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
15/*
16 * this file contains a new buffer I/O scheme implementing a coherent
17 * VM object and buffer cache scheme. Pains have been taken to make
18 * sure that the performance degradation associated with schemes such
19 * as this is not realized.
20 *
21 * Author: John S. Dyson
22 * Significant help during the development and debugging phases
23 * had been provided by David Greenman, also of the FreeBSD core team.
24 *
25 * see man buf(9) for more info.
26 */
27
28#include <sys/cdefs.h>
| 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
15/*
16 * this file contains a new buffer I/O scheme implementing a coherent
17 * VM object and buffer cache scheme. Pains have been taken to make
18 * sure that the performance degradation associated with schemes such
19 * as this is not realized.
20 *
21 * Author: John S. Dyson
22 * Significant help during the development and debugging phases
23 * had been provided by David Greenman, also of the FreeBSD core team.
24 *
25 * see man buf(9) for more info.
26 */
27
28#include <sys/cdefs.h>
|
29__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 130585 2004-06-16 09:47:26Z phk $");
| 29__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 130640 2004-06-17 17:16:53Z phk $");
|
30
31#include <sys/param.h>
32#include <sys/systm.h>
33#include <sys/bio.h>
34#include <sys/conf.h>
35#include <sys/buf.h>
36#include <sys/devicestat.h>
37#include <sys/eventhandler.h>
38#include <sys/lock.h>
39#include <sys/malloc.h>
40#include <sys/mount.h>
41#include <sys/mutex.h>
42#include <sys/kernel.h>
43#include <sys/kthread.h>
44#include <sys/proc.h>
45#include <sys/resourcevar.h>
46#include <sys/sysctl.h>
47#include <sys/vmmeter.h>
48#include <sys/vnode.h>
49#include <vm/vm.h>
50#include <vm/vm_param.h>
51#include <vm/vm_kern.h>
52#include <vm/vm_pageout.h>
53#include <vm/vm_page.h>
54#include <vm/vm_object.h>
55#include <vm/vm_extern.h>
56#include <vm/vm_map.h>
57#include "opt_directio.h"
58#include "opt_swap.h"
59
60static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
61
62struct bio_ops bioops; /* I/O operation notification */
63
64static int ibwrite(struct buf *);
65
66struct buf_ops buf_ops_bio = {
67 "buf_ops_bio",
68 ibwrite
69};
70
71/*
72 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
73 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
74 */
75struct buf *buf; /* buffer header pool */
76
77static struct proc *bufdaemonproc;
78
79static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
80 vm_offset_t to);
81static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
82 vm_offset_t to);
83static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
84 int pageno, vm_page_t m);
85static void vfs_clean_pages(struct buf * bp);
86static void vfs_setdirty(struct buf *bp);
87static void vfs_vmio_release(struct buf *bp);
88static void vfs_backgroundwritedone(struct buf *bp);
89static int vfs_bio_clcheck(struct vnode *vp, int size,
90 daddr_t lblkno, daddr_t blkno);
91static int flushbufqueues(int flushdeps);
92static void buf_daemon(void);
93void bremfreel(struct buf * bp);
94
95int vmiodirenable = TRUE;
96SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
97 "Use the VM system for directory writes");
98int runningbufspace;
99SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
100 "Amount of presently outstanding async buffer io");
101static int bufspace;
102SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
103 "KVA memory used for bufs");
104static int maxbufspace;
105SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
106 "Maximum allowed value of bufspace (including buf_daemon)");
107static int bufmallocspace;
108SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
109 "Amount of malloced memory for buffers");
110static int maxbufmallocspace;
111SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
112 "Maximum amount of malloced memory for buffers");
113static int lobufspace;
114SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
115 "Minimum amount of buffers we want to have");
116static int hibufspace;
117SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
118 "Maximum allowed value of bufspace (excluding buf_daemon)");
119static int bufreusecnt;
120SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
121 "Number of times we have reused a buffer");
122static int buffreekvacnt;
123SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
124 "Number of times we have freed the KVA space from some buffer");
125static int bufdefragcnt;
126SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
127 "Number of times we have had to repeat buffer allocation to defragment");
128static int lorunningspace;
129SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
130 "Minimum preferred space used for in-progress I/O");
131static int hirunningspace;
132SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
133 "Maximum amount of space to use for in-progress I/O");
134static int dirtybufferflushes;
135SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
136 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
137static int altbufferflushes;
138SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
139 0, "Number of fsync flushes to limit dirty buffers");
140static int recursiveflushes;
141SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
142 0, "Number of flushes skipped due to being recursive");
143static int numdirtybuffers;
144SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
145 "Number of buffers that are dirty (has unwritten changes) at the moment");
146static int lodirtybuffers;
147SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
148 "How many buffers we want to have free before bufdaemon can sleep");
149static int hidirtybuffers;
150SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
151 "When the number of dirty buffers is considered severe");
152static int dirtybufthresh;
153SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
154 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
155static int numfreebuffers;
156SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
157 "Number of free buffers");
158static int lofreebuffers;
159SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
160 "XXX Unused");
161static int hifreebuffers;
162SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
163 "XXX Complicatedly unused");
164static int getnewbufcalls;
165SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
166 "Number of calls to getnewbuf");
167static int getnewbufrestarts;
168SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
169 "Number of times getnewbuf has had to restart a buffer aquisition");
170static int dobkgrdwrite = 1;
171SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0,
172 "Do background writes (honoring the BV_BKGRDWRITE flag)?");
173
174/*
175 * Wakeup point for bufdaemon, as well as indicator of whether it is already
176 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
177 * is idling.
178 */
179static int bd_request;
180
181/*
182 * This lock synchronizes access to bd_request.
183 */
184static struct mtx bdlock;
185
186/*
187 * bogus page -- for I/O to/from partially complete buffers
188 * this is a temporary solution to the problem, but it is not
189 * really that bad. it would be better to split the buffer
190 * for input in the case of buffers partially already in memory,
191 * but the code is intricate enough already.
192 */
193vm_page_t bogus_page;
194
195/*
196 * Synchronization (sleep/wakeup) variable for active buffer space requests.
197 * Set when wait starts, cleared prior to wakeup().
198 * Used in runningbufwakeup() and waitrunningbufspace().
199 */
200static int runningbufreq;
201
202/*
203 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
204 * waitrunningbufspace().
205 */
206static struct mtx rbreqlock;
207
208/*
209 * Synchronization (sleep/wakeup) variable for buffer requests.
210 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
211 * by and/or.
212 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
213 * getnewbuf(), and getblk().
214 */
215static int needsbuffer;
216
217/*
218 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
219 */
220static struct mtx nblock;
221
222/*
223 * Lock that protects against bwait()/bdone()/B_DONE races.
224 */
225
226static struct mtx bdonelock;
227
228/*
229 * Definitions for the buffer free lists.
230 */
231#define BUFFER_QUEUES 5 /* number of free buffer queues */
232
233#define QUEUE_NONE 0 /* on no queue */
234#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
235#define QUEUE_DIRTY 2 /* B_DELWRI buffers */
236#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */
237#define QUEUE_EMPTY 4 /* empty buffer headers */
238
239/* Queues for free buffers with various properties */
240static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
241
242/* Lock for the bufqueues */
243static struct mtx bqlock;
244
245/*
246 * Single global constant for BUF_WMESG, to avoid getting multiple references.
247 * buf_wmesg is referred from macros.
248 */
249const char *buf_wmesg = BUF_WMESG;
250
251#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
252#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
253#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
254#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
255
256#ifdef DIRECTIO
257extern void ffs_rawread_setup(void);
258#endif /* DIRECTIO */
259/*
260 * numdirtywakeup:
261 *
262 * If someone is blocked due to there being too many dirty buffers,
263 * and numdirtybuffers is now reasonable, wake them up.
264 */
265
266static __inline void
267numdirtywakeup(int level)
268{
269 if (numdirtybuffers <= level) {
270 mtx_lock(&nblock);
271 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
272 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
273 wakeup(&needsbuffer);
274 }
275 mtx_unlock(&nblock);
276 }
277}
278
279/*
280 * bufspacewakeup:
281 *
282 * Called when buffer space is potentially available for recovery.
283 * getnewbuf() will block on this flag when it is unable to free
284 * sufficient buffer space. Buffer space becomes recoverable when
285 * bp's get placed back in the queues.
286 */
287
288static __inline void
289bufspacewakeup(void)
290{
291 /*
292 * If someone is waiting for BUF space, wake them up. Even
293 * though we haven't freed the kva space yet, the waiting
294 * process will be able to now.
295 */
296 mtx_lock(&nblock);
297 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
298 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
299 wakeup(&needsbuffer);
300 }
301 mtx_unlock(&nblock);
302}
303
304/*
305 * runningbufwakeup() - in-progress I/O accounting.
306 *
307 */
308static __inline void
309runningbufwakeup(struct buf *bp)
310{
311 if (bp->b_runningbufspace) {
312 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
313 bp->b_runningbufspace = 0;
314 mtx_lock(&rbreqlock);
315 if (runningbufreq && runningbufspace <= lorunningspace) {
316 runningbufreq = 0;
317 wakeup(&runningbufreq);
318 }
319 mtx_unlock(&rbreqlock);
320 }
321}
322
323/*
324 * bufcountwakeup:
325 *
326 * Called when a buffer has been added to one of the free queues to
327 * account for the buffer and to wakeup anyone waiting for free buffers.
328 * This typically occurs when large amounts of metadata are being handled
329 * by the buffer cache ( else buffer space runs out first, usually ).
330 */
331
332static __inline void
333bufcountwakeup(void)
334{
335 atomic_add_int(&numfreebuffers, 1);
336 mtx_lock(&nblock);
337 if (needsbuffer) {
338 needsbuffer &= ~VFS_BIO_NEED_ANY;
339 if (numfreebuffers >= hifreebuffers)
340 needsbuffer &= ~VFS_BIO_NEED_FREE;
341 wakeup(&needsbuffer);
342 }
343 mtx_unlock(&nblock);
344}
345
346/*
347 * waitrunningbufspace()
348 *
349 * runningbufspace is a measure of the amount of I/O currently
350 * running. This routine is used in async-write situations to
351 * prevent creating huge backups of pending writes to a device.
352 * Only asynchronous writes are governed by this function.
353 *
354 * Reads will adjust runningbufspace, but will not block based on it.
355 * The read load has a side effect of reducing the allowed write load.
356 *
357 * This does NOT turn an async write into a sync write. It waits
358 * for earlier writes to complete and generally returns before the
359 * caller's write has reached the device.
360 */
361static __inline void
362waitrunningbufspace(void)
363{
364 mtx_lock(&rbreqlock);
365 while (runningbufspace > hirunningspace) {
366 ++runningbufreq;
367 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
368 }
369 mtx_unlock(&rbreqlock);
370}
371
372
373/*
374 * vfs_buf_test_cache:
375 *
376 * Called when a buffer is extended. This function clears the B_CACHE
377 * bit if the newly extended portion of the buffer does not contain
378 * valid data.
379 */
380static __inline__
381void
382vfs_buf_test_cache(struct buf *bp,
383 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
384 vm_page_t m)
385{
386 GIANT_REQUIRED;
387
388 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
389 if (bp->b_flags & B_CACHE) {
390 int base = (foff + off) & PAGE_MASK;
391 if (vm_page_is_valid(m, base, size) == 0)
392 bp->b_flags &= ~B_CACHE;
393 }
394}
395
396/* Wake up the buffer deamon if necessary */
397static __inline__
398void
399bd_wakeup(int dirtybuflevel)
400{
401 mtx_lock(&bdlock);
402 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
403 bd_request = 1;
404 wakeup(&bd_request);
405 }
406 mtx_unlock(&bdlock);
407}
408
409/*
410 * bd_speedup - speedup the buffer cache flushing code
411 */
412
413static __inline__
414void
415bd_speedup(void)
416{
417 bd_wakeup(1);
418}
419
420/*
421 * Calculating buffer cache scaling values and reserve space for buffer
422 * headers. This is called during low level kernel initialization and
423 * may be called more then once. We CANNOT write to the memory area
424 * being reserved at this time.
425 */
426caddr_t
427kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
428{
429 /*
430 * physmem_est is in pages. Convert it to kilobytes (assumes
431 * PAGE_SIZE is >= 1K)
432 */
433 physmem_est = physmem_est * (PAGE_SIZE / 1024);
434
435 /*
436 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
437 * For the first 64MB of ram nominally allocate sufficient buffers to
438 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
439 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing
440 * the buffer cache we limit the eventual kva reservation to
441 * maxbcache bytes.
442 *
443 * factor represents the 1/4 x ram conversion.
444 */
445 if (nbuf == 0) {
446 int factor = 4 * BKVASIZE / 1024;
447
448 nbuf = 50;
449 if (physmem_est > 4096)
450 nbuf += min((physmem_est - 4096) / factor,
451 65536 / factor);
452 if (physmem_est > 65536)
453 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
454
455 if (maxbcache && nbuf > maxbcache / BKVASIZE)
456 nbuf = maxbcache / BKVASIZE;
457 }
458
459#if 0
460 /*
461 * Do not allow the buffer_map to be more then 1/2 the size of the
462 * kernel_map.
463 */
464 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
465 (BKVASIZE * 2)) {
466 nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
467 (BKVASIZE * 2);
468 printf("Warning: nbufs capped at %d\n", nbuf);
469 }
470#endif
471
472 /*
473 * swbufs are used as temporary holders for I/O, such as paging I/O.
474 * We have no less then 16 and no more then 256.
475 */
476 nswbuf = max(min(nbuf/4, 256), 16);
477#ifdef NSWBUF_MIN
478 if (nswbuf < NSWBUF_MIN)
479 nswbuf = NSWBUF_MIN;
480#endif
481#ifdef DIRECTIO
482 ffs_rawread_setup();
483#endif
484
485 /*
486 * Reserve space for the buffer cache buffers
487 */
488 swbuf = (void *)v;
489 v = (caddr_t)(swbuf + nswbuf);
490 buf = (void *)v;
491 v = (caddr_t)(buf + nbuf);
492
493 return(v);
494}
495
496/* Initialize the buffer subsystem. Called before use of any buffers. */
497void
498bufinit(void)
499{
500 struct buf *bp;
501 int i;
502
503 GIANT_REQUIRED;
504
505 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
506 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
507 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
508 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
509 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
510
511 /* next, make a null set of free lists */
512 for (i = 0; i < BUFFER_QUEUES; i++)
513 TAILQ_INIT(&bufqueues[i]);
514
515 /* finally, initialize each buffer header and stick on empty q */
516 for (i = 0; i < nbuf; i++) {
517 bp = &buf[i];
518 bzero(bp, sizeof *bp);
519 bp->b_flags = B_INVAL; /* we're just an empty header */
| 30
31#include <sys/param.h>
32#include <sys/systm.h>
33#include <sys/bio.h>
34#include <sys/conf.h>
35#include <sys/buf.h>
36#include <sys/devicestat.h>
37#include <sys/eventhandler.h>
38#include <sys/lock.h>
39#include <sys/malloc.h>
40#include <sys/mount.h>
41#include <sys/mutex.h>
42#include <sys/kernel.h>
43#include <sys/kthread.h>
44#include <sys/proc.h>
45#include <sys/resourcevar.h>
46#include <sys/sysctl.h>
47#include <sys/vmmeter.h>
48#include <sys/vnode.h>
49#include <vm/vm.h>
50#include <vm/vm_param.h>
51#include <vm/vm_kern.h>
52#include <vm/vm_pageout.h>
53#include <vm/vm_page.h>
54#include <vm/vm_object.h>
55#include <vm/vm_extern.h>
56#include <vm/vm_map.h>
57#include "opt_directio.h"
58#include "opt_swap.h"
59
60static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
61
62struct bio_ops bioops; /* I/O operation notification */
63
64static int ibwrite(struct buf *);
65
66struct buf_ops buf_ops_bio = {
67 "buf_ops_bio",
68 ibwrite
69};
70
71/*
72 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
73 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
74 */
75struct buf *buf; /* buffer header pool */
76
77static struct proc *bufdaemonproc;
78
79static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
80 vm_offset_t to);
81static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
82 vm_offset_t to);
83static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
84 int pageno, vm_page_t m);
85static void vfs_clean_pages(struct buf * bp);
86static void vfs_setdirty(struct buf *bp);
87static void vfs_vmio_release(struct buf *bp);
88static void vfs_backgroundwritedone(struct buf *bp);
89static int vfs_bio_clcheck(struct vnode *vp, int size,
90 daddr_t lblkno, daddr_t blkno);
91static int flushbufqueues(int flushdeps);
92static void buf_daemon(void);
93void bremfreel(struct buf * bp);
94
95int vmiodirenable = TRUE;
96SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
97 "Use the VM system for directory writes");
98int runningbufspace;
99SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
100 "Amount of presently outstanding async buffer io");
101static int bufspace;
102SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
103 "KVA memory used for bufs");
104static int maxbufspace;
105SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
106 "Maximum allowed value of bufspace (including buf_daemon)");
107static int bufmallocspace;
108SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
109 "Amount of malloced memory for buffers");
110static int maxbufmallocspace;
111SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
112 "Maximum amount of malloced memory for buffers");
113static int lobufspace;
114SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
115 "Minimum amount of buffers we want to have");
116static int hibufspace;
117SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
118 "Maximum allowed value of bufspace (excluding buf_daemon)");
119static int bufreusecnt;
120SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
121 "Number of times we have reused a buffer");
122static int buffreekvacnt;
123SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
124 "Number of times we have freed the KVA space from some buffer");
125static int bufdefragcnt;
126SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
127 "Number of times we have had to repeat buffer allocation to defragment");
128static int lorunningspace;
129SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
130 "Minimum preferred space used for in-progress I/O");
131static int hirunningspace;
132SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
133 "Maximum amount of space to use for in-progress I/O");
134static int dirtybufferflushes;
135SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
136 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
137static int altbufferflushes;
138SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
139 0, "Number of fsync flushes to limit dirty buffers");
140static int recursiveflushes;
141SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
142 0, "Number of flushes skipped due to being recursive");
143static int numdirtybuffers;
144SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
145 "Number of buffers that are dirty (has unwritten changes) at the moment");
146static int lodirtybuffers;
147SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
148 "How many buffers we want to have free before bufdaemon can sleep");
149static int hidirtybuffers;
150SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
151 "When the number of dirty buffers is considered severe");
152static int dirtybufthresh;
153SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
154 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
155static int numfreebuffers;
156SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
157 "Number of free buffers");
158static int lofreebuffers;
159SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
160 "XXX Unused");
161static int hifreebuffers;
162SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
163 "XXX Complicatedly unused");
164static int getnewbufcalls;
165SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
166 "Number of calls to getnewbuf");
167static int getnewbufrestarts;
168SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
169 "Number of times getnewbuf has had to restart a buffer aquisition");
170static int dobkgrdwrite = 1;
171SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0,
172 "Do background writes (honoring the BV_BKGRDWRITE flag)?");
173
174/*
175 * Wakeup point for bufdaemon, as well as indicator of whether it is already
176 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
177 * is idling.
178 */
179static int bd_request;
180
181/*
182 * This lock synchronizes access to bd_request.
183 */
184static struct mtx bdlock;
185
186/*
187 * bogus page -- for I/O to/from partially complete buffers
188 * this is a temporary solution to the problem, but it is not
189 * really that bad. it would be better to split the buffer
190 * for input in the case of buffers partially already in memory,
191 * but the code is intricate enough already.
192 */
193vm_page_t bogus_page;
194
195/*
196 * Synchronization (sleep/wakeup) variable for active buffer space requests.
197 * Set when wait starts, cleared prior to wakeup().
198 * Used in runningbufwakeup() and waitrunningbufspace().
199 */
200static int runningbufreq;
201
202/*
203 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
204 * waitrunningbufspace().
205 */
206static struct mtx rbreqlock;
207
208/*
209 * Synchronization (sleep/wakeup) variable for buffer requests.
210 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
211 * by and/or.
212 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
213 * getnewbuf(), and getblk().
214 */
215static int needsbuffer;
216
217/*
218 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
219 */
220static struct mtx nblock;
221
222/*
223 * Lock that protects against bwait()/bdone()/B_DONE races.
224 */
225
226static struct mtx bdonelock;
227
228/*
229 * Definitions for the buffer free lists.
230 */
231#define BUFFER_QUEUES 5 /* number of free buffer queues */
232
233#define QUEUE_NONE 0 /* on no queue */
234#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
235#define QUEUE_DIRTY 2 /* B_DELWRI buffers */
236#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */
237#define QUEUE_EMPTY 4 /* empty buffer headers */
238
239/* Queues for free buffers with various properties */
240static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
241
242/* Lock for the bufqueues */
243static struct mtx bqlock;
244
245/*
246 * Single global constant for BUF_WMESG, to avoid getting multiple references.
247 * buf_wmesg is referred from macros.
248 */
249const char *buf_wmesg = BUF_WMESG;
250
251#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
252#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
253#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
254#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
255
256#ifdef DIRECTIO
257extern void ffs_rawread_setup(void);
258#endif /* DIRECTIO */
259/*
260 * numdirtywakeup:
261 *
262 * If someone is blocked due to there being too many dirty buffers,
263 * and numdirtybuffers is now reasonable, wake them up.
264 */
265
266static __inline void
267numdirtywakeup(int level)
268{
269 if (numdirtybuffers <= level) {
270 mtx_lock(&nblock);
271 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
272 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
273 wakeup(&needsbuffer);
274 }
275 mtx_unlock(&nblock);
276 }
277}
278
279/*
280 * bufspacewakeup:
281 *
282 * Called when buffer space is potentially available for recovery.
283 * getnewbuf() will block on this flag when it is unable to free
284 * sufficient buffer space. Buffer space becomes recoverable when
285 * bp's get placed back in the queues.
286 */
287
288static __inline void
289bufspacewakeup(void)
290{
291 /*
292 * If someone is waiting for BUF space, wake them up. Even
293 * though we haven't freed the kva space yet, the waiting
294 * process will be able to now.
295 */
296 mtx_lock(&nblock);
297 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
298 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
299 wakeup(&needsbuffer);
300 }
301 mtx_unlock(&nblock);
302}
303
304/*
305 * runningbufwakeup() - in-progress I/O accounting.
306 *
307 */
308static __inline void
309runningbufwakeup(struct buf *bp)
310{
311 if (bp->b_runningbufspace) {
312 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace);
313 bp->b_runningbufspace = 0;
314 mtx_lock(&rbreqlock);
315 if (runningbufreq && runningbufspace <= lorunningspace) {
316 runningbufreq = 0;
317 wakeup(&runningbufreq);
318 }
319 mtx_unlock(&rbreqlock);
320 }
321}
322
323/*
324 * bufcountwakeup:
325 *
326 * Called when a buffer has been added to one of the free queues to
327 * account for the buffer and to wakeup anyone waiting for free buffers.
328 * This typically occurs when large amounts of metadata are being handled
329 * by the buffer cache ( else buffer space runs out first, usually ).
330 */
331
332static __inline void
333bufcountwakeup(void)
334{
335 atomic_add_int(&numfreebuffers, 1);
336 mtx_lock(&nblock);
337 if (needsbuffer) {
338 needsbuffer &= ~VFS_BIO_NEED_ANY;
339 if (numfreebuffers >= hifreebuffers)
340 needsbuffer &= ~VFS_BIO_NEED_FREE;
341 wakeup(&needsbuffer);
342 }
343 mtx_unlock(&nblock);
344}
345
346/*
347 * waitrunningbufspace()
348 *
349 * runningbufspace is a measure of the amount of I/O currently
350 * running. This routine is used in async-write situations to
351 * prevent creating huge backups of pending writes to a device.
352 * Only asynchronous writes are governed by this function.
353 *
354 * Reads will adjust runningbufspace, but will not block based on it.
355 * The read load has a side effect of reducing the allowed write load.
356 *
357 * This does NOT turn an async write into a sync write. It waits
358 * for earlier writes to complete and generally returns before the
359 * caller's write has reached the device.
360 */
361static __inline void
362waitrunningbufspace(void)
363{
364 mtx_lock(&rbreqlock);
365 while (runningbufspace > hirunningspace) {
366 ++runningbufreq;
367 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
368 }
369 mtx_unlock(&rbreqlock);
370}
371
372
373/*
374 * vfs_buf_test_cache:
375 *
376 * Called when a buffer is extended. This function clears the B_CACHE
377 * bit if the newly extended portion of the buffer does not contain
378 * valid data.
379 */
380static __inline__
381void
382vfs_buf_test_cache(struct buf *bp,
383 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
384 vm_page_t m)
385{
386 GIANT_REQUIRED;
387
388 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
389 if (bp->b_flags & B_CACHE) {
390 int base = (foff + off) & PAGE_MASK;
391 if (vm_page_is_valid(m, base, size) == 0)
392 bp->b_flags &= ~B_CACHE;
393 }
394}
395
396/* Wake up the buffer deamon if necessary */
397static __inline__
398void
399bd_wakeup(int dirtybuflevel)
400{
401 mtx_lock(&bdlock);
402 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
403 bd_request = 1;
404 wakeup(&bd_request);
405 }
406 mtx_unlock(&bdlock);
407}
408
409/*
410 * bd_speedup - speedup the buffer cache flushing code
411 */
412
413static __inline__
414void
415bd_speedup(void)
416{
417 bd_wakeup(1);
418}
419
420/*
421 * Calculating buffer cache scaling values and reserve space for buffer
422 * headers. This is called during low level kernel initialization and
423 * may be called more then once. We CANNOT write to the memory area
424 * being reserved at this time.
425 */
426caddr_t
427kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
428{
429 /*
430 * physmem_est is in pages. Convert it to kilobytes (assumes
431 * PAGE_SIZE is >= 1K)
432 */
433 physmem_est = physmem_est * (PAGE_SIZE / 1024);
434
435 /*
436 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
437 * For the first 64MB of ram nominally allocate sufficient buffers to
438 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
439 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing
440 * the buffer cache we limit the eventual kva reservation to
441 * maxbcache bytes.
442 *
443 * factor represents the 1/4 x ram conversion.
444 */
445 if (nbuf == 0) {
446 int factor = 4 * BKVASIZE / 1024;
447
448 nbuf = 50;
449 if (physmem_est > 4096)
450 nbuf += min((physmem_est - 4096) / factor,
451 65536 / factor);
452 if (physmem_est > 65536)
453 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
454
455 if (maxbcache && nbuf > maxbcache / BKVASIZE)
456 nbuf = maxbcache / BKVASIZE;
457 }
458
459#if 0
460 /*
461 * Do not allow the buffer_map to be more then 1/2 the size of the
462 * kernel_map.
463 */
464 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) /
465 (BKVASIZE * 2)) {
466 nbuf = (kernel_map->max_offset - kernel_map->min_offset) /
467 (BKVASIZE * 2);
468 printf("Warning: nbufs capped at %d\n", nbuf);
469 }
470#endif
471
472 /*
473 * swbufs are used as temporary holders for I/O, such as paging I/O.
474 * We have no less then 16 and no more then 256.
475 */
476 nswbuf = max(min(nbuf/4, 256), 16);
477#ifdef NSWBUF_MIN
478 if (nswbuf < NSWBUF_MIN)
479 nswbuf = NSWBUF_MIN;
480#endif
481#ifdef DIRECTIO
482 ffs_rawread_setup();
483#endif
484
485 /*
486 * Reserve space for the buffer cache buffers
487 */
488 swbuf = (void *)v;
489 v = (caddr_t)(swbuf + nswbuf);
490 buf = (void *)v;
491 v = (caddr_t)(buf + nbuf);
492
493 return(v);
494}
495
496/* Initialize the buffer subsystem. Called before use of any buffers. */
497void
498bufinit(void)
499{
500 struct buf *bp;
501 int i;
502
503 GIANT_REQUIRED;
504
505 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
506 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
507 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
508 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
509 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF);
510
511 /* next, make a null set of free lists */
512 for (i = 0; i < BUFFER_QUEUES; i++)
513 TAILQ_INIT(&bufqueues[i]);
514
515 /* finally, initialize each buffer header and stick on empty q */
516 for (i = 0; i < nbuf; i++) {
517 bp = &buf[i];
518 bzero(bp, sizeof *bp);
519 bp->b_flags = B_INVAL; /* we're just an empty header */
|
520 bp->b_dev = NODEV;
| 520 bp->b_dev = NULL;
|
521 bp->b_rcred = NOCRED;
522 bp->b_wcred = NOCRED;
523 bp->b_qindex = QUEUE_EMPTY;
524 bp->b_vflags = 0;
525 bp->b_xflags = 0;
526 LIST_INIT(&bp->b_dep);
527 BUF_LOCKINIT(bp);
528 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
529 }
530
531 /*
532 * maxbufspace is the absolute maximum amount of buffer space we are
533 * allowed to reserve in KVM and in real terms. The absolute maximum
534 * is nominally used by buf_daemon. hibufspace is the nominal maximum
535 * used by most other processes. The differential is required to
536 * ensure that buf_daemon is able to run when other processes might
537 * be blocked waiting for buffer space.
538 *
539 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
540 * this may result in KVM fragmentation which is not handled optimally
541 * by the system.
542 */
543 maxbufspace = nbuf * BKVASIZE;
544 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
545 lobufspace = hibufspace - MAXBSIZE;
546
547 lorunningspace = 512 * 1024;
548 hirunningspace = 1024 * 1024;
549
550/*
551 * Limit the amount of malloc memory since it is wired permanently into
552 * the kernel space. Even though this is accounted for in the buffer
553 * allocation, we don't want the malloced region to grow uncontrolled.
554 * The malloc scheme improves memory utilization significantly on average
555 * (small) directories.
556 */
557 maxbufmallocspace = hibufspace / 20;
558
559/*
560 * Reduce the chance of a deadlock occuring by limiting the number
561 * of delayed-write dirty buffers we allow to stack up.
562 */
563 hidirtybuffers = nbuf / 4 + 20;
564 dirtybufthresh = hidirtybuffers * 9 / 10;
565 numdirtybuffers = 0;
566/*
567 * To support extreme low-memory systems, make sure hidirtybuffers cannot
568 * eat up all available buffer space. This occurs when our minimum cannot
569 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
570 * BKVASIZE'd (8K) buffers.
571 */
572 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
573 hidirtybuffers >>= 1;
574 }
575 lodirtybuffers = hidirtybuffers / 2;
576
577/*
578 * Try to keep the number of free buffers in the specified range,
579 * and give special processes (e.g. like buf_daemon) access to an
580 * emergency reserve.
581 */
582 lofreebuffers = nbuf / 18 + 5;
583 hifreebuffers = 2 * lofreebuffers;
584 numfreebuffers = nbuf;
585
586/*
587 * Maximum number of async ops initiated per buf_daemon loop. This is
588 * somewhat of a hack at the moment, we really need to limit ourselves
589 * based on the number of bytes of I/O in-transit that were initiated
590 * from buf_daemon.
591 */
592
593 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
594 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
595}
596
597/*
598 * bfreekva() - free the kva allocation for a buffer.
599 *
600 * Must be called at splbio() or higher as this is the only locking for
601 * buffer_map.
602 *
603 * Since this call frees up buffer space, we call bufspacewakeup().
604 */
605static void
606bfreekva(struct buf * bp)
607{
608 GIANT_REQUIRED;
609
610 if (bp->b_kvasize) {
611 atomic_add_int(&buffreekvacnt, 1);
612 atomic_subtract_int(&bufspace, bp->b_kvasize);
613 vm_map_delete(buffer_map,
614 (vm_offset_t) bp->b_kvabase,
615 (vm_offset_t) bp->b_kvabase + bp->b_kvasize
616 );
617 bp->b_kvasize = 0;
618 bufspacewakeup();
619 }
620}
621
622/*
623 * bremfree:
624 *
625 * Remove the buffer from the appropriate free list.
626 */
627void
628bremfree(struct buf * bp)
629{
630 mtx_lock(&bqlock);
631 bremfreel(bp);
632 mtx_unlock(&bqlock);
633}
634
635void
636bremfreel(struct buf * bp)
637{
638 int s = splbio();
639 int old_qindex = bp->b_qindex;
640
641 GIANT_REQUIRED;
642
643 if (bp->b_qindex != QUEUE_NONE) {
644 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
645 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
646 bp->b_qindex = QUEUE_NONE;
647 } else {
648 if (BUF_REFCNT(bp) <= 1)
649 panic("bremfree: removing a buffer not on a queue");
650 }
651
652 /*
653 * Fixup numfreebuffers count. If the buffer is invalid or not
654 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
655 * the buffer was free and we must decrement numfreebuffers.
656 */
657 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
658 switch(old_qindex) {
659 case QUEUE_DIRTY:
660 case QUEUE_CLEAN:
661 case QUEUE_EMPTY:
662 case QUEUE_EMPTYKVA:
663 atomic_subtract_int(&numfreebuffers, 1);
664 break;
665 default:
666 break;
667 }
668 }
669 splx(s);
670}
671
672
673/*
674 * Get a buffer with the specified data. Look in the cache first. We
675 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
676 * is set, the buffer is valid and we do not have to do anything ( see
677 * getblk() ). This is really just a special case of breadn().
678 */
679int
680bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
681 struct buf ** bpp)
682{
683
684 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
685}
686
687/*
688 * Operates like bread, but also starts asynchronous I/O on
689 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior
690 * to initiating I/O . If B_CACHE is set, the buffer is valid
691 * and we do not have to do anything.
692 */
693int
694breadn(struct vnode * vp, daddr_t blkno, int size,
695 daddr_t * rablkno, int *rabsize,
696 int cnt, struct ucred * cred, struct buf ** bpp)
697{
698 struct buf *bp, *rabp;
699 int i;
700 int rv = 0, readwait = 0;
701
702 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
703
704 /* if not found in cache, do some I/O */
705 if ((bp->b_flags & B_CACHE) == 0) {
706 if (curthread != PCPU_GET(idlethread))
707 curthread->td_proc->p_stats->p_ru.ru_inblock++;
708 bp->b_iocmd = BIO_READ;
709 bp->b_flags &= ~B_INVAL;
710 bp->b_ioflags &= ~BIO_ERROR;
711 if (bp->b_rcred == NOCRED && cred != NOCRED)
712 bp->b_rcred = crhold(cred);
713 vfs_busy_pages(bp, 0);
714 bp->b_iooffset = dbtob(bp->b_blkno);
715 if (vp->v_type == VCHR)
716 VOP_SPECSTRATEGY(vp, bp);
717 else
718 VOP_STRATEGY(vp, bp);
719 ++readwait;
720 }
721
722 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
723 if (inmem(vp, *rablkno))
724 continue;
725 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
726
727 if ((rabp->b_flags & B_CACHE) == 0) {
728 if (curthread != PCPU_GET(idlethread))
729 curthread->td_proc->p_stats->p_ru.ru_inblock++;
730 rabp->b_flags |= B_ASYNC;
731 rabp->b_flags &= ~B_INVAL;
732 rabp->b_ioflags &= ~BIO_ERROR;
733 rabp->b_iocmd = BIO_READ;
734 if (rabp->b_rcred == NOCRED && cred != NOCRED)
735 rabp->b_rcred = crhold(cred);
736 vfs_busy_pages(rabp, 0);
737 BUF_KERNPROC(rabp);
738 rabp->b_iooffset = dbtob(rabp->b_blkno);
739 if (vp->v_type == VCHR)
740 VOP_SPECSTRATEGY(vp, rabp);
741 else
742 VOP_STRATEGY(vp, rabp);
743 } else {
744 brelse(rabp);
745 }
746 }
747
748 if (readwait) {
749 rv = bufwait(bp);
750 }
751 return (rv);
752}
753
754/*
755 * Write, release buffer on completion. (Done by iodone
756 * if async). Do not bother writing anything if the buffer
757 * is invalid.
758 *
759 * Note that we set B_CACHE here, indicating that buffer is
760 * fully valid and thus cacheable. This is true even of NFS
761 * now so we set it generally. This could be set either here
762 * or in biodone() since the I/O is synchronous. We put it
763 * here.
764 */
765int
766bwrite(struct buf * bp)
767{
768
769 KASSERT(bp->b_op != NULL && bp->b_op->bop_write != NULL,
770 ("Martian buffer %p in bwrite: nobody to write it.", bp));
771 return (bp->b_op->bop_write(bp));
772}
773
774static int
775ibwrite(struct buf * bp)
776{
777 int oldflags, s;
778 struct buf *newbp;
779
780 if (bp->b_flags & B_INVAL) {
781 brelse(bp);
782 return (0);
783 }
784
785 oldflags = bp->b_flags;
786
787 if (BUF_REFCNT(bp) == 0)
788 panic("ibwrite: buffer is not busy???");
789 s = splbio();
790 /*
791 * If a background write is already in progress, delay
792 * writing this block if it is asynchronous. Otherwise
793 * wait for the background write to complete.
794 */
795 VI_LOCK(bp->b_vp);
796 if (bp->b_vflags & BV_BKGRDINPROG) {
797 if (bp->b_flags & B_ASYNC) {
798 VI_UNLOCK(bp->b_vp);
799 splx(s);
800 bdwrite(bp);
801 return (0);
802 }
803 bp->b_vflags |= BV_BKGRDWAIT;
804 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0);
805 if (bp->b_vflags & BV_BKGRDINPROG)
806 panic("ibwrite: still writing");
807 }
808 VI_UNLOCK(bp->b_vp);
809
810 /* Mark the buffer clean */
811 bundirty(bp);
812
813 /*
814 * If this buffer is marked for background writing and we
815 * do not have to wait for it, make a copy and write the
816 * copy so as to leave this buffer ready for further use.
817 *
818 * This optimization eats a lot of memory. If we have a page
819 * or buffer shortfall we can't do it.
820 */
821 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) &&
822 (bp->b_flags & B_ASYNC) &&
823 !vm_page_count_severe() &&
824 !buf_dirty_count_severe()) {
825 if (bp->b_iodone != NULL) {
826 printf("bp->b_iodone = %p\n", bp->b_iodone);
827 panic("ibwrite: need chained iodone");
828 }
829
830 /* get a new block */
831 newbp = geteblk(bp->b_bufsize);
832
833 /*
834 * set it to be identical to the old block. We have to
835 * set b_lblkno and BKGRDMARKER before calling bgetvp()
836 * to avoid confusing the splay tree and gbincore().
837 */
838 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
839 newbp->b_lblkno = bp->b_lblkno;
840 newbp->b_xflags |= BX_BKGRDMARKER;
841 VI_LOCK(bp->b_vp);
842 bp->b_vflags |= BV_BKGRDINPROG;
843 bgetvp(bp->b_vp, newbp);
844 VI_UNLOCK(bp->b_vp);
845 newbp->b_blkno = bp->b_blkno;
846 newbp->b_offset = bp->b_offset;
847 newbp->b_iodone = vfs_backgroundwritedone;
848 newbp->b_flags |= B_ASYNC;
849 newbp->b_flags &= ~B_INVAL;
850
851 /* move over the dependencies */
852 if (LIST_FIRST(&bp->b_dep) != NULL)
853 buf_movedeps(bp, newbp);
854
855 /*
856 * Initiate write on the copy, release the original to
857 * the B_LOCKED queue so that it cannot go away until
858 * the background write completes. If not locked it could go
859 * away and then be reconstituted while it was being written.
860 * If the reconstituted buffer were written, we could end up
861 * with two background copies being written at the same time.
862 */
863 bqrelse(bp);
864 bp = newbp;
865 }
866
867 bp->b_flags &= ~B_DONE;
868 bp->b_ioflags &= ~BIO_ERROR;
869 bp->b_flags |= B_WRITEINPROG | B_CACHE;
870 bp->b_iocmd = BIO_WRITE;
871
872 VI_LOCK(bp->b_vp);
873 bp->b_vp->v_numoutput++;
874 VI_UNLOCK(bp->b_vp);
875 vfs_busy_pages(bp, 1);
876
877 /*
878 * Normal bwrites pipeline writes
879 */
880 bp->b_runningbufspace = bp->b_bufsize;
881 atomic_add_int(&runningbufspace, bp->b_runningbufspace);
882
883 if (curthread != PCPU_GET(idlethread))
884 curthread->td_proc->p_stats->p_ru.ru_oublock++;
885 splx(s);
886 if (oldflags & B_ASYNC)
887 BUF_KERNPROC(bp);
888 bp->b_iooffset = dbtob(bp->b_blkno);
889 if (bp->b_vp->v_type == VCHR) {
890 if (!buf_prewrite(bp->b_vp, bp))
891 VOP_SPECSTRATEGY(bp->b_vp, bp);
892 } else {
893 VOP_STRATEGY(bp->b_vp, bp);
894 }
895
896 if ((oldflags & B_ASYNC) == 0) {
897 int rtval = bufwait(bp);
898 brelse(bp);
899 return (rtval);
900 } else {
901 /*
902 * don't allow the async write to saturate the I/O
903 * system. We will not deadlock here because
904 * we are blocking waiting for I/O that is already in-progress
905 * to complete. We do not block here if it is the update
906 * or syncer daemon trying to clean up as that can lead
907 * to deadlock.
908 */
909 if (curthread->td_proc != bufdaemonproc &&
910 curthread->td_proc != updateproc)
911 waitrunningbufspace();
912 }
913
914 return (0);
915}
916
917/*
918 * Complete a background write started from bwrite.
919 */
920static void
921vfs_backgroundwritedone(bp)
922 struct buf *bp;
923{
924 struct buf *origbp;
925
926 /*
927 * Find the original buffer that we are writing.
928 */
929 VI_LOCK(bp->b_vp);
930 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
931 panic("backgroundwritedone: lost buffer");
932
933 /*
934 * Clear the BV_BKGRDINPROG flag in the original buffer
935 * and awaken it if it is waiting for the write to complete.
936 * If BV_BKGRDINPROG is not set in the original buffer it must
937 * have been released and re-instantiated - which is not legal.
938 */
939 KASSERT((origbp->b_vflags & BV_BKGRDINPROG),
940 ("backgroundwritedone: lost buffer2"));
941 origbp->b_vflags &= ~BV_BKGRDINPROG;
942 if (origbp->b_vflags & BV_BKGRDWAIT) {
943 origbp->b_vflags &= ~BV_BKGRDWAIT;
944 wakeup(&origbp->b_xflags);
945 }
946 VI_UNLOCK(bp->b_vp);
947 /*
948 * Process dependencies then return any unfinished ones.
949 */
950 if (LIST_FIRST(&bp->b_dep) != NULL)
951 buf_complete(bp);
952 if (LIST_FIRST(&bp->b_dep) != NULL)
953 buf_movedeps(bp, origbp);
954
955 /*
956 * This buffer is marked B_NOCACHE, so when it is released
957 * by biodone, it will be tossed. We mark it with BIO_READ
958 * to avoid biodone doing a second vwakeup.
959 */
960 bp->b_flags |= B_NOCACHE;
961 bp->b_iocmd = BIO_READ;
962 bp->b_flags &= ~(B_CACHE | B_DONE);
963 bp->b_iodone = 0;
964 bufdone(bp);
965}
966
967/*
968 * Delayed write. (Buffer is marked dirty). Do not bother writing
969 * anything if the buffer is marked invalid.
970 *
971 * Note that since the buffer must be completely valid, we can safely
972 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
973 * biodone() in order to prevent getblk from writing the buffer
974 * out synchronously.
975 */
976void
977bdwrite(struct buf * bp)
978{
979 struct thread *td = curthread;
980 struct vnode *vp;
981 struct buf *nbp;
982
983 GIANT_REQUIRED;
984
985 if (BUF_REFCNT(bp) == 0)
986 panic("bdwrite: buffer is not busy");
987
988 if (bp->b_flags & B_INVAL) {
989 brelse(bp);
990 return;
991 }
992
993 /*
994 * If we have too many dirty buffers, don't create any more.
995 * If we are wildly over our limit, then force a complete
996 * cleanup. Otherwise, just keep the situation from getting
997 * out of control. Note that we have to avoid a recursive
998 * disaster and not try to clean up after our own cleanup!
999 */
1000 vp = bp->b_vp;
1001 VI_LOCK(vp);
1002 if (td->td_pflags & TDP_COWINPROGRESS) {
1003 recursiveflushes++;
1004 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) {
1005 VI_UNLOCK(vp);
1006 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td);
1007 VI_LOCK(vp);
1008 altbufferflushes++;
1009 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) {
1010 /*
1011 * Try to find a buffer to flush.
1012 */
1013 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) {
1014 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1015 buf_countdeps(nbp, 0) ||
1016 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
1017 continue;
1018 if (bp == nbp)
1019 panic("bdwrite: found ourselves");
1020 VI_UNLOCK(vp);
1021 if (nbp->b_flags & B_CLUSTEROK) {
1022 vfs_bio_awrite(nbp);
1023 } else {
1024 bremfree(nbp);
1025 bawrite(nbp);
1026 }
1027 VI_LOCK(vp);
1028 dirtybufferflushes++;
1029 break;
1030 }
1031 }
1032 VI_UNLOCK(vp);
1033
1034 bdirty(bp);
1035 /*
1036 * Set B_CACHE, indicating that the buffer is fully valid. This is
1037 * true even of NFS now.
1038 */
1039 bp->b_flags |= B_CACHE;
1040
1041 /*
1042 * This bmap keeps the system from needing to do the bmap later,
1043 * perhaps when the system is attempting to do a sync. Since it
1044 * is likely that the indirect block -- or whatever other datastructure
1045 * that the filesystem needs is still in memory now, it is a good
1046 * thing to do this. Note also, that if the pageout daemon is
1047 * requesting a sync -- there might not be enough memory to do
1048 * the bmap then... So, this is important to do.
1049 */
1050 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1051 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1052 }
1053
1054 /*
1055 * Set the *dirty* buffer range based upon the VM system dirty pages.
1056 */
1057 vfs_setdirty(bp);
1058
1059 /*
1060 * We need to do this here to satisfy the vnode_pager and the
1061 * pageout daemon, so that it thinks that the pages have been
1062 * "cleaned". Note that since the pages are in a delayed write
1063 * buffer -- the VFS layer "will" see that the pages get written
1064 * out on the next sync, or perhaps the cluster will be completed.
1065 */
1066 vfs_clean_pages(bp);
1067 bqrelse(bp);
1068
1069 /*
1070 * Wakeup the buffer flushing daemon if we have a lot of dirty
1071 * buffers (midpoint between our recovery point and our stall
1072 * point).
1073 */
1074 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1075
1076 /*
1077 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1078 * due to the softdep code.
1079 */
1080}
1081
1082/*
1083 * bdirty:
1084 *
1085 * Turn buffer into delayed write request. We must clear BIO_READ and
1086 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1087 * itself to properly update it in the dirty/clean lists. We mark it
1088 * B_DONE to ensure that any asynchronization of the buffer properly
1089 * clears B_DONE ( else a panic will occur later ).
1090 *
1091 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1092 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1093 * should only be called if the buffer is known-good.
1094 *
1095 * Since the buffer is not on a queue, we do not update the numfreebuffers
1096 * count.
1097 *
1098 * Must be called at splbio().
1099 * The buffer must be on QUEUE_NONE.
1100 */
1101void
1102bdirty(bp)
1103 struct buf *bp;
1104{
1105 KASSERT(bp->b_qindex == QUEUE_NONE,
1106 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1107 bp->b_flags &= ~(B_RELBUF);
1108 bp->b_iocmd = BIO_WRITE;
1109
1110 if ((bp->b_flags & B_DELWRI) == 0) {
1111 bp->b_flags |= B_DONE | B_DELWRI;
1112 reassignbuf(bp, bp->b_vp);
1113 atomic_add_int(&numdirtybuffers, 1);
1114 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1115 }
1116}
1117
1118/*
1119 * bundirty:
1120 *
1121 * Clear B_DELWRI for buffer.
1122 *
1123 * Since the buffer is not on a queue, we do not update the numfreebuffers
1124 * count.
1125 *
1126 * Must be called at splbio().
1127 * The buffer must be on QUEUE_NONE.
1128 */
1129
1130void
1131bundirty(bp)
1132 struct buf *bp;
1133{
1134 KASSERT(bp->b_qindex == QUEUE_NONE,
1135 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1136
1137 if (bp->b_flags & B_DELWRI) {
1138 bp->b_flags &= ~B_DELWRI;
1139 reassignbuf(bp, bp->b_vp);
1140 atomic_subtract_int(&numdirtybuffers, 1);
1141 numdirtywakeup(lodirtybuffers);
1142 }
1143 /*
1144 * Since it is now being written, we can clear its deferred write flag.
1145 */
1146 bp->b_flags &= ~B_DEFERRED;
1147}
1148
1149/*
1150 * bawrite:
1151 *
1152 * Asynchronous write. Start output on a buffer, but do not wait for
1153 * it to complete. The buffer is released when the output completes.
1154 *
1155 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1156 * B_INVAL buffers. Not us.
1157 */
1158void
1159bawrite(struct buf * bp)
1160{
1161 bp->b_flags |= B_ASYNC;
1162 (void) bwrite(bp);
1163}
1164
1165/*
1166 * bwillwrite:
1167 *
1168 * Called prior to the locking of any vnodes when we are expecting to
1169 * write. We do not want to starve the buffer cache with too many
1170 * dirty buffers so we block here. By blocking prior to the locking
1171 * of any vnodes we attempt to avoid the situation where a locked vnode
1172 * prevents the various system daemons from flushing related buffers.
1173 */
1174
1175void
1176bwillwrite(void)
1177{
1178 if (numdirtybuffers >= hidirtybuffers) {
1179 int s;
1180
1181 mtx_lock(&Giant);
1182 s = splbio();
1183 mtx_lock(&nblock);
1184 while (numdirtybuffers >= hidirtybuffers) {
1185 bd_wakeup(1);
1186 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1187 msleep(&needsbuffer, &nblock,
1188 (PRIBIO + 4), "flswai", 0);
1189 }
1190 splx(s);
1191 mtx_unlock(&nblock);
1192 mtx_unlock(&Giant);
1193 }
1194}
1195
1196/*
1197 * Return true if we have too many dirty buffers.
1198 */
1199int
1200buf_dirty_count_severe(void)
1201{
1202 return(numdirtybuffers >= hidirtybuffers);
1203}
1204
1205/*
1206 * brelse:
1207 *
1208 * Release a busy buffer and, if requested, free its resources. The
1209 * buffer will be stashed in the appropriate bufqueue[] allowing it
1210 * to be accessed later as a cache entity or reused for other purposes.
1211 */
1212void
1213brelse(struct buf * bp)
1214{
1215 int s;
1216
1217 GIANT_REQUIRED;
1218
1219 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1220 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1221
1222 s = splbio();
1223
1224 if (bp->b_iocmd == BIO_WRITE &&
1225 (bp->b_ioflags & BIO_ERROR) &&
1226 !(bp->b_flags & B_INVAL)) {
1227 /*
1228 * Failed write, redirty. Must clear BIO_ERROR to prevent
1229 * pages from being scrapped. If B_INVAL is set then
1230 * this case is not run and the next case is run to
1231 * destroy the buffer. B_INVAL can occur if the buffer
1232 * is outside the range supported by the underlying device.
1233 */
1234 bp->b_ioflags &= ~BIO_ERROR;
1235 bdirty(bp);
1236 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1237 (bp->b_ioflags & BIO_ERROR) ||
1238 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) {
1239 /*
1240 * Either a failed I/O or we were asked to free or not
1241 * cache the buffer.
1242 */
1243 bp->b_flags |= B_INVAL;
1244 if (LIST_FIRST(&bp->b_dep) != NULL)
1245 buf_deallocate(bp);
1246 if (bp->b_flags & B_DELWRI) {
1247 atomic_subtract_int(&numdirtybuffers, 1);
1248 numdirtywakeup(lodirtybuffers);
1249 }
1250 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1251 if ((bp->b_flags & B_VMIO) == 0) {
1252 if (bp->b_bufsize)
1253 allocbuf(bp, 0);
1254 if (bp->b_vp)
1255 brelvp(bp);
1256 }
1257 }
1258
1259 /*
1260 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1261 * is called with B_DELWRI set, the underlying pages may wind up
1262 * getting freed causing a previous write (bdwrite()) to get 'lost'
1263 * because pages associated with a B_DELWRI bp are marked clean.
1264 *
1265 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1266 * if B_DELWRI is set.
1267 *
1268 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1269 * on pages to return pages to the VM page queues.
1270 */
1271 if (bp->b_flags & B_DELWRI)
1272 bp->b_flags &= ~B_RELBUF;
1273 else if (vm_page_count_severe()) {
1274 /*
1275 * XXX This lock may not be necessary since BKGRDINPROG
1276 * cannot be set while we hold the buf lock, it can only be
1277 * cleared if it is already pending.
1278 */
1279 if (bp->b_vp) {
1280 VI_LOCK(bp->b_vp);
1281 if (!(bp->b_vflags & BV_BKGRDINPROG))
1282 bp->b_flags |= B_RELBUF;
1283 VI_UNLOCK(bp->b_vp);
1284 } else
1285 bp->b_flags |= B_RELBUF;
1286 }
1287
1288 /*
1289 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1290 * constituted, not even NFS buffers now. Two flags effect this. If
1291 * B_INVAL, the struct buf is invalidated but the VM object is kept
1292 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1293 *
1294 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1295 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1296 * buffer is also B_INVAL because it hits the re-dirtying code above.
1297 *
1298 * Normally we can do this whether a buffer is B_DELWRI or not. If
1299 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1300 * the commit state and we cannot afford to lose the buffer. If the
1301 * buffer has a background write in progress, we need to keep it
1302 * around to prevent it from being reconstituted and starting a second
1303 * background write.
1304 */
1305 if ((bp->b_flags & B_VMIO)
1306 && !(bp->b_vp->v_mount != NULL &&
1307 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1308 !vn_isdisk(bp->b_vp, NULL) &&
1309 (bp->b_flags & B_DELWRI))
1310 ) {
1311
1312 int i, j, resid;
1313 vm_page_t m;
1314 off_t foff;
1315 vm_pindex_t poff;
1316 vm_object_t obj;
1317
1318 obj = bp->b_object;
1319
1320 /*
1321 * Get the base offset and length of the buffer. Note that
1322 * in the VMIO case if the buffer block size is not
1323 * page-aligned then b_data pointer may not be page-aligned.
1324 * But our b_pages[] array *IS* page aligned.
1325 *
1326 * block sizes less then DEV_BSIZE (usually 512) are not
1327 * supported due to the page granularity bits (m->valid,
1328 * m->dirty, etc...).
1329 *
1330 * See man buf(9) for more information
1331 */
1332 resid = bp->b_bufsize;
1333 foff = bp->b_offset;
1334 VM_OBJECT_LOCK(obj);
1335 for (i = 0; i < bp->b_npages; i++) {
1336 int had_bogus = 0;
1337
1338 m = bp->b_pages[i];
1339
1340 /*
1341 * If we hit a bogus page, fixup *all* the bogus pages
1342 * now.
1343 */
1344 if (m == bogus_page) {
1345 poff = OFF_TO_IDX(bp->b_offset);
1346 had_bogus = 1;
1347
1348 for (j = i; j < bp->b_npages; j++) {
1349 vm_page_t mtmp;
1350 mtmp = bp->b_pages[j];
1351 if (mtmp == bogus_page) {
1352 mtmp = vm_page_lookup(obj, poff + j);
1353 if (!mtmp) {
1354 panic("brelse: page missing\n");
1355 }
1356 bp->b_pages[j] = mtmp;
1357 }
1358 }
1359
1360 if ((bp->b_flags & B_INVAL) == 0) {
1361 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1362 }
1363 m = bp->b_pages[i];
1364 }
1365 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) {
1366 int poffset = foff & PAGE_MASK;
1367 int presid = resid > (PAGE_SIZE - poffset) ?
1368 (PAGE_SIZE - poffset) : resid;
1369
1370 KASSERT(presid >= 0, ("brelse: extra page"));
1371 vm_page_lock_queues();
1372 vm_page_set_invalid(m, poffset, presid);
1373 vm_page_unlock_queues();
1374 if (had_bogus)
1375 printf("avoided corruption bug in bogus_page/brelse code\n");
1376 }
1377 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1378 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1379 }
1380 VM_OBJECT_UNLOCK(obj);
1381 if (bp->b_flags & (B_INVAL | B_RELBUF))
1382 vfs_vmio_release(bp);
1383
1384 } else if (bp->b_flags & B_VMIO) {
1385
1386 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1387 vfs_vmio_release(bp);
1388 }
1389
1390 }
1391
1392 if (bp->b_qindex != QUEUE_NONE)
1393 panic("brelse: free buffer onto another queue???");
1394 if (BUF_REFCNT(bp) > 1) {
1395 /* do not release to free list */
1396 BUF_UNLOCK(bp);
1397 splx(s);
1398 return;
1399 }
1400
1401 /* enqueue */
1402 mtx_lock(&bqlock);
1403
1404 /* buffers with no memory */
1405 if (bp->b_bufsize == 0) {
1406 bp->b_flags |= B_INVAL;
1407 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1408 if (bp->b_vflags & BV_BKGRDINPROG)
1409 panic("losing buffer 1");
1410 if (bp->b_kvasize) {
1411 bp->b_qindex = QUEUE_EMPTYKVA;
1412 } else {
1413 bp->b_qindex = QUEUE_EMPTY;
1414 }
1415 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
| 521 bp->b_rcred = NOCRED;
522 bp->b_wcred = NOCRED;
523 bp->b_qindex = QUEUE_EMPTY;
524 bp->b_vflags = 0;
525 bp->b_xflags = 0;
526 LIST_INIT(&bp->b_dep);
527 BUF_LOCKINIT(bp);
528 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
529 }
530
531 /*
532 * maxbufspace is the absolute maximum amount of buffer space we are
533 * allowed to reserve in KVM and in real terms. The absolute maximum
534 * is nominally used by buf_daemon. hibufspace is the nominal maximum
535 * used by most other processes. The differential is required to
536 * ensure that buf_daemon is able to run when other processes might
537 * be blocked waiting for buffer space.
538 *
539 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
540 * this may result in KVM fragmentation which is not handled optimally
541 * by the system.
542 */
543 maxbufspace = nbuf * BKVASIZE;
544 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
545 lobufspace = hibufspace - MAXBSIZE;
546
547 lorunningspace = 512 * 1024;
548 hirunningspace = 1024 * 1024;
549
550/*
551 * Limit the amount of malloc memory since it is wired permanently into
552 * the kernel space. Even though this is accounted for in the buffer
553 * allocation, we don't want the malloced region to grow uncontrolled.
554 * The malloc scheme improves memory utilization significantly on average
555 * (small) directories.
556 */
557 maxbufmallocspace = hibufspace / 20;
558
559/*
560 * Reduce the chance of a deadlock occuring by limiting the number
561 * of delayed-write dirty buffers we allow to stack up.
562 */
563 hidirtybuffers = nbuf / 4 + 20;
564 dirtybufthresh = hidirtybuffers * 9 / 10;
565 numdirtybuffers = 0;
566/*
567 * To support extreme low-memory systems, make sure hidirtybuffers cannot
568 * eat up all available buffer space. This occurs when our minimum cannot
569 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
570 * BKVASIZE'd (8K) buffers.
571 */
572 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
573 hidirtybuffers >>= 1;
574 }
575 lodirtybuffers = hidirtybuffers / 2;
576
577/*
578 * Try to keep the number of free buffers in the specified range,
579 * and give special processes (e.g. like buf_daemon) access to an
580 * emergency reserve.
581 */
582 lofreebuffers = nbuf / 18 + 5;
583 hifreebuffers = 2 * lofreebuffers;
584 numfreebuffers = nbuf;
585
586/*
587 * Maximum number of async ops initiated per buf_daemon loop. This is
588 * somewhat of a hack at the moment, we really need to limit ourselves
589 * based on the number of bytes of I/O in-transit that were initiated
590 * from buf_daemon.
591 */
592
593 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
594 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
595}
596
597/*
598 * bfreekva() - free the kva allocation for a buffer.
599 *
600 * Must be called at splbio() or higher as this is the only locking for
601 * buffer_map.
602 *
603 * Since this call frees up buffer space, we call bufspacewakeup().
604 */
605static void
606bfreekva(struct buf * bp)
607{
608 GIANT_REQUIRED;
609
610 if (bp->b_kvasize) {
611 atomic_add_int(&buffreekvacnt, 1);
612 atomic_subtract_int(&bufspace, bp->b_kvasize);
613 vm_map_delete(buffer_map,
614 (vm_offset_t) bp->b_kvabase,
615 (vm_offset_t) bp->b_kvabase + bp->b_kvasize
616 );
617 bp->b_kvasize = 0;
618 bufspacewakeup();
619 }
620}
621
622/*
623 * bremfree:
624 *
625 * Remove the buffer from the appropriate free list.
626 */
627void
628bremfree(struct buf * bp)
629{
630 mtx_lock(&bqlock);
631 bremfreel(bp);
632 mtx_unlock(&bqlock);
633}
634
635void
636bremfreel(struct buf * bp)
637{
638 int s = splbio();
639 int old_qindex = bp->b_qindex;
640
641 GIANT_REQUIRED;
642
643 if (bp->b_qindex != QUEUE_NONE) {
644 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
645 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
646 bp->b_qindex = QUEUE_NONE;
647 } else {
648 if (BUF_REFCNT(bp) <= 1)
649 panic("bremfree: removing a buffer not on a queue");
650 }
651
652 /*
653 * Fixup numfreebuffers count. If the buffer is invalid or not
654 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
655 * the buffer was free and we must decrement numfreebuffers.
656 */
657 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
658 switch(old_qindex) {
659 case QUEUE_DIRTY:
660 case QUEUE_CLEAN:
661 case QUEUE_EMPTY:
662 case QUEUE_EMPTYKVA:
663 atomic_subtract_int(&numfreebuffers, 1);
664 break;
665 default:
666 break;
667 }
668 }
669 splx(s);
670}
671
672
673/*
674 * Get a buffer with the specified data. Look in the cache first. We
675 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
676 * is set, the buffer is valid and we do not have to do anything ( see
677 * getblk() ). This is really just a special case of breadn().
678 */
679int
680bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
681 struct buf ** bpp)
682{
683
684 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
685}
686
687/*
688 * Operates like bread, but also starts asynchronous I/O on
689 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior
690 * to initiating I/O . If B_CACHE is set, the buffer is valid
691 * and we do not have to do anything.
692 */
693int
694breadn(struct vnode * vp, daddr_t blkno, int size,
695 daddr_t * rablkno, int *rabsize,
696 int cnt, struct ucred * cred, struct buf ** bpp)
697{
698 struct buf *bp, *rabp;
699 int i;
700 int rv = 0, readwait = 0;
701
702 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
703
704 /* if not found in cache, do some I/O */
705 if ((bp->b_flags & B_CACHE) == 0) {
706 if (curthread != PCPU_GET(idlethread))
707 curthread->td_proc->p_stats->p_ru.ru_inblock++;
708 bp->b_iocmd = BIO_READ;
709 bp->b_flags &= ~B_INVAL;
710 bp->b_ioflags &= ~BIO_ERROR;
711 if (bp->b_rcred == NOCRED && cred != NOCRED)
712 bp->b_rcred = crhold(cred);
713 vfs_busy_pages(bp, 0);
714 bp->b_iooffset = dbtob(bp->b_blkno);
715 if (vp->v_type == VCHR)
716 VOP_SPECSTRATEGY(vp, bp);
717 else
718 VOP_STRATEGY(vp, bp);
719 ++readwait;
720 }
721
722 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
723 if (inmem(vp, *rablkno))
724 continue;
725 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
726
727 if ((rabp->b_flags & B_CACHE) == 0) {
728 if (curthread != PCPU_GET(idlethread))
729 curthread->td_proc->p_stats->p_ru.ru_inblock++;
730 rabp->b_flags |= B_ASYNC;
731 rabp->b_flags &= ~B_INVAL;
732 rabp->b_ioflags &= ~BIO_ERROR;
733 rabp->b_iocmd = BIO_READ;
734 if (rabp->b_rcred == NOCRED && cred != NOCRED)
735 rabp->b_rcred = crhold(cred);
736 vfs_busy_pages(rabp, 0);
737 BUF_KERNPROC(rabp);
738 rabp->b_iooffset = dbtob(rabp->b_blkno);
739 if (vp->v_type == VCHR)
740 VOP_SPECSTRATEGY(vp, rabp);
741 else
742 VOP_STRATEGY(vp, rabp);
743 } else {
744 brelse(rabp);
745 }
746 }
747
748 if (readwait) {
749 rv = bufwait(bp);
750 }
751 return (rv);
752}
753
754/*
755 * Write, release buffer on completion. (Done by iodone
756 * if async). Do not bother writing anything if the buffer
757 * is invalid.
758 *
759 * Note that we set B_CACHE here, indicating that buffer is
760 * fully valid and thus cacheable. This is true even of NFS
761 * now so we set it generally. This could be set either here
762 * or in biodone() since the I/O is synchronous. We put it
763 * here.
764 */
765int
766bwrite(struct buf * bp)
767{
768
769 KASSERT(bp->b_op != NULL && bp->b_op->bop_write != NULL,
770 ("Martian buffer %p in bwrite: nobody to write it.", bp));
771 return (bp->b_op->bop_write(bp));
772}
773
774static int
775ibwrite(struct buf * bp)
776{
777 int oldflags, s;
778 struct buf *newbp;
779
780 if (bp->b_flags & B_INVAL) {
781 brelse(bp);
782 return (0);
783 }
784
785 oldflags = bp->b_flags;
786
787 if (BUF_REFCNT(bp) == 0)
788 panic("ibwrite: buffer is not busy???");
789 s = splbio();
790 /*
791 * If a background write is already in progress, delay
792 * writing this block if it is asynchronous. Otherwise
793 * wait for the background write to complete.
794 */
795 VI_LOCK(bp->b_vp);
796 if (bp->b_vflags & BV_BKGRDINPROG) {
797 if (bp->b_flags & B_ASYNC) {
798 VI_UNLOCK(bp->b_vp);
799 splx(s);
800 bdwrite(bp);
801 return (0);
802 }
803 bp->b_vflags |= BV_BKGRDWAIT;
804 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0);
805 if (bp->b_vflags & BV_BKGRDINPROG)
806 panic("ibwrite: still writing");
807 }
808 VI_UNLOCK(bp->b_vp);
809
810 /* Mark the buffer clean */
811 bundirty(bp);
812
813 /*
814 * If this buffer is marked for background writing and we
815 * do not have to wait for it, make a copy and write the
816 * copy so as to leave this buffer ready for further use.
817 *
818 * This optimization eats a lot of memory. If we have a page
819 * or buffer shortfall we can't do it.
820 */
821 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) &&
822 (bp->b_flags & B_ASYNC) &&
823 !vm_page_count_severe() &&
824 !buf_dirty_count_severe()) {
825 if (bp->b_iodone != NULL) {
826 printf("bp->b_iodone = %p\n", bp->b_iodone);
827 panic("ibwrite: need chained iodone");
828 }
829
830 /* get a new block */
831 newbp = geteblk(bp->b_bufsize);
832
833 /*
834 * set it to be identical to the old block. We have to
835 * set b_lblkno and BKGRDMARKER before calling bgetvp()
836 * to avoid confusing the splay tree and gbincore().
837 */
838 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
839 newbp->b_lblkno = bp->b_lblkno;
840 newbp->b_xflags |= BX_BKGRDMARKER;
841 VI_LOCK(bp->b_vp);
842 bp->b_vflags |= BV_BKGRDINPROG;
843 bgetvp(bp->b_vp, newbp);
844 VI_UNLOCK(bp->b_vp);
845 newbp->b_blkno = bp->b_blkno;
846 newbp->b_offset = bp->b_offset;
847 newbp->b_iodone = vfs_backgroundwritedone;
848 newbp->b_flags |= B_ASYNC;
849 newbp->b_flags &= ~B_INVAL;
850
851 /* move over the dependencies */
852 if (LIST_FIRST(&bp->b_dep) != NULL)
853 buf_movedeps(bp, newbp);
854
855 /*
856 * Initiate write on the copy, release the original to
857 * the B_LOCKED queue so that it cannot go away until
858 * the background write completes. If not locked it could go
859 * away and then be reconstituted while it was being written.
860 * If the reconstituted buffer were written, we could end up
861 * with two background copies being written at the same time.
862 */
863 bqrelse(bp);
864 bp = newbp;
865 }
866
867 bp->b_flags &= ~B_DONE;
868 bp->b_ioflags &= ~BIO_ERROR;
869 bp->b_flags |= B_WRITEINPROG | B_CACHE;
870 bp->b_iocmd = BIO_WRITE;
871
872 VI_LOCK(bp->b_vp);
873 bp->b_vp->v_numoutput++;
874 VI_UNLOCK(bp->b_vp);
875 vfs_busy_pages(bp, 1);
876
877 /*
878 * Normal bwrites pipeline writes
879 */
880 bp->b_runningbufspace = bp->b_bufsize;
881 atomic_add_int(&runningbufspace, bp->b_runningbufspace);
882
883 if (curthread != PCPU_GET(idlethread))
884 curthread->td_proc->p_stats->p_ru.ru_oublock++;
885 splx(s);
886 if (oldflags & B_ASYNC)
887 BUF_KERNPROC(bp);
888 bp->b_iooffset = dbtob(bp->b_blkno);
889 if (bp->b_vp->v_type == VCHR) {
890 if (!buf_prewrite(bp->b_vp, bp))
891 VOP_SPECSTRATEGY(bp->b_vp, bp);
892 } else {
893 VOP_STRATEGY(bp->b_vp, bp);
894 }
895
896 if ((oldflags & B_ASYNC) == 0) {
897 int rtval = bufwait(bp);
898 brelse(bp);
899 return (rtval);
900 } else {
901 /*
902 * don't allow the async write to saturate the I/O
903 * system. We will not deadlock here because
904 * we are blocking waiting for I/O that is already in-progress
905 * to complete. We do not block here if it is the update
906 * or syncer daemon trying to clean up as that can lead
907 * to deadlock.
908 */
909 if (curthread->td_proc != bufdaemonproc &&
910 curthread->td_proc != updateproc)
911 waitrunningbufspace();
912 }
913
914 return (0);
915}
916
917/*
918 * Complete a background write started from bwrite.
919 */
920static void
921vfs_backgroundwritedone(bp)
922 struct buf *bp;
923{
924 struct buf *origbp;
925
926 /*
927 * Find the original buffer that we are writing.
928 */
929 VI_LOCK(bp->b_vp);
930 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
931 panic("backgroundwritedone: lost buffer");
932
933 /*
934 * Clear the BV_BKGRDINPROG flag in the original buffer
935 * and awaken it if it is waiting for the write to complete.
936 * If BV_BKGRDINPROG is not set in the original buffer it must
937 * have been released and re-instantiated - which is not legal.
938 */
939 KASSERT((origbp->b_vflags & BV_BKGRDINPROG),
940 ("backgroundwritedone: lost buffer2"));
941 origbp->b_vflags &= ~BV_BKGRDINPROG;
942 if (origbp->b_vflags & BV_BKGRDWAIT) {
943 origbp->b_vflags &= ~BV_BKGRDWAIT;
944 wakeup(&origbp->b_xflags);
945 }
946 VI_UNLOCK(bp->b_vp);
947 /*
948 * Process dependencies then return any unfinished ones.
949 */
950 if (LIST_FIRST(&bp->b_dep) != NULL)
951 buf_complete(bp);
952 if (LIST_FIRST(&bp->b_dep) != NULL)
953 buf_movedeps(bp, origbp);
954
955 /*
956 * This buffer is marked B_NOCACHE, so when it is released
957 * by biodone, it will be tossed. We mark it with BIO_READ
958 * to avoid biodone doing a second vwakeup.
959 */
960 bp->b_flags |= B_NOCACHE;
961 bp->b_iocmd = BIO_READ;
962 bp->b_flags &= ~(B_CACHE | B_DONE);
963 bp->b_iodone = 0;
964 bufdone(bp);
965}
966
967/*
968 * Delayed write. (Buffer is marked dirty). Do not bother writing
969 * anything if the buffer is marked invalid.
970 *
971 * Note that since the buffer must be completely valid, we can safely
972 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
973 * biodone() in order to prevent getblk from writing the buffer
974 * out synchronously.
975 */
976void
977bdwrite(struct buf * bp)
978{
979 struct thread *td = curthread;
980 struct vnode *vp;
981 struct buf *nbp;
982
983 GIANT_REQUIRED;
984
985 if (BUF_REFCNT(bp) == 0)
986 panic("bdwrite: buffer is not busy");
987
988 if (bp->b_flags & B_INVAL) {
989 brelse(bp);
990 return;
991 }
992
993 /*
994 * If we have too many dirty buffers, don't create any more.
995 * If we are wildly over our limit, then force a complete
996 * cleanup. Otherwise, just keep the situation from getting
997 * out of control. Note that we have to avoid a recursive
998 * disaster and not try to clean up after our own cleanup!
999 */
1000 vp = bp->b_vp;
1001 VI_LOCK(vp);
1002 if (td->td_pflags & TDP_COWINPROGRESS) {
1003 recursiveflushes++;
1004 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) {
1005 VI_UNLOCK(vp);
1006 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td);
1007 VI_LOCK(vp);
1008 altbufferflushes++;
1009 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) {
1010 /*
1011 * Try to find a buffer to flush.
1012 */
1013 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) {
1014 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1015 buf_countdeps(nbp, 0) ||
1016 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
1017 continue;
1018 if (bp == nbp)
1019 panic("bdwrite: found ourselves");
1020 VI_UNLOCK(vp);
1021 if (nbp->b_flags & B_CLUSTEROK) {
1022 vfs_bio_awrite(nbp);
1023 } else {
1024 bremfree(nbp);
1025 bawrite(nbp);
1026 }
1027 VI_LOCK(vp);
1028 dirtybufferflushes++;
1029 break;
1030 }
1031 }
1032 VI_UNLOCK(vp);
1033
1034 bdirty(bp);
1035 /*
1036 * Set B_CACHE, indicating that the buffer is fully valid. This is
1037 * true even of NFS now.
1038 */
1039 bp->b_flags |= B_CACHE;
1040
1041 /*
1042 * This bmap keeps the system from needing to do the bmap later,
1043 * perhaps when the system is attempting to do a sync. Since it
1044 * is likely that the indirect block -- or whatever other datastructure
1045 * that the filesystem needs is still in memory now, it is a good
1046 * thing to do this. Note also, that if the pageout daemon is
1047 * requesting a sync -- there might not be enough memory to do
1048 * the bmap then... So, this is important to do.
1049 */
1050 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1051 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1052 }
1053
1054 /*
1055 * Set the *dirty* buffer range based upon the VM system dirty pages.
1056 */
1057 vfs_setdirty(bp);
1058
1059 /*
1060 * We need to do this here to satisfy the vnode_pager and the
1061 * pageout daemon, so that it thinks that the pages have been
1062 * "cleaned". Note that since the pages are in a delayed write
1063 * buffer -- the VFS layer "will" see that the pages get written
1064 * out on the next sync, or perhaps the cluster will be completed.
1065 */
1066 vfs_clean_pages(bp);
1067 bqrelse(bp);
1068
1069 /*
1070 * Wakeup the buffer flushing daemon if we have a lot of dirty
1071 * buffers (midpoint between our recovery point and our stall
1072 * point).
1073 */
1074 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1075
1076 /*
1077 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1078 * due to the softdep code.
1079 */
1080}
1081
1082/*
1083 * bdirty:
1084 *
1085 * Turn buffer into delayed write request. We must clear BIO_READ and
1086 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1087 * itself to properly update it in the dirty/clean lists. We mark it
1088 * B_DONE to ensure that any asynchronization of the buffer properly
1089 * clears B_DONE ( else a panic will occur later ).
1090 *
1091 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1092 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1093 * should only be called if the buffer is known-good.
1094 *
1095 * Since the buffer is not on a queue, we do not update the numfreebuffers
1096 * count.
1097 *
1098 * Must be called at splbio().
1099 * The buffer must be on QUEUE_NONE.
1100 */
1101void
1102bdirty(bp)
1103 struct buf *bp;
1104{
1105 KASSERT(bp->b_qindex == QUEUE_NONE,
1106 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1107 bp->b_flags &= ~(B_RELBUF);
1108 bp->b_iocmd = BIO_WRITE;
1109
1110 if ((bp->b_flags & B_DELWRI) == 0) {
1111 bp->b_flags |= B_DONE | B_DELWRI;
1112 reassignbuf(bp, bp->b_vp);
1113 atomic_add_int(&numdirtybuffers, 1);
1114 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1115 }
1116}
1117
1118/*
1119 * bundirty:
1120 *
1121 * Clear B_DELWRI for buffer.
1122 *
1123 * Since the buffer is not on a queue, we do not update the numfreebuffers
1124 * count.
1125 *
1126 * Must be called at splbio().
1127 * The buffer must be on QUEUE_NONE.
1128 */
1129
1130void
1131bundirty(bp)
1132 struct buf *bp;
1133{
1134 KASSERT(bp->b_qindex == QUEUE_NONE,
1135 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1136
1137 if (bp->b_flags & B_DELWRI) {
1138 bp->b_flags &= ~B_DELWRI;
1139 reassignbuf(bp, bp->b_vp);
1140 atomic_subtract_int(&numdirtybuffers, 1);
1141 numdirtywakeup(lodirtybuffers);
1142 }
1143 /*
1144 * Since it is now being written, we can clear its deferred write flag.
1145 */
1146 bp->b_flags &= ~B_DEFERRED;
1147}
1148
1149/*
1150 * bawrite:
1151 *
1152 * Asynchronous write. Start output on a buffer, but do not wait for
1153 * it to complete. The buffer is released when the output completes.
1154 *
1155 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1156 * B_INVAL buffers. Not us.
1157 */
1158void
1159bawrite(struct buf * bp)
1160{
1161 bp->b_flags |= B_ASYNC;
1162 (void) bwrite(bp);
1163}
1164
1165/*
1166 * bwillwrite:
1167 *
1168 * Called prior to the locking of any vnodes when we are expecting to
1169 * write. We do not want to starve the buffer cache with too many
1170 * dirty buffers so we block here. By blocking prior to the locking
1171 * of any vnodes we attempt to avoid the situation where a locked vnode
1172 * prevents the various system daemons from flushing related buffers.
1173 */
1174
1175void
1176bwillwrite(void)
1177{
1178 if (numdirtybuffers >= hidirtybuffers) {
1179 int s;
1180
1181 mtx_lock(&Giant);
1182 s = splbio();
1183 mtx_lock(&nblock);
1184 while (numdirtybuffers >= hidirtybuffers) {
1185 bd_wakeup(1);
1186 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1187 msleep(&needsbuffer, &nblock,
1188 (PRIBIO + 4), "flswai", 0);
1189 }
1190 splx(s);
1191 mtx_unlock(&nblock);
1192 mtx_unlock(&Giant);
1193 }
1194}
1195
1196/*
1197 * Return true if we have too many dirty buffers.
1198 */
1199int
1200buf_dirty_count_severe(void)
1201{
1202 return(numdirtybuffers >= hidirtybuffers);
1203}
1204
1205/*
1206 * brelse:
1207 *
1208 * Release a busy buffer and, if requested, free its resources. The
1209 * buffer will be stashed in the appropriate bufqueue[] allowing it
1210 * to be accessed later as a cache entity or reused for other purposes.
1211 */
1212void
1213brelse(struct buf * bp)
1214{
1215 int s;
1216
1217 GIANT_REQUIRED;
1218
1219 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1220 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1221
1222 s = splbio();
1223
1224 if (bp->b_iocmd == BIO_WRITE &&
1225 (bp->b_ioflags & BIO_ERROR) &&
1226 !(bp->b_flags & B_INVAL)) {
1227 /*
1228 * Failed write, redirty. Must clear BIO_ERROR to prevent
1229 * pages from being scrapped. If B_INVAL is set then
1230 * this case is not run and the next case is run to
1231 * destroy the buffer. B_INVAL can occur if the buffer
1232 * is outside the range supported by the underlying device.
1233 */
1234 bp->b_ioflags &= ~BIO_ERROR;
1235 bdirty(bp);
1236 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1237 (bp->b_ioflags & BIO_ERROR) ||
1238 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) {
1239 /*
1240 * Either a failed I/O or we were asked to free or not
1241 * cache the buffer.
1242 */
1243 bp->b_flags |= B_INVAL;
1244 if (LIST_FIRST(&bp->b_dep) != NULL)
1245 buf_deallocate(bp);
1246 if (bp->b_flags & B_DELWRI) {
1247 atomic_subtract_int(&numdirtybuffers, 1);
1248 numdirtywakeup(lodirtybuffers);
1249 }
1250 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1251 if ((bp->b_flags & B_VMIO) == 0) {
1252 if (bp->b_bufsize)
1253 allocbuf(bp, 0);
1254 if (bp->b_vp)
1255 brelvp(bp);
1256 }
1257 }
1258
1259 /*
1260 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1261 * is called with B_DELWRI set, the underlying pages may wind up
1262 * getting freed causing a previous write (bdwrite()) to get 'lost'
1263 * because pages associated with a B_DELWRI bp are marked clean.
1264 *
1265 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1266 * if B_DELWRI is set.
1267 *
1268 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1269 * on pages to return pages to the VM page queues.
1270 */
1271 if (bp->b_flags & B_DELWRI)
1272 bp->b_flags &= ~B_RELBUF;
1273 else if (vm_page_count_severe()) {
1274 /*
1275 * XXX This lock may not be necessary since BKGRDINPROG
1276 * cannot be set while we hold the buf lock, it can only be
1277 * cleared if it is already pending.
1278 */
1279 if (bp->b_vp) {
1280 VI_LOCK(bp->b_vp);
1281 if (!(bp->b_vflags & BV_BKGRDINPROG))
1282 bp->b_flags |= B_RELBUF;
1283 VI_UNLOCK(bp->b_vp);
1284 } else
1285 bp->b_flags |= B_RELBUF;
1286 }
1287
1288 /*
1289 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1290 * constituted, not even NFS buffers now. Two flags effect this. If
1291 * B_INVAL, the struct buf is invalidated but the VM object is kept
1292 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1293 *
1294 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1295 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1296 * buffer is also B_INVAL because it hits the re-dirtying code above.
1297 *
1298 * Normally we can do this whether a buffer is B_DELWRI or not. If
1299 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1300 * the commit state and we cannot afford to lose the buffer. If the
1301 * buffer has a background write in progress, we need to keep it
1302 * around to prevent it from being reconstituted and starting a second
1303 * background write.
1304 */
1305 if ((bp->b_flags & B_VMIO)
1306 && !(bp->b_vp->v_mount != NULL &&
1307 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1308 !vn_isdisk(bp->b_vp, NULL) &&
1309 (bp->b_flags & B_DELWRI))
1310 ) {
1311
1312 int i, j, resid;
1313 vm_page_t m;
1314 off_t foff;
1315 vm_pindex_t poff;
1316 vm_object_t obj;
1317
1318 obj = bp->b_object;
1319
1320 /*
1321 * Get the base offset and length of the buffer. Note that
1322 * in the VMIO case if the buffer block size is not
1323 * page-aligned then b_data pointer may not be page-aligned.
1324 * But our b_pages[] array *IS* page aligned.
1325 *
1326 * block sizes less then DEV_BSIZE (usually 512) are not
1327 * supported due to the page granularity bits (m->valid,
1328 * m->dirty, etc...).
1329 *
1330 * See man buf(9) for more information
1331 */
1332 resid = bp->b_bufsize;
1333 foff = bp->b_offset;
1334 VM_OBJECT_LOCK(obj);
1335 for (i = 0; i < bp->b_npages; i++) {
1336 int had_bogus = 0;
1337
1338 m = bp->b_pages[i];
1339
1340 /*
1341 * If we hit a bogus page, fixup *all* the bogus pages
1342 * now.
1343 */
1344 if (m == bogus_page) {
1345 poff = OFF_TO_IDX(bp->b_offset);
1346 had_bogus = 1;
1347
1348 for (j = i; j < bp->b_npages; j++) {
1349 vm_page_t mtmp;
1350 mtmp = bp->b_pages[j];
1351 if (mtmp == bogus_page) {
1352 mtmp = vm_page_lookup(obj, poff + j);
1353 if (!mtmp) {
1354 panic("brelse: page missing\n");
1355 }
1356 bp->b_pages[j] = mtmp;
1357 }
1358 }
1359
1360 if ((bp->b_flags & B_INVAL) == 0) {
1361 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1362 }
1363 m = bp->b_pages[i];
1364 }
1365 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) {
1366 int poffset = foff & PAGE_MASK;
1367 int presid = resid > (PAGE_SIZE - poffset) ?
1368 (PAGE_SIZE - poffset) : resid;
1369
1370 KASSERT(presid >= 0, ("brelse: extra page"));
1371 vm_page_lock_queues();
1372 vm_page_set_invalid(m, poffset, presid);
1373 vm_page_unlock_queues();
1374 if (had_bogus)
1375 printf("avoided corruption bug in bogus_page/brelse code\n");
1376 }
1377 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1378 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1379 }
1380 VM_OBJECT_UNLOCK(obj);
1381 if (bp->b_flags & (B_INVAL | B_RELBUF))
1382 vfs_vmio_release(bp);
1383
1384 } else if (bp->b_flags & B_VMIO) {
1385
1386 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1387 vfs_vmio_release(bp);
1388 }
1389
1390 }
1391
1392 if (bp->b_qindex != QUEUE_NONE)
1393 panic("brelse: free buffer onto another queue???");
1394 if (BUF_REFCNT(bp) > 1) {
1395 /* do not release to free list */
1396 BUF_UNLOCK(bp);
1397 splx(s);
1398 return;
1399 }
1400
1401 /* enqueue */
1402 mtx_lock(&bqlock);
1403
1404 /* buffers with no memory */
1405 if (bp->b_bufsize == 0) {
1406 bp->b_flags |= B_INVAL;
1407 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1408 if (bp->b_vflags & BV_BKGRDINPROG)
1409 panic("losing buffer 1");
1410 if (bp->b_kvasize) {
1411 bp->b_qindex = QUEUE_EMPTYKVA;
1412 } else {
1413 bp->b_qindex = QUEUE_EMPTY;
1414 }
1415 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
|
1416 bp->b_dev = NODEV;
| 1416 bp->b_dev = NULL;
|
1417 /* buffers with junk contents */
1418 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1419 (bp->b_ioflags & BIO_ERROR)) {
1420 bp->b_flags |= B_INVAL;
1421 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1422 if (bp->b_vflags & BV_BKGRDINPROG)
1423 panic("losing buffer 2");
1424 bp->b_qindex = QUEUE_CLEAN;
1425 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
| 1417 /* buffers with junk contents */
1418 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1419 (bp->b_ioflags & BIO_ERROR)) {
1420 bp->b_flags |= B_INVAL;
1421 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1422 if (bp->b_vflags & BV_BKGRDINPROG)
1423 panic("losing buffer 2");
1424 bp->b_qindex = QUEUE_CLEAN;
1425 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
|
1426 bp->b_dev = NODEV;
| 1426 bp->b_dev = NULL;
|
1427 /* remaining buffers */
1428 } else {
1429 if (bp->b_flags & B_DELWRI)
1430 bp->b_qindex = QUEUE_DIRTY;
1431 else
1432 bp->b_qindex = QUEUE_CLEAN;
1433 if (bp->b_flags & B_AGE)
1434 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1435 else
1436 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1437 }
1438 mtx_unlock(&bqlock);
1439
1440 /*
1441 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already
1442 * placed the buffer on the correct queue. We must also disassociate
1443 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1444 * find it.
1445 */
1446 if (bp->b_flags & B_INVAL) {
1447 if (bp->b_flags & B_DELWRI)
1448 bundirty(bp);
1449 if (bp->b_vp)
1450 brelvp(bp);
1451 }
1452
1453 /*
1454 * Fixup numfreebuffers count. The bp is on an appropriate queue
1455 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1456 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1457 * if B_INVAL is set ).
1458 */
1459
1460 if (!(bp->b_flags & B_DELWRI))
1461 bufcountwakeup();
1462
1463 /*
1464 * Something we can maybe free or reuse
1465 */
1466 if (bp->b_bufsize || bp->b_kvasize)
1467 bufspacewakeup();
1468
1469 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1470 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1471 panic("brelse: not dirty");
1472 /* unlock */
1473 BUF_UNLOCK(bp);
1474 splx(s);
1475}
1476
1477/*
1478 * Release a buffer back to the appropriate queue but do not try to free
1479 * it. The buffer is expected to be used again soon.
1480 *
1481 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1482 * biodone() to requeue an async I/O on completion. It is also used when
1483 * known good buffers need to be requeued but we think we may need the data
1484 * again soon.
1485 *
1486 * XXX we should be able to leave the B_RELBUF hint set on completion.
1487 */
1488void
1489bqrelse(struct buf * bp)
1490{
1491 int s;
1492
1493 s = splbio();
1494
1495 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1496
1497 if (bp->b_qindex != QUEUE_NONE)
1498 panic("bqrelse: free buffer onto another queue???");
1499 if (BUF_REFCNT(bp) > 1) {
1500 /* do not release to free list */
1501 BUF_UNLOCK(bp);
1502 splx(s);
1503 return;
1504 }
1505 mtx_lock(&bqlock);
1506 /* buffers with stale but valid contents */
1507 if (bp->b_flags & B_DELWRI) {
1508 bp->b_qindex = QUEUE_DIRTY;
1509 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1510 } else {
1511 /*
1512 * XXX This lock may not be necessary since BKGRDINPROG
1513 * cannot be set while we hold the buf lock, it can only be
1514 * cleared if it is already pending.
1515 */
1516 VI_LOCK(bp->b_vp);
1517 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) {
1518 VI_UNLOCK(bp->b_vp);
1519 bp->b_qindex = QUEUE_CLEAN;
1520 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1521 b_freelist);
1522 } else {
1523 /*
1524 * We are too low on memory, we have to try to free
1525 * the buffer (most importantly: the wired pages
1526 * making up its backing store) *now*.
1527 */
1528 VI_UNLOCK(bp->b_vp);
1529 mtx_unlock(&bqlock);
1530 splx(s);
1531 brelse(bp);
1532 return;
1533 }
1534 }
1535 mtx_unlock(&bqlock);
1536
1537 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1538 bufcountwakeup();
1539
1540 /*
1541 * Something we can maybe free or reuse.
1542 */
1543 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1544 bufspacewakeup();
1545
1546 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1547 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1548 panic("bqrelse: not dirty");
1549 /* unlock */
1550 BUF_UNLOCK(bp);
1551 splx(s);
1552}
1553
1554/* Give pages used by the bp back to the VM system (where possible) */
1555static void
1556vfs_vmio_release(bp)
1557 struct buf *bp;
1558{
1559 int i;
1560 vm_page_t m;
1561
1562 GIANT_REQUIRED;
1563 VM_OBJECT_LOCK(bp->b_object);
1564 vm_page_lock_queues();
1565 for (i = 0; i < bp->b_npages; i++) {
1566 m = bp->b_pages[i];
1567 bp->b_pages[i] = NULL;
1568 /*
1569 * In order to keep page LRU ordering consistent, put
1570 * everything on the inactive queue.
1571 */
1572 vm_page_unwire(m, 0);
1573 /*
1574 * We don't mess with busy pages, it is
1575 * the responsibility of the process that
1576 * busied the pages to deal with them.
1577 */
1578 if ((m->flags & PG_BUSY) || (m->busy != 0))
1579 continue;
1580
1581 if (m->wire_count == 0) {
1582 /*
1583 * Might as well free the page if we can and it has
1584 * no valid data. We also free the page if the
1585 * buffer was used for direct I/O
1586 */
1587 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1588 m->hold_count == 0) {
1589 vm_page_busy(m);
1590 pmap_remove_all(m);
1591 vm_page_free(m);
1592 } else if (bp->b_flags & B_DIRECT) {
1593 vm_page_try_to_free(m);
1594 } else if (vm_page_count_severe()) {
1595 vm_page_try_to_cache(m);
1596 }
1597 }
1598 }
1599 vm_page_unlock_queues();
1600 VM_OBJECT_UNLOCK(bp->b_object);
1601 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1602
1603 if (bp->b_bufsize) {
1604 bufspacewakeup();
1605 bp->b_bufsize = 0;
1606 }
1607 bp->b_npages = 0;
1608 bp->b_flags &= ~B_VMIO;
1609 if (bp->b_vp)
1610 brelvp(bp);
1611}
1612
1613/*
1614 * Check to see if a block at a particular lbn is available for a clustered
1615 * write.
1616 */
1617static int
1618vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1619{
1620 struct buf *bpa;
1621 int match;
1622
1623 match = 0;
1624
1625 /* If the buf isn't in core skip it */
1626 if ((bpa = gbincore(vp, lblkno)) == NULL)
1627 return (0);
1628
1629 /* If the buf is busy we don't want to wait for it */
1630 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1631 return (0);
1632
1633 /* Only cluster with valid clusterable delayed write buffers */
1634 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1635 (B_DELWRI | B_CLUSTEROK))
1636 goto done;
1637
1638 if (bpa->b_bufsize != size)
1639 goto done;
1640
1641 /*
1642 * Check to see if it is in the expected place on disk and that the
1643 * block has been mapped.
1644 */
1645 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1646 match = 1;
1647done:
1648 BUF_UNLOCK(bpa);
1649 return (match);
1650}
1651
1652/*
1653 * vfs_bio_awrite:
1654 *
1655 * Implement clustered async writes for clearing out B_DELWRI buffers.
1656 * This is much better then the old way of writing only one buffer at
1657 * a time. Note that we may not be presented with the buffers in the
1658 * correct order, so we search for the cluster in both directions.
1659 */
1660int
1661vfs_bio_awrite(struct buf * bp)
1662{
1663 int i;
1664 int j;
1665 daddr_t lblkno = bp->b_lblkno;
1666 struct vnode *vp = bp->b_vp;
1667 int s;
1668 int ncl;
1669 int nwritten;
1670 int size;
1671 int maxcl;
1672
1673 s = splbio();
1674 /*
1675 * right now we support clustered writing only to regular files. If
1676 * we find a clusterable block we could be in the middle of a cluster
1677 * rather then at the beginning.
1678 */
1679 if ((vp->v_type == VREG) &&
1680 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1681 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1682
1683 size = vp->v_mount->mnt_stat.f_iosize;
1684 maxcl = MAXPHYS / size;
1685
1686 VI_LOCK(vp);
1687 for (i = 1; i < maxcl; i++)
1688 if (vfs_bio_clcheck(vp, size, lblkno + i,
1689 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1690 break;
1691
1692 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1693 if (vfs_bio_clcheck(vp, size, lblkno - j,
1694 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1695 break;
1696
1697 VI_UNLOCK(vp);
1698 --j;
1699 ncl = i + j;
1700 /*
1701 * this is a possible cluster write
1702 */
1703 if (ncl != 1) {
1704 BUF_UNLOCK(bp);
1705 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1706 splx(s);
1707 return nwritten;
1708 }
1709 }
1710
1711 bremfree(bp);
1712 bp->b_flags |= B_ASYNC;
1713
1714 splx(s);
1715 /*
1716 * default (old) behavior, writing out only one block
1717 *
1718 * XXX returns b_bufsize instead of b_bcount for nwritten?
1719 */
1720 nwritten = bp->b_bufsize;
1721 (void) bwrite(bp);
1722
1723 return nwritten;
1724}
1725
1726/*
1727 * getnewbuf:
1728 *
1729 * Find and initialize a new buffer header, freeing up existing buffers
1730 * in the bufqueues as necessary. The new buffer is returned locked.
1731 *
1732 * Important: B_INVAL is not set. If the caller wishes to throw the
1733 * buffer away, the caller must set B_INVAL prior to calling brelse().
1734 *
1735 * We block if:
1736 * We have insufficient buffer headers
1737 * We have insufficient buffer space
1738 * buffer_map is too fragmented ( space reservation fails )
1739 * If we have to flush dirty buffers ( but we try to avoid this )
1740 *
1741 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1742 * Instead we ask the buf daemon to do it for us. We attempt to
1743 * avoid piecemeal wakeups of the pageout daemon.
1744 */
1745
1746static struct buf *
1747getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1748{
1749 struct buf *bp;
1750 struct buf *nbp;
1751 int defrag = 0;
1752 int nqindex;
1753 static int flushingbufs;
1754
1755 GIANT_REQUIRED;
1756
1757 /*
1758 * We can't afford to block since we might be holding a vnode lock,
1759 * which may prevent system daemons from running. We deal with
1760 * low-memory situations by proactively returning memory and running
1761 * async I/O rather then sync I/O.
1762 */
1763
1764 atomic_add_int(&getnewbufcalls, 1);
1765 atomic_subtract_int(&getnewbufrestarts, 1);
1766restart:
1767 atomic_add_int(&getnewbufrestarts, 1);
1768
1769 /*
1770 * Setup for scan. If we do not have enough free buffers,
1771 * we setup a degenerate case that immediately fails. Note
1772 * that if we are specially marked process, we are allowed to
1773 * dip into our reserves.
1774 *
1775 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1776 *
1777 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1778 * However, there are a number of cases (defragging, reusing, ...)
1779 * where we cannot backup.
1780 */
1781 mtx_lock(&bqlock);
1782 nqindex = QUEUE_EMPTYKVA;
1783 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1784
1785 if (nbp == NULL) {
1786 /*
1787 * If no EMPTYKVA buffers and we are either
1788 * defragging or reusing, locate a CLEAN buffer
1789 * to free or reuse. If bufspace useage is low
1790 * skip this step so we can allocate a new buffer.
1791 */
1792 if (defrag || bufspace >= lobufspace) {
1793 nqindex = QUEUE_CLEAN;
1794 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1795 }
1796
1797 /*
1798 * If we could not find or were not allowed to reuse a
1799 * CLEAN buffer, check to see if it is ok to use an EMPTY
1800 * buffer. We can only use an EMPTY buffer if allocating
1801 * its KVA would not otherwise run us out of buffer space.
1802 */
1803 if (nbp == NULL && defrag == 0 &&
1804 bufspace + maxsize < hibufspace) {
1805 nqindex = QUEUE_EMPTY;
1806 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1807 }
1808 }
1809
1810 /*
1811 * Run scan, possibly freeing data and/or kva mappings on the fly
1812 * depending.
1813 */
1814
1815 while ((bp = nbp) != NULL) {
1816 int qindex = nqindex;
1817
1818 /*
1819 * Calculate next bp ( we can only use it if we do not block
1820 * or do other fancy things ).
1821 */
1822 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1823 switch(qindex) {
1824 case QUEUE_EMPTY:
1825 nqindex = QUEUE_EMPTYKVA;
1826 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1827 break;
1828 /* FALLTHROUGH */
1829 case QUEUE_EMPTYKVA:
1830 nqindex = QUEUE_CLEAN;
1831 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1832 break;
1833 /* FALLTHROUGH */
1834 case QUEUE_CLEAN:
1835 /*
1836 * nbp is NULL.
1837 */
1838 break;
1839 }
1840 }
1841 if (bp->b_vp) {
1842 VI_LOCK(bp->b_vp);
1843 if (bp->b_vflags & BV_BKGRDINPROG) {
1844 VI_UNLOCK(bp->b_vp);
1845 continue;
1846 }
1847 VI_UNLOCK(bp->b_vp);
1848 }
1849
1850 /*
1851 * Sanity Checks
1852 */
1853 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1854
1855 /*
1856 * Note: we no longer distinguish between VMIO and non-VMIO
1857 * buffers.
1858 */
1859
1860 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1861
1862 /*
1863 * If we are defragging then we need a buffer with
1864 * b_kvasize != 0. XXX this situation should no longer
1865 * occur, if defrag is non-zero the buffer's b_kvasize
1866 * should also be non-zero at this point. XXX
1867 */
1868 if (defrag && bp->b_kvasize == 0) {
1869 printf("Warning: defrag empty buffer %p\n", bp);
1870 continue;
1871 }
1872
1873 /*
1874 * Start freeing the bp. This is somewhat involved. nbp
1875 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1876 */
1877
1878 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1879 panic("getnewbuf: locked buf");
1880 bremfreel(bp);
1881 mtx_unlock(&bqlock);
1882
1883 if (qindex == QUEUE_CLEAN) {
1884 if (bp->b_flags & B_VMIO) {
1885 bp->b_flags &= ~B_ASYNC;
1886 vfs_vmio_release(bp);
1887 }
1888 if (bp->b_vp)
1889 brelvp(bp);
1890 }
1891
1892 /*
1893 * NOTE: nbp is now entirely invalid. We can only restart
1894 * the scan from this point on.
1895 *
1896 * Get the rest of the buffer freed up. b_kva* is still
1897 * valid after this operation.
1898 */
1899
1900 if (bp->b_rcred != NOCRED) {
1901 crfree(bp->b_rcred);
1902 bp->b_rcred = NOCRED;
1903 }
1904 if (bp->b_wcred != NOCRED) {
1905 crfree(bp->b_wcred);
1906 bp->b_wcred = NOCRED;
1907 }
1908 if (LIST_FIRST(&bp->b_dep) != NULL)
1909 buf_deallocate(bp);
1910 if (bp->b_vflags & BV_BKGRDINPROG)
1911 panic("losing buffer 3");
1912
1913 if (bp->b_bufsize)
1914 allocbuf(bp, 0);
1915
1916 bp->b_flags = 0;
1917 bp->b_ioflags = 0;
1918 bp->b_xflags = 0;
1919 bp->b_vflags = 0;
| 1427 /* remaining buffers */
1428 } else {
1429 if (bp->b_flags & B_DELWRI)
1430 bp->b_qindex = QUEUE_DIRTY;
1431 else
1432 bp->b_qindex = QUEUE_CLEAN;
1433 if (bp->b_flags & B_AGE)
1434 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1435 else
1436 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1437 }
1438 mtx_unlock(&bqlock);
1439
1440 /*
1441 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already
1442 * placed the buffer on the correct queue. We must also disassociate
1443 * the device and vnode for a B_INVAL buffer so gbincore() doesn't
1444 * find it.
1445 */
1446 if (bp->b_flags & B_INVAL) {
1447 if (bp->b_flags & B_DELWRI)
1448 bundirty(bp);
1449 if (bp->b_vp)
1450 brelvp(bp);
1451 }
1452
1453 /*
1454 * Fixup numfreebuffers count. The bp is on an appropriate queue
1455 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1456 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1457 * if B_INVAL is set ).
1458 */
1459
1460 if (!(bp->b_flags & B_DELWRI))
1461 bufcountwakeup();
1462
1463 /*
1464 * Something we can maybe free or reuse
1465 */
1466 if (bp->b_bufsize || bp->b_kvasize)
1467 bufspacewakeup();
1468
1469 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1470 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1471 panic("brelse: not dirty");
1472 /* unlock */
1473 BUF_UNLOCK(bp);
1474 splx(s);
1475}
1476
1477/*
1478 * Release a buffer back to the appropriate queue but do not try to free
1479 * it. The buffer is expected to be used again soon.
1480 *
1481 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1482 * biodone() to requeue an async I/O on completion. It is also used when
1483 * known good buffers need to be requeued but we think we may need the data
1484 * again soon.
1485 *
1486 * XXX we should be able to leave the B_RELBUF hint set on completion.
1487 */
1488void
1489bqrelse(struct buf * bp)
1490{
1491 int s;
1492
1493 s = splbio();
1494
1495 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1496
1497 if (bp->b_qindex != QUEUE_NONE)
1498 panic("bqrelse: free buffer onto another queue???");
1499 if (BUF_REFCNT(bp) > 1) {
1500 /* do not release to free list */
1501 BUF_UNLOCK(bp);
1502 splx(s);
1503 return;
1504 }
1505 mtx_lock(&bqlock);
1506 /* buffers with stale but valid contents */
1507 if (bp->b_flags & B_DELWRI) {
1508 bp->b_qindex = QUEUE_DIRTY;
1509 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1510 } else {
1511 /*
1512 * XXX This lock may not be necessary since BKGRDINPROG
1513 * cannot be set while we hold the buf lock, it can only be
1514 * cleared if it is already pending.
1515 */
1516 VI_LOCK(bp->b_vp);
1517 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) {
1518 VI_UNLOCK(bp->b_vp);
1519 bp->b_qindex = QUEUE_CLEAN;
1520 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1521 b_freelist);
1522 } else {
1523 /*
1524 * We are too low on memory, we have to try to free
1525 * the buffer (most importantly: the wired pages
1526 * making up its backing store) *now*.
1527 */
1528 VI_UNLOCK(bp->b_vp);
1529 mtx_unlock(&bqlock);
1530 splx(s);
1531 brelse(bp);
1532 return;
1533 }
1534 }
1535 mtx_unlock(&bqlock);
1536
1537 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1538 bufcountwakeup();
1539
1540 /*
1541 * Something we can maybe free or reuse.
1542 */
1543 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1544 bufspacewakeup();
1545
1546 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1547 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1548 panic("bqrelse: not dirty");
1549 /* unlock */
1550 BUF_UNLOCK(bp);
1551 splx(s);
1552}
1553
1554/* Give pages used by the bp back to the VM system (where possible) */
1555static void
1556vfs_vmio_release(bp)
1557 struct buf *bp;
1558{
1559 int i;
1560 vm_page_t m;
1561
1562 GIANT_REQUIRED;
1563 VM_OBJECT_LOCK(bp->b_object);
1564 vm_page_lock_queues();
1565 for (i = 0; i < bp->b_npages; i++) {
1566 m = bp->b_pages[i];
1567 bp->b_pages[i] = NULL;
1568 /*
1569 * In order to keep page LRU ordering consistent, put
1570 * everything on the inactive queue.
1571 */
1572 vm_page_unwire(m, 0);
1573 /*
1574 * We don't mess with busy pages, it is
1575 * the responsibility of the process that
1576 * busied the pages to deal with them.
1577 */
1578 if ((m->flags & PG_BUSY) || (m->busy != 0))
1579 continue;
1580
1581 if (m->wire_count == 0) {
1582 /*
1583 * Might as well free the page if we can and it has
1584 * no valid data. We also free the page if the
1585 * buffer was used for direct I/O
1586 */
1587 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1588 m->hold_count == 0) {
1589 vm_page_busy(m);
1590 pmap_remove_all(m);
1591 vm_page_free(m);
1592 } else if (bp->b_flags & B_DIRECT) {
1593 vm_page_try_to_free(m);
1594 } else if (vm_page_count_severe()) {
1595 vm_page_try_to_cache(m);
1596 }
1597 }
1598 }
1599 vm_page_unlock_queues();
1600 VM_OBJECT_UNLOCK(bp->b_object);
1601 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1602
1603 if (bp->b_bufsize) {
1604 bufspacewakeup();
1605 bp->b_bufsize = 0;
1606 }
1607 bp->b_npages = 0;
1608 bp->b_flags &= ~B_VMIO;
1609 if (bp->b_vp)
1610 brelvp(bp);
1611}
1612
1613/*
1614 * Check to see if a block at a particular lbn is available for a clustered
1615 * write.
1616 */
1617static int
1618vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1619{
1620 struct buf *bpa;
1621 int match;
1622
1623 match = 0;
1624
1625 /* If the buf isn't in core skip it */
1626 if ((bpa = gbincore(vp, lblkno)) == NULL)
1627 return (0);
1628
1629 /* If the buf is busy we don't want to wait for it */
1630 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1631 return (0);
1632
1633 /* Only cluster with valid clusterable delayed write buffers */
1634 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1635 (B_DELWRI | B_CLUSTEROK))
1636 goto done;
1637
1638 if (bpa->b_bufsize != size)
1639 goto done;
1640
1641 /*
1642 * Check to see if it is in the expected place on disk and that the
1643 * block has been mapped.
1644 */
1645 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1646 match = 1;
1647done:
1648 BUF_UNLOCK(bpa);
1649 return (match);
1650}
1651
1652/*
1653 * vfs_bio_awrite:
1654 *
1655 * Implement clustered async writes for clearing out B_DELWRI buffers.
1656 * This is much better then the old way of writing only one buffer at
1657 * a time. Note that we may not be presented with the buffers in the
1658 * correct order, so we search for the cluster in both directions.
1659 */
1660int
1661vfs_bio_awrite(struct buf * bp)
1662{
1663 int i;
1664 int j;
1665 daddr_t lblkno = bp->b_lblkno;
1666 struct vnode *vp = bp->b_vp;
1667 int s;
1668 int ncl;
1669 int nwritten;
1670 int size;
1671 int maxcl;
1672
1673 s = splbio();
1674 /*
1675 * right now we support clustered writing only to regular files. If
1676 * we find a clusterable block we could be in the middle of a cluster
1677 * rather then at the beginning.
1678 */
1679 if ((vp->v_type == VREG) &&
1680 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1681 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1682
1683 size = vp->v_mount->mnt_stat.f_iosize;
1684 maxcl = MAXPHYS / size;
1685
1686 VI_LOCK(vp);
1687 for (i = 1; i < maxcl; i++)
1688 if (vfs_bio_clcheck(vp, size, lblkno + i,
1689 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1690 break;
1691
1692 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1693 if (vfs_bio_clcheck(vp, size, lblkno - j,
1694 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1695 break;
1696
1697 VI_UNLOCK(vp);
1698 --j;
1699 ncl = i + j;
1700 /*
1701 * this is a possible cluster write
1702 */
1703 if (ncl != 1) {
1704 BUF_UNLOCK(bp);
1705 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1706 splx(s);
1707 return nwritten;
1708 }
1709 }
1710
1711 bremfree(bp);
1712 bp->b_flags |= B_ASYNC;
1713
1714 splx(s);
1715 /*
1716 * default (old) behavior, writing out only one block
1717 *
1718 * XXX returns b_bufsize instead of b_bcount for nwritten?
1719 */
1720 nwritten = bp->b_bufsize;
1721 (void) bwrite(bp);
1722
1723 return nwritten;
1724}
1725
1726/*
1727 * getnewbuf:
1728 *
1729 * Find and initialize a new buffer header, freeing up existing buffers
1730 * in the bufqueues as necessary. The new buffer is returned locked.
1731 *
1732 * Important: B_INVAL is not set. If the caller wishes to throw the
1733 * buffer away, the caller must set B_INVAL prior to calling brelse().
1734 *
1735 * We block if:
1736 * We have insufficient buffer headers
1737 * We have insufficient buffer space
1738 * buffer_map is too fragmented ( space reservation fails )
1739 * If we have to flush dirty buffers ( but we try to avoid this )
1740 *
1741 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1742 * Instead we ask the buf daemon to do it for us. We attempt to
1743 * avoid piecemeal wakeups of the pageout daemon.
1744 */
1745
1746static struct buf *
1747getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1748{
1749 struct buf *bp;
1750 struct buf *nbp;
1751 int defrag = 0;
1752 int nqindex;
1753 static int flushingbufs;
1754
1755 GIANT_REQUIRED;
1756
1757 /*
1758 * We can't afford to block since we might be holding a vnode lock,
1759 * which may prevent system daemons from running. We deal with
1760 * low-memory situations by proactively returning memory and running
1761 * async I/O rather then sync I/O.
1762 */
1763
1764 atomic_add_int(&getnewbufcalls, 1);
1765 atomic_subtract_int(&getnewbufrestarts, 1);
1766restart:
1767 atomic_add_int(&getnewbufrestarts, 1);
1768
1769 /*
1770 * Setup for scan. If we do not have enough free buffers,
1771 * we setup a degenerate case that immediately fails. Note
1772 * that if we are specially marked process, we are allowed to
1773 * dip into our reserves.
1774 *
1775 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1776 *
1777 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1778 * However, there are a number of cases (defragging, reusing, ...)
1779 * where we cannot backup.
1780 */
1781 mtx_lock(&bqlock);
1782 nqindex = QUEUE_EMPTYKVA;
1783 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1784
1785 if (nbp == NULL) {
1786 /*
1787 * If no EMPTYKVA buffers and we are either
1788 * defragging or reusing, locate a CLEAN buffer
1789 * to free or reuse. If bufspace useage is low
1790 * skip this step so we can allocate a new buffer.
1791 */
1792 if (defrag || bufspace >= lobufspace) {
1793 nqindex = QUEUE_CLEAN;
1794 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1795 }
1796
1797 /*
1798 * If we could not find or were not allowed to reuse a
1799 * CLEAN buffer, check to see if it is ok to use an EMPTY
1800 * buffer. We can only use an EMPTY buffer if allocating
1801 * its KVA would not otherwise run us out of buffer space.
1802 */
1803 if (nbp == NULL && defrag == 0 &&
1804 bufspace + maxsize < hibufspace) {
1805 nqindex = QUEUE_EMPTY;
1806 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1807 }
1808 }
1809
1810 /*
1811 * Run scan, possibly freeing data and/or kva mappings on the fly
1812 * depending.
1813 */
1814
1815 while ((bp = nbp) != NULL) {
1816 int qindex = nqindex;
1817
1818 /*
1819 * Calculate next bp ( we can only use it if we do not block
1820 * or do other fancy things ).
1821 */
1822 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1823 switch(qindex) {
1824 case QUEUE_EMPTY:
1825 nqindex = QUEUE_EMPTYKVA;
1826 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1827 break;
1828 /* FALLTHROUGH */
1829 case QUEUE_EMPTYKVA:
1830 nqindex = QUEUE_CLEAN;
1831 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1832 break;
1833 /* FALLTHROUGH */
1834 case QUEUE_CLEAN:
1835 /*
1836 * nbp is NULL.
1837 */
1838 break;
1839 }
1840 }
1841 if (bp->b_vp) {
1842 VI_LOCK(bp->b_vp);
1843 if (bp->b_vflags & BV_BKGRDINPROG) {
1844 VI_UNLOCK(bp->b_vp);
1845 continue;
1846 }
1847 VI_UNLOCK(bp->b_vp);
1848 }
1849
1850 /*
1851 * Sanity Checks
1852 */
1853 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1854
1855 /*
1856 * Note: we no longer distinguish between VMIO and non-VMIO
1857 * buffers.
1858 */
1859
1860 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1861
1862 /*
1863 * If we are defragging then we need a buffer with
1864 * b_kvasize != 0. XXX this situation should no longer
1865 * occur, if defrag is non-zero the buffer's b_kvasize
1866 * should also be non-zero at this point. XXX
1867 */
1868 if (defrag && bp->b_kvasize == 0) {
1869 printf("Warning: defrag empty buffer %p\n", bp);
1870 continue;
1871 }
1872
1873 /*
1874 * Start freeing the bp. This is somewhat involved. nbp
1875 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1876 */
1877
1878 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1879 panic("getnewbuf: locked buf");
1880 bremfreel(bp);
1881 mtx_unlock(&bqlock);
1882
1883 if (qindex == QUEUE_CLEAN) {
1884 if (bp->b_flags & B_VMIO) {
1885 bp->b_flags &= ~B_ASYNC;
1886 vfs_vmio_release(bp);
1887 }
1888 if (bp->b_vp)
1889 brelvp(bp);
1890 }
1891
1892 /*
1893 * NOTE: nbp is now entirely invalid. We can only restart
1894 * the scan from this point on.
1895 *
1896 * Get the rest of the buffer freed up. b_kva* is still
1897 * valid after this operation.
1898 */
1899
1900 if (bp->b_rcred != NOCRED) {
1901 crfree(bp->b_rcred);
1902 bp->b_rcred = NOCRED;
1903 }
1904 if (bp->b_wcred != NOCRED) {
1905 crfree(bp->b_wcred);
1906 bp->b_wcred = NOCRED;
1907 }
1908 if (LIST_FIRST(&bp->b_dep) != NULL)
1909 buf_deallocate(bp);
1910 if (bp->b_vflags & BV_BKGRDINPROG)
1911 panic("losing buffer 3");
1912
1913 if (bp->b_bufsize)
1914 allocbuf(bp, 0);
1915
1916 bp->b_flags = 0;
1917 bp->b_ioflags = 0;
1918 bp->b_xflags = 0;
1919 bp->b_vflags = 0;
|
1920 bp->b_dev = NODEV;
| 1920 bp->b_dev = NULL;
|
1921 bp->b_vp = NULL;
1922 bp->b_blkno = bp->b_lblkno = 0;
1923 bp->b_offset = NOOFFSET;
1924 bp->b_iodone = 0;
1925 bp->b_error = 0;
1926 bp->b_resid = 0;
1927 bp->b_bcount = 0;
1928 bp->b_npages = 0;
1929 bp->b_dirtyoff = bp->b_dirtyend = 0;
1930 bp->b_magic = B_MAGIC_BIO;
1931 bp->b_op = &buf_ops_bio;
1932 bp->b_object = NULL;
1933
1934 LIST_INIT(&bp->b_dep);
1935
1936 /*
1937 * If we are defragging then free the buffer.
1938 */
1939 if (defrag) {
1940 bp->b_flags |= B_INVAL;
1941 bfreekva(bp);
1942 brelse(bp);
1943 defrag = 0;
1944 goto restart;
1945 }
1946
1947 /*
1948 * If we are overcomitted then recover the buffer and its
1949 * KVM space. This occurs in rare situations when multiple
1950 * processes are blocked in getnewbuf() or allocbuf().
1951 */
1952 if (bufspace >= hibufspace)
1953 flushingbufs = 1;
1954 if (flushingbufs && bp->b_kvasize != 0) {
1955 bp->b_flags |= B_INVAL;
1956 bfreekva(bp);
1957 brelse(bp);
1958 goto restart;
1959 }
1960 if (bufspace < lobufspace)
1961 flushingbufs = 0;
1962 break;
1963 }
1964
1965 /*
1966 * If we exhausted our list, sleep as appropriate. We may have to
1967 * wakeup various daemons and write out some dirty buffers.
1968 *
1969 * Generally we are sleeping due to insufficient buffer space.
1970 */
1971
1972 if (bp == NULL) {
1973 int flags;
1974 char *waitmsg;
1975
1976 mtx_unlock(&bqlock);
1977 if (defrag) {
1978 flags = VFS_BIO_NEED_BUFSPACE;
1979 waitmsg = "nbufkv";
1980 } else if (bufspace >= hibufspace) {
1981 waitmsg = "nbufbs";
1982 flags = VFS_BIO_NEED_BUFSPACE;
1983 } else {
1984 waitmsg = "newbuf";
1985 flags = VFS_BIO_NEED_ANY;
1986 }
1987
1988 bd_speedup(); /* heeeelp */
1989
1990 mtx_lock(&nblock);
1991 needsbuffer |= flags;
1992 while (needsbuffer & flags) {
1993 if (msleep(&needsbuffer, &nblock,
1994 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1995 mtx_unlock(&nblock);
1996 return (NULL);
1997 }
1998 }
1999 mtx_unlock(&nblock);
2000 } else {
2001 /*
2002 * We finally have a valid bp. We aren't quite out of the
2003 * woods, we still have to reserve kva space. In order
2004 * to keep fragmentation sane we only allocate kva in
2005 * BKVASIZE chunks.
2006 */
2007 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2008
2009 if (maxsize != bp->b_kvasize) {
2010 vm_offset_t addr = 0;
2011
2012 bfreekva(bp);
2013
2014 if (vm_map_findspace(buffer_map,
2015 vm_map_min(buffer_map), maxsize, &addr)) {
2016 /*
2017 * Uh oh. Buffer map is to fragmented. We
2018 * must defragment the map.
2019 */
2020 atomic_add_int(&bufdefragcnt, 1);
2021 defrag = 1;
2022 bp->b_flags |= B_INVAL;
2023 brelse(bp);
2024 goto restart;
2025 }
2026 if (addr) {
2027 vm_map_insert(buffer_map, NULL, 0,
2028 addr, addr + maxsize,
2029 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2030
2031 bp->b_kvabase = (caddr_t) addr;
2032 bp->b_kvasize = maxsize;
2033 atomic_add_int(&bufspace, bp->b_kvasize);
2034 atomic_add_int(&bufreusecnt, 1);
2035 }
2036 }
2037 bp->b_saveaddr = bp->b_kvabase;
2038 bp->b_data = bp->b_saveaddr;
2039 }
2040 return(bp);
2041}
2042
2043/*
2044 * buf_daemon:
2045 *
2046 * buffer flushing daemon. Buffers are normally flushed by the
2047 * update daemon but if it cannot keep up this process starts to
2048 * take the load in an attempt to prevent getnewbuf() from blocking.
2049 */
2050
2051static struct kproc_desc buf_kp = {
2052 "bufdaemon",
2053 buf_daemon,
2054 &bufdaemonproc
2055};
2056SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
2057
2058static void
2059buf_daemon()
2060{
2061 int s;
2062
2063 mtx_lock(&Giant);
2064
2065 /*
2066 * This process needs to be suspended prior to shutdown sync.
2067 */
2068 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2069 SHUTDOWN_PRI_LAST);
2070
2071 /*
2072 * This process is allowed to take the buffer cache to the limit
2073 */
2074 s = splbio();
2075 mtx_lock(&bdlock);
2076
2077 for (;;) {
2078 bd_request = 0;
2079 mtx_unlock(&bdlock);
2080
2081 kthread_suspend_check(bufdaemonproc);
2082
2083 /*
2084 * Do the flush. Limit the amount of in-transit I/O we
2085 * allow to build up, otherwise we would completely saturate
2086 * the I/O system. Wakeup any waiting processes before we
2087 * normally would so they can run in parallel with our drain.
2088 */
2089 while (numdirtybuffers > lodirtybuffers) {
2090 if (flushbufqueues(0) == 0) {
2091 /*
2092 * Could not find any buffers without rollback
2093 * dependencies, so just write the first one
2094 * in the hopes of eventually making progress.
2095 */
2096 flushbufqueues(1);
2097 break;
2098 }
2099 waitrunningbufspace();
2100 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2101 }
2102
2103 /*
2104 * Only clear bd_request if we have reached our low water
2105 * mark. The buf_daemon normally waits 1 second and
2106 * then incrementally flushes any dirty buffers that have
2107 * built up, within reason.
2108 *
2109 * If we were unable to hit our low water mark and couldn't
2110 * find any flushable buffers, we sleep half a second.
2111 * Otherwise we loop immediately.
2112 */
2113 mtx_lock(&bdlock);
2114 if (numdirtybuffers <= lodirtybuffers) {
2115 /*
2116 * We reached our low water mark, reset the
2117 * request and sleep until we are needed again.
2118 * The sleep is just so the suspend code works.
2119 */
2120 bd_request = 0;
2121 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2122 } else {
2123 /*
2124 * We couldn't find any flushable dirty buffers but
2125 * still have too many dirty buffers, we
2126 * have to sleep and try again. (rare)
2127 */
2128 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2129 }
2130 }
2131}
2132
2133/*
2134 * flushbufqueues:
2135 *
2136 * Try to flush a buffer in the dirty queue. We must be careful to
2137 * free up B_INVAL buffers instead of write them, which NFS is
2138 * particularly sensitive to.
2139 */
2140int flushwithdeps = 0;
2141SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2142 0, "Number of buffers flushed with dependecies that require rollbacks");
2143static int
2144flushbufqueues(int flushdeps)
2145{
2146 struct thread *td = curthread;
2147 struct vnode *vp;
2148 struct mount *mp;
2149 struct buf *bp;
2150 int hasdeps;
2151
2152 mtx_lock(&bqlock);
2153 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) {
2154 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2155 continue;
2156 KASSERT((bp->b_flags & B_DELWRI),
2157 ("unexpected clean buffer %p", bp));
2158 VI_LOCK(bp->b_vp);
2159 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
2160 VI_UNLOCK(bp->b_vp);
2161 BUF_UNLOCK(bp);
2162 continue;
2163 }
2164 VI_UNLOCK(bp->b_vp);
2165 if (bp->b_flags & B_INVAL) {
2166 bremfreel(bp);
2167 mtx_unlock(&bqlock);
2168 brelse(bp);
2169 return (1);
2170 }
2171
2172 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) {
2173 if (flushdeps == 0) {
2174 BUF_UNLOCK(bp);
2175 continue;
2176 }
2177 hasdeps = 1;
2178 } else
2179 hasdeps = 0;
2180 /*
2181 * We must hold the lock on a vnode before writing
2182 * one of its buffers. Otherwise we may confuse, or
2183 * in the case of a snapshot vnode, deadlock the
2184 * system.
2185 *
2186 * The lock order here is the reverse of the normal
2187 * of vnode followed by buf lock. This is ok because
2188 * the NOWAIT will prevent deadlock.
2189 */
2190 vp = bp->b_vp;
2191 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2192 BUF_UNLOCK(bp);
2193 continue;
2194 }
2195 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2196 mtx_unlock(&bqlock);
2197 vfs_bio_awrite(bp);
2198 vn_finished_write(mp);
2199 VOP_UNLOCK(vp, 0, td);
2200 flushwithdeps += hasdeps;
2201 return (1);
2202 }
2203 vn_finished_write(mp);
2204 BUF_UNLOCK(bp);
2205 }
2206 mtx_unlock(&bqlock);
2207 return (0);
2208}
2209
2210/*
2211 * Check to see if a block is currently memory resident.
2212 */
2213struct buf *
2214incore(struct vnode * vp, daddr_t blkno)
2215{
2216 struct buf *bp;
2217
2218 int s = splbio();
2219 VI_LOCK(vp);
2220 bp = gbincore(vp, blkno);
2221 VI_UNLOCK(vp);
2222 splx(s);
2223 return (bp);
2224}
2225
2226/*
2227 * Returns true if no I/O is needed to access the
2228 * associated VM object. This is like incore except
2229 * it also hunts around in the VM system for the data.
2230 */
2231
2232int
2233inmem(struct vnode * vp, daddr_t blkno)
2234{
2235 vm_object_t obj;
2236 vm_offset_t toff, tinc, size;
2237 vm_page_t m;
2238 vm_ooffset_t off;
2239
2240 GIANT_REQUIRED;
2241 ASSERT_VOP_LOCKED(vp, "inmem");
2242
2243 if (incore(vp, blkno))
2244 return 1;
2245 if (vp->v_mount == NULL)
2246 return 0;
2247 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0)
2248 return 0;
2249
2250 size = PAGE_SIZE;
2251 if (size > vp->v_mount->mnt_stat.f_iosize)
2252 size = vp->v_mount->mnt_stat.f_iosize;
2253 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2254
2255 VM_OBJECT_LOCK(obj);
2256 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2257 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2258 if (!m)
2259 goto notinmem;
2260 tinc = size;
2261 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2262 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2263 if (vm_page_is_valid(m,
2264 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2265 goto notinmem;
2266 }
2267 VM_OBJECT_UNLOCK(obj);
2268 return 1;
2269
2270notinmem:
2271 VM_OBJECT_UNLOCK(obj);
2272 return (0);
2273}
2274
2275/*
2276 * vfs_setdirty:
2277 *
2278 * Sets the dirty range for a buffer based on the status of the dirty
2279 * bits in the pages comprising the buffer.
2280 *
2281 * The range is limited to the size of the buffer.
2282 *
2283 * This routine is primarily used by NFS, but is generalized for the
2284 * B_VMIO case.
2285 */
2286static void
2287vfs_setdirty(struct buf *bp)
2288{
2289 int i;
2290 vm_object_t object;
2291
2292 GIANT_REQUIRED;
2293 /*
2294 * Degenerate case - empty buffer
2295 */
2296
2297 if (bp->b_bufsize == 0)
2298 return;
2299
2300 /*
2301 * We qualify the scan for modified pages on whether the
2302 * object has been flushed yet. The OBJ_WRITEABLE flag
2303 * is not cleared simply by protecting pages off.
2304 */
2305
2306 if ((bp->b_flags & B_VMIO) == 0)
2307 return;
2308
2309 object = bp->b_pages[0]->object;
2310 VM_OBJECT_LOCK(object);
2311 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2312 printf("Warning: object %p writeable but not mightbedirty\n", object);
2313 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2314 printf("Warning: object %p mightbedirty but not writeable\n", object);
2315
2316 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2317 vm_offset_t boffset;
2318 vm_offset_t eoffset;
2319
2320 vm_page_lock_queues();
2321 /*
2322 * test the pages to see if they have been modified directly
2323 * by users through the VM system.
2324 */
2325 for (i = 0; i < bp->b_npages; i++)
2326 vm_page_test_dirty(bp->b_pages[i]);
2327
2328 /*
2329 * Calculate the encompassing dirty range, boffset and eoffset,
2330 * (eoffset - boffset) bytes.
2331 */
2332
2333 for (i = 0; i < bp->b_npages; i++) {
2334 if (bp->b_pages[i]->dirty)
2335 break;
2336 }
2337 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2338
2339 for (i = bp->b_npages - 1; i >= 0; --i) {
2340 if (bp->b_pages[i]->dirty) {
2341 break;
2342 }
2343 }
2344 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2345
2346 vm_page_unlock_queues();
2347 /*
2348 * Fit it to the buffer.
2349 */
2350
2351 if (eoffset > bp->b_bcount)
2352 eoffset = bp->b_bcount;
2353
2354 /*
2355 * If we have a good dirty range, merge with the existing
2356 * dirty range.
2357 */
2358
2359 if (boffset < eoffset) {
2360 if (bp->b_dirtyoff > boffset)
2361 bp->b_dirtyoff = boffset;
2362 if (bp->b_dirtyend < eoffset)
2363 bp->b_dirtyend = eoffset;
2364 }
2365 }
2366 VM_OBJECT_UNLOCK(object);
2367}
2368
2369/*
2370 * getblk:
2371 *
2372 * Get a block given a specified block and offset into a file/device.
2373 * The buffers B_DONE bit will be cleared on return, making it almost
2374 * ready for an I/O initiation. B_INVAL may or may not be set on
2375 * return. The caller should clear B_INVAL prior to initiating a
2376 * READ.
2377 *
2378 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2379 * an existing buffer.
2380 *
2381 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2382 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2383 * and then cleared based on the backing VM. If the previous buffer is
2384 * non-0-sized but invalid, B_CACHE will be cleared.
2385 *
2386 * If getblk() must create a new buffer, the new buffer is returned with
2387 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2388 * case it is returned with B_INVAL clear and B_CACHE set based on the
2389 * backing VM.
2390 *
2391 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2392 * B_CACHE bit is clear.
2393 *
2394 * What this means, basically, is that the caller should use B_CACHE to
2395 * determine whether the buffer is fully valid or not and should clear
2396 * B_INVAL prior to issuing a read. If the caller intends to validate
2397 * the buffer by loading its data area with something, the caller needs
2398 * to clear B_INVAL. If the caller does this without issuing an I/O,
2399 * the caller should set B_CACHE ( as an optimization ), else the caller
2400 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2401 * a write attempt or if it was a successfull read. If the caller
2402 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2403 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2404 */
2405struct buf *
2406getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2407 int flags)
2408{
2409 struct buf *bp;
2410 int s;
2411 int error;
2412 ASSERT_VOP_LOCKED(vp, "getblk");
2413
2414 if (size > MAXBSIZE)
2415 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2416
2417 s = splbio();
2418loop:
2419 /*
2420 * Block if we are low on buffers. Certain processes are allowed
2421 * to completely exhaust the buffer cache.
2422 *
2423 * If this check ever becomes a bottleneck it may be better to
2424 * move it into the else, when gbincore() fails. At the moment
2425 * it isn't a problem.
2426 *
2427 * XXX remove if 0 sections (clean this up after its proven)
2428 */
2429 if (numfreebuffers == 0) {
2430 if (curthread == PCPU_GET(idlethread))
2431 return NULL;
2432 mtx_lock(&nblock);
2433 needsbuffer |= VFS_BIO_NEED_ANY;
2434 mtx_unlock(&nblock);
2435 }
2436
2437 VI_LOCK(vp);
2438 if ((bp = gbincore(vp, blkno))) {
2439 int lockflags;
2440 /*
2441 * Buffer is in-core. If the buffer is not busy, it must
2442 * be on a queue.
2443 */
2444 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2445
2446 if (flags & GB_LOCK_NOWAIT)
2447 lockflags |= LK_NOWAIT;
2448
2449 error = BUF_TIMELOCK(bp, lockflags,
2450 VI_MTX(vp), "getblk", slpflag, slptimeo);
2451
2452 /*
2453 * If we slept and got the lock we have to restart in case
2454 * the buffer changed identities.
2455 */
2456 if (error == ENOLCK)
2457 goto loop;
2458 /* We timed out or were interrupted. */
2459 else if (error)
2460 return (NULL);
2461
2462 /*
2463 * The buffer is locked. B_CACHE is cleared if the buffer is
2464 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2465 * and for a VMIO buffer B_CACHE is adjusted according to the
2466 * backing VM cache.
2467 */
2468 if (bp->b_flags & B_INVAL)
2469 bp->b_flags &= ~B_CACHE;
2470 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2471 bp->b_flags |= B_CACHE;
2472 bremfree(bp);
2473
2474 /*
2475 * check for size inconsistancies for non-VMIO case.
2476 */
2477
2478 if (bp->b_bcount != size) {
2479 if ((bp->b_flags & B_VMIO) == 0 ||
2480 (size > bp->b_kvasize)) {
2481 if (bp->b_flags & B_DELWRI) {
2482 bp->b_flags |= B_NOCACHE;
2483 bwrite(bp);
2484 } else {
2485 if ((bp->b_flags & B_VMIO) &&
2486 (LIST_FIRST(&bp->b_dep) == NULL)) {
2487 bp->b_flags |= B_RELBUF;
2488 brelse(bp);
2489 } else {
2490 bp->b_flags |= B_NOCACHE;
2491 bwrite(bp);
2492 }
2493 }
2494 goto loop;
2495 }
2496 }
2497
2498 /*
2499 * If the size is inconsistant in the VMIO case, we can resize
2500 * the buffer. This might lead to B_CACHE getting set or
2501 * cleared. If the size has not changed, B_CACHE remains
2502 * unchanged from its previous state.
2503 */
2504
2505 if (bp->b_bcount != size)
2506 allocbuf(bp, size);
2507
2508 KASSERT(bp->b_offset != NOOFFSET,
2509 ("getblk: no buffer offset"));
2510
2511 /*
2512 * A buffer with B_DELWRI set and B_CACHE clear must
2513 * be committed before we can return the buffer in
2514 * order to prevent the caller from issuing a read
2515 * ( due to B_CACHE not being set ) and overwriting
2516 * it.
2517 *
2518 * Most callers, including NFS and FFS, need this to
2519 * operate properly either because they assume they
2520 * can issue a read if B_CACHE is not set, or because
2521 * ( for example ) an uncached B_DELWRI might loop due
2522 * to softupdates re-dirtying the buffer. In the latter
2523 * case, B_CACHE is set after the first write completes,
2524 * preventing further loops.
2525 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2526 * above while extending the buffer, we cannot allow the
2527 * buffer to remain with B_CACHE set after the write
2528 * completes or it will represent a corrupt state. To
2529 * deal with this we set B_NOCACHE to scrap the buffer
2530 * after the write.
2531 *
2532 * We might be able to do something fancy, like setting
2533 * B_CACHE in bwrite() except if B_DELWRI is already set,
2534 * so the below call doesn't set B_CACHE, but that gets real
2535 * confusing. This is much easier.
2536 */
2537
2538 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2539 bp->b_flags |= B_NOCACHE;
2540 bwrite(bp);
2541 goto loop;
2542 }
2543
2544 splx(s);
2545 bp->b_flags &= ~B_DONE;
2546 } else {
2547 int bsize, maxsize, vmio;
2548 off_t offset;
2549
2550 /*
2551 * Buffer is not in-core, create new buffer. The buffer
2552 * returned by getnewbuf() is locked. Note that the returned
2553 * buffer is also considered valid (not marked B_INVAL).
2554 */
2555 VI_UNLOCK(vp);
2556 /*
2557 * If the user does not want us to create the buffer, bail out
2558 * here.
2559 */
2560 if (flags & GB_NOCREAT) {
2561 splx(s);
2562 return NULL;
2563 }
2564 if (vn_isdisk(vp, NULL))
2565 bsize = DEV_BSIZE;
2566 else if (vp->v_mountedhere)
2567 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2568 else if (vp->v_mount)
2569 bsize = vp->v_mount->mnt_stat.f_iosize;
2570 else
2571 bsize = size;
2572
2573 offset = blkno * bsize;
2574 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) &&
2575 (vp->v_vflag & VV_OBJBUF);
2576 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2577 maxsize = imax(maxsize, bsize);
2578
2579 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2580 if (slpflag || slptimeo) {
2581 splx(s);
2582 return NULL;
2583 }
2584 goto loop;
2585 }
2586
2587 /*
2588 * This code is used to make sure that a buffer is not
2589 * created while the getnewbuf routine is blocked.
2590 * This can be a problem whether the vnode is locked or not.
2591 * If the buffer is created out from under us, we have to
2592 * throw away the one we just created. There is now window
2593 * race because we are safely running at splbio() from the
2594 * point of the duplicate buffer creation through to here,
2595 * and we've locked the buffer.
2596 *
2597 * Note: this must occur before we associate the buffer
2598 * with the vp especially considering limitations in
2599 * the splay tree implementation when dealing with duplicate
2600 * lblkno's.
2601 */
2602 VI_LOCK(vp);
2603 if (gbincore(vp, blkno)) {
2604 VI_UNLOCK(vp);
2605 bp->b_flags |= B_INVAL;
2606 brelse(bp);
2607 goto loop;
2608 }
2609
2610 /*
2611 * Insert the buffer into the hash, so that it can
2612 * be found by incore.
2613 */
2614 bp->b_blkno = bp->b_lblkno = blkno;
2615 bp->b_offset = offset;
2616
2617 bgetvp(vp, bp);
2618 VI_UNLOCK(vp);
2619
2620 /*
2621 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2622 * buffer size starts out as 0, B_CACHE will be set by
2623 * allocbuf() for the VMIO case prior to it testing the
2624 * backing store for validity.
2625 */
2626
2627 if (vmio) {
2628 bp->b_flags |= B_VMIO;
2629#if defined(VFS_BIO_DEBUG)
2630 if (vn_canvmio(vp) != TRUE)
2631 printf("getblk: VMIO on vnode type %d\n",
2632 vp->v_type);
2633#endif
2634 VOP_GETVOBJECT(vp, &bp->b_object);
2635 } else {
2636 bp->b_flags &= ~B_VMIO;
2637 bp->b_object = NULL;
2638 }
2639
2640 allocbuf(bp, size);
2641
2642 splx(s);
2643 bp->b_flags &= ~B_DONE;
2644 }
2645 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2646 return (bp);
2647}
2648
2649/*
2650 * Get an empty, disassociated buffer of given size. The buffer is initially
2651 * set to B_INVAL.
2652 */
2653struct buf *
2654geteblk(int size)
2655{
2656 struct buf *bp;
2657 int s;
2658 int maxsize;
2659
2660 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2661
2662 s = splbio();
2663 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2664 continue;
2665 splx(s);
2666 allocbuf(bp, size);
2667 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2668 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2669 return (bp);
2670}
2671
2672
2673/*
2674 * This code constitutes the buffer memory from either anonymous system
2675 * memory (in the case of non-VMIO operations) or from an associated
2676 * VM object (in the case of VMIO operations). This code is able to
2677 * resize a buffer up or down.
2678 *
2679 * Note that this code is tricky, and has many complications to resolve
2680 * deadlock or inconsistant data situations. Tread lightly!!!
2681 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2682 * the caller. Calling this code willy nilly can result in the loss of data.
2683 *
2684 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2685 * B_CACHE for the non-VMIO case.
2686 */
2687
2688int
2689allocbuf(struct buf *bp, int size)
2690{
2691 int newbsize, mbsize;
2692 int i;
2693
2694 GIANT_REQUIRED;
2695
2696 if (BUF_REFCNT(bp) == 0)
2697 panic("allocbuf: buffer not busy");
2698
2699 if (bp->b_kvasize < size)
2700 panic("allocbuf: buffer too small");
2701
2702 if ((bp->b_flags & B_VMIO) == 0) {
2703 caddr_t origbuf;
2704 int origbufsize;
2705 /*
2706 * Just get anonymous memory from the kernel. Don't
2707 * mess with B_CACHE.
2708 */
2709 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2710 if (bp->b_flags & B_MALLOC)
2711 newbsize = mbsize;
2712 else
2713 newbsize = round_page(size);
2714
2715 if (newbsize < bp->b_bufsize) {
2716 /*
2717 * malloced buffers are not shrunk
2718 */
2719 if (bp->b_flags & B_MALLOC) {
2720 if (newbsize) {
2721 bp->b_bcount = size;
2722 } else {
2723 free(bp->b_data, M_BIOBUF);
2724 if (bp->b_bufsize) {
2725 atomic_subtract_int(
2726 &bufmallocspace,
2727 bp->b_bufsize);
2728 bufspacewakeup();
2729 bp->b_bufsize = 0;
2730 }
2731 bp->b_saveaddr = bp->b_kvabase;
2732 bp->b_data = bp->b_saveaddr;
2733 bp->b_bcount = 0;
2734 bp->b_flags &= ~B_MALLOC;
2735 }
2736 return 1;
2737 }
2738 vm_hold_free_pages(
2739 bp,
2740 (vm_offset_t) bp->b_data + newbsize,
2741 (vm_offset_t) bp->b_data + bp->b_bufsize);
2742 } else if (newbsize > bp->b_bufsize) {
2743 /*
2744 * We only use malloced memory on the first allocation.
2745 * and revert to page-allocated memory when the buffer
2746 * grows.
2747 */
2748 /*
2749 * There is a potential smp race here that could lead
2750 * to bufmallocspace slightly passing the max. It
2751 * is probably extremely rare and not worth worrying
2752 * over.
2753 */
2754 if ( (bufmallocspace < maxbufmallocspace) &&
2755 (bp->b_bufsize == 0) &&
2756 (mbsize <= PAGE_SIZE/2)) {
2757
2758 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2759 bp->b_bufsize = mbsize;
2760 bp->b_bcount = size;
2761 bp->b_flags |= B_MALLOC;
2762 atomic_add_int(&bufmallocspace, mbsize);
2763 return 1;
2764 }
2765 origbuf = NULL;
2766 origbufsize = 0;
2767 /*
2768 * If the buffer is growing on its other-than-first allocation,
2769 * then we revert to the page-allocation scheme.
2770 */
2771 if (bp->b_flags & B_MALLOC) {
2772 origbuf = bp->b_data;
2773 origbufsize = bp->b_bufsize;
2774 bp->b_data = bp->b_kvabase;
2775 if (bp->b_bufsize) {
2776 atomic_subtract_int(&bufmallocspace,
2777 bp->b_bufsize);
2778 bufspacewakeup();
2779 bp->b_bufsize = 0;
2780 }
2781 bp->b_flags &= ~B_MALLOC;
2782 newbsize = round_page(newbsize);
2783 }
2784 vm_hold_load_pages(
2785 bp,
2786 (vm_offset_t) bp->b_data + bp->b_bufsize,
2787 (vm_offset_t) bp->b_data + newbsize);
2788 if (origbuf) {
2789 bcopy(origbuf, bp->b_data, origbufsize);
2790 free(origbuf, M_BIOBUF);
2791 }
2792 }
2793 } else {
2794 int desiredpages;
2795
2796 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2797 desiredpages = (size == 0) ? 0 :
2798 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2799
2800 if (bp->b_flags & B_MALLOC)
2801 panic("allocbuf: VMIO buffer can't be malloced");
2802 /*
2803 * Set B_CACHE initially if buffer is 0 length or will become
2804 * 0-length.
2805 */
2806 if (size == 0 || bp->b_bufsize == 0)
2807 bp->b_flags |= B_CACHE;
2808
2809 if (newbsize < bp->b_bufsize) {
2810 /*
2811 * DEV_BSIZE aligned new buffer size is less then the
2812 * DEV_BSIZE aligned existing buffer size. Figure out
2813 * if we have to remove any pages.
2814 */
2815 if (desiredpages < bp->b_npages) {
2816 vm_page_t m;
2817
2818 vm_page_lock_queues();
2819 for (i = desiredpages; i < bp->b_npages; i++) {
2820 /*
2821 * the page is not freed here -- it
2822 * is the responsibility of
2823 * vnode_pager_setsize
2824 */
2825 m = bp->b_pages[i];
2826 KASSERT(m != bogus_page,
2827 ("allocbuf: bogus page found"));
2828 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2829 vm_page_lock_queues();
2830
2831 bp->b_pages[i] = NULL;
2832 vm_page_unwire(m, 0);
2833 }
2834 vm_page_unlock_queues();
2835 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2836 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2837 bp->b_npages = desiredpages;
2838 }
2839 } else if (size > bp->b_bcount) {
2840 /*
2841 * We are growing the buffer, possibly in a
2842 * byte-granular fashion.
2843 */
2844 struct vnode *vp;
2845 vm_object_t obj;
2846 vm_offset_t toff;
2847 vm_offset_t tinc;
2848
2849 /*
2850 * Step 1, bring in the VM pages from the object,
2851 * allocating them if necessary. We must clear
2852 * B_CACHE if these pages are not valid for the
2853 * range covered by the buffer.
2854 */
2855
2856 vp = bp->b_vp;
2857 obj = bp->b_object;
2858
2859 VM_OBJECT_LOCK(obj);
2860 while (bp->b_npages < desiredpages) {
2861 vm_page_t m;
2862 vm_pindex_t pi;
2863
2864 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2865 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2866 /*
2867 * note: must allocate system pages
2868 * since blocking here could intefere
2869 * with paging I/O, no matter which
2870 * process we are.
2871 */
2872 m = vm_page_alloc(obj, pi,
2873 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
2874 if (m == NULL) {
2875 atomic_add_int(&vm_pageout_deficit,
2876 desiredpages - bp->b_npages);
2877 VM_OBJECT_UNLOCK(obj);
2878 VM_WAIT;
2879 VM_OBJECT_LOCK(obj);
2880 } else {
2881 vm_page_lock_queues();
2882 vm_page_wakeup(m);
2883 vm_page_unlock_queues();
2884 bp->b_flags &= ~B_CACHE;
2885 bp->b_pages[bp->b_npages] = m;
2886 ++bp->b_npages;
2887 }
2888 continue;
2889 }
2890
2891 /*
2892 * We found a page. If we have to sleep on it,
2893 * retry because it might have gotten freed out
2894 * from under us.
2895 *
2896 * We can only test PG_BUSY here. Blocking on
2897 * m->busy might lead to a deadlock:
2898 *
2899 * vm_fault->getpages->cluster_read->allocbuf
2900 *
2901 */
2902 vm_page_lock_queues();
2903 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2904 continue;
2905
2906 /*
2907 * We have a good page. Should we wakeup the
2908 * page daemon?
2909 */
2910 if ((curproc != pageproc) &&
2911 ((m->queue - m->pc) == PQ_CACHE) &&
2912 ((cnt.v_free_count + cnt.v_cache_count) <
2913 (cnt.v_free_min + cnt.v_cache_min))) {
2914 pagedaemon_wakeup();
2915 }
2916 vm_page_wire(m);
2917 vm_page_unlock_queues();
2918 bp->b_pages[bp->b_npages] = m;
2919 ++bp->b_npages;
2920 }
2921
2922 /*
2923 * Step 2. We've loaded the pages into the buffer,
2924 * we have to figure out if we can still have B_CACHE
2925 * set. Note that B_CACHE is set according to the
2926 * byte-granular range ( bcount and size ), new the
2927 * aligned range ( newbsize ).
2928 *
2929 * The VM test is against m->valid, which is DEV_BSIZE
2930 * aligned. Needless to say, the validity of the data
2931 * needs to also be DEV_BSIZE aligned. Note that this
2932 * fails with NFS if the server or some other client
2933 * extends the file's EOF. If our buffer is resized,
2934 * B_CACHE may remain set! XXX
2935 */
2936
2937 toff = bp->b_bcount;
2938 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2939
2940 while ((bp->b_flags & B_CACHE) && toff < size) {
2941 vm_pindex_t pi;
2942
2943 if (tinc > (size - toff))
2944 tinc = size - toff;
2945
2946 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2947 PAGE_SHIFT;
2948
2949 vfs_buf_test_cache(
2950 bp,
2951 bp->b_offset,
2952 toff,
2953 tinc,
2954 bp->b_pages[pi]
2955 );
2956 toff += tinc;
2957 tinc = PAGE_SIZE;
2958 }
2959 VM_OBJECT_UNLOCK(obj);
2960
2961 /*
2962 * Step 3, fixup the KVM pmap. Remember that
2963 * bp->b_data is relative to bp->b_offset, but
2964 * bp->b_offset may be offset into the first page.
2965 */
2966
2967 bp->b_data = (caddr_t)
2968 trunc_page((vm_offset_t)bp->b_data);
2969 pmap_qenter(
2970 (vm_offset_t)bp->b_data,
2971 bp->b_pages,
2972 bp->b_npages
2973 );
2974
2975 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2976 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2977 }
2978 }
2979 if (newbsize < bp->b_bufsize)
2980 bufspacewakeup();
2981 bp->b_bufsize = newbsize; /* actual buffer allocation */
2982 bp->b_bcount = size; /* requested buffer size */
2983 return 1;
2984}
2985
2986void
2987biodone(struct bio *bp)
2988{
2989 mtx_lock(&bdonelock);
2990 bp->bio_flags |= BIO_DONE;
2991 if (bp->bio_done == NULL)
2992 wakeup(bp);
2993 mtx_unlock(&bdonelock);
2994 if (bp->bio_done != NULL)
2995 bp->bio_done(bp);
2996}
2997
2998/*
2999 * Wait for a BIO to finish.
3000 *
3001 * XXX: resort to a timeout for now. The optimal locking (if any) for this
3002 * case is not yet clear.
3003 */
3004int
3005biowait(struct bio *bp, const char *wchan)
3006{
3007
3008 mtx_lock(&bdonelock);
3009 while ((bp->bio_flags & BIO_DONE) == 0)
3010 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
3011 mtx_unlock(&bdonelock);
3012 if (bp->bio_error != 0)
3013 return (bp->bio_error);
3014 if (!(bp->bio_flags & BIO_ERROR))
3015 return (0);
3016 return (EIO);
3017}
3018
3019void
3020biofinish(struct bio *bp, struct devstat *stat, int error)
3021{
3022
3023 if (error) {
3024 bp->bio_error = error;
3025 bp->bio_flags |= BIO_ERROR;
3026 }
3027 if (stat != NULL)
3028 devstat_end_transaction_bio(stat, bp);
3029 biodone(bp);
3030}
3031
3032/*
3033 * bufwait:
3034 *
3035 * Wait for buffer I/O completion, returning error status. The buffer
3036 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3037 * error and cleared.
3038 */
3039int
3040bufwait(register struct buf * bp)
3041{
3042 int s;
3043
3044 s = splbio();
3045 if (bp->b_iocmd == BIO_READ)
3046 bwait(bp, PRIBIO, "biord");
3047 else
3048 bwait(bp, PRIBIO, "biowr");
3049 splx(s);
3050 if (bp->b_flags & B_EINTR) {
3051 bp->b_flags &= ~B_EINTR;
3052 return (EINTR);
3053 }
3054 if (bp->b_ioflags & BIO_ERROR) {
3055 return (bp->b_error ? bp->b_error : EIO);
3056 } else {
3057 return (0);
3058 }
3059}
3060
3061 /*
3062 * Call back function from struct bio back up to struct buf.
3063 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY().
3064 */
3065static void
3066bufdonebio(struct bio *bp)
3067{
3068
3069 /* Device drivers may or may not hold giant, hold it here. */
3070 mtx_lock(&Giant);
3071 bufdone(bp->bio_caller2);
3072 mtx_unlock(&Giant);
3073}
3074
3075void
3076dev_strategy(struct buf *bp)
3077{
3078 struct cdevsw *csw;
3079
3080 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3081 panic("b_iocmd botch");
3082 bp->b_io.bio_done = bufdonebio;
3083 bp->b_io.bio_caller2 = bp;
3084 csw = devsw(bp->b_io.bio_dev);
3085 KASSERT(bp->b_io.bio_dev->si_refcount > 0,
3086 ("dev_strategy on un-referenced struct cdev *(%s)",
3087 devtoname(bp->b_io.bio_dev)));
3088 cdevsw_ref(csw);
3089 (*devsw(bp->b_io.bio_dev)->d_strategy)(&bp->b_io);
3090 cdevsw_rel(csw);
3091}
3092
3093/*
3094 * bufdone:
3095 *
3096 * Finish I/O on a buffer, optionally calling a completion function.
3097 * This is usually called from an interrupt so process blocking is
3098 * not allowed.
3099 *
3100 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3101 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3102 * assuming B_INVAL is clear.
3103 *
3104 * For the VMIO case, we set B_CACHE if the op was a read and no
3105 * read error occured, or if the op was a write. B_CACHE is never
3106 * set if the buffer is invalid or otherwise uncacheable.
3107 *
3108 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3109 * initiator to leave B_INVAL set to brelse the buffer out of existance
3110 * in the biodone routine.
3111 */
3112void
3113bufdone(struct buf *bp)
3114{
3115 int s;
3116 void (*biodone)(struct buf *);
3117
3118
3119 s = splbio();
3120
3121 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
3122 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3123
3124 bp->b_flags |= B_DONE;
3125 runningbufwakeup(bp);
3126
3127 if (bp->b_iocmd == BIO_DELETE) {
3128 brelse(bp);
3129 splx(s);
3130 return;
3131 }
3132
3133 if (bp->b_iocmd == BIO_WRITE) {
3134 vwakeup(bp);
3135 }
3136
3137 /* call optional completion function if requested */
3138 if (bp->b_iodone != NULL) {
3139 biodone = bp->b_iodone;
3140 bp->b_iodone = NULL;
3141 (*biodone) (bp);
3142 splx(s);
3143 return;
3144 }
3145 if (LIST_FIRST(&bp->b_dep) != NULL)
3146 buf_complete(bp);
3147
3148 if (bp->b_flags & B_VMIO) {
3149 int i;
3150 vm_ooffset_t foff;
3151 vm_page_t m;
3152 vm_object_t obj;
3153 int iosize;
3154 struct vnode *vp = bp->b_vp;
3155
3156 obj = bp->b_object;
3157
3158#if defined(VFS_BIO_DEBUG)
3159 mp_fixme("usecount and vflag accessed without locks.");
3160 if (vp->v_usecount == 0) {
3161 panic("biodone: zero vnode ref count");
3162 }
3163
3164 if ((vp->v_vflag & VV_OBJBUF) == 0) {
3165 panic("biodone: vnode is not setup for merged cache");
3166 }
3167#endif
3168
3169 foff = bp->b_offset;
3170 KASSERT(bp->b_offset != NOOFFSET,
3171 ("biodone: no buffer offset"));
3172
3173 VM_OBJECT_LOCK(obj);
3174#if defined(VFS_BIO_DEBUG)
3175 if (obj->paging_in_progress < bp->b_npages) {
3176 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3177 obj->paging_in_progress, bp->b_npages);
3178 }
3179#endif
3180
3181 /*
3182 * Set B_CACHE if the op was a normal read and no error
3183 * occured. B_CACHE is set for writes in the b*write()
3184 * routines.
3185 */
3186 iosize = bp->b_bcount - bp->b_resid;
3187 if (bp->b_iocmd == BIO_READ &&
3188 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3189 !(bp->b_ioflags & BIO_ERROR)) {
3190 bp->b_flags |= B_CACHE;
3191 }
3192 vm_page_lock_queues();
3193 for (i = 0; i < bp->b_npages; i++) {
3194 int bogusflag = 0;
3195 int resid;
3196
3197 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3198 if (resid > iosize)
3199 resid = iosize;
3200
3201 /*
3202 * cleanup bogus pages, restoring the originals
3203 */
3204 m = bp->b_pages[i];
3205 if (m == bogus_page) {
3206 bogusflag = 1;
3207 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3208 if (m == NULL)
3209 panic("biodone: page disappeared!");
3210 bp->b_pages[i] = m;
3211 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3212 }
3213#if defined(VFS_BIO_DEBUG)
3214 if (OFF_TO_IDX(foff) != m->pindex) {
3215 printf(
3216"biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3217 (intmax_t)foff, (uintmax_t)m->pindex);
3218 }
3219#endif
3220
3221 /*
3222 * In the write case, the valid and clean bits are
3223 * already changed correctly ( see bdwrite() ), so we
3224 * only need to do this here in the read case.
3225 */
3226 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3227 vfs_page_set_valid(bp, foff, i, m);
3228 }
3229
3230 /*
3231 * when debugging new filesystems or buffer I/O methods, this
3232 * is the most common error that pops up. if you see this, you
3233 * have not set the page busy flag correctly!!!
3234 */
3235 if (m->busy == 0) {
3236 printf("biodone: page busy < 0, "
3237 "pindex: %d, foff: 0x(%x,%x), "
3238 "resid: %d, index: %d\n",
3239 (int) m->pindex, (int)(foff >> 32),
3240 (int) foff & 0xffffffff, resid, i);
3241 if (!vn_isdisk(vp, NULL))
3242 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3243 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3244 (intmax_t) bp->b_lblkno,
3245 bp->b_flags, bp->b_npages);
3246 else
3247 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3248 (intmax_t) bp->b_lblkno,
3249 bp->b_flags, bp->b_npages);
3250 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3251 (u_long)m->valid, (u_long)m->dirty,
3252 m->wire_count);
3253 panic("biodone: page busy < 0\n");
3254 }
3255 vm_page_io_finish(m);
3256 vm_object_pip_subtract(obj, 1);
3257 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3258 iosize -= resid;
3259 }
3260 vm_page_unlock_queues();
3261 vm_object_pip_wakeupn(obj, 0);
3262 VM_OBJECT_UNLOCK(obj);
3263 }
3264
3265 /*
3266 * For asynchronous completions, release the buffer now. The brelse
3267 * will do a wakeup there if necessary - so no need to do a wakeup
3268 * here in the async case. The sync case always needs to do a wakeup.
3269 */
3270
3271 if (bp->b_flags & B_ASYNC) {
3272 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3273 brelse(bp);
3274 else
3275 bqrelse(bp);
3276 } else {
3277 bdone(bp);
3278 }
3279 splx(s);
3280}
3281
3282/*
3283 * This routine is called in lieu of iodone in the case of
3284 * incomplete I/O. This keeps the busy status for pages
3285 * consistant.
3286 */
3287void
3288vfs_unbusy_pages(struct buf * bp)
3289{
3290 int i;
3291
3292 runningbufwakeup(bp);
3293 if (bp->b_flags & B_VMIO) {
3294 vm_object_t obj;
3295
3296 obj = bp->b_object;
3297 VM_OBJECT_LOCK(obj);
3298 vm_page_lock_queues();
3299 for (i = 0; i < bp->b_npages; i++) {
3300 vm_page_t m = bp->b_pages[i];
3301
3302 if (m == bogus_page) {
3303 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3304 if (!m) {
3305 panic("vfs_unbusy_pages: page missing\n");
3306 }
3307 bp->b_pages[i] = m;
3308 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3309 }
3310 vm_object_pip_subtract(obj, 1);
3311 vm_page_io_finish(m);
3312 }
3313 vm_page_unlock_queues();
3314 vm_object_pip_wakeupn(obj, 0);
3315 VM_OBJECT_UNLOCK(obj);
3316 }
3317}
3318
3319/*
3320 * vfs_page_set_valid:
3321 *
3322 * Set the valid bits in a page based on the supplied offset. The
3323 * range is restricted to the buffer's size.
3324 *
3325 * This routine is typically called after a read completes.
3326 */
3327static void
3328vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3329{
3330 vm_ooffset_t soff, eoff;
3331
3332 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3333 /*
3334 * Start and end offsets in buffer. eoff - soff may not cross a
3335 * page boundry or cross the end of the buffer. The end of the
3336 * buffer, in this case, is our file EOF, not the allocation size
3337 * of the buffer.
3338 */
3339 soff = off;
3340 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3341 if (eoff > bp->b_offset + bp->b_bcount)
3342 eoff = bp->b_offset + bp->b_bcount;
3343
3344 /*
3345 * Set valid range. This is typically the entire buffer and thus the
3346 * entire page.
3347 */
3348 if (eoff > soff) {
3349 vm_page_set_validclean(
3350 m,
3351 (vm_offset_t) (soff & PAGE_MASK),
3352 (vm_offset_t) (eoff - soff)
3353 );
3354 }
3355}
3356
3357/*
3358 * This routine is called before a device strategy routine.
3359 * It is used to tell the VM system that paging I/O is in
3360 * progress, and treat the pages associated with the buffer
3361 * almost as being PG_BUSY. Also the object paging_in_progress
3362 * flag is handled to make sure that the object doesn't become
3363 * inconsistant.
3364 *
3365 * Since I/O has not been initiated yet, certain buffer flags
3366 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3367 * and should be ignored.
3368 */
3369void
3370vfs_busy_pages(struct buf * bp, int clear_modify)
3371{
3372 int i, bogus;
3373
3374 if (bp->b_flags & B_VMIO) {
3375 vm_object_t obj;
3376 vm_ooffset_t foff;
3377
3378 obj = bp->b_object;
3379 foff = bp->b_offset;
3380 KASSERT(bp->b_offset != NOOFFSET,
3381 ("vfs_busy_pages: no buffer offset"));
3382 vfs_setdirty(bp);
3383 VM_OBJECT_LOCK(obj);
3384retry:
3385 vm_page_lock_queues();
3386 for (i = 0; i < bp->b_npages; i++) {
3387 vm_page_t m = bp->b_pages[i];
3388
3389 if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3390 goto retry;
3391 }
3392 bogus = 0;
3393 for (i = 0; i < bp->b_npages; i++) {
3394 vm_page_t m = bp->b_pages[i];
3395
3396 if ((bp->b_flags & B_CLUSTER) == 0) {
3397 vm_object_pip_add(obj, 1);
3398 vm_page_io_start(m);
3399 }
3400 /*
3401 * When readying a buffer for a read ( i.e
3402 * clear_modify == 0 ), it is important to do
3403 * bogus_page replacement for valid pages in
3404 * partially instantiated buffers. Partially
3405 * instantiated buffers can, in turn, occur when
3406 * reconstituting a buffer from its VM backing store
3407 * base. We only have to do this if B_CACHE is
3408 * clear ( which causes the I/O to occur in the
3409 * first place ). The replacement prevents the read
3410 * I/O from overwriting potentially dirty VM-backed
3411 * pages. XXX bogus page replacement is, uh, bogus.
3412 * It may not work properly with small-block devices.
3413 * We need to find a better way.
3414 */
3415 pmap_remove_all(m);
3416 if (clear_modify)
3417 vfs_page_set_valid(bp, foff, i, m);
3418 else if (m->valid == VM_PAGE_BITS_ALL &&
3419 (bp->b_flags & B_CACHE) == 0) {
3420 bp->b_pages[i] = bogus_page;
3421 bogus++;
3422 }
3423 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3424 }
3425 vm_page_unlock_queues();
3426 VM_OBJECT_UNLOCK(obj);
3427 if (bogus)
3428 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3429 }
3430}
3431
3432/*
3433 * Tell the VM system that the pages associated with this buffer
3434 * are clean. This is used for delayed writes where the data is
3435 * going to go to disk eventually without additional VM intevention.
3436 *
3437 * Note that while we only really need to clean through to b_bcount, we
3438 * just go ahead and clean through to b_bufsize.
3439 */
3440static void
3441vfs_clean_pages(struct buf * bp)
3442{
3443 int i;
3444
3445 if (bp->b_flags & B_VMIO) {
3446 vm_ooffset_t foff;
3447
3448 foff = bp->b_offset;
3449 KASSERT(bp->b_offset != NOOFFSET,
3450 ("vfs_clean_pages: no buffer offset"));
3451 VM_OBJECT_LOCK(bp->b_object);
3452 vm_page_lock_queues();
3453 for (i = 0; i < bp->b_npages; i++) {
3454 vm_page_t m = bp->b_pages[i];
3455 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3456 vm_ooffset_t eoff = noff;
3457
3458 if (eoff > bp->b_offset + bp->b_bufsize)
3459 eoff = bp->b_offset + bp->b_bufsize;
3460 vfs_page_set_valid(bp, foff, i, m);
3461 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3462 foff = noff;
3463 }
3464 vm_page_unlock_queues();
3465 VM_OBJECT_UNLOCK(bp->b_object);
3466 }
3467}
3468
3469/*
3470 * vfs_bio_set_validclean:
3471 *
3472 * Set the range within the buffer to valid and clean. The range is
3473 * relative to the beginning of the buffer, b_offset. Note that b_offset
3474 * itself may be offset from the beginning of the first page.
3475 *
3476 */
3477
3478void
3479vfs_bio_set_validclean(struct buf *bp, int base, int size)
3480{
3481 if (bp->b_flags & B_VMIO) {
3482 int i;
3483 int n;
3484
3485 /*
3486 * Fixup base to be relative to beginning of first page.
3487 * Set initial n to be the maximum number of bytes in the
3488 * first page that can be validated.
3489 */
3490
3491 base += (bp->b_offset & PAGE_MASK);
3492 n = PAGE_SIZE - (base & PAGE_MASK);
3493
3494 VM_OBJECT_LOCK(bp->b_object);
3495 vm_page_lock_queues();
3496 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3497 vm_page_t m = bp->b_pages[i];
3498
3499 if (n > size)
3500 n = size;
3501
3502 vm_page_set_validclean(m, base & PAGE_MASK, n);
3503 base += n;
3504 size -= n;
3505 n = PAGE_SIZE;
3506 }
3507 vm_page_unlock_queues();
3508 VM_OBJECT_UNLOCK(bp->b_object);
3509 }
3510}
3511
3512/*
3513 * vfs_bio_clrbuf:
3514 *
3515 * clear a buffer. This routine essentially fakes an I/O, so we need
3516 * to clear BIO_ERROR and B_INVAL.
3517 *
3518 * Note that while we only theoretically need to clear through b_bcount,
3519 * we go ahead and clear through b_bufsize.
3520 */
3521
3522void
3523vfs_bio_clrbuf(struct buf *bp)
3524{
3525 int i, j, mask = 0;
3526 caddr_t sa, ea;
3527
3528 GIANT_REQUIRED;
3529
3530 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3531 bp->b_flags &= ~B_INVAL;
3532 bp->b_ioflags &= ~BIO_ERROR;
3533 VM_OBJECT_LOCK(bp->b_object);
3534 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3535 (bp->b_offset & PAGE_MASK) == 0) {
3536 if (bp->b_pages[0] == bogus_page)
3537 goto unlock;
3538 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3539 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3540 if ((bp->b_pages[0]->valid & mask) == mask)
3541 goto unlock;
3542 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3543 ((bp->b_pages[0]->valid & mask) == 0)) {
3544 bzero(bp->b_data, bp->b_bufsize);
3545 bp->b_pages[0]->valid |= mask;
3546 goto unlock;
3547 }
3548 }
3549 ea = sa = bp->b_data;
3550 for(i=0;i<bp->b_npages;i++,sa=ea) {
3551 if (bp->b_pages[i] == bogus_page)
3552 continue;
3553 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3554 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3555 ea = (caddr_t)(vm_offset_t)ulmin(
3556 (u_long)(vm_offset_t)ea,
3557 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3558 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3559 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3560 if ((bp->b_pages[i]->valid & mask) == mask)
3561 continue;
3562 if ((bp->b_pages[i]->valid & mask) == 0) {
3563 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3564 bzero(sa, ea - sa);
3565 }
3566 } else {
3567 for (; sa < ea; sa += DEV_BSIZE, j++) {
3568 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3569 (bp->b_pages[i]->valid & (1<<j)) == 0)
3570 bzero(sa, DEV_BSIZE);
3571 }
3572 }
3573 bp->b_pages[i]->valid |= mask;
3574 }
3575unlock:
3576 VM_OBJECT_UNLOCK(bp->b_object);
3577 bp->b_resid = 0;
3578 } else {
3579 clrbuf(bp);
3580 }
3581}
3582
3583/*
3584 * vm_hold_load_pages and vm_hold_free_pages get pages into
3585 * a buffers address space. The pages are anonymous and are
3586 * not associated with a file object.
3587 */
3588static void
3589vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3590{
3591 vm_offset_t pg;
3592 vm_page_t p;
3593 int index;
3594
3595 to = round_page(to);
3596 from = round_page(from);
3597 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3598
3599 VM_OBJECT_LOCK(kernel_object);
3600 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3601tryagain:
3602 /*
3603 * note: must allocate system pages since blocking here
3604 * could intefere with paging I/O, no matter which
3605 * process we are.
3606 */
3607 p = vm_page_alloc(kernel_object,
3608 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3609 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3610 if (!p) {
3611 atomic_add_int(&vm_pageout_deficit,
3612 (to - pg) >> PAGE_SHIFT);
3613 VM_OBJECT_UNLOCK(kernel_object);
3614 VM_WAIT;
3615 VM_OBJECT_LOCK(kernel_object);
3616 goto tryagain;
3617 }
3618 p->valid = VM_PAGE_BITS_ALL;
3619 pmap_qenter(pg, &p, 1);
3620 bp->b_pages[index] = p;
3621 vm_page_lock_queues();
3622 vm_page_wakeup(p);
3623 vm_page_unlock_queues();
3624 }
3625 VM_OBJECT_UNLOCK(kernel_object);
3626 bp->b_npages = index;
3627}
3628
3629/* Return pages associated with this buf to the vm system */
3630static void
3631vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3632{
3633 vm_offset_t pg;
3634 vm_page_t p;
3635 int index, newnpages;
3636
3637 GIANT_REQUIRED;
3638
3639 from = round_page(from);
3640 to = round_page(to);
3641 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3642
3643 VM_OBJECT_LOCK(kernel_object);
3644 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3645 p = bp->b_pages[index];
3646 if (p && (index < bp->b_npages)) {
3647 if (p->busy) {
3648 printf(
3649 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3650 (intmax_t)bp->b_blkno,
3651 (intmax_t)bp->b_lblkno);
3652 }
3653 bp->b_pages[index] = NULL;
3654 pmap_qremove(pg, 1);
3655 vm_page_lock_queues();
3656 vm_page_busy(p);
3657 vm_page_unwire(p, 0);
3658 vm_page_free(p);
3659 vm_page_unlock_queues();
3660 }
3661 }
3662 VM_OBJECT_UNLOCK(kernel_object);
3663 bp->b_npages = newnpages;
3664}
3665
3666/*
3667 * Map an IO request into kernel virtual address space.
3668 *
3669 * All requests are (re)mapped into kernel VA space.
3670 * Notice that we use b_bufsize for the size of the buffer
3671 * to be mapped. b_bcount might be modified by the driver.
3672 *
3673 * Note that even if the caller determines that the address space should
3674 * be valid, a race or a smaller-file mapped into a larger space may
3675 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3676 * check the return value.
3677 */
3678int
3679vmapbuf(struct buf *bp)
3680{
3681 caddr_t addr, kva;
3682 vm_prot_t prot;
3683 int pidx, i;
3684 struct vm_page *m;
3685 struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3686
3687 GIANT_REQUIRED;
3688
3689 if (bp->b_bufsize < 0)
3690 return (-1);
3691 prot = (bp->b_iocmd == BIO_READ) ? VM_PROT_READ | VM_PROT_WRITE :
3692 VM_PROT_READ;
3693 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3694 addr < bp->b_data + bp->b_bufsize;
3695 addr += PAGE_SIZE, pidx++) {
3696 /*
3697 * Do the vm_fault if needed; do the copy-on-write thing
3698 * when reading stuff off device into memory.
3699 *
3700 * NOTE! Must use pmap_extract() because addr may be in
3701 * the userland address space, and kextract is only guarenteed
3702 * to work for the kernland address space (see: sparc64 port).
3703 */
3704retry:
3705 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3706 prot) < 0) {
3707 vm_page_lock_queues();
3708 for (i = 0; i < pidx; ++i) {
3709 vm_page_unhold(bp->b_pages[i]);
3710 bp->b_pages[i] = NULL;
3711 }
3712 vm_page_unlock_queues();
3713 return(-1);
3714 }
3715 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3716 if (m == NULL)
3717 goto retry;
3718 bp->b_pages[pidx] = m;
3719 }
3720 if (pidx > btoc(MAXPHYS))
3721 panic("vmapbuf: mapped more than MAXPHYS");
3722 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3723
3724 kva = bp->b_saveaddr;
3725 bp->b_npages = pidx;
3726 bp->b_saveaddr = bp->b_data;
3727 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3728 return(0);
3729}
3730
3731/*
3732 * Free the io map PTEs associated with this IO operation.
3733 * We also invalidate the TLB entries and restore the original b_addr.
3734 */
3735void
3736vunmapbuf(struct buf *bp)
3737{
3738 int pidx;
3739 int npages;
3740
3741 npages = bp->b_npages;
3742 pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3743 npages);
3744 vm_page_lock_queues();
3745 for (pidx = 0; pidx < npages; pidx++)
3746 vm_page_unhold(bp->b_pages[pidx]);
3747 vm_page_unlock_queues();
3748
3749 bp->b_data = bp->b_saveaddr;
3750}
3751
3752void
3753bdone(struct buf *bp)
3754{
3755 mtx_lock(&bdonelock);
3756 bp->b_flags |= B_DONE;
3757 wakeup(bp);
3758 mtx_unlock(&bdonelock);
3759}
3760
3761void
3762bwait(struct buf *bp, u_char pri, const char *wchan)
3763{
3764 mtx_lock(&bdonelock);
3765 while ((bp->b_flags & B_DONE) == 0)
3766 msleep(bp, &bdonelock, pri, wchan, 0);
3767 mtx_unlock(&bdonelock);
3768}
3769
3770#include "opt_ddb.h"
3771#ifdef DDB
3772#include <ddb/ddb.h>
3773
3774/* DDB command to show buffer data */
3775DB_SHOW_COMMAND(buffer, db_show_buffer)
3776{
3777 /* get args */
3778 struct buf *bp = (struct buf *)addr;
3779
3780 if (!have_addr) {
3781 db_printf("usage: show buffer <addr>\n");
3782 return;
3783 }
3784
3785 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3786 db_printf(
3787 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3788 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n",
3789 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3790 major(bp->b_dev), minor(bp->b_dev), bp->b_data,
3791 (intmax_t)bp->b_blkno);
3792 if (bp->b_npages) {
3793 int i;
3794 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3795 for (i = 0; i < bp->b_npages; i++) {
3796 vm_page_t m;
3797 m = bp->b_pages[i];
3798 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3799 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3800 if ((i + 1) < bp->b_npages)
3801 db_printf(",");
3802 }
3803 db_printf("\n");
3804 }
3805}
3806#endif /* DDB */
| 1921 bp->b_vp = NULL;
1922 bp->b_blkno = bp->b_lblkno = 0;
1923 bp->b_offset = NOOFFSET;
1924 bp->b_iodone = 0;
1925 bp->b_error = 0;
1926 bp->b_resid = 0;
1927 bp->b_bcount = 0;
1928 bp->b_npages = 0;
1929 bp->b_dirtyoff = bp->b_dirtyend = 0;
1930 bp->b_magic = B_MAGIC_BIO;
1931 bp->b_op = &buf_ops_bio;
1932 bp->b_object = NULL;
1933
1934 LIST_INIT(&bp->b_dep);
1935
1936 /*
1937 * If we are defragging then free the buffer.
1938 */
1939 if (defrag) {
1940 bp->b_flags |= B_INVAL;
1941 bfreekva(bp);
1942 brelse(bp);
1943 defrag = 0;
1944 goto restart;
1945 }
1946
1947 /*
1948 * If we are overcomitted then recover the buffer and its
1949 * KVM space. This occurs in rare situations when multiple
1950 * processes are blocked in getnewbuf() or allocbuf().
1951 */
1952 if (bufspace >= hibufspace)
1953 flushingbufs = 1;
1954 if (flushingbufs && bp->b_kvasize != 0) {
1955 bp->b_flags |= B_INVAL;
1956 bfreekva(bp);
1957 brelse(bp);
1958 goto restart;
1959 }
1960 if (bufspace < lobufspace)
1961 flushingbufs = 0;
1962 break;
1963 }
1964
1965 /*
1966 * If we exhausted our list, sleep as appropriate. We may have to
1967 * wakeup various daemons and write out some dirty buffers.
1968 *
1969 * Generally we are sleeping due to insufficient buffer space.
1970 */
1971
1972 if (bp == NULL) {
1973 int flags;
1974 char *waitmsg;
1975
1976 mtx_unlock(&bqlock);
1977 if (defrag) {
1978 flags = VFS_BIO_NEED_BUFSPACE;
1979 waitmsg = "nbufkv";
1980 } else if (bufspace >= hibufspace) {
1981 waitmsg = "nbufbs";
1982 flags = VFS_BIO_NEED_BUFSPACE;
1983 } else {
1984 waitmsg = "newbuf";
1985 flags = VFS_BIO_NEED_ANY;
1986 }
1987
1988 bd_speedup(); /* heeeelp */
1989
1990 mtx_lock(&nblock);
1991 needsbuffer |= flags;
1992 while (needsbuffer & flags) {
1993 if (msleep(&needsbuffer, &nblock,
1994 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
1995 mtx_unlock(&nblock);
1996 return (NULL);
1997 }
1998 }
1999 mtx_unlock(&nblock);
2000 } else {
2001 /*
2002 * We finally have a valid bp. We aren't quite out of the
2003 * woods, we still have to reserve kva space. In order
2004 * to keep fragmentation sane we only allocate kva in
2005 * BKVASIZE chunks.
2006 */
2007 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2008
2009 if (maxsize != bp->b_kvasize) {
2010 vm_offset_t addr = 0;
2011
2012 bfreekva(bp);
2013
2014 if (vm_map_findspace(buffer_map,
2015 vm_map_min(buffer_map), maxsize, &addr)) {
2016 /*
2017 * Uh oh. Buffer map is to fragmented. We
2018 * must defragment the map.
2019 */
2020 atomic_add_int(&bufdefragcnt, 1);
2021 defrag = 1;
2022 bp->b_flags |= B_INVAL;
2023 brelse(bp);
2024 goto restart;
2025 }
2026 if (addr) {
2027 vm_map_insert(buffer_map, NULL, 0,
2028 addr, addr + maxsize,
2029 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2030
2031 bp->b_kvabase = (caddr_t) addr;
2032 bp->b_kvasize = maxsize;
2033 atomic_add_int(&bufspace, bp->b_kvasize);
2034 atomic_add_int(&bufreusecnt, 1);
2035 }
2036 }
2037 bp->b_saveaddr = bp->b_kvabase;
2038 bp->b_data = bp->b_saveaddr;
2039 }
2040 return(bp);
2041}
2042
2043/*
2044 * buf_daemon:
2045 *
2046 * buffer flushing daemon. Buffers are normally flushed by the
2047 * update daemon but if it cannot keep up this process starts to
2048 * take the load in an attempt to prevent getnewbuf() from blocking.
2049 */
2050
2051static struct kproc_desc buf_kp = {
2052 "bufdaemon",
2053 buf_daemon,
2054 &bufdaemonproc
2055};
2056SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
2057
2058static void
2059buf_daemon()
2060{
2061 int s;
2062
2063 mtx_lock(&Giant);
2064
2065 /*
2066 * This process needs to be suspended prior to shutdown sync.
2067 */
2068 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2069 SHUTDOWN_PRI_LAST);
2070
2071 /*
2072 * This process is allowed to take the buffer cache to the limit
2073 */
2074 s = splbio();
2075 mtx_lock(&bdlock);
2076
2077 for (;;) {
2078 bd_request = 0;
2079 mtx_unlock(&bdlock);
2080
2081 kthread_suspend_check(bufdaemonproc);
2082
2083 /*
2084 * Do the flush. Limit the amount of in-transit I/O we
2085 * allow to build up, otherwise we would completely saturate
2086 * the I/O system. Wakeup any waiting processes before we
2087 * normally would so they can run in parallel with our drain.
2088 */
2089 while (numdirtybuffers > lodirtybuffers) {
2090 if (flushbufqueues(0) == 0) {
2091 /*
2092 * Could not find any buffers without rollback
2093 * dependencies, so just write the first one
2094 * in the hopes of eventually making progress.
2095 */
2096 flushbufqueues(1);
2097 break;
2098 }
2099 waitrunningbufspace();
2100 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2101 }
2102
2103 /*
2104 * Only clear bd_request if we have reached our low water
2105 * mark. The buf_daemon normally waits 1 second and
2106 * then incrementally flushes any dirty buffers that have
2107 * built up, within reason.
2108 *
2109 * If we were unable to hit our low water mark and couldn't
2110 * find any flushable buffers, we sleep half a second.
2111 * Otherwise we loop immediately.
2112 */
2113 mtx_lock(&bdlock);
2114 if (numdirtybuffers <= lodirtybuffers) {
2115 /*
2116 * We reached our low water mark, reset the
2117 * request and sleep until we are needed again.
2118 * The sleep is just so the suspend code works.
2119 */
2120 bd_request = 0;
2121 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2122 } else {
2123 /*
2124 * We couldn't find any flushable dirty buffers but
2125 * still have too many dirty buffers, we
2126 * have to sleep and try again. (rare)
2127 */
2128 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2129 }
2130 }
2131}
2132
2133/*
2134 * flushbufqueues:
2135 *
2136 * Try to flush a buffer in the dirty queue. We must be careful to
2137 * free up B_INVAL buffers instead of write them, which NFS is
2138 * particularly sensitive to.
2139 */
2140int flushwithdeps = 0;
2141SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2142 0, "Number of buffers flushed with dependecies that require rollbacks");
2143static int
2144flushbufqueues(int flushdeps)
2145{
2146 struct thread *td = curthread;
2147 struct vnode *vp;
2148 struct mount *mp;
2149 struct buf *bp;
2150 int hasdeps;
2151
2152 mtx_lock(&bqlock);
2153 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) {
2154 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2155 continue;
2156 KASSERT((bp->b_flags & B_DELWRI),
2157 ("unexpected clean buffer %p", bp));
2158 VI_LOCK(bp->b_vp);
2159 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
2160 VI_UNLOCK(bp->b_vp);
2161 BUF_UNLOCK(bp);
2162 continue;
2163 }
2164 VI_UNLOCK(bp->b_vp);
2165 if (bp->b_flags & B_INVAL) {
2166 bremfreel(bp);
2167 mtx_unlock(&bqlock);
2168 brelse(bp);
2169 return (1);
2170 }
2171
2172 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) {
2173 if (flushdeps == 0) {
2174 BUF_UNLOCK(bp);
2175 continue;
2176 }
2177 hasdeps = 1;
2178 } else
2179 hasdeps = 0;
2180 /*
2181 * We must hold the lock on a vnode before writing
2182 * one of its buffers. Otherwise we may confuse, or
2183 * in the case of a snapshot vnode, deadlock the
2184 * system.
2185 *
2186 * The lock order here is the reverse of the normal
2187 * of vnode followed by buf lock. This is ok because
2188 * the NOWAIT will prevent deadlock.
2189 */
2190 vp = bp->b_vp;
2191 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2192 BUF_UNLOCK(bp);
2193 continue;
2194 }
2195 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) {
2196 mtx_unlock(&bqlock);
2197 vfs_bio_awrite(bp);
2198 vn_finished_write(mp);
2199 VOP_UNLOCK(vp, 0, td);
2200 flushwithdeps += hasdeps;
2201 return (1);
2202 }
2203 vn_finished_write(mp);
2204 BUF_UNLOCK(bp);
2205 }
2206 mtx_unlock(&bqlock);
2207 return (0);
2208}
2209
2210/*
2211 * Check to see if a block is currently memory resident.
2212 */
2213struct buf *
2214incore(struct vnode * vp, daddr_t blkno)
2215{
2216 struct buf *bp;
2217
2218 int s = splbio();
2219 VI_LOCK(vp);
2220 bp = gbincore(vp, blkno);
2221 VI_UNLOCK(vp);
2222 splx(s);
2223 return (bp);
2224}
2225
2226/*
2227 * Returns true if no I/O is needed to access the
2228 * associated VM object. This is like incore except
2229 * it also hunts around in the VM system for the data.
2230 */
2231
2232int
2233inmem(struct vnode * vp, daddr_t blkno)
2234{
2235 vm_object_t obj;
2236 vm_offset_t toff, tinc, size;
2237 vm_page_t m;
2238 vm_ooffset_t off;
2239
2240 GIANT_REQUIRED;
2241 ASSERT_VOP_LOCKED(vp, "inmem");
2242
2243 if (incore(vp, blkno))
2244 return 1;
2245 if (vp->v_mount == NULL)
2246 return 0;
2247 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0)
2248 return 0;
2249
2250 size = PAGE_SIZE;
2251 if (size > vp->v_mount->mnt_stat.f_iosize)
2252 size = vp->v_mount->mnt_stat.f_iosize;
2253 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2254
2255 VM_OBJECT_LOCK(obj);
2256 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2257 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2258 if (!m)
2259 goto notinmem;
2260 tinc = size;
2261 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2262 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2263 if (vm_page_is_valid(m,
2264 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2265 goto notinmem;
2266 }
2267 VM_OBJECT_UNLOCK(obj);
2268 return 1;
2269
2270notinmem:
2271 VM_OBJECT_UNLOCK(obj);
2272 return (0);
2273}
2274
2275/*
2276 * vfs_setdirty:
2277 *
2278 * Sets the dirty range for a buffer based on the status of the dirty
2279 * bits in the pages comprising the buffer.
2280 *
2281 * The range is limited to the size of the buffer.
2282 *
2283 * This routine is primarily used by NFS, but is generalized for the
2284 * B_VMIO case.
2285 */
2286static void
2287vfs_setdirty(struct buf *bp)
2288{
2289 int i;
2290 vm_object_t object;
2291
2292 GIANT_REQUIRED;
2293 /*
2294 * Degenerate case - empty buffer
2295 */
2296
2297 if (bp->b_bufsize == 0)
2298 return;
2299
2300 /*
2301 * We qualify the scan for modified pages on whether the
2302 * object has been flushed yet. The OBJ_WRITEABLE flag
2303 * is not cleared simply by protecting pages off.
2304 */
2305
2306 if ((bp->b_flags & B_VMIO) == 0)
2307 return;
2308
2309 object = bp->b_pages[0]->object;
2310 VM_OBJECT_LOCK(object);
2311 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2312 printf("Warning: object %p writeable but not mightbedirty\n", object);
2313 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2314 printf("Warning: object %p mightbedirty but not writeable\n", object);
2315
2316 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2317 vm_offset_t boffset;
2318 vm_offset_t eoffset;
2319
2320 vm_page_lock_queues();
2321 /*
2322 * test the pages to see if they have been modified directly
2323 * by users through the VM system.
2324 */
2325 for (i = 0; i < bp->b_npages; i++)
2326 vm_page_test_dirty(bp->b_pages[i]);
2327
2328 /*
2329 * Calculate the encompassing dirty range, boffset and eoffset,
2330 * (eoffset - boffset) bytes.
2331 */
2332
2333 for (i = 0; i < bp->b_npages; i++) {
2334 if (bp->b_pages[i]->dirty)
2335 break;
2336 }
2337 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2338
2339 for (i = bp->b_npages - 1; i >= 0; --i) {
2340 if (bp->b_pages[i]->dirty) {
2341 break;
2342 }
2343 }
2344 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2345
2346 vm_page_unlock_queues();
2347 /*
2348 * Fit it to the buffer.
2349 */
2350
2351 if (eoffset > bp->b_bcount)
2352 eoffset = bp->b_bcount;
2353
2354 /*
2355 * If we have a good dirty range, merge with the existing
2356 * dirty range.
2357 */
2358
2359 if (boffset < eoffset) {
2360 if (bp->b_dirtyoff > boffset)
2361 bp->b_dirtyoff = boffset;
2362 if (bp->b_dirtyend < eoffset)
2363 bp->b_dirtyend = eoffset;
2364 }
2365 }
2366 VM_OBJECT_UNLOCK(object);
2367}
2368
2369/*
2370 * getblk:
2371 *
2372 * Get a block given a specified block and offset into a file/device.
2373 * The buffers B_DONE bit will be cleared on return, making it almost
2374 * ready for an I/O initiation. B_INVAL may or may not be set on
2375 * return. The caller should clear B_INVAL prior to initiating a
2376 * READ.
2377 *
2378 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2379 * an existing buffer.
2380 *
2381 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2382 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2383 * and then cleared based on the backing VM. If the previous buffer is
2384 * non-0-sized but invalid, B_CACHE will be cleared.
2385 *
2386 * If getblk() must create a new buffer, the new buffer is returned with
2387 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2388 * case it is returned with B_INVAL clear and B_CACHE set based on the
2389 * backing VM.
2390 *
2391 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2392 * B_CACHE bit is clear.
2393 *
2394 * What this means, basically, is that the caller should use B_CACHE to
2395 * determine whether the buffer is fully valid or not and should clear
2396 * B_INVAL prior to issuing a read. If the caller intends to validate
2397 * the buffer by loading its data area with something, the caller needs
2398 * to clear B_INVAL. If the caller does this without issuing an I/O,
2399 * the caller should set B_CACHE ( as an optimization ), else the caller
2400 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2401 * a write attempt or if it was a successfull read. If the caller
2402 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2403 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2404 */
2405struct buf *
2406getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2407 int flags)
2408{
2409 struct buf *bp;
2410 int s;
2411 int error;
2412 ASSERT_VOP_LOCKED(vp, "getblk");
2413
2414 if (size > MAXBSIZE)
2415 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2416
2417 s = splbio();
2418loop:
2419 /*
2420 * Block if we are low on buffers. Certain processes are allowed
2421 * to completely exhaust the buffer cache.
2422 *
2423 * If this check ever becomes a bottleneck it may be better to
2424 * move it into the else, when gbincore() fails. At the moment
2425 * it isn't a problem.
2426 *
2427 * XXX remove if 0 sections (clean this up after its proven)
2428 */
2429 if (numfreebuffers == 0) {
2430 if (curthread == PCPU_GET(idlethread))
2431 return NULL;
2432 mtx_lock(&nblock);
2433 needsbuffer |= VFS_BIO_NEED_ANY;
2434 mtx_unlock(&nblock);
2435 }
2436
2437 VI_LOCK(vp);
2438 if ((bp = gbincore(vp, blkno))) {
2439 int lockflags;
2440 /*
2441 * Buffer is in-core. If the buffer is not busy, it must
2442 * be on a queue.
2443 */
2444 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2445
2446 if (flags & GB_LOCK_NOWAIT)
2447 lockflags |= LK_NOWAIT;
2448
2449 error = BUF_TIMELOCK(bp, lockflags,
2450 VI_MTX(vp), "getblk", slpflag, slptimeo);
2451
2452 /*
2453 * If we slept and got the lock we have to restart in case
2454 * the buffer changed identities.
2455 */
2456 if (error == ENOLCK)
2457 goto loop;
2458 /* We timed out or were interrupted. */
2459 else if (error)
2460 return (NULL);
2461
2462 /*
2463 * The buffer is locked. B_CACHE is cleared if the buffer is
2464 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2465 * and for a VMIO buffer B_CACHE is adjusted according to the
2466 * backing VM cache.
2467 */
2468 if (bp->b_flags & B_INVAL)
2469 bp->b_flags &= ~B_CACHE;
2470 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2471 bp->b_flags |= B_CACHE;
2472 bremfree(bp);
2473
2474 /*
2475 * check for size inconsistancies for non-VMIO case.
2476 */
2477
2478 if (bp->b_bcount != size) {
2479 if ((bp->b_flags & B_VMIO) == 0 ||
2480 (size > bp->b_kvasize)) {
2481 if (bp->b_flags & B_DELWRI) {
2482 bp->b_flags |= B_NOCACHE;
2483 bwrite(bp);
2484 } else {
2485 if ((bp->b_flags & B_VMIO) &&
2486 (LIST_FIRST(&bp->b_dep) == NULL)) {
2487 bp->b_flags |= B_RELBUF;
2488 brelse(bp);
2489 } else {
2490 bp->b_flags |= B_NOCACHE;
2491 bwrite(bp);
2492 }
2493 }
2494 goto loop;
2495 }
2496 }
2497
2498 /*
2499 * If the size is inconsistant in the VMIO case, we can resize
2500 * the buffer. This might lead to B_CACHE getting set or
2501 * cleared. If the size has not changed, B_CACHE remains
2502 * unchanged from its previous state.
2503 */
2504
2505 if (bp->b_bcount != size)
2506 allocbuf(bp, size);
2507
2508 KASSERT(bp->b_offset != NOOFFSET,
2509 ("getblk: no buffer offset"));
2510
2511 /*
2512 * A buffer with B_DELWRI set and B_CACHE clear must
2513 * be committed before we can return the buffer in
2514 * order to prevent the caller from issuing a read
2515 * ( due to B_CACHE not being set ) and overwriting
2516 * it.
2517 *
2518 * Most callers, including NFS and FFS, need this to
2519 * operate properly either because they assume they
2520 * can issue a read if B_CACHE is not set, or because
2521 * ( for example ) an uncached B_DELWRI might loop due
2522 * to softupdates re-dirtying the buffer. In the latter
2523 * case, B_CACHE is set after the first write completes,
2524 * preventing further loops.
2525 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2526 * above while extending the buffer, we cannot allow the
2527 * buffer to remain with B_CACHE set after the write
2528 * completes or it will represent a corrupt state. To
2529 * deal with this we set B_NOCACHE to scrap the buffer
2530 * after the write.
2531 *
2532 * We might be able to do something fancy, like setting
2533 * B_CACHE in bwrite() except if B_DELWRI is already set,
2534 * so the below call doesn't set B_CACHE, but that gets real
2535 * confusing. This is much easier.
2536 */
2537
2538 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2539 bp->b_flags |= B_NOCACHE;
2540 bwrite(bp);
2541 goto loop;
2542 }
2543
2544 splx(s);
2545 bp->b_flags &= ~B_DONE;
2546 } else {
2547 int bsize, maxsize, vmio;
2548 off_t offset;
2549
2550 /*
2551 * Buffer is not in-core, create new buffer. The buffer
2552 * returned by getnewbuf() is locked. Note that the returned
2553 * buffer is also considered valid (not marked B_INVAL).
2554 */
2555 VI_UNLOCK(vp);
2556 /*
2557 * If the user does not want us to create the buffer, bail out
2558 * here.
2559 */
2560 if (flags & GB_NOCREAT) {
2561 splx(s);
2562 return NULL;
2563 }
2564 if (vn_isdisk(vp, NULL))
2565 bsize = DEV_BSIZE;
2566 else if (vp->v_mountedhere)
2567 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2568 else if (vp->v_mount)
2569 bsize = vp->v_mount->mnt_stat.f_iosize;
2570 else
2571 bsize = size;
2572
2573 offset = blkno * bsize;
2574 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) &&
2575 (vp->v_vflag & VV_OBJBUF);
2576 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2577 maxsize = imax(maxsize, bsize);
2578
2579 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2580 if (slpflag || slptimeo) {
2581 splx(s);
2582 return NULL;
2583 }
2584 goto loop;
2585 }
2586
2587 /*
2588 * This code is used to make sure that a buffer is not
2589 * created while the getnewbuf routine is blocked.
2590 * This can be a problem whether the vnode is locked or not.
2591 * If the buffer is created out from under us, we have to
2592 * throw away the one we just created. There is now window
2593 * race because we are safely running at splbio() from the
2594 * point of the duplicate buffer creation through to here,
2595 * and we've locked the buffer.
2596 *
2597 * Note: this must occur before we associate the buffer
2598 * with the vp especially considering limitations in
2599 * the splay tree implementation when dealing with duplicate
2600 * lblkno's.
2601 */
2602 VI_LOCK(vp);
2603 if (gbincore(vp, blkno)) {
2604 VI_UNLOCK(vp);
2605 bp->b_flags |= B_INVAL;
2606 brelse(bp);
2607 goto loop;
2608 }
2609
2610 /*
2611 * Insert the buffer into the hash, so that it can
2612 * be found by incore.
2613 */
2614 bp->b_blkno = bp->b_lblkno = blkno;
2615 bp->b_offset = offset;
2616
2617 bgetvp(vp, bp);
2618 VI_UNLOCK(vp);
2619
2620 /*
2621 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2622 * buffer size starts out as 0, B_CACHE will be set by
2623 * allocbuf() for the VMIO case prior to it testing the
2624 * backing store for validity.
2625 */
2626
2627 if (vmio) {
2628 bp->b_flags |= B_VMIO;
2629#if defined(VFS_BIO_DEBUG)
2630 if (vn_canvmio(vp) != TRUE)
2631 printf("getblk: VMIO on vnode type %d\n",
2632 vp->v_type);
2633#endif
2634 VOP_GETVOBJECT(vp, &bp->b_object);
2635 } else {
2636 bp->b_flags &= ~B_VMIO;
2637 bp->b_object = NULL;
2638 }
2639
2640 allocbuf(bp, size);
2641
2642 splx(s);
2643 bp->b_flags &= ~B_DONE;
2644 }
2645 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp));
2646 return (bp);
2647}
2648
2649/*
2650 * Get an empty, disassociated buffer of given size. The buffer is initially
2651 * set to B_INVAL.
2652 */
2653struct buf *
2654geteblk(int size)
2655{
2656 struct buf *bp;
2657 int s;
2658 int maxsize;
2659
2660 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2661
2662 s = splbio();
2663 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2664 continue;
2665 splx(s);
2666 allocbuf(bp, size);
2667 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2668 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp));
2669 return (bp);
2670}
2671
2672
2673/*
2674 * This code constitutes the buffer memory from either anonymous system
2675 * memory (in the case of non-VMIO operations) or from an associated
2676 * VM object (in the case of VMIO operations). This code is able to
2677 * resize a buffer up or down.
2678 *
2679 * Note that this code is tricky, and has many complications to resolve
2680 * deadlock or inconsistant data situations. Tread lightly!!!
2681 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2682 * the caller. Calling this code willy nilly can result in the loss of data.
2683 *
2684 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2685 * B_CACHE for the non-VMIO case.
2686 */
2687
2688int
2689allocbuf(struct buf *bp, int size)
2690{
2691 int newbsize, mbsize;
2692 int i;
2693
2694 GIANT_REQUIRED;
2695
2696 if (BUF_REFCNT(bp) == 0)
2697 panic("allocbuf: buffer not busy");
2698
2699 if (bp->b_kvasize < size)
2700 panic("allocbuf: buffer too small");
2701
2702 if ((bp->b_flags & B_VMIO) == 0) {
2703 caddr_t origbuf;
2704 int origbufsize;
2705 /*
2706 * Just get anonymous memory from the kernel. Don't
2707 * mess with B_CACHE.
2708 */
2709 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2710 if (bp->b_flags & B_MALLOC)
2711 newbsize = mbsize;
2712 else
2713 newbsize = round_page(size);
2714
2715 if (newbsize < bp->b_bufsize) {
2716 /*
2717 * malloced buffers are not shrunk
2718 */
2719 if (bp->b_flags & B_MALLOC) {
2720 if (newbsize) {
2721 bp->b_bcount = size;
2722 } else {
2723 free(bp->b_data, M_BIOBUF);
2724 if (bp->b_bufsize) {
2725 atomic_subtract_int(
2726 &bufmallocspace,
2727 bp->b_bufsize);
2728 bufspacewakeup();
2729 bp->b_bufsize = 0;
2730 }
2731 bp->b_saveaddr = bp->b_kvabase;
2732 bp->b_data = bp->b_saveaddr;
2733 bp->b_bcount = 0;
2734 bp->b_flags &= ~B_MALLOC;
2735 }
2736 return 1;
2737 }
2738 vm_hold_free_pages(
2739 bp,
2740 (vm_offset_t) bp->b_data + newbsize,
2741 (vm_offset_t) bp->b_data + bp->b_bufsize);
2742 } else if (newbsize > bp->b_bufsize) {
2743 /*
2744 * We only use malloced memory on the first allocation.
2745 * and revert to page-allocated memory when the buffer
2746 * grows.
2747 */
2748 /*
2749 * There is a potential smp race here that could lead
2750 * to bufmallocspace slightly passing the max. It
2751 * is probably extremely rare and not worth worrying
2752 * over.
2753 */
2754 if ( (bufmallocspace < maxbufmallocspace) &&
2755 (bp->b_bufsize == 0) &&
2756 (mbsize <= PAGE_SIZE/2)) {
2757
2758 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2759 bp->b_bufsize = mbsize;
2760 bp->b_bcount = size;
2761 bp->b_flags |= B_MALLOC;
2762 atomic_add_int(&bufmallocspace, mbsize);
2763 return 1;
2764 }
2765 origbuf = NULL;
2766 origbufsize = 0;
2767 /*
2768 * If the buffer is growing on its other-than-first allocation,
2769 * then we revert to the page-allocation scheme.
2770 */
2771 if (bp->b_flags & B_MALLOC) {
2772 origbuf = bp->b_data;
2773 origbufsize = bp->b_bufsize;
2774 bp->b_data = bp->b_kvabase;
2775 if (bp->b_bufsize) {
2776 atomic_subtract_int(&bufmallocspace,
2777 bp->b_bufsize);
2778 bufspacewakeup();
2779 bp->b_bufsize = 0;
2780 }
2781 bp->b_flags &= ~B_MALLOC;
2782 newbsize = round_page(newbsize);
2783 }
2784 vm_hold_load_pages(
2785 bp,
2786 (vm_offset_t) bp->b_data + bp->b_bufsize,
2787 (vm_offset_t) bp->b_data + newbsize);
2788 if (origbuf) {
2789 bcopy(origbuf, bp->b_data, origbufsize);
2790 free(origbuf, M_BIOBUF);
2791 }
2792 }
2793 } else {
2794 int desiredpages;
2795
2796 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2797 desiredpages = (size == 0) ? 0 :
2798 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2799
2800 if (bp->b_flags & B_MALLOC)
2801 panic("allocbuf: VMIO buffer can't be malloced");
2802 /*
2803 * Set B_CACHE initially if buffer is 0 length or will become
2804 * 0-length.
2805 */
2806 if (size == 0 || bp->b_bufsize == 0)
2807 bp->b_flags |= B_CACHE;
2808
2809 if (newbsize < bp->b_bufsize) {
2810 /*
2811 * DEV_BSIZE aligned new buffer size is less then the
2812 * DEV_BSIZE aligned existing buffer size. Figure out
2813 * if we have to remove any pages.
2814 */
2815 if (desiredpages < bp->b_npages) {
2816 vm_page_t m;
2817
2818 vm_page_lock_queues();
2819 for (i = desiredpages; i < bp->b_npages; i++) {
2820 /*
2821 * the page is not freed here -- it
2822 * is the responsibility of
2823 * vnode_pager_setsize
2824 */
2825 m = bp->b_pages[i];
2826 KASSERT(m != bogus_page,
2827 ("allocbuf: bogus page found"));
2828 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2829 vm_page_lock_queues();
2830
2831 bp->b_pages[i] = NULL;
2832 vm_page_unwire(m, 0);
2833 }
2834 vm_page_unlock_queues();
2835 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2836 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2837 bp->b_npages = desiredpages;
2838 }
2839 } else if (size > bp->b_bcount) {
2840 /*
2841 * We are growing the buffer, possibly in a
2842 * byte-granular fashion.
2843 */
2844 struct vnode *vp;
2845 vm_object_t obj;
2846 vm_offset_t toff;
2847 vm_offset_t tinc;
2848
2849 /*
2850 * Step 1, bring in the VM pages from the object,
2851 * allocating them if necessary. We must clear
2852 * B_CACHE if these pages are not valid for the
2853 * range covered by the buffer.
2854 */
2855
2856 vp = bp->b_vp;
2857 obj = bp->b_object;
2858
2859 VM_OBJECT_LOCK(obj);
2860 while (bp->b_npages < desiredpages) {
2861 vm_page_t m;
2862 vm_pindex_t pi;
2863
2864 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2865 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2866 /*
2867 * note: must allocate system pages
2868 * since blocking here could intefere
2869 * with paging I/O, no matter which
2870 * process we are.
2871 */
2872 m = vm_page_alloc(obj, pi,
2873 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
2874 if (m == NULL) {
2875 atomic_add_int(&vm_pageout_deficit,
2876 desiredpages - bp->b_npages);
2877 VM_OBJECT_UNLOCK(obj);
2878 VM_WAIT;
2879 VM_OBJECT_LOCK(obj);
2880 } else {
2881 vm_page_lock_queues();
2882 vm_page_wakeup(m);
2883 vm_page_unlock_queues();
2884 bp->b_flags &= ~B_CACHE;
2885 bp->b_pages[bp->b_npages] = m;
2886 ++bp->b_npages;
2887 }
2888 continue;
2889 }
2890
2891 /*
2892 * We found a page. If we have to sleep on it,
2893 * retry because it might have gotten freed out
2894 * from under us.
2895 *
2896 * We can only test PG_BUSY here. Blocking on
2897 * m->busy might lead to a deadlock:
2898 *
2899 * vm_fault->getpages->cluster_read->allocbuf
2900 *
2901 */
2902 vm_page_lock_queues();
2903 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
2904 continue;
2905
2906 /*
2907 * We have a good page. Should we wakeup the
2908 * page daemon?
2909 */
2910 if ((curproc != pageproc) &&
2911 ((m->queue - m->pc) == PQ_CACHE) &&
2912 ((cnt.v_free_count + cnt.v_cache_count) <
2913 (cnt.v_free_min + cnt.v_cache_min))) {
2914 pagedaemon_wakeup();
2915 }
2916 vm_page_wire(m);
2917 vm_page_unlock_queues();
2918 bp->b_pages[bp->b_npages] = m;
2919 ++bp->b_npages;
2920 }
2921
2922 /*
2923 * Step 2. We've loaded the pages into the buffer,
2924 * we have to figure out if we can still have B_CACHE
2925 * set. Note that B_CACHE is set according to the
2926 * byte-granular range ( bcount and size ), new the
2927 * aligned range ( newbsize ).
2928 *
2929 * The VM test is against m->valid, which is DEV_BSIZE
2930 * aligned. Needless to say, the validity of the data
2931 * needs to also be DEV_BSIZE aligned. Note that this
2932 * fails with NFS if the server or some other client
2933 * extends the file's EOF. If our buffer is resized,
2934 * B_CACHE may remain set! XXX
2935 */
2936
2937 toff = bp->b_bcount;
2938 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2939
2940 while ((bp->b_flags & B_CACHE) && toff < size) {
2941 vm_pindex_t pi;
2942
2943 if (tinc > (size - toff))
2944 tinc = size - toff;
2945
2946 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2947 PAGE_SHIFT;
2948
2949 vfs_buf_test_cache(
2950 bp,
2951 bp->b_offset,
2952 toff,
2953 tinc,
2954 bp->b_pages[pi]
2955 );
2956 toff += tinc;
2957 tinc = PAGE_SIZE;
2958 }
2959 VM_OBJECT_UNLOCK(obj);
2960
2961 /*
2962 * Step 3, fixup the KVM pmap. Remember that
2963 * bp->b_data is relative to bp->b_offset, but
2964 * bp->b_offset may be offset into the first page.
2965 */
2966
2967 bp->b_data = (caddr_t)
2968 trunc_page((vm_offset_t)bp->b_data);
2969 pmap_qenter(
2970 (vm_offset_t)bp->b_data,
2971 bp->b_pages,
2972 bp->b_npages
2973 );
2974
2975 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2976 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2977 }
2978 }
2979 if (newbsize < bp->b_bufsize)
2980 bufspacewakeup();
2981 bp->b_bufsize = newbsize; /* actual buffer allocation */
2982 bp->b_bcount = size; /* requested buffer size */
2983 return 1;
2984}
2985
2986void
2987biodone(struct bio *bp)
2988{
2989 mtx_lock(&bdonelock);
2990 bp->bio_flags |= BIO_DONE;
2991 if (bp->bio_done == NULL)
2992 wakeup(bp);
2993 mtx_unlock(&bdonelock);
2994 if (bp->bio_done != NULL)
2995 bp->bio_done(bp);
2996}
2997
2998/*
2999 * Wait for a BIO to finish.
3000 *
3001 * XXX: resort to a timeout for now. The optimal locking (if any) for this
3002 * case is not yet clear.
3003 */
3004int
3005biowait(struct bio *bp, const char *wchan)
3006{
3007
3008 mtx_lock(&bdonelock);
3009 while ((bp->bio_flags & BIO_DONE) == 0)
3010 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10);
3011 mtx_unlock(&bdonelock);
3012 if (bp->bio_error != 0)
3013 return (bp->bio_error);
3014 if (!(bp->bio_flags & BIO_ERROR))
3015 return (0);
3016 return (EIO);
3017}
3018
3019void
3020biofinish(struct bio *bp, struct devstat *stat, int error)
3021{
3022
3023 if (error) {
3024 bp->bio_error = error;
3025 bp->bio_flags |= BIO_ERROR;
3026 }
3027 if (stat != NULL)
3028 devstat_end_transaction_bio(stat, bp);
3029 biodone(bp);
3030}
3031
3032/*
3033 * bufwait:
3034 *
3035 * Wait for buffer I/O completion, returning error status. The buffer
3036 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3037 * error and cleared.
3038 */
3039int
3040bufwait(register struct buf * bp)
3041{
3042 int s;
3043
3044 s = splbio();
3045 if (bp->b_iocmd == BIO_READ)
3046 bwait(bp, PRIBIO, "biord");
3047 else
3048 bwait(bp, PRIBIO, "biowr");
3049 splx(s);
3050 if (bp->b_flags & B_EINTR) {
3051 bp->b_flags &= ~B_EINTR;
3052 return (EINTR);
3053 }
3054 if (bp->b_ioflags & BIO_ERROR) {
3055 return (bp->b_error ? bp->b_error : EIO);
3056 } else {
3057 return (0);
3058 }
3059}
3060
3061 /*
3062 * Call back function from struct bio back up to struct buf.
3063 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY().
3064 */
3065static void
3066bufdonebio(struct bio *bp)
3067{
3068
3069 /* Device drivers may or may not hold giant, hold it here. */
3070 mtx_lock(&Giant);
3071 bufdone(bp->bio_caller2);
3072 mtx_unlock(&Giant);
3073}
3074
3075void
3076dev_strategy(struct buf *bp)
3077{
3078 struct cdevsw *csw;
3079
3080 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3081 panic("b_iocmd botch");
3082 bp->b_io.bio_done = bufdonebio;
3083 bp->b_io.bio_caller2 = bp;
3084 csw = devsw(bp->b_io.bio_dev);
3085 KASSERT(bp->b_io.bio_dev->si_refcount > 0,
3086 ("dev_strategy on un-referenced struct cdev *(%s)",
3087 devtoname(bp->b_io.bio_dev)));
3088 cdevsw_ref(csw);
3089 (*devsw(bp->b_io.bio_dev)->d_strategy)(&bp->b_io);
3090 cdevsw_rel(csw);
3091}
3092
3093/*
3094 * bufdone:
3095 *
3096 * Finish I/O on a buffer, optionally calling a completion function.
3097 * This is usually called from an interrupt so process blocking is
3098 * not allowed.
3099 *
3100 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3101 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3102 * assuming B_INVAL is clear.
3103 *
3104 * For the VMIO case, we set B_CACHE if the op was a read and no
3105 * read error occured, or if the op was a write. B_CACHE is never
3106 * set if the buffer is invalid or otherwise uncacheable.
3107 *
3108 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3109 * initiator to leave B_INVAL set to brelse the buffer out of existance
3110 * in the biodone routine.
3111 */
3112void
3113bufdone(struct buf *bp)
3114{
3115 int s;
3116 void (*biodone)(struct buf *);
3117
3118
3119 s = splbio();
3120
3121 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
3122 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3123
3124 bp->b_flags |= B_DONE;
3125 runningbufwakeup(bp);
3126
3127 if (bp->b_iocmd == BIO_DELETE) {
3128 brelse(bp);
3129 splx(s);
3130 return;
3131 }
3132
3133 if (bp->b_iocmd == BIO_WRITE) {
3134 vwakeup(bp);
3135 }
3136
3137 /* call optional completion function if requested */
3138 if (bp->b_iodone != NULL) {
3139 biodone = bp->b_iodone;
3140 bp->b_iodone = NULL;
3141 (*biodone) (bp);
3142 splx(s);
3143 return;
3144 }
3145 if (LIST_FIRST(&bp->b_dep) != NULL)
3146 buf_complete(bp);
3147
3148 if (bp->b_flags & B_VMIO) {
3149 int i;
3150 vm_ooffset_t foff;
3151 vm_page_t m;
3152 vm_object_t obj;
3153 int iosize;
3154 struct vnode *vp = bp->b_vp;
3155
3156 obj = bp->b_object;
3157
3158#if defined(VFS_BIO_DEBUG)
3159 mp_fixme("usecount and vflag accessed without locks.");
3160 if (vp->v_usecount == 0) {
3161 panic("biodone: zero vnode ref count");
3162 }
3163
3164 if ((vp->v_vflag & VV_OBJBUF) == 0) {
3165 panic("biodone: vnode is not setup for merged cache");
3166 }
3167#endif
3168
3169 foff = bp->b_offset;
3170 KASSERT(bp->b_offset != NOOFFSET,
3171 ("biodone: no buffer offset"));
3172
3173 VM_OBJECT_LOCK(obj);
3174#if defined(VFS_BIO_DEBUG)
3175 if (obj->paging_in_progress < bp->b_npages) {
3176 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3177 obj->paging_in_progress, bp->b_npages);
3178 }
3179#endif
3180
3181 /*
3182 * Set B_CACHE if the op was a normal read and no error
3183 * occured. B_CACHE is set for writes in the b*write()
3184 * routines.
3185 */
3186 iosize = bp->b_bcount - bp->b_resid;
3187 if (bp->b_iocmd == BIO_READ &&
3188 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3189 !(bp->b_ioflags & BIO_ERROR)) {
3190 bp->b_flags |= B_CACHE;
3191 }
3192 vm_page_lock_queues();
3193 for (i = 0; i < bp->b_npages; i++) {
3194 int bogusflag = 0;
3195 int resid;
3196
3197 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3198 if (resid > iosize)
3199 resid = iosize;
3200
3201 /*
3202 * cleanup bogus pages, restoring the originals
3203 */
3204 m = bp->b_pages[i];
3205 if (m == bogus_page) {
3206 bogusflag = 1;
3207 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3208 if (m == NULL)
3209 panic("biodone: page disappeared!");
3210 bp->b_pages[i] = m;
3211 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3212 }
3213#if defined(VFS_BIO_DEBUG)
3214 if (OFF_TO_IDX(foff) != m->pindex) {
3215 printf(
3216"biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3217 (intmax_t)foff, (uintmax_t)m->pindex);
3218 }
3219#endif
3220
3221 /*
3222 * In the write case, the valid and clean bits are
3223 * already changed correctly ( see bdwrite() ), so we
3224 * only need to do this here in the read case.
3225 */
3226 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3227 vfs_page_set_valid(bp, foff, i, m);
3228 }
3229
3230 /*
3231 * when debugging new filesystems or buffer I/O methods, this
3232 * is the most common error that pops up. if you see this, you
3233 * have not set the page busy flag correctly!!!
3234 */
3235 if (m->busy == 0) {
3236 printf("biodone: page busy < 0, "
3237 "pindex: %d, foff: 0x(%x,%x), "
3238 "resid: %d, index: %d\n",
3239 (int) m->pindex, (int)(foff >> 32),
3240 (int) foff & 0xffffffff, resid, i);
3241 if (!vn_isdisk(vp, NULL))
3242 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3243 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3244 (intmax_t) bp->b_lblkno,
3245 bp->b_flags, bp->b_npages);
3246 else
3247 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3248 (intmax_t) bp->b_lblkno,
3249 bp->b_flags, bp->b_npages);
3250 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3251 (u_long)m->valid, (u_long)m->dirty,
3252 m->wire_count);
3253 panic("biodone: page busy < 0\n");
3254 }
3255 vm_page_io_finish(m);
3256 vm_object_pip_subtract(obj, 1);
3257 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3258 iosize -= resid;
3259 }
3260 vm_page_unlock_queues();
3261 vm_object_pip_wakeupn(obj, 0);
3262 VM_OBJECT_UNLOCK(obj);
3263 }
3264
3265 /*
3266 * For asynchronous completions, release the buffer now. The brelse
3267 * will do a wakeup there if necessary - so no need to do a wakeup
3268 * here in the async case. The sync case always needs to do a wakeup.
3269 */
3270
3271 if (bp->b_flags & B_ASYNC) {
3272 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3273 brelse(bp);
3274 else
3275 bqrelse(bp);
3276 } else {
3277 bdone(bp);
3278 }
3279 splx(s);
3280}
3281
3282/*
3283 * This routine is called in lieu of iodone in the case of
3284 * incomplete I/O. This keeps the busy status for pages
3285 * consistant.
3286 */
3287void
3288vfs_unbusy_pages(struct buf * bp)
3289{
3290 int i;
3291
3292 runningbufwakeup(bp);
3293 if (bp->b_flags & B_VMIO) {
3294 vm_object_t obj;
3295
3296 obj = bp->b_object;
3297 VM_OBJECT_LOCK(obj);
3298 vm_page_lock_queues();
3299 for (i = 0; i < bp->b_npages; i++) {
3300 vm_page_t m = bp->b_pages[i];
3301
3302 if (m == bogus_page) {
3303 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3304 if (!m) {
3305 panic("vfs_unbusy_pages: page missing\n");
3306 }
3307 bp->b_pages[i] = m;
3308 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3309 }
3310 vm_object_pip_subtract(obj, 1);
3311 vm_page_io_finish(m);
3312 }
3313 vm_page_unlock_queues();
3314 vm_object_pip_wakeupn(obj, 0);
3315 VM_OBJECT_UNLOCK(obj);
3316 }
3317}
3318
3319/*
3320 * vfs_page_set_valid:
3321 *
3322 * Set the valid bits in a page based on the supplied offset. The
3323 * range is restricted to the buffer's size.
3324 *
3325 * This routine is typically called after a read completes.
3326 */
3327static void
3328vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3329{
3330 vm_ooffset_t soff, eoff;
3331
3332 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3333 /*
3334 * Start and end offsets in buffer. eoff - soff may not cross a
3335 * page boundry or cross the end of the buffer. The end of the
3336 * buffer, in this case, is our file EOF, not the allocation size
3337 * of the buffer.
3338 */
3339 soff = off;
3340 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3341 if (eoff > bp->b_offset + bp->b_bcount)
3342 eoff = bp->b_offset + bp->b_bcount;
3343
3344 /*
3345 * Set valid range. This is typically the entire buffer and thus the
3346 * entire page.
3347 */
3348 if (eoff > soff) {
3349 vm_page_set_validclean(
3350 m,
3351 (vm_offset_t) (soff & PAGE_MASK),
3352 (vm_offset_t) (eoff - soff)
3353 );
3354 }
3355}
3356
3357/*
3358 * This routine is called before a device strategy routine.
3359 * It is used to tell the VM system that paging I/O is in
3360 * progress, and treat the pages associated with the buffer
3361 * almost as being PG_BUSY. Also the object paging_in_progress
3362 * flag is handled to make sure that the object doesn't become
3363 * inconsistant.
3364 *
3365 * Since I/O has not been initiated yet, certain buffer flags
3366 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3367 * and should be ignored.
3368 */
3369void
3370vfs_busy_pages(struct buf * bp, int clear_modify)
3371{
3372 int i, bogus;
3373
3374 if (bp->b_flags & B_VMIO) {
3375 vm_object_t obj;
3376 vm_ooffset_t foff;
3377
3378 obj = bp->b_object;
3379 foff = bp->b_offset;
3380 KASSERT(bp->b_offset != NOOFFSET,
3381 ("vfs_busy_pages: no buffer offset"));
3382 vfs_setdirty(bp);
3383 VM_OBJECT_LOCK(obj);
3384retry:
3385 vm_page_lock_queues();
3386 for (i = 0; i < bp->b_npages; i++) {
3387 vm_page_t m = bp->b_pages[i];
3388
3389 if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3390 goto retry;
3391 }
3392 bogus = 0;
3393 for (i = 0; i < bp->b_npages; i++) {
3394 vm_page_t m = bp->b_pages[i];
3395
3396 if ((bp->b_flags & B_CLUSTER) == 0) {
3397 vm_object_pip_add(obj, 1);
3398 vm_page_io_start(m);
3399 }
3400 /*
3401 * When readying a buffer for a read ( i.e
3402 * clear_modify == 0 ), it is important to do
3403 * bogus_page replacement for valid pages in
3404 * partially instantiated buffers. Partially
3405 * instantiated buffers can, in turn, occur when
3406 * reconstituting a buffer from its VM backing store
3407 * base. We only have to do this if B_CACHE is
3408 * clear ( which causes the I/O to occur in the
3409 * first place ). The replacement prevents the read
3410 * I/O from overwriting potentially dirty VM-backed
3411 * pages. XXX bogus page replacement is, uh, bogus.
3412 * It may not work properly with small-block devices.
3413 * We need to find a better way.
3414 */
3415 pmap_remove_all(m);
3416 if (clear_modify)
3417 vfs_page_set_valid(bp, foff, i, m);
3418 else if (m->valid == VM_PAGE_BITS_ALL &&
3419 (bp->b_flags & B_CACHE) == 0) {
3420 bp->b_pages[i] = bogus_page;
3421 bogus++;
3422 }
3423 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3424 }
3425 vm_page_unlock_queues();
3426 VM_OBJECT_UNLOCK(obj);
3427 if (bogus)
3428 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3429 }
3430}
3431
3432/*
3433 * Tell the VM system that the pages associated with this buffer
3434 * are clean. This is used for delayed writes where the data is
3435 * going to go to disk eventually without additional VM intevention.
3436 *
3437 * Note that while we only really need to clean through to b_bcount, we
3438 * just go ahead and clean through to b_bufsize.
3439 */
3440static void
3441vfs_clean_pages(struct buf * bp)
3442{
3443 int i;
3444
3445 if (bp->b_flags & B_VMIO) {
3446 vm_ooffset_t foff;
3447
3448 foff = bp->b_offset;
3449 KASSERT(bp->b_offset != NOOFFSET,
3450 ("vfs_clean_pages: no buffer offset"));
3451 VM_OBJECT_LOCK(bp->b_object);
3452 vm_page_lock_queues();
3453 for (i = 0; i < bp->b_npages; i++) {
3454 vm_page_t m = bp->b_pages[i];
3455 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3456 vm_ooffset_t eoff = noff;
3457
3458 if (eoff > bp->b_offset + bp->b_bufsize)
3459 eoff = bp->b_offset + bp->b_bufsize;
3460 vfs_page_set_valid(bp, foff, i, m);
3461 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3462 foff = noff;
3463 }
3464 vm_page_unlock_queues();
3465 VM_OBJECT_UNLOCK(bp->b_object);
3466 }
3467}
3468
3469/*
3470 * vfs_bio_set_validclean:
3471 *
3472 * Set the range within the buffer to valid and clean. The range is
3473 * relative to the beginning of the buffer, b_offset. Note that b_offset
3474 * itself may be offset from the beginning of the first page.
3475 *
3476 */
3477
3478void
3479vfs_bio_set_validclean(struct buf *bp, int base, int size)
3480{
3481 if (bp->b_flags & B_VMIO) {
3482 int i;
3483 int n;
3484
3485 /*
3486 * Fixup base to be relative to beginning of first page.
3487 * Set initial n to be the maximum number of bytes in the
3488 * first page that can be validated.
3489 */
3490
3491 base += (bp->b_offset & PAGE_MASK);
3492 n = PAGE_SIZE - (base & PAGE_MASK);
3493
3494 VM_OBJECT_LOCK(bp->b_object);
3495 vm_page_lock_queues();
3496 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3497 vm_page_t m = bp->b_pages[i];
3498
3499 if (n > size)
3500 n = size;
3501
3502 vm_page_set_validclean(m, base & PAGE_MASK, n);
3503 base += n;
3504 size -= n;
3505 n = PAGE_SIZE;
3506 }
3507 vm_page_unlock_queues();
3508 VM_OBJECT_UNLOCK(bp->b_object);
3509 }
3510}
3511
3512/*
3513 * vfs_bio_clrbuf:
3514 *
3515 * clear a buffer. This routine essentially fakes an I/O, so we need
3516 * to clear BIO_ERROR and B_INVAL.
3517 *
3518 * Note that while we only theoretically need to clear through b_bcount,
3519 * we go ahead and clear through b_bufsize.
3520 */
3521
3522void
3523vfs_bio_clrbuf(struct buf *bp)
3524{
3525 int i, j, mask = 0;
3526 caddr_t sa, ea;
3527
3528 GIANT_REQUIRED;
3529
3530 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3531 bp->b_flags &= ~B_INVAL;
3532 bp->b_ioflags &= ~BIO_ERROR;
3533 VM_OBJECT_LOCK(bp->b_object);
3534 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3535 (bp->b_offset & PAGE_MASK) == 0) {
3536 if (bp->b_pages[0] == bogus_page)
3537 goto unlock;
3538 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3539 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3540 if ((bp->b_pages[0]->valid & mask) == mask)
3541 goto unlock;
3542 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3543 ((bp->b_pages[0]->valid & mask) == 0)) {
3544 bzero(bp->b_data, bp->b_bufsize);
3545 bp->b_pages[0]->valid |= mask;
3546 goto unlock;
3547 }
3548 }
3549 ea = sa = bp->b_data;
3550 for(i=0;i<bp->b_npages;i++,sa=ea) {
3551 if (bp->b_pages[i] == bogus_page)
3552 continue;
3553 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3554 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3555 ea = (caddr_t)(vm_offset_t)ulmin(
3556 (u_long)(vm_offset_t)ea,
3557 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3558 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3559 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3560 if ((bp->b_pages[i]->valid & mask) == mask)
3561 continue;
3562 if ((bp->b_pages[i]->valid & mask) == 0) {
3563 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3564 bzero(sa, ea - sa);
3565 }
3566 } else {
3567 for (; sa < ea; sa += DEV_BSIZE, j++) {
3568 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3569 (bp->b_pages[i]->valid & (1<<j)) == 0)
3570 bzero(sa, DEV_BSIZE);
3571 }
3572 }
3573 bp->b_pages[i]->valid |= mask;
3574 }
3575unlock:
3576 VM_OBJECT_UNLOCK(bp->b_object);
3577 bp->b_resid = 0;
3578 } else {
3579 clrbuf(bp);
3580 }
3581}
3582
3583/*
3584 * vm_hold_load_pages and vm_hold_free_pages get pages into
3585 * a buffers address space. The pages are anonymous and are
3586 * not associated with a file object.
3587 */
3588static void
3589vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3590{
3591 vm_offset_t pg;
3592 vm_page_t p;
3593 int index;
3594
3595 to = round_page(to);
3596 from = round_page(from);
3597 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3598
3599 VM_OBJECT_LOCK(kernel_object);
3600 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3601tryagain:
3602 /*
3603 * note: must allocate system pages since blocking here
3604 * could intefere with paging I/O, no matter which
3605 * process we are.
3606 */
3607 p = vm_page_alloc(kernel_object,
3608 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3609 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3610 if (!p) {
3611 atomic_add_int(&vm_pageout_deficit,
3612 (to - pg) >> PAGE_SHIFT);
3613 VM_OBJECT_UNLOCK(kernel_object);
3614 VM_WAIT;
3615 VM_OBJECT_LOCK(kernel_object);
3616 goto tryagain;
3617 }
3618 p->valid = VM_PAGE_BITS_ALL;
3619 pmap_qenter(pg, &p, 1);
3620 bp->b_pages[index] = p;
3621 vm_page_lock_queues();
3622 vm_page_wakeup(p);
3623 vm_page_unlock_queues();
3624 }
3625 VM_OBJECT_UNLOCK(kernel_object);
3626 bp->b_npages = index;
3627}
3628
3629/* Return pages associated with this buf to the vm system */
3630static void
3631vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3632{
3633 vm_offset_t pg;
3634 vm_page_t p;
3635 int index, newnpages;
3636
3637 GIANT_REQUIRED;
3638
3639 from = round_page(from);
3640 to = round_page(to);
3641 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3642
3643 VM_OBJECT_LOCK(kernel_object);
3644 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3645 p = bp->b_pages[index];
3646 if (p && (index < bp->b_npages)) {
3647 if (p->busy) {
3648 printf(
3649 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3650 (intmax_t)bp->b_blkno,
3651 (intmax_t)bp->b_lblkno);
3652 }
3653 bp->b_pages[index] = NULL;
3654 pmap_qremove(pg, 1);
3655 vm_page_lock_queues();
3656 vm_page_busy(p);
3657 vm_page_unwire(p, 0);
3658 vm_page_free(p);
3659 vm_page_unlock_queues();
3660 }
3661 }
3662 VM_OBJECT_UNLOCK(kernel_object);
3663 bp->b_npages = newnpages;
3664}
3665
3666/*
3667 * Map an IO request into kernel virtual address space.
3668 *
3669 * All requests are (re)mapped into kernel VA space.
3670 * Notice that we use b_bufsize for the size of the buffer
3671 * to be mapped. b_bcount might be modified by the driver.
3672 *
3673 * Note that even if the caller determines that the address space should
3674 * be valid, a race or a smaller-file mapped into a larger space may
3675 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3676 * check the return value.
3677 */
3678int
3679vmapbuf(struct buf *bp)
3680{
3681 caddr_t addr, kva;
3682 vm_prot_t prot;
3683 int pidx, i;
3684 struct vm_page *m;
3685 struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3686
3687 GIANT_REQUIRED;
3688
3689 if (bp->b_bufsize < 0)
3690 return (-1);
3691 prot = (bp->b_iocmd == BIO_READ) ? VM_PROT_READ | VM_PROT_WRITE :
3692 VM_PROT_READ;
3693 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3694 addr < bp->b_data + bp->b_bufsize;
3695 addr += PAGE_SIZE, pidx++) {
3696 /*
3697 * Do the vm_fault if needed; do the copy-on-write thing
3698 * when reading stuff off device into memory.
3699 *
3700 * NOTE! Must use pmap_extract() because addr may be in
3701 * the userland address space, and kextract is only guarenteed
3702 * to work for the kernland address space (see: sparc64 port).
3703 */
3704retry:
3705 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3706 prot) < 0) {
3707 vm_page_lock_queues();
3708 for (i = 0; i < pidx; ++i) {
3709 vm_page_unhold(bp->b_pages[i]);
3710 bp->b_pages[i] = NULL;
3711 }
3712 vm_page_unlock_queues();
3713 return(-1);
3714 }
3715 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3716 if (m == NULL)
3717 goto retry;
3718 bp->b_pages[pidx] = m;
3719 }
3720 if (pidx > btoc(MAXPHYS))
3721 panic("vmapbuf: mapped more than MAXPHYS");
3722 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3723
3724 kva = bp->b_saveaddr;
3725 bp->b_npages = pidx;
3726 bp->b_saveaddr = bp->b_data;
3727 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3728 return(0);
3729}
3730
3731/*
3732 * Free the io map PTEs associated with this IO operation.
3733 * We also invalidate the TLB entries and restore the original b_addr.
3734 */
3735void
3736vunmapbuf(struct buf *bp)
3737{
3738 int pidx;
3739 int npages;
3740
3741 npages = bp->b_npages;
3742 pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3743 npages);
3744 vm_page_lock_queues();
3745 for (pidx = 0; pidx < npages; pidx++)
3746 vm_page_unhold(bp->b_pages[pidx]);
3747 vm_page_unlock_queues();
3748
3749 bp->b_data = bp->b_saveaddr;
3750}
3751
3752void
3753bdone(struct buf *bp)
3754{
3755 mtx_lock(&bdonelock);
3756 bp->b_flags |= B_DONE;
3757 wakeup(bp);
3758 mtx_unlock(&bdonelock);
3759}
3760
3761void
3762bwait(struct buf *bp, u_char pri, const char *wchan)
3763{
3764 mtx_lock(&bdonelock);
3765 while ((bp->b_flags & B_DONE) == 0)
3766 msleep(bp, &bdonelock, pri, wchan, 0);
3767 mtx_unlock(&bdonelock);
3768}
3769
3770#include "opt_ddb.h"
3771#ifdef DDB
3772#include <ddb/ddb.h>
3773
3774/* DDB command to show buffer data */
3775DB_SHOW_COMMAND(buffer, db_show_buffer)
3776{
3777 /* get args */
3778 struct buf *bp = (struct buf *)addr;
3779
3780 if (!have_addr) {
3781 db_printf("usage: show buffer <addr>\n");
3782 return;
3783 }
3784
3785 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3786 db_printf(
3787 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3788 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n",
3789 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3790 major(bp->b_dev), minor(bp->b_dev), bp->b_data,
3791 (intmax_t)bp->b_blkno);
3792 if (bp->b_npages) {
3793 int i;
3794 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3795 for (i = 0; i < bp->b_npages; i++) {
3796 vm_page_t m;
3797 m = bp->b_pages[i];
3798 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3799 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3800 if ((i + 1) < bp->b_npages)
3801 db_printf(",");
3802 }
3803 db_printf("\n");
3804 }
3805}
3806#endif /* DDB */
|