffer were written, we could end up
847 * with two background copies being written at the same time.
848 */
849 bp->b_xflags |= BX_BKGRDINPROG;
850 bp->b_flags |= B_LOCKED;
851 bqrelse(bp);
852 bp = newbp;
853 }
854
855 bp->b_flags &= ~B_DONE;
856 bp->b_ioflags &= ~BIO_ERROR;
857 bp->b_flags |= B_WRITEINPROG | B_CACHE;
858 bp->b_iocmd = BIO_WRITE;
859
860 VI_LOCK(bp->b_vp);
861 bp->b_vp->v_numoutput++;
862 VI_UNLOCK(bp->b_vp);
863 vfs_busy_pages(bp, 1);
864
865 /*
866 * Normal bwrites pipeline writes
867 */
868 bp->b_runningbufspace = bp->b_bufsize;
869 atomic_add_int(&runningbufspace, bp->b_runningbufspace);
870
871 if (curthread != PCPU_GET(idlethread))
872 curthread->td_proc->p_stats->p_ru.ru_oublock++;
873 splx(s);
874 if (oldflags & B_ASYNC)
875 BUF_KERNPROC(bp);
876 if (bp->b_vp->v_type == VCHR)
877 VOP_SPECSTRATEGY(bp->b_vp, bp);
878 else
879 VOP_STRATEGY(bp->b_vp, bp);
880
881 if ((oldflags & B_ASYNC) == 0) {
882 int rtval = bufwait(bp);
883 brelse(bp);
884 return (rtval);
885 } else if ((oldflags & B_NOWDRAIN) == 0) {
886 /*
887 * don't allow the async write to saturate the I/O
888 * system. Deadlocks can occur only if a device strategy
889 * routine (like in MD) turns around and issues another
890 * high-level write, in which case B_NOWDRAIN is expected
891 * to be set. Otherwise we will not deadlock here because
892 * we are blocking waiting for I/O that is already in-progress
893 * to complete.
894 */
895 waitrunningbufspace();
896 }
897
898 return (0);
899 }
900
901 /*
902 * Complete a background write started from bwrite.
903 */
904 static void
905 vfs_backgroundwritedone(bp)
906 struct buf *bp;
907 {
908 struct buf *origbp;
909
910 /*
911 * Find the original buffer that we are writing.
912 */
913 VI_LOCK(bp->b_vp);
914 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
915 panic("backgroundwritedone: lost buffer");
916 VI_UNLOCK(bp->b_vp);
917 /*
918 * Process dependencies then return any unfinished ones.
919 */
920 if (LIST_FIRST(&bp->b_dep) != NULL)
921 buf_complete(bp);
922 if (LIST_FIRST(&bp->b_dep) != NULL)
923 buf_movedeps(bp, origbp);
924
925 /* XXX Find out if origbp can disappear or get inconsistent */
926 /*
927 * Clear the BX_BKGRDINPROG flag in the original buffer
928 * and awaken it if it is waiting for the write to complete.
929 * If BX_BKGRDINPROG is not set in the original buffer it must
930 * have been released and re-instantiated - which is not legal.
931 */
932 KASSERT((origbp->b_xflags & BX_BKGRDINPROG),
933 ("backgroundwritedone: lost buffer2"));
934 origbp->b_xflags &= ~BX_BKGRDINPROG;
935 if (origbp->b_xflags & BX_BKGRDWAIT) {
936 origbp->b_xflags &= ~BX_BKGRDWAIT;
937 wakeup(&origbp->b_xflags);
938 }
939 /*
940 * Clear the B_LOCKED flag and remove it from the locked
941 * queue if it currently resides there.
942 */
943 origbp->b_flags &= ~B_LOCKED;
944 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) {
945 bremfree(origbp);
946 bqrelse(origbp);
947 }
948 /*
949 * This buffer is marked B_NOCACHE, so when it is released
950 * by biodone, it will be tossed. We mark it with BIO_READ
951 * to avoid biodone doing a second vwakeup.
952 */
953 bp->b_flags |= B_NOCACHE;
954 bp->b_iocmd = BIO_READ;
955 bp->b_flags &= ~(B_CACHE | B_DONE);
956 bp->b_iodone = 0;
957 bufdone(bp);
958 }
959
960 /*
961 * Delayed write. (Buffer is marked dirty). Do not bother writing
962 * anything if the buffer is marked invalid.
963 *
964 * Note that since the buffer must be completely valid, we can safely
965 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
966 * biodone() in order to prevent getblk from writing the buffer
967 * out synchronously.
968 */
969 void
970 bdwrite(struct buf * bp)
971 {
972 struct thread *td = curthread;
973 struct vnode *vp;
974 struct buf *nbp;
975
976 GIANT_REQUIRED;
977
978 if (BUF_REFCNT(bp) == 0)
979 panic("bdwrite: buffer is not busy");
980
981 if (bp->b_flags & B_INVAL) {
982 brelse(bp);
983 return;
984 }
985
986 /*
987 * If we have too many dirty buffers, don't create any more.
988 * If we are wildly over our limit, then force a complete
989 * cleanup. Otherwise, just keep the situation from getting
990 * out of control. Note that we have to avoid a recursive
991 * disaster and not try to clean up after our own cleanup!
992 */
993 vp = bp->b_vp;
994 VI_LOCK(vp);
995 if (td->td_proc->p_flag & P_COWINPROGRESS) {
996 recursiveflushes++;
997 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) {
998 VI_UNLOCK(vp);
999 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td);
1000 VI_LOCK(vp);
1001 altbufferflushes++;
1002 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) {
1003 /*
1004 * Try to find a buffer to flush.
1005 */
1006 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) {
1007 if ((nbp->b_xflags & BX_BKGRDINPROG) ||
1008 buf_countdeps(nbp, 0) ||
1009 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
1010 continue;
1011 if (bp == nbp)
1012 panic("bdwrite: found ourselves");
1013 VI_UNLOCK(vp);
1014 if (nbp->b_flags & B_CLUSTEROK) {
1015 vfs_bio_awrite(nbp);
1016 } else {
1017 bremfree(nbp);
1018 bawrite(nbp);
1019 }
1020 VI_LOCK(vp);
1021 dirtybufferflushes++;
1022 break;
1023 }
1024 }
1025 VI_UNLOCK(vp);
1026
1027 bdirty(bp);
1028 /*
1029 * Set B_CACHE, indicating that the buffer is fully valid. This is
1030 * true even of NFS now.
1031 */
1032 bp->b_flags |= B_CACHE;
1033
1034 /*
1035 * This bmap keeps the system from needing to do the bmap later,
1036 * perhaps when the system is attempting to do a sync. Since it
1037 * is likely that the indirect block -- or whatever other datastructure
1038 * that the filesystem needs is still in memory now, it is a good
1039 * thing to do this. Note also, that if the pageout daemon is
1040 * requesting a sync -- there might not be enough memory to do
1041 * the bmap then... So, this is important to do.
1042 */
1043 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1044 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1045 }
1046
1047 /*
1048 * Set the *dirty* buffer range based upon the VM system dirty pages.
1049 */
1050 vfs_setdirty(bp);
1051
1052 /*
1053 * We need to do this here to satisfy the vnode_pager and the
1054 * pageout daemon, so that it thinks that the pages have been
1055 * "cleaned". Note that since the pages are in a delayed write
1056 * buffer -- the VFS layer "will" see that the pages get written
1057 * out on the next sync, or perhaps the cluster will be completed.
1058 */
1059 vfs_clean_pages(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 * Must be called at splbio().
1092 * The buffer must be on QUEUE_NONE.
1093 */
1094 void
1095 bdirty(bp)
1096 struct buf *bp;
1097 {
1098 KASSERT(bp->b_qindex == QUEUE_NONE,
1099 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1100 bp->b_flags &= ~(B_RELBUF);
1101 bp->b_iocmd = BIO_WRITE;
1102
1103 if ((bp->b_flags & B_DELWRI) == 0) {
1104 bp->b_flags |= B_DONE | B_DELWRI;
1105 reassignbuf(bp, bp->b_vp);
1106 atomic_add_int(&numdirtybuffers, 1);
1107 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1108 }
1109 }
1110
1111 /*
1112 * bundirty:
1113 *
1114 * Clear B_DELWRI for buffer.
1115 *
1116 * Since the buffer is not on a queue, we do not update the numfreebuffers
1117 * count.
1118 *
1119 * Must be called at splbio().
1120 * The buffer must be on QUEUE_NONE.
1121 */
1122
1123 void
1124 bundirty(bp)
1125 struct buf *bp;
1126 {
1127 KASSERT(bp->b_qindex == QUEUE_NONE,
1128 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1129
1130 if (bp->b_flags & B_DELWRI) {
1131 bp->b_flags &= ~B_DELWRI;
1132 reassignbuf(bp, bp->b_vp);
1133 atomic_subtract_int(&numdirtybuffers, 1);
1134 numdirtywakeup(lodirtybuffers);
1135 }
1136 /*
1137 * Since it is now being written, we can clear its deferred write flag.
1138 */
1139 bp->b_flags &= ~B_DEFERRED;
1140 }
1141
1142 /*
1143 * bawrite:
1144 *
1145 * Asynchronous write. Start output on a buffer, but do not wait for
1146 * it to complete. The buffer is released when the output completes.
1147 *
1148 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1149 * B_INVAL buffers. Not us.
1150 */
1151 void
1152 bawrite(struct buf * bp)
1153 {
1154 bp->b_flags |= B_ASYNC;
1155 (void) BUF_WRITE(bp);
1156 }
1157
1158 /*
1159 * bwillwrite:
1160 *
1161 * Called prior to the locking of any vnodes when we are expecting to
1162 * write. We do not want to starve the buffer cache with too many
1163 * dirty buffers so we block here. By blocking prior to the locking
1164 * of any vnodes we attempt to avoid the situation where a locked vnode
1165 * prevents the various system daemons from flushing related buffers.
1166 */
1167
1168 void
1169 bwillwrite(void)
1170 {
1171 if (numdirtybuffers >= hidirtybuffers) {
1172 int s;
1173
1174 mtx_lock(&Giant);
1175 s = splbio();
1176 mtx_lock(&nblock);
1177 while (numdirtybuffers >= hidirtybuffers) {
1178 bd_wakeup(1);
1179 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1180 msleep(&needsbuffer, &nblock,
1181 (PRIBIO + 4), "flswai", 0);
1182 }
1183 splx(s);
1184 mtx_unlock(&nblock);
1185 mtx_unlock(&Giant);
1186 }
1187 }
1188
1189 /*
1190 * Return true if we have too many dirty buffers.
1191 */
1192 int
1193 buf_dirty_count_severe(void)
1194 {
1195 return(numdirtybuffers >= hidirtybuffers);
1196 }
1197
1198 /*
1199 * brelse:
1200 *
1201 * Release a busy buffer and, if requested, free its resources. The
1202 * buffer will be stashed in the appropriate bufqueue[] allowing it
1203 * to be accessed later as a cache entity or reused for other purposes.
1204 */
1205 void
1206 brelse(struct buf * bp)
1207 {
1208 int s;
1209
1210 GIANT_REQUIRED;
1211
1212 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1213 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1214
1215 s = splbio();
1216
1217 if (bp->b_flags & B_LOCKED)
1218 bp->b_ioflags &= ~BIO_ERROR;
1219
1220 if (bp->b_iocmd == BIO_WRITE &&
1221 (bp->b_ioflags & BIO_ERROR) &&
1222 !(bp->b_flags & B_INVAL)) {
1223 /*
1224 * Failed write, redirty. Must clear BIO_ERROR to prevent
1225 * pages from being scrapped. If B_INVAL is set then
1226 * this case is not run and the next case is run to
1227 * destroy the buffer. B_INVAL can occur if the buffer
1228 * is outside the range supported by the underlying device.
1229 */
1230 bp->b_ioflags &= ~BIO_ERROR;
1231 bdirty(bp);
1232 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1233 (bp->b_ioflags & BIO_ERROR) ||
1234 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) {
1235 /*
1236 * Either a failed I/O or we were asked to free or not
1237 * cache the buffer.
1238 */
1239 bp->b_flags |= B_INVAL;
1240 if (LIST_FIRST(&bp->b_dep) != NULL)
1241 buf_deallocate(bp);
1242 if (bp->b_flags & B_DELWRI) {
1243 atomic_subtract_int(&numdirtybuffers, 1);
1244 numdirtywakeup(lodirtybuffers);
1245 }
1246 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1247 if ((bp->b_flags & B_VMIO) == 0) {
1248 if (bp->b_bufsize)
1249 allocbuf(bp, 0);
1250 if (bp->b_vp)
1251 brelvp(bp);
1252 }
1253 }
1254
1255 /*
1256 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1257 * is called with B_DELWRI set, the underlying pages may wind up
1258 * getting freed causing a previous write (bdwrite()) to get 'lost'
1259 * because pages associated with a B_DELWRI bp are marked clean.
1260 *
1261 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1262 * if B_DELWRI is set.
1263 *
1264 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1265 * on pages to return pages to the VM page queues.
1266 */
1267 if (bp->b_flags & B_DELWRI)
1268 bp->b_flags &= ~B_RELBUF;
1269 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG))
1270 bp->b_flags |= B_RELBUF;
1271
1272 /*
1273 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1274 * constituted, not even NFS buffers now. Two flags effect this. If
1275 * B_INVAL, the struct buf is invalidated but the VM object is kept
1276 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1277 *
1278 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1279 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1280 * buffer is also B_INVAL because it hits the re-dirtying code above.
1281 *
1282 * Normally we can do this whether a buffer is B_DELWRI or not. If
1283 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1284 * the commit state and we cannot afford to lose the buffer. If the
1285 * buffer has a background write in progress, we need to keep it
1286 * around to prevent it from being reconstituted and starting a second
1287 * background write.
1288 */
1289 if ((bp->b_flags & B_VMIO)
1290 && !(bp->b_vp->v_mount != NULL &&
1291 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1292 !vn_isdisk(bp->b_vp, NULL) &&
1293 (bp->b_flags & B_DELWRI))
1294 ) {
1295
1296 int i, j, resid;
1297 vm_page_t m;
1298 off_t foff;
1299 vm_pindex_t poff;
1300 vm_object_t obj;
1301 struct vnode *vp;
1302
1303 vp = bp->b_vp;
1304 obj = bp->b_object;
1305
1306 /*
1307 * Get the base offset and length of the buffer. Note that
1308 * in the VMIO case if the buffer block size is not
1309 * page-aligned then b_data pointer may not be page-aligned.
1310 * But our b_pages[] array *IS* page aligned.
1311 *
1312 * block sizes less then DEV_BSIZE (usually 512) are not
1313 * supported due to the page granularity bits (m->valid,
1314 * m->dirty, etc...).
1315 *
1316 * See man buf(9) for more information
1317 */
1318 resid = bp->b_bufsize;
1319 foff = bp->b_offset;
1320 if (obj != NULL)
1321 VM_OBJECT_LOCK(obj);
1322 for (i = 0; i < bp->b_npages; i++) {
1323 int had_bogus = 0;
1324
1325 m = bp->b_pages[i];
1326 vm_page_lock_queues();
1327 vm_page_flag_clear(m, PG_ZERO);
1328 vm_page_unlock_queues();
1329
1330 /*
1331 * If we hit a bogus page, fixup *all* the bogus pages
1332 * now.
1333 */
1334 if (m == bogus_page) {
1335 poff = OFF_TO_IDX(bp->b_offset);
1336 had_bogus = 1;
1337
1338 for (j = i; j < bp->b_npages; j++) {
1339 vm_page_t mtmp;
1340 mtmp = bp->b_pages[j];
1341 if (mtmp == bogus_page) {
1342 mtmp = vm_page_lookup(obj, poff + j);
1343 if (!mtmp) {
1344 panic("brelse: page missing\n");
1345 }
1346 bp->b_pages[j] = mtmp;
1347 }
1348 }
1349
1350 if ((bp->b_flags & B_INVAL) == 0) {
1351 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1352 }
1353 m = bp->b_pages[i];
1354 }
1355 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) {
1356 int poffset = foff & PAGE_MASK;
1357 int presid = resid > (PAGE_SIZE - poffset) ?
1358 (PAGE_SIZE - poffset) : resid;
1359
1360 KASSERT(presid >= 0, ("brelse: extra page"));
1361 vm_page_set_invalid(m, poffset, presid);
1362 if (had_bogus)
1363 printf("avoided corruption bug in bogus_page/brelse code\n");
1364 }
1365 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1366 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1367 }
1368 if (obj != NULL)
1369 VM_OBJECT_UNLOCK(obj);
1370 if (bp->b_flags & (B_INVAL | B_RELBUF))
1371 vfs_vmio_release(bp);
1372
1373 } else if (bp->b_flags & B_VMIO) {
1374
1375 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1376 vfs_vmio_release(bp);
1377 }
1378
1379 }
1380
1381 if (bp->b_qindex != QUEUE_NONE)
1382 panic("brelse: free buffer onto another queue???");
1383 if (BUF_REFCNT(bp) > 1) {
1384 /* do not release to free list */
1385 BUF_UNLOCK(bp);
1386 splx(s);
1387 return;
1388 }
1389
1390 /* enqueue */
1391 mtx_lock(&bqlock);
1392
1393 /* buffers with no memory */
1394 if (bp->b_bufsize == 0) {
1395 bp->b_flags |= B_INVAL;
1396 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1397 if (bp->b_xflags & BX_BKGRDINPROG)
1398 panic("losing buffer 1");
1399 if (bp->b_kvasize) {
1400 bp->b_qindex = QUEUE_EMPTYKVA;
1401 } else {
1402 bp->b_qindex = QUEUE_EMPTY;
1403 }
1404 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1405 bp->b_dev = NODEV;
1406 /* buffers with junk contents */
1407 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1408 (bp->b_ioflags & BIO_ERROR)) {
1409 bp->b_flags |= B_INVAL;
1410 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1411 if (bp->b_xflags & BX_BKGRDINPROG)
1412 panic("losing buffer 2");
1413 bp->b_qindex = QUEUE_CLEAN;
1414 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1415 bp->b_dev = NODEV;
1416
1417 /* buffers that are locked */
1418 } else if (bp->b_flags & B_LOCKED) {
1419 bp->b_qindex = QUEUE_LOCKED;
1420 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1421
1422 /* remaining buffers */
1423 } else {
1424 if (bp->b_flags & B_DELWRI)
1425 bp->b_qindex = QUEUE_DIRTY;
1426 else
1427 bp->b_qindex = QUEUE_CLEAN;
1428 if (bp->b_flags & B_AGE)
1429 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1430 else
1431 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1432 }
1433 mtx_unlock(&bqlock);
1434
1435 /*
1436 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already
1437 * placed the buffer on the correct queue. We must also disassociate
1438 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1439 * find it.
1440 */
1441 if (bp->b_flags & B_INVAL) {
1442 if (bp->b_flags & B_DELWRI)
1443 bundirty(bp);
1444 if (bp->b_vp)
1445 brelvp(bp);
1446 }
1447
1448 /*
1449 * Fixup numfreebuffers count. The bp is on an appropriate queue
1450 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1451 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1452 * if B_INVAL is set ).
1453 */
1454
1455 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1456 bufcountwakeup();
1457
1458 /*
1459 * Something we can maybe free or reuse
1460 */
1461 if (bp->b_bufsize || bp->b_kvasize)
1462 bufspacewakeup();
1463
1464 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1465 B_DIRECT | B_NOWDRAIN);
1466 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1467 panic("brelse: not dirty");
1468 /* unlock */
1469 BUF_UNLOCK(bp);
1470 splx(s);
1471 }
1472
1473 /*
1474 * Release a buffer back to the appropriate queue but do not try to free
1475 * it. The buffer is expected to be used again soon.
1476 *
1477 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1478 * biodone() to requeue an async I/O on completion. It is also used when
1479 * known good buffers need to be requeued but we think we may need the data
1480 * again soon.
1481 *
1482 * XXX we should be able to leave the B_RELBUF hint set on completion.
1483 */
1484 void
1485 bqrelse(struct buf * bp)
1486 {
1487 int s;
1488
1489 s = splbio();
1490
1491 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1492
1493 if (bp->b_qindex != QUEUE_NONE)
1494 panic("bqrelse: free buffer onto another queue???");
1495 if (BUF_REFCNT(bp) > 1) {
1496 /* do not release to free list */
1497 BUF_UNLOCK(bp);
1498 splx(s);
1499 return;
1500 }
1501 mtx_lock(&bqlock);
1502 if (bp->b_flags & B_LOCKED) {
1503 bp->b_ioflags &= ~BIO_ERROR;
1504 bp->b_qindex = QUEUE_LOCKED;
1505 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1506 /* buffers with stale but valid contents */
1507 } else if (bp->b_flags & B_DELWRI) {
1508 bp->b_qindex = QUEUE_DIRTY;
1509 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1510 } else if (vm_page_count_severe()) {
1511 /*
1512 * We are too low on memory, we have to try to free the
1513 * buffer (most importantly: the wired pages making up its
1514 * backing store) *now*.
1515 */
1516 mtx_unlock(&bqlock);
1517 splx(s);
1518 brelse(bp);
1519 return;
1520 } else {
1521 bp->b_qindex = QUEUE_CLEAN;
1522 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1523 }
1524 mtx_unlock(&bqlock);
1525
1526 if ((bp->b_flags & B_LOCKED) == 0 &&
1527 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1528 bufcountwakeup();
1529 }
1530
1531 /*
1532 * Something we can maybe free or reuse.
1533 */
1534 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1535 bufspacewakeup();
1536
1537 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1538 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1539 panic("bqrelse: not dirty");
1540 /* unlock */
1541 BUF_UNLOCK(bp);
1542 splx(s);
1543 }
1544
1545 /* Give pages used by the bp back to the VM system (where possible) */
1546 static void
1547 vfs_vmio_release(bp)
1548 struct buf *bp;
1549 {
1550 int i;
1551 vm_page_t m;
1552
1553 GIANT_REQUIRED;
1554 vm_page_lock_queues();
1555 for (i = 0; i < bp->b_npages; i++) {
1556 m = bp->b_pages[i];
1557 bp->b_pages[i] = NULL;
1558 /*
1559 * In order to keep page LRU ordering consistent, put
1560 * everything on the inactive queue.
1561 */
1562 vm_page_unwire(m, 0);
1563 /*
1564 * We don't mess with busy pages, it is
1565 * the responsibility of the process that
1566 * busied the pages to deal with them.
1567 */
1568 if ((m->flags & PG_BUSY) || (m->busy != 0))
1569 continue;
1570
1571 if (m->wire_count == 0) {
1572 vm_page_flag_clear(m, PG_ZERO);
1573 /*
1574 * Might as well free the page if we can and it has
1575 * no valid data. We also free the page if the
1576 * buffer was used for direct I/O
1577 */
1578 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1579 m->hold_count == 0) {
1580 vm_page_busy(m);
1581 pmap_remove_all(m);
1582 vm_page_free(m);
1583 } else if (bp->b_flags & B_DIRECT) {
1584 vm_page_try_to_free(m);
1585 } else if (vm_page_count_severe()) {
1586 vm_page_try_to_cache(m);
1587 }
1588 }
1589 }
1590 vm_page_unlock_queues();
1591 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1592
1593 if (bp->b_bufsize) {
1594 bufspacewakeup();
1595 bp->b_bufsize = 0;
1596 }
1597 bp->b_npages = 0;
1598 bp->b_flags &= ~B_VMIO;
1599 if (bp->b_vp)
1600 brelvp(bp);
1601 }
1602
1603 /*
1604 * Check to see if a block at a particular lbn is available for a clustered
1605 * write.
1606 */
1607 static int
1608 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1609 {
1610 struct buf *bpa;
1611 int match;
1612
1613 match = 0;
1614
1615 /* If the buf isn't in core skip it */
1616 if ((bpa = gbincore(vp, lblkno)) == NULL)
1617 return (0);
1618
1619 /* If the buf is busy we don't want to wait for it */
1620 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1621 return (0);
1622
1623 /* Only cluster with valid clusterable delayed write buffers */
1624 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1625 (B_DELWRI | B_CLUSTEROK))
1626 goto done;
1627
1628 if (bpa->b_bufsize != size)
1629 goto done;
1630
1631 /*
1632 * Check to see if it is in the expected place on disk and that the
1633 * block has been mapped.
1634 */
1635 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1636 match = 1;
1637 done:
1638 BUF_UNLOCK(bpa);
1639 return (match);
1640 }
1641
1642 /*
1643 * vfs_bio_awrite:
1644 *
1645 * Implement clustered async writes for clearing out B_DELWRI buffers.
1646 * This is much better then the old way of writing only one buffer at
1647 * a time. Note that we may not be presented with the buffers in the
1648 * correct order, so we search for the cluster in both directions.
1649 */
1650 int
1651 vfs_bio_awrite(struct buf * bp)
1652 {
1653 int i;
1654 int j;
1655 daddr_t lblkno = bp->b_lblkno;
1656 struct vnode *vp = bp->b_vp;
1657 int s;
1658 int ncl;
1659 int nwritten;
1660 int size;
1661 int maxcl;
1662
1663 s = splbio();
1664 /*
1665 * right now we support clustered writing only to regular files. If
1666 * we find a clusterable block we could be in the middle of a cluster
1667 * rather then at the beginning.
1668 */
1669 if ((vp->v_type == VREG) &&
1670 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1671 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1672
1673 size = vp->v_mount->mnt_stat.f_iosize;
1674 maxcl = MAXPHYS / size;
1675
1676 VI_LOCK(vp);
1677 for (i = 1; i < maxcl; i++)
1678 if (vfs_bio_clcheck(vp, size, lblkno + i,
1679 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1680 break;
1681
1682 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1683 if (vfs_bio_clcheck(vp, size, lblkno - j,
1684 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1685 break;
1686
1687 VI_UNLOCK(vp);
1688 --j;
1689 ncl = i + j;
1690 /*
1691 * this is a possible cluster write
1692 */
1693 if (ncl != 1) {
1694 BUF_UNLOCK(bp);
1695 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1696 splx(s);
1697 return nwritten;
1698 }
1699 }
1700
1701 bremfree(bp);
1702 bp->b_flags |= B_ASYNC;
1703
1704 splx(s);
1705 /*
1706 * default (old) behavior, writing out only one block
1707 *
1708 * XXX returns b_bufsize instead of b_bcount for nwritten?
1709 */
1710 nwritten = bp->b_bufsize;
1711 (void) BUF_WRITE(bp);
1712
1713 return nwritten;
1714 }
1715
1716 /*
1717 * getnewbuf:
1718 *
1719 * Find and initialize a new buffer header, freeing up existing buffers
1720 * in the bufqueues as necessary. The new buffer is returned locked.
1721 *
1722 * Important: B_INVAL is not set. If the caller wishes to throw the
1723 * buffer away, the caller must set B_INVAL prior to calling brelse().
1724 *
1725 * We block if:
1726 * We have insufficient buffer headers
1727 * We have insufficient buffer space
1728 * buffer_map is too fragmented ( space reservation fails )
1729 * If we have to flush dirty buffers ( but we try to avoid this )
1730 *
1731 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1732 * Instead we ask the buf daemon to do it for us. We attempt to
1733 * avoid piecemeal wakeups of the pageout daemon.
1734 */
1735
1736 static struct buf *
1737 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1738 {
1739 struct buf *bp;
1740 struct buf *nbp;
1741 int defrag = 0;
1742 int nqindex;
1743 static int flushingbufs;
1744
1745 GIANT_REQUIRED;
1746
1747 /*
1748 * We can't afford to block since we might be holding a vnode lock,
1749 * which may prevent system daemons from running. We deal with
1750 * low-memory situations by proactively returning memory and running
1751 * async I/O rather then sync I/O.
1752 */
1753
1754 atomic_add_int(&getnewbufcalls, 1);
1755 atomic_subtract_int(&getnewbufrestarts, 1);
1756 restart:
1757 atomic_add_int(&getnewbufrestarts, 1);
1758
1759 /*
1760 * Setup for scan. If we do not have enough free buffers,
1761 * we setup a degenerate case that immediately fails. Note
1762 * that if we are specially marked process, we are allowed to
1763 * dip into our reserves.
1764 *
1765 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1766 *
1767 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1768 * However, there are a number of cases (defragging, reusing, ...)
1769 * where we cannot backup.
1770 */
1771 mtx_lock(&bqlock);
1772 nqindex = QUEUE_EMPTYKVA;
1773 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1774
1775 if (nbp == NULL) {
1776 /*
1777 * If no EMPTYKVA buffers and we are either
1778 * defragging or reusing, locate a CLEAN buffer
1779 * to free or reuse. If bufspace useage is low
1780 * skip this step so we can allocate a new buffer.
1781 */
1782 if (defrag || bufspace >= lobufspace) {
1783 nqindex = QUEUE_CLEAN;
1784 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1785 }
1786
1787 /*
1788 * If we could not find or were not allowed to reuse a
1789 * CLEAN buffer, check to see if it is ok to use an EMPTY
1790 * buffer. We can only use an EMPTY buffer if allocating
1791 * its KVA would not otherwise run us out of buffer space.
1792 */
1793 if (nbp == NULL && defrag == 0 &&
1794 bufspace + maxsize < hibufspace) {
1795 nqindex = QUEUE_EMPTY;
1796 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1797 }
1798 }
1799
1800 /*
1801 * Run scan, possibly freeing data and/or kva mappings on the fly
1802 * depending.
1803 */
1804
1805 while ((bp = nbp) != NULL) {
1806 int qindex = nqindex;
1807
1808 /*
1809 * Calculate next bp ( we can only use it if we do not block
1810 * or do other fancy things ).
1811 */
1812 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1813 switch(qindex) {
1814 case QUEUE_EMPTY:
1815 nqindex = QUEUE_EMPTYKVA;
1816 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1817 break;
1818 /* FALLTHROUGH */
1819 case QUEUE_EMPTYKVA:
1820 nqindex = QUEUE_CLEAN;
1821 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1822 break;
1823 /* FALLTHROUGH */
1824 case QUEUE_CLEAN:
1825 /*
1826 * nbp is NULL.
1827 */
1828 break;
1829 }
1830 }
1831
1832 /*
1833 * Sanity Checks
1834 */
1835 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1836
1837 /*
1838 * Note: we no longer distinguish between VMIO and non-VMIO
1839 * buffers.
1840 */
1841
1842 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1843
1844 /*
1845 * If we are defragging then we need a buffer with
1846 * b_kvasize != 0. XXX this situation should no longer
1847 * occur, if defrag is non-zero the buffer's b_kvasize
1848 * should also be non-zero at this point. XXX
1849 */
1850 if (defrag && bp->b_kvasize == 0) {
1851 printf("Warning: defrag empty buffer %p\n", bp);
1852 continue;
1853 }
1854
1855 /*
1856 * Start freeing the bp. This is somewhat involved. nbp
1857 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1858 */
1859
1860 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1861 panic("getnewbuf: locked buf");
1862 bremfreel(bp);
1863 mtx_unlock(&bqlock);
1864
1865 if (qindex == QUEUE_CLEAN) {
1866 if (bp->b_flags & B_VMIO) {
1867 bp->b_flags &= ~B_ASYNC;
1868 vfs_vmio_release(bp);
1869 }
1870 if (bp->b_vp)
1871 brelvp(bp);
1872 }
1873
1874 /*
1875 * NOTE: nbp is now entirely invalid. We can only restart
1876 * the scan from this point on.
1877 *
1878 * Get the rest of the buffer freed up. b_kva* is still
1879 * valid after this operation.
1880 */
1881
1882 if (bp->b_rcred != NOCRED) {
1883 crfree(bp->b_rcred);
1884 bp->b_rcred = NOCRED;
1885 }
1886 if (bp->b_wcred != NOCRED) {
1887 crfree(bp->b_wcred);
1888 bp->b_wcred = NOCRED;
1889 }
1890 if (LIST_FIRST(&bp->b_dep) != NULL)
1891 buf_deallocate(bp);
1892 if (bp->b_xflags & BX_BKGRDINPROG)
1893 panic("losing buffer 3");
1894
1895 if (bp->b_bufsize)
1896 allocbuf(bp, 0);
1897
1898 bp->b_flags = 0;
1899 bp->b_ioflags = 0;
1900 bp->b_xflags = 0;
1901 bp->b_vflags = 0;
1902 bp->b_dev = NODEV;
1903 bp->b_vp = NULL;
1904 bp->b_blkno = bp->b_lblkno = 0;
1905 bp->b_offset = NOOFFSET;
1906 bp->b_iodone = 0;
1907 bp->b_error = 0;
1908 bp->b_resid = 0;
1909 bp->b_bcount = 0;
1910 bp->b_npages = 0;
1911 bp->b_dirtyoff = bp->b_dirtyend = 0;
1912 bp->b_magic = B_MAGIC_BIO;
1913 bp->b_op = &buf_ops_bio;
1914 bp->b_object = NULL;
1915
1916 LIST_INIT(&bp->b_dep);
1917
1918 /*
1919 * If we are defragging then free the buffer.
1920 */
1921 if (defrag) {
1922 bp->b_flags |= B_INVAL;
1923 bfreekva(bp);
1924 brelse(bp);
1925 defrag = 0;
1926 goto restart;
1927 }
1928
1929 /*
1930 * If we are overcomitted then recover the buffer and its
1931 * KVM space. This occurs in rare situations when multiple
1932 * processes are blocked in getnewbuf() or allocbuf().
1933 */
1934 if (bufspace >= hibufspace)
1935 flushingbufs = 1;
1936 if (flushingbufs && bp->b_kvasize != 0) {
1937 bp->b_flags |= B_INVAL;
1938 bfreekva(bp);
1939 brelse(bp);
1940 goto restart;
1941 }
1942 if (bufspace < lobufspace)
1943 flushingbufs = 0;
1944 break;
1945 }
1946
1947 /*
1948 * If we exhausted our list, sleep as appropriate. We may have to
1949 * wakeup various daemons and write out some dirty buffers.
1950 *
1951 * Generally we are sleeping due to insufficient buffer space.
1952 */
1953
1954 if (bp == NULL) {
1955 int flags;
1956 char *waitmsg;
1957
1958 mtx_unlock(&bqlock);
1959 if (defrag) {
1960 flags = VFS_BIO_NEED_BUFSPACE;
1961 waitmsg = "nbufkv";
1962 } else if (bufspace >= hibufspace) {
1963 waitmsg = "nbufbs";
1964 flags = VFS_BIO_NEED_BUFSPACE;
1965 } else {
1966 waitmsg = "newbuf";
1967 flags = VFS_BIO_NEED_ANY;
1968 }
1969
1970 bd_speedup(); /* heeeelp */
1971
1972 mtx_lock(&nblock);
1973 needsbuffer |= flags;
1974 while (needsbuffer & flags) {
1975 if (msleep(&needsbuffer, &nblock,
1976 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1977 mtx_unlock(&nblock);
1978 return (NULL);
1979 }
1980 }
1981 mtx_unlock(&nblock);
1982 } else {
1983 /*
1984 * We finally have a valid bp. We aren't quite out of the
1985 * woods, we still have to reserve kva space. In order
1986 * to keep fragmentation sane we only allocate kva in
1987 * BKVASIZE chunks.
1988 */
1989 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1990
1991 if (maxsize != bp->b_kvasize) {
1992 vm_offset_t addr = 0;
1993
1994 bfreekva(bp);
1995
1996 if (vm_map_findspace(buffer_map,
1997 vm_map_min(buffer_map), maxsize, &addr)) {
1998 /*
1999 * Uh oh. Buffer map is to fragmented. We
2000 * must defragment the map.
2001 */
2002 atomic_add_int(&bufdefragcnt, 1);
2003 defrag = 1;
2004 bp->b_flags |= B_INVAL;
2005 brelse(bp);
2006 goto restart;
2007 }
2008 if (addr) {
2009 vm_map_insert(buffer_map, NULL, 0,
2010 addr, addr + maxsize,
2011 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2012
2013 bp->b_kvabase = (caddr_t) addr;
2014 bp->b_kvasize = maxsize;
2015 atomic_add_int(&bufspace, bp->b_kvasize);
2016 atomic_add_int(&bufreusecnt, 1);
2017 }
2018 }
2019 bp->b_data = bp->b_kvabase;
2020 }
2021 return(bp);
2022 }
2023
2024 /*
2025 * buf_daemon:
2026 *
2027 * buffer flushing daemon. Buffers are normally flushed by the
2028 * update daemon but if it cannot keep up this process starts to
2029 * take the load in an attempt to prevent getnewbuf() from blocking.
2030 */
2031
2032 static struct proc *bufdaemonproc;
2033
2034 static struct kproc_desc buf_kp = {
2035 "bufdaemon",
2036 buf_daemon,
2037 &bufdaemonproc
2038 };
2039 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
2040
2041 static void
2042 buf_daemon()
2043 {
2044 int s;
2045
2046 mtx_lock(&Giant);
2047
2048 /*
2049 * This process needs to be suspended prior to shutdown sync.
2050 */
2051 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2052 SHUTDOWN_PRI_LAST);
2053
2054 /*
2055 * This process is allowed to take the buffer cache to the limit
2056 */
2057 s = splbio();
2058 mtx_lock(&bdlock);
2059
2060 for (;;) {
2061 bd_request = 0;
2062 mtx_unlock(&bdlock);
2063
2064 kthread_suspend_check(bufdaemonproc);
2065
2066 /*
2067 * Do the flush. Limit the amount of in-transit I/O we
2068 * allow to build up, otherwise we would completely saturate
2069 * the I/O system. Wakeup any waiting processes before we
2070 * normally would so they can run in parallel with our drain.
2071 */
2072 while (numdirtybuffers > lodirtybuffers) {
2073 if (flushbufqueues(0) == 0) {
2074 /*
2075 * Could not find any buffers without rollback
2076 * dependencies, so just write the first one
2077 * in the hopes of eventually making progress.
2078 */
2079 flushbufqueues(1);
2080 break;
2081 }
2082 waitrunningbufspace();
2083 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2084 }
2085
2086 /*
2087 * Only clear bd_request if we have reached our low water
2088 * mark. The buf_daemon normally waits 1 second and
2089 * then incrementally flushes any dirty buffers that have
2090 * built up, within reason.
2091 *
2092 * If we were unable to hit our low water mark and couldn't
2093 * find any flushable buffers, we sleep half a second.
2094 * Otherwise we loop immediately.
2095 */
2096 mtx_lock(&bdlock);
2097 if (numdirtybuffers <= lodirtybuffers) {
2098 /*
2099 * We reached our low water mark, reset the
2100 * request and sleep until we are needed again.
2101 * The sleep is just so the suspend code works.
2102 */
2103 bd_request = 0;
2104 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2105 } else {
2106 /*
2107 * We couldn't find any flushable dirty buffers but
2108 * still have too many dirty buffers, we
2109 * have to sleep and try again. (rare)
2110 */
2111 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2112 }
2113 }
2114 }
2115
2116 /*
2117 * flushbufqueues:
2118 *
2119 * Try to flush a buffer in the dirty queue. We must be careful to
2120 * free up B_INVAL buffers instead of write them, which NFS is
2121 * particularly sensitive to.
2122 */
2123 int flushwithdeps = 0;
2124 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2125 0, "Number of buffers flushed with dependecies that require rollbacks");
2126 static int
2127 flushbufqueues(int flushdeps)
2128 {
2129 struct thread *td = curthread;
2130 struct vnode *vp;
2131 struct buf *bp;
2132 int hasdeps;
2133
2134 mtx_lock(&bqlock);
2135 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) {
2136 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2137 continue;
2138 KASSERT((bp->b_flags & B_DELWRI),
2139 ("unexpected clean buffer %p", bp));
2140 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) {
2141 BUF_UNLOCK(bp);
2142 continue;
2143 }
2144 if (bp->b_flags & B_INVAL) {
2145 bremfreel(bp);
2146 mtx_unlock(&bqlock);
2147 brelse(bp);
2148 return (1);
2149 }
2150
2151 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) {
2152 if (flushdeps == 0) {
2153 BUF_UNLOCK(bp);
2154 continue;
2155 }
2156 hasdeps = 1;
2157 } else
2158 hasdeps = 0;
2159 /*
2160 * We must hold the lock on a vnode before writing
2161 * one of its buffers. Otherwise we may confuse, or
2162 * in the case of a snapshot vnode, deadlock the
2163 * system.
2164 *
2165 * The lock order here is the reverse of the normal
2166 * of vnode followed by buf lock. This is ok because
2167 * the NOWAIT will prevent deadlock.
2168 */
2169 if ((vp = bp->b_vp) == NULL ||
2170 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2171 mtx_unlock(&bqlock);
2172 vfs_bio_awrite(bp);
2173 if (vp != NULL)
2174 VOP_UNLOCK(vp, 0, td);
2175 flushwithdeps += hasdeps;
2176 return (1);
2177 }
2178 BUF_UNLOCK(bp);
2179 }
2180 mtx_unlock(&bqlock);
2181 return (0);
2182 }
2183
2184 /*
2185 * Check to see if a block is currently memory resident.
2186 */
2187 struct buf *
2188 incore(struct vnode * vp, daddr_t blkno)
2189 {
2190 struct buf *bp;
2191
2192 int s = splbio();
2193 VI_LOCK(vp);
2194 bp = gbincore(vp, blkno);
2195 VI_UNLOCK(vp);
2196 splx(s);
2197 return (bp);
2198 }
2199
2200 /*
2201 * Returns true if no I/O is needed to access the
2202 * associated VM object. This is like incore except
2203 * it also hunts around in the VM system for the data.
2204 */
2205
2206 int
2207 inmem(struct vnode * vp, daddr_t blkno)
2208 {
2209 vm_object_t obj;
2210 vm_offset_t toff, tinc, size;
2211 vm_page_t m;
2212 vm_ooffset_t off;
2213
2214 GIANT_REQUIRED;
2215 ASSERT_VOP_LOCKED(vp, "inmem");
2216
2217 if (incore(vp, blkno))
2218 return 1;
2219 if (vp->v_mount == NULL)
2220 return 0;
2221 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0)
2222 return 0;
2223
2224 size = PAGE_SIZE;
2225 if (size > vp->v_mount->mnt_stat.f_iosize)
2226 size = vp->v_mount->mnt_stat.f_iosize;
2227 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2228
2229 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2230 VM_OBJECT_LOCK(obj);
2231 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2232 VM_OBJECT_UNLOCK(obj);
2233 if (!m)
2234 goto notinmem;
2235 tinc = size;
2236 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2237 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2238 if (vm_page_is_valid(m,
2239 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2240 goto notinmem;
2241 }
2242 return 1;
2243
2244 notinmem:
2245 return (0);
2246 }
2247
2248 /*
2249 * vfs_setdirty:
2250 *
2251 * Sets the dirty range for a buffer based on the status of the dirty
2252 * bits in the pages comprising the buffer.
2253 *
2254 * The range is limited to the size of the buffer.
2255 *
2256 * This routine is primarily used by NFS, but is generalized for the
2257 * B_VMIO case.
2258 */
2259 static void
2260 vfs_setdirty(struct buf *bp)
2261 {
2262 int i;
2263 vm_object_t object;
2264
2265 GIANT_REQUIRED;
2266 /*
2267 * Degenerate case - empty buffer
2268 */
2269
2270 if (bp->b_bufsize == 0)
2271 return;
2272
2273 /*
2274 * We qualify the scan for modified pages on whether the
2275 * object has been flushed yet. The OBJ_WRITEABLE flag
2276 * is not cleared simply by protecting pages off.
2277 */
2278
2279 if ((bp->b_flags & B_VMIO) == 0)
2280 return;
2281
2282 object = bp->b_pages[0]->object;
2283 VM_OBJECT_LOCK(object);
2284 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2285 printf("Warning: object %p writeable but not mightbedirty\n", object);
2286 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2287 printf("Warning: object %p mightbedirty but not writeable\n", object);
2288
2289 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2290 vm_offset_t boffset;
2291 vm_offset_t eoffset;
2292
2293 vm_page_lock_queues();
2294 /*
2295 * test the pages to see if they have been modified directly
2296 * by users through the VM system.
2297 */
2298 for (i = 0; i < bp->b_npages; i++) {
2299 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
2300 vm_page_test_dirty(bp->b_pages[i]);
2301 }
2302
2303 /*
2304 * Calculate the encompassing dirty range, boffset and eoffset,
2305 * (eoffset - boffset) bytes.
2306 */
2307
2308 for (i = 0; i < bp->b_npages; i++) {
2309 if (bp->b_pages[i]->dirty)
2310 break;
2311 }
2312 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2313
2314 for (i = bp->b_npages - 1; i >= 0; --i) {
2315 if (bp->b_pages[i]->dirty) {
2316 break;
2317 }
2318 }
2319 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2320
2321 vm_page_unlock_queues();
2322 /*
2323 * Fit it to the buffer.
2324 */
2325
2326 if (eoffset > bp->b_bcount)
2327 eoffset = bp->b_bcount;
2328
2329 /*
2330 * If we have a good dirty range, merge with the existing
2331 * dirty range.
2332 */
2333
2334 if (boffset < eoffset) {
2335 if (bp->b_dirtyoff > boffset)
2336 bp->b_dirtyoff = boffset;
2337 if (bp->b_dirtyend < eoffset)
2338 bp->b_dirtyend = eoffset;
2339 }
2340 }
2341 VM_OBJECT_UNLOCK(object);
2342 }
2343
2344 /*
2345 * getblk:
2346 *
2347 * Get a block given a specified block and offset into a file/device.
2348 * The buffers B_DONE bit will be cleared on return, making it almost
2349 * ready for an I/O initiation. B_INVAL may or may not be set on
2350 * return. The caller should clear B_INVAL prior to initiating a
2351 * READ.
2352 *
2353 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2354 * an existing buffer.
2355 *
2356 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2357 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2358 * and then cleared based on the backing VM. If the previous buffer is
2359 * non-0-sized but invalid, B_CACHE will be cleared.
2360 *
2361 * If getblk() must create a new buffer, the new buffer is returned with
2362 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2363 * case it is returned with B_INVAL clear and B_CACHE set based on the
2364 * backing VM.
2365 *
2366 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos
2367 * B_CACHE bit is clear.
2368 *
2369 * What this means, basically, is that the caller should use B_CACHE to
2370 * determine whether the buffer is fully valid or not and should clear
2371 * B_INVAL prior to issuing a read. If the caller intends to validate
2372 * the buffer by loading its data area with something, the caller needs
2373 * to clear B_INVAL. If the caller does this without issuing an I/O,
2374 * the caller should set B_CACHE ( as an optimization ), else the caller
2375 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2376 * a write attempt or if it was a successfull read. If the caller
2377 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2378 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2379 */
2380 struct buf *
2381 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2382 int flags)
2383 {
2384 struct buf *bp;
2385 int s;
2386 int error;
2387 ASSERT_VOP_LOCKED(vp, "getblk");
2388
2389 if (size > MAXBSIZE)
2390 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2391
2392 s = splbio();
2393 loop:
2394 /*
2395 * Block if we are low on buffers. Certain processes are allowed
2396 * to completely exhaust the buffer cache.
2397 *
2398 * If this check ever becomes a bottleneck it may be better to
2399 * move it into the else, when gbincore() fails. At the moment
2400 * it isn't a problem.
2401 *
2402 * XXX remove if 0 sections (clean this up after its proven)
2403 */
2404 if (numfreebuffers == 0) {
2405 if (curthread == PCPU_GET(idlethread))
2406 return NULL;
2407 mtx_lock(&nblock);
2408 needsbuffer |= VFS_BIO_NEED_ANY;
2409 mtx_unlock(&nblock);
2410 }
2411
2412 VI_LOCK(vp);
2413 if ((bp = gbincore(vp, blkno))) {
2414 int lockflags;
2415 /*
2416 * Buffer is in-core. If the buffer is not busy, it must
2417 * be on a queue.
2418 */
2419 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2420
2421 if (flags & GB_LOCK_NOWAIT)
2422 lockflags |= LK_NOWAIT;
2423
2424 error = BUF_TIMELOCK(bp, lockflags,
2425 VI_MTX(vp), "getblk", slpflag, slptimeo);
2426
2427 /*
2428 * If we slept and got the lock we have to restart in case
2429 * the buffer changed identities.
2430 */
2431 if (error == ENOLCK)
2432 goto loop;
2433 /* We timed out or were interrupted. */
2434 else if (error)
2435 return (NULL);
2436
2437 /*
2438 * The buffer is locked. B_CACHE is cleared if the buffer is
2439 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2440 * and for a VMIO buffer B_CACHE is adjusted according to the
2441 * backing VM cache.
2442 */
2443 if (bp->b_flags & B_INVAL)
2444 bp->b_flags &= ~B_CACHE;
2445 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2446 bp->b_flags |= B_CACHE;
2447 bremfree(bp);
2448
2449 /*
2450 * check for size inconsistancies for non-VMIO case.
2451 */
2452
2453 if (bp->b_bcount != size) {
2454 if ((bp->b_flags & B_VMIO) == 0 ||
2455 (size > bp->b_kvasize)) {
2456 if (bp->b_flags & B_DELWRI) {
2457 bp->b_flags |= B_NOCACHE;
2458 BUF_WRITE(bp);
2459 } else {
2460 if ((bp->b_flags & B_VMIO) &&
2461 (LIST_FIRST(&bp->b_dep) == NULL)) {
2462 bp->b_flags |= B_RELBUF;
2463 brelse(bp);
2464 } else {
2465 bp->b_flags |= B_NOCACHE;
2466 BUF_WRITE(bp);
2467 }
2468 }
2469 goto loop;
2470 }
2471 }
2472
2473 /*
2474 * If the size is inconsistant in the VMIO case, we can resize
2475 * the buffer. This might lead to B_CACHE getting set or
2476 * cleared. If the size has not changed, B_CACHE remains
2477 * unchanged from its previous state.
2478 */
2479
2480 if (bp->b_bcount != size)
2481 allocbuf(bp, size);
2482
2483 KASSERT(bp->b_offset != NOOFFSET,
2484 ("getblk: no buffer offset"));
2485
2486 /*
2487 * A buffer with B_DELWRI set and B_CACHE clear must
2488 * be committed before we can return the buffer in
2489 * order to prevent the caller from issuing a read
2490 * ( due to B_CACHE not being set ) and overwriting
2491 * it.
2492 *
2493 * Most callers, including NFS and FFS, need this to
2494 * operate properly either because they assume they
2495 * can issue a read if B_CACHE is not set, or because
2496 * ( for example ) an uncached B_DELWRI might loop due
2497 * to softupdates re-dirtying the buffer. In the latter
2498 * case, B_CACHE is set after the first write completes,
2499 * preventing further loops.
2500 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2501 * above while extending the buffer, we cannot allow the
2502 * buffer to remain with B_CACHE set after the write
2503 * completes or it will represent a corrupt state. To
2504 * deal with this we set B_NOCACHE to scrap the buffer
2505 * after the write.
2506 *
2507 * We might be able to do something fancy, like setting
2508 * B_CACHE in bwrite() except if B_DELWRI is already set,
2509 * so the below call doesn't set B_CACHE, but that gets real
2510 * confusing. This is much easier.
2511 */
2512
2513 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2514 bp->b_flags |= B_NOCACHE;
2515 BUF_WRITE(bp);
2516 goto loop;
2517 }
2518
2519 splx(s);
2520 bp->b_flags &= ~B_DONE;
2521 } else {
2522 int bsize, maxsize, vmio;
2523 off_t offset;
2524
2525 /*
2526 * Buffer is not in-core, create new buffer. The buffer
2527 * returned by getnewbuf() is locked. Note that the returned
2528 * buffer is also considered valid (not marked B_INVAL).
2529 */
2530 VI_UNLOCK(vp);
2531 if (vn_isdisk(vp, NULL))
2532 bsize = DEV_BSIZE;
2533 else if (vp->v_mountedhere)
2534 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2535 else if (vp->v_mount)
2536 bsize = vp->v_mount->mnt_stat.f_iosize;
2537 else
2538 bsize = size;
2539
2540 offset = blkno * bsize;
2541 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) &&
2542 (vp->v_vflag & VV_OBJBUF);
2543 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2544 maxsize = imax(maxsize, bsize);
2545
2546 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2547 if (slpflag || slptimeo) {
2548 splx(s);
2549 return NULL;
2550 }
2551 goto loop;
2552 }
2553
2554 /*
2555 * This code is used to make sure that a buffer is not
2556 * created while the getnewbuf routine is blocked.
2557 * This can be a problem whether the vnode is locked or not.
2558 * If the buffer is created out from under us, we have to
2559 * throw away the one we just created. There is now window
2560 * race because we are safely running at splbio() from the
2561 * point of the duplicate buffer creation through to here,
2562 * and we've locked the buffer.
2563 *
2564 * Note: this must occur before we associate the buffer
2565 * with the vp especially considering limitations in
2566 * the splay tree implementation when dealing with duplicate
2567 * lblkno's.
2568 */
2569 VI_LOCK(vp);
2570 if (gbincore(vp, blkno)) {
2571 VI_UNLOCK(vp);
2572 bp->b_flags |= B_INVAL;
2573 brelse(bp);
2574 goto loop;
2575 }
2576
2577 /*
2578 * Insert the buffer into the hash, so that it can
2579 * be found by incore.
2580 */
2581 bp->b_blkno = bp->b_lblkno = blkno;
2582 bp->b_offset = offset;
2583
2584 bgetvp(vp, bp);
2585 VI_UNLOCK(vp);
2586
2587 /*
2588 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2589 * buffer size starts out as 0, B_CACHE will be set by
2590 * allocbuf() for the VMIO case prior to it testing the
2591 * backing store for validity.
2592 */
2593
2594 if (vmio) {
2595 bp->b_flags |= B_VMIO;
2596 #if defined(VFS_BIO_DEBUG)
2597 if (vp->v_type != VREG)
2598 printf("getblk: vmioing file type %d???\n", vp->v_type);
2599 #endif
2600 VOP_GETVOBJECT(vp, &bp->b_object);
2601 } else {
2602 bp->b_flags &= ~B_VMIO;
2603 bp->b_object = NULL;
2604 }
2605
2606 allocbuf(bp, size);
2607
2608 splx(s);
2609 bp->b_flags &= ~B_DONE;
2610 }
2611 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2612 return (bp);
2613 }
2614
2615 /*
2616 * Get an empty, disassociated buffer of given size. The buffer is initially
2617 * set to B_INVAL.
2618 */
2619 struct buf *
2620 geteblk(int size)
2621 {
2622 struct buf *bp;
2623 int s;
2624 int maxsize;
2625
2626 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2627
2628 s = splbio();
2629 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2630 continue;
2631 splx(s);
2632 allocbuf(bp, size);
2633 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2634 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2635 return (bp);
2636 }
2637
2638
2639 /*
2640 * This code constitutes the buffer memory from either anonymous system
2641 * memory (in the case of non-VMIO operations) or from an associated
2642 * VM object (in the case of VMIO operations). This code is able to
2643 * resize a buffer up or down.
2644 *
2645 * Note that this code is tricky, and has many complications to resolve
2646 * deadlock or inconsistant data situations. Tread lightly!!!
2647 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2648 * the caller. Calling this code willy nilly can result in the loss of data.
2649 *
2650 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2651 * B_CACHE for the non-VMIO case.
2652 */
2653
2654 int
2655 allocbuf(struct buf *bp, int size)
2656 {
2657 int newbsize, mbsize;
2658 int i;
2659
2660 GIANT_REQUIRED;
2661
2662 if (BUF_REFCNT(bp) == 0)
2663 panic("allocbuf: buffer not busy");
2664
2665 if (bp->b_kvasize < size)
2666 panic("allocbuf: buffer too small");
2667
2668 if ((bp->b_flags & B_VMIO) == 0) {
2669 caddr_t origbuf;
2670 int origbufsize;
2671 /*
2672 * Just get anonymous memory from the kernel. Don't
2673 * mess with B_CACHE.
2674 */
2675 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2676 if (bp->b_flags & B_MALLOC)
2677 newbsize = mbsize;
2678 else
2679 newbsize = round_page(size);
2680
2681 if (newbsize < bp->b_bufsize) {
2682 /*
2683 * malloced buffers are not shrunk
2684 */
2685 if (bp->b_flags & B_MALLOC) {
2686 if (newbsize) {
2687 bp->b_bcount = size;
2688 } else {
2689 free(bp->b_data, M_BIOBUF);
2690 if (bp->b_bufsize) {
2691 atomic_subtract_int(
2692 &bufmallocspace,
2693 bp->b_bufsize);
2694 bufspacewakeup();
2695 bp->b_bufsize = 0;
2696 }
2697 bp->b_data = bp->b_kvabase;
2698 bp->b_bcount = 0;
2699 bp->b_flags &= ~B_MALLOC;
2700 }
2701 return 1;
2702 }
2703 vm_hold_free_pages(
2704 bp,
2705 (vm_offset_t) bp->b_data + newbsize,
2706 (vm_offset_t) bp->b_data + bp->b_bufsize);
2707 } else if (newbsize > bp->b_bufsize) {
2708 /*
2709 * We only use malloced memory on the first allocation.
2710 * and revert to page-allocated memory when the buffer
2711 * grows.
2712 */
2713 /*
2714 * There is a potential smp race here that could lead
2715 * to bufmallocspace slightly passing the max. It
2716 * is probably extremely rare and not worth worrying
2717 * over.
2718 */
2719 if ( (bufmallocspace < maxbufmallocspace) &&
2720 (bp->b_bufsize == 0) &&
2721 (mbsize <= PAGE_SIZE/2)) {
2722
2723 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2724 bp->b_bufsize = mbsize;
2725 bp->b_bcount = size;
2726 bp->b_flags |= B_MALLOC;
2727 atomic_add_int(&bufmallocspace, mbsize);
2728 return 1;
2729 }
2730 origbuf = NULL;
2731 origbufsize = 0;
2732 /*
2733 * If the buffer is growing on its other-than-first allocation,
2734 * then we revert to the page-allocation scheme.
2735 */
2736 if (bp->b_flags & B_MALLOC) {
2737 origbuf = bp->b_data;
2738 origbufsize = bp->b_bufsize;
2739 bp->b_data = bp->b_kvabase;
2740 if (bp->b_bufsize) {
2741 atomic_subtract_int(&bufmallocspace,
2742 bp->b_bufsize);
2743 bufspacewakeup();
2744 bp->b_bufsize = 0;
2745 }
2746 bp->b_flags &= ~B_MALLOC;
2747 newbsize = round_page(newbsize);
2748 }
2749 vm_hold_load_pages(
2750 bp,
2751 (vm_offset_t) bp->b_data + bp->b_bufsize,
2752 (vm_offset_t) bp->b_data + newbsize);
2753 if (origbuf) {
2754 bcopy(origbuf, bp->b_data, origbufsize);
2755 free(origbuf, M_BIOBUF);
2756 }
2757 }
2758 } else {
2759 int desiredpages;
2760
2761 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2762 desiredpages = (size == 0) ? 0 :
2763 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2764
2765 if (bp->b_flags & B_MALLOC)
2766 panic("allocbuf: VMIO buffer can't be malloced");
2767 /*
2768 * Set B_CACHE initially if buffer is 0 length or will become
2769 * 0-length.
2770 */
2771 if (size == 0 || bp->b_bufsize == 0)
2772 bp->b_flags |= B_CACHE;
2773
2774 if (newbsize < bp->b_bufsize) {
2775 /*
2776 * DEV_BSIZE aligned new buffer size is less then the
2777 * DEV_BSIZE aligned existing buffer size. Figure out
2778 * if we have to remove any pages.
2779 */
2780 if (desiredpages < bp->b_npages) {
2781 vm_page_t m;
2782
2783 vm_page_lock_queues();
2784 for (i = desiredpages; i < bp->b_npages; i++) {
2785 /*
2786 * the page is not freed here -- it
2787 * is the responsibility of
2788 * vnode_pager_setsize
2789 */
2790 m = bp->b_pages[i];
2791 KASSERT(m != bogus_page,
2792 ("allocbuf: bogus page found"));
2793 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2794 vm_page_lock_queues();
2795
2796 bp->b_pages[i] = NULL;
2797 vm_page_unwire(m, 0);
2798 }
2799 vm_page_unlock_queues();
2800 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2801 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2802 bp->b_npages = desiredpages;
2803 }
2804 } else if (size > bp->b_bcount) {
2805 /*
2806 * We are growing the buffer, possibly in a
2807 * byte-granular fashion.
2808 */
2809 struct vnode *vp;
2810 vm_object_t obj;
2811 vm_offset_t toff;
2812 vm_offset_t tinc;
2813
2814 /*
2815 * Step 1, bring in the VM pages from the object,
2816 * allocating them if necessary. We must clear
2817 * B_CACHE if these pages are not valid for the
2818 * range covered by the buffer.
2819 */
2820
2821 vp = bp->b_vp;
2822 obj = bp->b_object;
2823
2824 VM_OBJECT_LOCK(obj);
2825 while (bp->b_npages < desiredpages) {
2826 vm_page_t m;
2827 vm_pindex_t pi;
2828
2829 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2830 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2831 /*
2832 * note: must allocate system pages
2833 * since blocking here could intefere
2834 * with paging I/O, no matter which
2835 * process we are.
2836 */
2837 m = vm_page_alloc(obj, pi,
2838 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
2839 if (m == NULL) {
2840 atomic_add_int(&vm_pageout_deficit,
2841 desiredpages - bp->b_npages);
2842 VM_OBJECT_UNLOCK(obj);
2843 VM_WAIT;
2844 VM_OBJECT_LOCK(obj);
2845 } else {
2846 vm_page_lock_queues();
2847 vm_page_wakeup(m);
2848 vm_page_unlock_queues();
2849 bp->b_flags &= ~B_CACHE;
2850 bp->b_pages[bp->b_npages] = m;
2851 ++bp->b_npages;
2852 }
2853 continue;
2854 }
2855
2856 /*
2857 * We found a page. If we have to sleep on it,
2858 * retry because it might have gotten freed out
2859 * from under us.
2860 *
2861 * We can only test PG_BUSY here. Blocking on
2862 * m->busy might lead to a deadlock:
2863 *
2864 * vm_fault->getpages->cluster_read->allocbuf
2865 *
2866 */
2867 vm_page_lock_queues();
2868 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2869 continue;
2870
2871 /*
2872 * We have a good page. Should we wakeup the
2873 * page daemon?
2874 */
2875 if ((curproc != pageproc) &&
2876 ((m->queue - m->pc) == PQ_CACHE) &&
2877 ((cnt.v_free_count + cnt.v_cache_count) <
2878 (cnt.v_free_min + cnt.v_cache_min))) {
2879 pagedaemon_wakeup();
2880 }
2881 vm_page_flag_clear(m, PG_ZERO);
2882 vm_page_wire(m);
2883 vm_page_unlock_queues();
2884 bp->b_pages[bp->b_npages] = m;
2885 ++bp->b_npages;
2886 }
2887 VM_OBJECT_UNLOCK(obj);
2888
2889 /*
2890 * Step 2. We've loaded the pages into the buffer,
2891 * we have to figure out if we can still have B_CACHE
2892 * set. Note that B_CACHE is set according to the
2893 * byte-granular range ( bcount and size ), new the
2894 * aligned range ( newbsize ).
2895 *
2896 * The VM test is against m->valid, which is DEV_BSIZE
2897 * aligned. Needless to say, the validity of the data
2898 * needs to also be DEV_BSIZE aligned. Note that this
2899 * fails with NFS if the server or some other client
2900 * extends the file's EOF. If our buffer is resized,
2901 * B_CACHE may remain set! XXX
2902 */
2903
2904 toff = bp->b_bcount;
2905 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2906
2907 while ((bp->b_flags & B_CACHE) && toff < size) {
2908 vm_pindex_t pi;
2909
2910 if (tinc > (size - toff))
2911 tinc = size - toff;
2912
2913 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2914 PAGE_SHIFT;
2915
2916 vfs_buf_test_cache(
2917 bp,
2918 bp->b_offset,
2919 toff,
2920 tinc,
2921 bp->b_pages[pi]
2922 );
2923 toff += tinc;
2924 tinc = PAGE_SIZE;
2925 }
2926
2927 /*
2928 * Step 3, fixup the KVM pmap. Remember that
2929 * bp->b_data is relative to bp->b_offset, but
2930 * bp->b_offset may be offset into the first page.
2931 */
2932
2933 bp->b_data = (caddr_t)
2934 trunc_page((vm_offset_t)bp->b_data);
2935 pmap_qenter(
2936 (vm_offset_t)bp->b_data,
2937 bp->b_pages,
2938 bp->b_npages
2939 );
2940
2941 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2942 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2943 }
2944 }
2945 if (newbsize < bp->b_bufsize)
2946 bufspacewakeup();
2947 bp->b_bufsize = newbsize; /* actual buffer allocation */
2948 bp->b_bcount = size; /* requested buffer size */
2949 return 1;
2950 }
2951
2952 void
2953 biodone(struct bio *bp)
2954 {
2955 mtx_lock(&bdonelock);
2956 bp->bio_flags |= BIO_DONE;
2957 if (bp->bio_done == NULL)
2958 wakeup(bp);
2959 mtx_unlock(&bdonelock);
2960 if (bp->bio_done != NULL)
2961 bp->bio_done(bp);
2962 }
2963
2964 /*
2965 * Wait for a BIO to finish.
2966 *
2967 * XXX: resort to a timeout for now. The optimal locking (if any) for this
2968 * case is not yet clear.
2969 */
2970 int
2971 biowait(struct bio *bp, const char *wchan)
2972 {
2973
2974 mtx_lock(&bdonelock);
2975 while ((bp->bio_flags & BIO_DONE) == 0)
2976 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
2977 mtx_unlock(&bdonelock);
2978 if (bp->bio_error != 0)
2979 return (bp->bio_error);
2980 if (!(bp->bio_flags & BIO_ERROR))
2981 return (0);
2982 return (EIO);
2983 }
2984
2985 void
2986 biofinish(struct bio *bp, struct devstat *stat, int error)
2987 {
2988
2989 if (error) {
2990 bp->bio_error = error;
2991 bp->bio_flags |= BIO_ERROR;
2992 }
2993 if (stat != NULL)
2994 devstat_end_transaction_bio(stat, bp);
2995 biodone(bp);
2996 }
2997
2998 /*
2999 * bufwait:
3000 *
3001 * Wait for buffer I/O completion, returning error status. The buffer
3002 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3003 * error and cleared.
3004 */
3005 int
3006 bufwait(register struct buf * bp)
3007 {
3008 int s;
3009
3010 s = splbio();
3011 if (bp->b_iocmd == BIO_READ)
3012 bwait(bp, PRIBIO, "biord");
3013 else
3014 bwait(bp, PRIBIO, "biowr");
3015 splx(s);
3016 if (bp->b_flags & B_EINTR) {
3017 bp->b_flags &= ~B_EINTR;
3018 return (EINTR);
3019 }
3020 if (bp->b_ioflags & BIO_ERROR) {
3021 return (bp->b_error ? bp->b_error : EIO);
3022 } else {
3023 return (0);
3024 }
3025 }
3026
3027 /*
3028 * Call back function from struct bio back up to struct buf.
3029 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY().
3030 */
3031 void
3032 bufdonebio(struct bio *bp)
3033 {
3034 bufdone(bp->bio_caller2);
3035 }
3036
3037 /*
3038 * bufdone:
3039 *
3040 * Finish I/O on a buffer, optionally calling a completion function.
3041 * This is usually called from an interrupt so process blocking is
3042 * not allowed.
3043 *
3044 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3045 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3046 * assuming B_INVAL is clear.
3047 *
3048 * For the VMIO case, we set B_CACHE if the op was a read and no
3049 * read error occured, or if the op was a write. B_CACHE is never
3050 * set if the buffer is invalid or otherwise uncacheable.
3051 *
3052 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3053 * initiator to leave B_INVAL set to brelse the buffer out of existance
3054 * in the biodone routine.
3055 */
3056 void
3057 bufdone(struct buf *bp)
3058 {
3059 int s;
3060 void (*biodone)(struct buf *);
3061
3062 GIANT_REQUIRED;
3063
3064 s = splbio();
3065
3066 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
3067 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3068
3069 bp->b_flags |= B_DONE;
3070 runningbufwakeup(bp);
3071
3072 if (bp->b_iocmd == BIO_DELETE) {
3073 brelse(bp);
3074 splx(s);
3075 return;
3076 }
3077
3078 if (bp->b_iocmd == BIO_WRITE) {
3079 vwakeup(bp);
3080 }
3081
3082 /* call optional completion function if requested */
3083 if (bp->b_iodone != NULL) {
3084 biodone = bp->b_iodone;
3085 bp->b_iodone = NULL;
3086 (*biodone) (bp);
3087 splx(s);
3088 return;
3089 }
3090 if (LIST_FIRST(&bp->b_dep) != NULL)
3091 buf_complete(bp);
3092
3093 if (bp->b_flags & B_VMIO) {
3094 int i;
3095 vm_ooffset_t foff;
3096 vm_page_t m;
3097 vm_object_t obj;
3098 int iosize;
3099 struct vnode *vp = bp->b_vp;
3100
3101 obj = bp->b_object;
3102
3103 #if defined(VFS_BIO_DEBUG)
3104 mp_fixme("usecount and vflag accessed without locks.");
3105 if (vp->v_usecount == 0) {
3106 panic("biodone: zero vnode ref count");
3107 }
3108
3109 if ((vp->v_vflag & VV_OBJBUF) == 0) {
3110 panic("biodone: vnode is not setup for merged cache");
3111 }
3112 #endif
3113
3114 foff = bp->b_offset;
3115 KASSERT(bp->b_offset != NOOFFSET,
3116 ("biodone: no buffer offset"));
3117
3118 if (obj != NULL)
3119 VM_OBJECT_LOCK(obj);
3120 #if defined(VFS_BIO_DEBUG)
3121 if (obj->paging_in_progress < bp->b_npages) {
3122 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3123 obj->paging_in_progress, bp->b_npages);
3124 }
3125 #endif
3126
3127 /*
3128 * Set B_CACHE if the op was a normal read and no error
3129 * occured. B_CACHE is set for writes in the b*write()
3130 * routines.
3131 */
3132 iosize = bp->b_bcount - bp->b_resid;
3133 if (bp->b_iocmd == BIO_READ &&
3134 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3135 !(bp->b_ioflags & BIO_ERROR)) {
3136 bp->b_flags |= B_CACHE;
3137 }
3138 vm_page_lock_queues();
3139 for (i = 0; i < bp->b_npages; i++) {
3140 int bogusflag = 0;
3141 int resid;
3142
3143 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3144 if (resid > iosize)
3145 resid = iosize;
3146
3147 /*
3148 * cleanup bogus pages, restoring the originals
3149 */
3150 m = bp->b_pages[i];
3151 if (m == bogus_page) {
3152 bogusflag = 1;
3153 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3154 if (m == NULL)
3155 panic("biodone: page disappeared!");
3156 bp->b_pages[i] = m;
3157 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3158 }
3159 #if defined(VFS_BIO_DEBUG)
3160 if (OFF_TO_IDX(foff) != m->pindex) {
3161 printf(
3162 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3163 (intmax_t)foff, (uintmax_t)m->pindex);
3164 }
3165 #endif
3166
3167 /*
3168 * In the write case, the valid and clean bits are
3169 * already changed correctly ( see bdwrite() ), so we
3170 * only need to do this here in the read case.
3171 */
3172 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3173 vfs_page_set_valid(bp, foff, i, m);
3174 }
3175 vm_page_flag_clear(m, PG_ZERO);
3176
3177 /*
3178 * when debugging new filesystems or buffer I/O methods, this
3179 * is the most common error that pops up. if you see this, you
3180 * have not set the page busy flag correctly!!!
3181 */
3182 if (m->busy == 0) {
3183 printf("biodone: page busy < 0, "
3184 "pindex: %d, foff: 0x(%x,%x), "
3185 "resid: %d, index: %d\n",
3186 (int) m->pindex, (int)(foff >> 32),
3187 (int) foff & 0xffffffff, resid, i);
3188 if (!vn_isdisk(vp, NULL))
3189 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n",
3190 bp->b_vp->v_mount->mnt_stat.f_iosize,
3191 (intmax_t) bp->b_lblkno,
3192 bp->b_flags, bp->b_npages);
3193 else
3194 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3195 (intmax_t) bp->b_lblkno,
3196 bp->b_flags, bp->b_npages);
3197 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3198 m->valid, m->dirty, m->wire_count);
3199 panic("biodone: page busy < 0\n");
3200 }
3201 vm_page_io_finish(m);
3202 vm_object_pip_subtract(obj, 1);
3203 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3204 iosize -= resid;
3205 }
3206 vm_page_unlock_queues();
3207 if (obj != NULL) {
3208 vm_object_pip_wakeupn(obj, 0);
3209 VM_OBJECT_UNLOCK(obj);
3210 }
3211 }
3212
3213 /*
3214 * For asynchronous completions, release the buffer now. The brelse
3215 * will do a wakeup there if necessary - so no need to do a wakeup
3216 * here in the async case. The sync case always needs to do a wakeup.
3217 */
3218
3219 if (bp->b_flags & B_ASYNC) {
3220 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3221 brelse(bp);
3222 else
3223 bqrelse(bp);
3224 } else {
3225 bdone(bp);
3226 }
3227 splx(s);
3228 }
3229
3230 /*
3231 * This routine is called in lieu of iodone in the case of
3232 * incomplete I/O. This keeps the busy status for pages
3233 * consistant.
3234 */
3235 void
3236 vfs_unbusy_pages(struct buf * bp)
3237 {
3238 int i;
3239
3240 GIANT_REQUIRED;
3241
3242 runningbufwakeup(bp);
3243 if (bp->b_flags & B_VMIO) {
3244 vm_object_t obj;
3245
3246 obj = bp->b_object;
3247 VM_OBJECT_LOCK(obj);
3248 vm_page_lock_queues();
3249 for (i = 0; i < bp->b_npages; i++) {
3250 vm_page_t m = bp->b_pages[i];
3251
3252 if (m == bogus_page) {
3253 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3254 if (!m) {
3255 panic("vfs_unbusy_pages: page missing\n");
3256 }
3257 bp->b_pages[i] = m;
3258 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3259 }
3260 vm_object_pip_subtract(obj, 1);
3261 vm_page_flag_clear(m, PG_ZERO);
3262 vm_page_io_finish(m);
3263 }
3264 vm_page_unlock_queues();
3265 vm_object_pip_wakeupn(obj, 0);
3266 VM_OBJECT_UNLOCK(obj);
3267 }
3268 }
3269
3270 /*
3271 * vfs_page_set_valid:
3272 *
3273 * Set the valid bits in a page based on the supplied offset. The
3274 * range is restricted to the buffer's size.
3275 *
3276 * This routine is typically called after a read completes.
3277 */
3278 static void
3279 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3280 {
3281 vm_ooffset_t soff, eoff;
3282
3283 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3284 /*
3285 * Start and end offsets in buffer. eoff - soff may not cross a
3286 * page boundry or cross the end of the buffer. The end of the
3287 * buffer, in this case, is our file EOF, not the allocation size
3288 * of the buffer.
3289 */
3290 soff = off;
3291 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3292 if (eoff > bp->b_offset + bp->b_bcount)
3293 eoff = bp->b_offset + bp->b_bcount;
3294
3295 /*
3296 * Set valid range. This is typically the entire buffer and thus the
3297 * entire page.
3298 */
3299 if (eoff > soff) {
3300 vm_page_set_validclean(
3301 m,
3302 (vm_offset_t) (soff & PAGE_MASK),
3303 (vm_offset_t) (eoff - soff)
3304 );
3305 }
3306 }
3307
3308 /*
3309 * This routine is called before a device strategy routine.
3310 * It is used to tell the VM system that paging I/O is in
3311 * progress, and treat the pages associated with the buffer
3312 * almost as being PG_BUSY. Also the object paging_in_progress
3313 * flag is handled to make sure that the object doesn't become
3314 * inconsistant.
3315 *
3316 * Since I/O has not been initiated yet, certain buffer flags
3317 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3318 * and should be ignored.
3319 */
3320 void
3321 vfs_busy_pages(struct buf * bp, int clear_modify)
3322 {
3323 int i, bogus;
3324
3325 if (bp->b_flags & B_VMIO) {
3326 vm_object_t obj;
3327 vm_ooffset_t foff;
3328
3329 obj = bp->b_object;
3330 foff = bp->b_offset;
3331 KASSERT(bp->b_offset != NOOFFSET,
3332 ("vfs_busy_pages: no buffer offset"));
3333 vfs_setdirty(bp);
3334 if (obj != NULL)
3335 VM_OBJECT_LOCK(obj);
3336 retry:
3337 vm_page_lock_queues();
3338 for (i = 0; i < bp->b_npages; i++) {
3339 vm_page_t m = bp->b_pages[i];
3340
3341 if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3342 goto retry;
3343 }
3344 bogus = 0;
3345 for (i = 0; i < bp->b_npages; i++) {
3346 vm_page_t m = bp->b_pages[i];
3347
3348 vm_page_flag_clear(m, PG_ZERO);
3349 if ((bp->b_flags & B_CLUSTER) == 0) {
3350 vm_object_pip_add(obj, 1);
3351 vm_page_io_start(m);
3352 }
3353 /*
3354 * When readying a buffer for a read ( i.e
3355 * clear_modify == 0 ), it is important to do
3356 * bogus_page replacement for valid pages in
3357 * partially instantiated buffers. Partially
3358 * instantiated buffers can, in turn, occur when
3359 * reconstituting a buffer from its VM backing store
3360 * base. We only have to do this if B_CACHE is
3361 * clear ( which causes the I/O to occur in the
3362 * first place ). The replacement prevents the read
3363 * I/O from overwriting potentially dirty VM-backed
3364 * pages. XXX bogus page replacement is, uh, bogus.
3365 * It may not work properly with small-block devices.
3366 * We need to find a better way.
3367 */
3368 pmap_remove_all(m);
3369 if (clear_modify)
3370 vfs_page_set_valid(bp, foff, i, m);
3371 else if (m->valid == VM_PAGE_BITS_ALL &&
3372 (bp->b_flags & B_CACHE) == 0) {
3373 bp->b_pages[i] = bogus_page;
3374 bogus++;
3375 }
3376 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3377 }
3378 vm_page_unlock_queues();
3379 if (obj != NULL)
3380 VM_OBJECT_UNLOCK(obj);
3381 if (bogus)
3382 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3383 }
3384 }
3385
3386 /*
3387 * Tell the VM system that the pages associated with this buffer
3388 * are clean. This is used for delayed writes where the data is
3389 * going to go to disk eventually without additional VM intevention.
3390 *
3391 * Note that while we only really need to clean through to b_bcount, we
3392 * just go ahead and clean through to b_bufsize.
3393 */
3394 static void
3395 vfs_clean_pages(struct buf * bp)
3396 {
3397 int i;
3398
3399 GIANT_REQUIRED;
3400
3401 if (bp->b_flags & B_VMIO) {
3402 vm_ooffset_t foff;
3403
3404 foff = bp->b_offset;
3405 KASSERT(bp->b_offset != NOOFFSET,
3406 ("vfs_clean_pages: no buffer offset"));
3407 vm_page_lock_queues();
3408 for (i = 0; i < bp->b_npages; i++) {
3409 vm_page_t m = bp->b_pages[i];
3410 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3411 vm_ooffset_t eoff = noff;
3412
3413 if (eoff > bp->b_offset + bp->b_bufsize)
3414 eoff = bp->b_offset + bp->b_bufsize;
3415 vfs_page_set_valid(bp, foff, i, m);
3416 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3417 foff = noff;
3418 }
3419 vm_page_unlock_queues();
3420 }
3421 }
3422
3423 /*
3424 * vfs_bio_set_validclean:
3425 *
3426 * Set the range within the buffer to valid and clean. The range is
3427 * relative to the beginning of the buffer, b_offset. Note that b_offset
3428 * itself may be offset from the beginning of the first page.
3429 *
3430 */
3431
3432 void
3433 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3434 {
3435 if (bp->b_flags & B_VMIO) {
3436 int i;
3437 int n;
3438
3439 /*
3440 * Fixup base to be relative to beginning of first page.
3441 * Set initial n to be the maximum number of bytes in the
3442 * first page that can be validated.
3443 */
3444
3445 base += (bp->b_offset & PAGE_MASK);
3446 n = PAGE_SIZE - (base & PAGE_MASK);
3447
3448 vm_page_lock_queues();
3449 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3450 vm_page_t m = bp->b_pages[i];
3451
3452 if (n > size)
3453 n = size;
3454
3455 vm_page_set_validclean(m, base & PAGE_MASK, n);
3456 base += n;
3457 size -= n;
3458 n = PAGE_SIZE;
3459 }
3460 vm_page_unlock_queues();
3461 }
3462 }
3463
3464 /*
3465 * vfs_bio_clrbuf:
3466 *
3467 * clear a buffer. This routine essentially fakes an I/O, so we need
3468 * to clear BIO_ERROR and B_INVAL.
3469 *
3470 * Note that while we only theoretically need to clear through b_bcount,
3471 * we go ahead and clear through b_bufsize.
3472 */
3473
3474 void
3475 vfs_bio_clrbuf(struct buf *bp)
3476 {
3477 int i, mask = 0;
3478 caddr_t sa, ea;
3479
3480 GIANT_REQUIRED;
3481
3482 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3483 bp->b_flags &= ~B_INVAL;
3484 bp->b_ioflags &= ~BIO_ERROR;
3485 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3486 (bp->b_offset & PAGE_MASK) == 0) {
3487 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3488 if ((bp->b_pages[0]->valid & mask) == mask) {
3489 bp->b_resid = 0;
3490 return;
3491 }
3492 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3493 ((bp->b_pages[0]->valid & mask) == 0)) {
3494 bzero(bp->b_data, bp->b_bufsize);
3495 bp->b_pages[0]->valid |= mask;
3496 bp->b_resid = 0;
3497 return;
3498 }
3499 }
3500 ea = sa = bp->b_data;
3501 for(i=0;i<bp->b_npages;i++,sa=ea) {
3502 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3503 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3504 ea = (caddr_t)(vm_offset_t)ulmin(
3505 (u_long)(vm_offset_t)ea,
3506 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3507 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3508 if ((bp->b_pages[i]->valid & mask) == mask)
3509 continue;
3510 if ((bp->b_pages[i]->valid & mask) == 0) {
3511 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3512 bzero(sa, ea - sa);
3513 }
3514 } else {
3515 for (; sa < ea; sa += DEV_BSIZE, j++) {
3516 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3517 (bp->b_pages[i]->valid & (1<<j)) == 0)
3518 bzero(sa, DEV_BSIZE);
3519 }
3520 }
3521 bp->b_pages[i]->valid |= mask;
3522 vm_page_lock_queues();
3523 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
3524 vm_page_unlock_queues();
3525 }
3526 bp->b_resid = 0;
3527 } else {
3528 clrbuf(bp);
3529 }
3530 }
3531
3532 /*
3533 * vm_hold_load_pages and vm_hold_free_pages get pages into
3534 * a buffers address space. The pages are anonymous and are
3535 * not associated with a file object.
3536 */
3537 static void
3538 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3539 {
3540 vm_offset_t pg;
3541 vm_page_t p;
3542 int index;
3543
3544 GIANT_REQUIRED;
3545
3546 to = round_page(to);
3547 from = round_page(from);
3548 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3549
3550 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3551 tryagain:
3552 /*
3553 * note: must allocate system pages since blocking here
3554 * could intefere with paging I/O, no matter which
3555 * process we are.
3556 */
3557 VM_OBJECT_LOCK(kernel_object);
3558 p = vm_page_alloc(kernel_object,
3559 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3560 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3561 VM_OBJECT_UNLOCK(kernel_object);
3562 if (!p) {
3563 atomic_add_int(&vm_pageout_deficit,
3564 (to - pg) >> PAGE_SHIFT);
3565 VM_WAIT;
3566 goto tryagain;
3567 }
3568 vm_page_lock_queues();
3569 p->valid = VM_PAGE_BITS_ALL;
3570 vm_page_unlock_queues();
3571 pmap_qenter(pg, &p, 1);
3572 bp->b_pages[index] = p;
3573 vm_page_lock_queues();
3574 vm_page_wakeup(p);
3575 vm_page_unlock_queues();
3576 }
3577 bp->b_npages = index;
3578 }
3579
3580 /* Return pages associated with this buf to the vm system */
3581 static void
3582 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3583 {
3584 vm_offset_t pg;
3585 vm_page_t p;
3586 int index, newnpages;
3587
3588 GIANT_REQUIRED;
3589
3590 from = round_page(from);
3591 to = round_page(to);
3592 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3593
3594 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3595 p = bp->b_pages[index];
3596 if (p && (index < bp->b_npages)) {
3597 if (p->busy) {
3598 printf(
3599 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3600 (intmax_t)bp->b_blkno,
3601 (intmax_t)bp->b_lblkno);
3602 }
3603 bp->b_pages[index] = NULL;
3604 pmap_qremove(pg, 1);
3605 vm_page_lock_queues();
3606 vm_page_busy(p);
3607 vm_page_unwire(p, 0);
3608 vm_page_free(p);
3609 vm_page_unlock_queues();
3610 }
3611 }
3612 bp->b_npages = newnpages;
3613 }
3614
3615 /*
3616 * Map an IO request into kernel virtual address space.
3617 *
3618 * All requests are (re)mapped into kernel VA space.
3619 * Notice that we use b_bufsize for the size of the buffer
3620 * to be mapped. b_bcount might be modified by the driver.
3621 *
3622 * Note that even if the caller determines that the address space should
3623 * be valid, a race or a smaller-file mapped into a larger space may
3624 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3625 * check the return value.
3626 */
3627 int
3628 vmapbuf(struct buf *bp)
3629 {
3630 caddr_t addr, kva;
3631 vm_paddr_t pa;
3632 int pidx, i;
3633 struct vm_page *m;
3634 struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3635
3636 GIANT_REQUIRED;
3637
3638 if ((bp->b_flags & B_PHYS) == 0)
3639 panic("vmapbuf");
3640 if (bp->b_bufsize < 0)
3641 return (-1);
3642 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3643 addr < bp->b_data + bp->b_bufsize;
3644 addr += PAGE_SIZE, pidx++) {
3645 /*
3646 * Do the vm_fault if needed; do the copy-on-write thing
3647 * when reading stuff off device into memory.
3648 *
3649 * NOTE! Must use pmap_extract() because addr may be in
3650 * the userland address space, and kextract is only guarenteed
3651 * to work for the kernland address space (see: sparc64 port).
3652 */
3653 retry:
3654 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data,
3655 (bp->b_iocmd == BIO_READ) ?
3656 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ);
3657 if (i < 0) {
3658 vm_page_lock_queues();
3659 for (i = 0; i < pidx; ++i) {
3660 vm_page_unhold(bp->b_pages[i]);
3661 bp->b_pages[i] = NULL;
3662 }
3663 vm_page_unlock_queues();
3664 return(-1);
3665 }
3666 pa = pmap_extract(pmap, (vm_offset_t)addr);
3667 if (pa == 0) {
3668 printf("vmapbuf: warning, race against user address during I/O");
3669 goto retry;
3670 }
3671 m = PHYS_TO_VM_PAGE(pa);
3672 vm_page_lock_queues();
3673 vm_page_hold(m);
3674 vm_page_unlock_queues();
3675 bp->b_pages[pidx] = m;
3676 }
3677 if (pidx > btoc(MAXPHYS))
3678 panic("vmapbuf: mapped more than MAXPHYS");
3679 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3680
3681 kva = bp->b_saveaddr;
3682 bp->b_npages = pidx;
3683 bp->b_saveaddr = bp->b_data;
3684 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3685 return(0);
3686 }
3687
3688 /*
3689 * Free the io map PTEs associated with this IO operation.
3690 * We also invalidate the TLB entries and restore the original b_addr.
3691 */
3692 void
3693 vunmapbuf(struct buf *bp)
3694 {
3695 int pidx;
3696 int npages;
3697
3698 GIANT_REQUIRED;
3699
3700 if ((bp->b_flags & B_PHYS) == 0)
3701 panic("vunmapbuf");
3702
3703 npages = bp->b_npages;
3704 pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3705 npages);
3706 vm_page_lock_queues();
3707 for (pidx = 0; pidx < npages; pidx++)
3708 vm_page_unhold(bp->b_pages[pidx]);
3709 vm_page_unlock_queues();
3710
3711 bp->b_data = bp->b_saveaddr;
3712 }
3713
3714 void
3715 bdone(struct buf *bp)
3716 {
3717 mtx_lock(&bdonelock);
3718 bp->b_flags |= B_DONE;
3719 wakeup(bp);
3720 mtx_unlock(&bdonelock);
3721 }
3722
3723 void
3724 bwait(struct buf *bp, u_char pri, const char *wchan)
3725 {
3726 mtx_lock(&bdonelock);
3727 while ((bp->b_flags & B_DONE) == 0)
3728 msleep(bp, &bdonelock, pri, wchan, 0);
3729 mtx_unlock(&bdonelock);
3730 }
3731
3732 #include "opt_ddb.h"
3733 #ifdef DDB
3734 #include <ddb/ddb.h>
3735
3736 /* DDB command to show buffer data */
3737 DB_SHOW_COMMAND(buffer, db_show_buffer)
3738 {
3739 /* get args */
3740 struct buf *bp = (struct buf *)addr;
3741
3742 if (!have_addr) {
3743 db_printf("usage: show buffer <addr>\n");
3744 return;
3745 }
3746
3747 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3748 db_printf(
3749 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3750 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n",
3751 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3752 major(bp->b_dev), minor(bp->b_dev), bp->b_data,
3753 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno);
3754 if (bp->b_npages) {
3755 int i;
3756 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3757 for (i = 0; i < bp->b_npages; i++) {
3758 vm_page_t m;
3759 m = bp->b_pages[i];
3760 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3761 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3762 if ((i + 1) < bp->b_npages)
3763 db_printf(",");
3764 }
3765 db_printf("\n");
3766 }
3767 }
3768 #endif /* DDB */
Cache object: 47af01950a12bc5acf16b2d6257cdcc6
|