kernel - lwkt_token revamp
[dragonfly.git] / sys / kern / vfs_bio.c
CommitLineData
984263bc
MD
1/*
2 * Copyright (c) 1994,1997 John S. Dyson
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
12 * John S. Dyson.
13 *
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
ab2200aa 15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $
984263bc
MD
16 */
17
18/*
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
23 *
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
27 *
28 * see man buf(9) for more info.
29 */
30
31#include <sys/param.h>
32#include <sys/systm.h>
33#include <sys/buf.h>
34#include <sys/conf.h>
35#include <sys/eventhandler.h>
36#include <sys/lock.h>
37#include <sys/malloc.h>
38#include <sys/mount.h>
39#include <sys/kernel.h>
40#include <sys/kthread.h>
41#include <sys/proc.h>
42#include <sys/reboot.h>
43#include <sys/resourcevar.h>
44#include <sys/sysctl.h>
45#include <sys/vmmeter.h>
46#include <sys/vnode.h>
8c72e3d5 47#include <sys/dsched.h>
3020e3be 48#include <sys/proc.h>
984263bc
MD
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>
c504e38e
MD
57#include <vm/vm_pager.h>
58#include <vm/swap_pager.h>
654a39f0 59
3020e3be 60#include <sys/buf2.h>
654a39f0 61#include <sys/thread2.h>
f832287e 62#include <sys/spinlock2.h>
684a93c4 63#include <sys/mplock2.h>
12e4aaff 64#include <vm/vm_page2.h>
984263bc 65
135bd6a8
MD
66#include "opt_ddb.h"
67#ifdef DDB
68#include <ddb/ddb.h>
69#endif
70
b3098c79
HP
71/*
72 * Buffer queues.
73 */
b3098c79
HP
74enum bufq_type {
75 BQUEUE_NONE, /* not on any queue */
76 BQUEUE_LOCKED, /* locked buffers */
77 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
78 BQUEUE_DIRTY, /* B_DELWRI buffers */
4b958e7b 79 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
b3098c79 80 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
4b958e7b
MD
81 BQUEUE_EMPTY, /* empty buffer headers */
82
83 BUFFER_QUEUES /* number of buffer queues */
b3098c79 84};
4b958e7b
MD
85
86typedef enum bufq_type bufq_type_t;
87
79eae878 88#define BD_WAKE_SIZE 16384
c4df9635
MD
89#define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
90
b3098c79 91TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
c3d1e862 92struct spinlock bufspin = SPINLOCK_INITIALIZER(&bufspin);
b3098c79 93
984263bc
MD
94static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
95
984263bc 96struct buf *buf; /* buffer header pool */
984263bc 97
c8e4131d 98static void vfs_clean_pages(struct buf *bp);
cb1cf930 99static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
0a8aee15 100static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
984263bc 101static void vfs_vmio_release(struct buf *bp);
4b958e7b 102static int flushbufqueues(bufq_type_t q);
4ecf7cc9 103static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
984263bc 104
868d24af 105static void bd_signal(int totalspace);
4b958e7b
MD
106static void buf_daemon(void);
107static void buf_daemon_hw(void);
c4df9635 108
984263bc
MD
109/*
110 * bogus page -- for I/O to/from partially complete buffers
111 * this is a temporary solution to the problem, but it is not
112 * really that bad. it would be better to split the buffer
113 * for input in the case of buffers partially already in memory,
114 * but the code is intricate enough already.
115 */
116vm_page_t bogus_page;
a0c36a34 117
460426e6
MD
118/*
119 * These are all static, but make the ones we export globals so we do
120 * not need to use compiler magic.
121 */
122int bufspace, maxbufspace,
984263bc
MD
123 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
124static int bufreusecnt, bufdefragcnt, buffreekvacnt;
984263bc 125static int lorunningspace, hirunningspace, runningbufreq;
868d24af 126int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
70ac7d6c 127int dirtybufcount, dirtybufcounthw;
1b30fbcc 128int runningbufspace, runningbufcount;
984263bc
MD
129static int getnewbufcalls;
130static int getnewbufrestarts;
4ecf7cc9 131static int recoverbufcalls;
f832287e
MD
132static int needsbuffer; /* locked by needsbuffer_spin */
133static int bd_request; /* locked by needsbuffer_spin */
4b958e7b 134static int bd_request_hw; /* locked by needsbuffer_spin */
c4df9635
MD
135static u_int bd_wake_ary[BD_WAKE_SIZE];
136static u_int bd_wake_index;
d300946f 137static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
f832287e 138static struct spinlock needsbuffer_spin;
8ae5c7e0 139static int debug_commit;
f832287e 140
4ecf7cc9
MD
141static struct thread *bufdaemon_td;
142static struct thread *bufdaemonhw_td;
143
144
3f779080
HP
145/*
146 * Sysctls for operational control of the buffer cache.
147 */
868d24af 148SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
3f779080 149 "Number of dirty buffers to flush before bufdaemon becomes inactive");
868d24af 150SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
bb606263 151 "High watermark used to trigger explicit flushing of dirty buffers");
3f779080
HP
152SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
153 "Minimum amount of buffer space required for active I/O");
154SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
155 "Maximum amount of buffer space to usable for active I/O");
0e8bd897
MD
156SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
157 "Recycle pages to active or inactive queue transition pt 0-64");
3f779080
HP
158/*
159 * Sysctls determining current state of the buffer cache.
160 */
17cde63e
MD
161SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
162 "Total number of buffers in buffer cache");
868d24af 163SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
70ac7d6c 164 "Pending bytes of dirty buffers (all)");
868d24af 165SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
70ac7d6c
MD
166 "Pending bytes of dirty buffers (heavy weight)");
167SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
168 "Pending number of dirty buffers");
169SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
4b958e7b 170 "Pending number of dirty buffers (heavy weight)");
3f779080 171SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
bb606263 172 "I/O bytes currently in progress due to asynchronous writes");
1b30fbcc
MD
173SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
174 "I/O buffers currently in progress due to asynchronous writes");
3f779080
HP
175SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
176 "Hard limit on maximum amount of memory usable for buffer space");
177SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
178 "Soft limit on maximum amount of memory usable for buffer space");
179SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
180 "Minimum amount of memory to reserve for system buffer space");
181SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
182 "Amount of memory available for buffers");
183SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
bb606263 184 0, "Maximum amount of memory reserved for buffers using malloc");
3f779080
HP
185SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
186 "Amount of memory left for buffers using malloc-scheme");
187SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
188 "New buffer header acquisition requests");
189SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
190 0, "New buffer header acquisition restarts");
4ecf7cc9
MD
191SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
192 "Recover VM space in an emergency");
3f779080 193SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
bb606263 194 "Buffer acquisition restarts due to fragmented buffer map");
3f779080
HP
195SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
196 "Amount of time KVA space was deallocated in an arbitrary buffer");
197SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
198 "Amount of time buffer re-use operations were successful");
8ae5c7e0 199SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
306ab3cb
HP
200SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
201 "sizeof(struct buf)");
984263bc 202
984263bc
MD
203char *buf_wmesg = BUF_WMESG;
204
984263bc 205#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
c4df9635 206#define VFS_BIO_NEED_UNUSED02 0x02
868d24af 207#define VFS_BIO_NEED_UNUSED04 0x04
984263bc
MD
208#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
209
210/*
3f779080 211 * bufspacewakeup:
984263bc
MD
212 *
213 * Called when buffer space is potentially available for recovery.
214 * getnewbuf() will block on this flag when it is unable to free
215 * sufficient buffer space. Buffer space becomes recoverable when
216 * bp's get placed back in the queues.
217 */
218
219static __inline void
220bufspacewakeup(void)
221{
222 /*
223 * If someone is waiting for BUF space, wake them up. Even
224 * though we haven't freed the kva space yet, the waiting
225 * process will be able to now.
226 */
227 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
f832287e 228 spin_lock_wr(&needsbuffer_spin);
984263bc 229 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
f832287e 230 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
231 wakeup(&needsbuffer);
232 }
233}
234
235/*
3f779080
HP
236 * runningbufwakeup:
237 *
238 * Accounting for I/O in progress.
984263bc
MD
239 *
240 */
241static __inline void
242runningbufwakeup(struct buf *bp)
243{
868d24af 244 int totalspace;
4afeea0d 245 int limit;
868d24af
MD
246
247 if ((totalspace = bp->b_runningbufspace) != 0) {
ae8e83e6
MD
248 atomic_subtract_int(&runningbufspace, totalspace);
249 atomic_subtract_int(&runningbufcount, 1);
984263bc 250 bp->b_runningbufspace = 0;
4afeea0d
MD
251
252 /*
253 * see waitrunningbufspace() for limit test.
254 */
255 limit = hirunningspace * 2 / 3;
256 if (runningbufreq && runningbufspace <= limit) {
984263bc
MD
257 runningbufreq = 0;
258 wakeup(&runningbufreq);
259 }
868d24af 260 bd_signal(totalspace);
984263bc
MD
261 }
262}
263
264/*
3f779080 265 * bufcountwakeup:
984263bc
MD
266 *
267 * Called when a buffer has been added to one of the free queues to
268 * account for the buffer and to wakeup anyone waiting for free buffers.
269 * This typically occurs when large amounts of metadata are being handled
270 * by the buffer cache ( else buffer space runs out first, usually ).
b1c20cfa
MD
271 *
272 * MPSAFE
984263bc 273 */
984263bc
MD
274static __inline void
275bufcountwakeup(void)
276{
984263bc 277 if (needsbuffer) {
f832287e 278 spin_lock_wr(&needsbuffer_spin);
984263bc 279 needsbuffer &= ~VFS_BIO_NEED_ANY;
f832287e 280 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
281 wakeup(&needsbuffer);
282 }
283}
284
285/*
3f779080 286 * waitrunningbufspace()
984263bc 287 *
4afeea0d
MD
288 * Wait for the amount of running I/O to drop to hirunningspace * 2 / 3.
289 * This is the point where write bursting stops so we don't want to wait
290 * for the running amount to drop below it (at least if we still want bioq
291 * to burst writes).
984263bc 292 *
cd083340 293 * The caller may be using this function to block in a tight loop, we
4afeea0d
MD
294 * must block while runningbufspace is greater then or equal to
295 * hirunningspace * 2 / 3.
296 *
cd083340
MD
297 * And even with that it may not be enough, due to the presence of
298 * B_LOCKED dirty buffers, so also wait for at least one running buffer
299 * to complete.
984263bc
MD
300 */
301static __inline void
4afeea0d 302waitrunningbufspace(void)
984263bc 303{
4afeea0d 304 int limit = hirunningspace * 2 / 3;
cd083340
MD
305
306 crit_enter();
4afeea0d
MD
307 if (runningbufspace > limit) {
308 while (runningbufspace > limit) {
e43a034f 309 ++runningbufreq;
4afeea0d 310 tsleep(&runningbufreq, 0, "wdrn1", 0);
e43a034f 311 }
cd083340
MD
312 } else if (runningbufspace) {
313 ++runningbufreq;
4afeea0d 314 tsleep(&runningbufreq, 0, "wdrn2", 1);
984263bc 315 }
cd083340 316 crit_exit();
984263bc
MD
317}
318
319/*
ae8e83e6
MD
320 * buf_dirty_count_severe:
321 *
322 * Return true if we have too many dirty buffers.
323 */
324int
325buf_dirty_count_severe(void)
326{
327 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
328 dirtybufcount >= nbuf / 2);
329}
330
331/*
4afeea0d
MD
332 * Return true if the amount of running I/O is severe and BIOQ should
333 * start bursting.
334 */
335int
336buf_runningbufspace_severe(void)
337{
338 return (runningbufspace >= hirunningspace * 2 / 3);
339}
340
341/*
3f779080 342 * vfs_buf_test_cache:
984263bc 343 *
cb1cf930
MD
344 * Called when a buffer is extended. This function clears the B_CACHE
345 * bit if the newly extended portion of the buffer does not contain
346 * valid data.
347 *
348 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
349 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
350 * them while a clean buffer was present.
984263bc
MD
351 */
352static __inline__
353void
354vfs_buf_test_cache(struct buf *bp,
355 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
356 vm_page_t m)
357{
358 if (bp->b_flags & B_CACHE) {
359 int base = (foff + off) & PAGE_MASK;
360 if (vm_page_is_valid(m, base, size) == 0)
361 bp->b_flags &= ~B_CACHE;
362 }
363}
364
3f779080 365/*
cd083340 366 * bd_speedup()
4b958e7b 367 *
cd083340
MD
368 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
369 * low water mark.
b1c20cfa
MD
370 *
371 * MPSAFE
3f779080 372 */
984263bc
MD
373static __inline__
374void
c4df9635 375bd_speedup(void)
984263bc 376{
70ac7d6c 377 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
cd083340
MD
378 return;
379
380 if (bd_request == 0 &&
70ac7d6c
MD
381 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
382 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
f832287e 383 spin_lock_wr(&needsbuffer_spin);
984263bc 384 bd_request = 1;
f832287e 385 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
386 wakeup(&bd_request);
387 }
cd083340 388 if (bd_request_hw == 0 &&
70ac7d6c
MD
389 (dirtybufspacehw > lodirtybufspace / 2 ||
390 dirtybufcounthw >= nbuf / 2)) {
4b958e7b
MD
391 spin_lock_wr(&needsbuffer_spin);
392 bd_request_hw = 1;
393 spin_unlock_wr(&needsbuffer_spin);
394 wakeup(&bd_request_hw);
395 }
984263bc
MD
396}
397
398/*
c4df9635 399 * bd_heatup()
3f779080 400 *
c4df9635
MD
401 * Get the buf_daemon heated up when the number of running and dirty
402 * buffers exceeds the mid-point.
b1c20cfa 403 *
79eae878
MD
404 * Return the total number of dirty bytes past the second mid point
405 * as a measure of how much excess dirty data there is in the system.
406 *
b1c20cfa 407 * MPSAFE
984263bc 408 */
c4df9635
MD
409int
410bd_heatup(void)
411{
412 int mid1;
413 int mid2;
868d24af 414 int totalspace;
984263bc 415
868d24af 416 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
c4df9635 417
868d24af 418 totalspace = runningbufspace + dirtybufspace;
70ac7d6c 419 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
c4df9635 420 bd_speedup();
868d24af
MD
421 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
422 if (totalspace >= mid2)
423 return(totalspace - mid2);
c4df9635
MD
424 }
425 return(0);
426}
427
428/*
429 * bd_wait()
430 *
868d24af
MD
431 * Wait for the buffer cache to flush (totalspace) bytes worth of
432 * buffers, then return.
c4df9635
MD
433 *
434 * Regardless this function blocks while the number of dirty buffers
868d24af 435 * exceeds hidirtybufspace.
b1c20cfa
MD
436 *
437 * MPSAFE
c4df9635 438 */
984263bc 439void
868d24af 440bd_wait(int totalspace)
984263bc 441{
c4df9635 442 u_int i;
868d24af 443 int count;
c4df9635 444
4ecf7cc9
MD
445 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
446 return;
447
868d24af 448 while (totalspace > 0) {
c4df9635 449 bd_heatup();
868d24af
MD
450 if (totalspace > runningbufspace + dirtybufspace)
451 totalspace = runningbufspace + dirtybufspace;
452 count = totalspace / BKVASIZE;
c4df9635
MD
453 if (count >= BD_WAKE_SIZE)
454 count = BD_WAKE_SIZE - 1;
b1c20cfa
MD
455
456 spin_lock_wr(&needsbuffer_spin);
c4df9635
MD
457 i = (bd_wake_index + count) & BD_WAKE_MASK;
458 ++bd_wake_ary[i];
ae8e83e6 459 tsleep_interlock(&bd_wake_ary[i], 0);
b1c20cfa 460 spin_unlock_wr(&needsbuffer_spin);
d9345d3a 461 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
c4df9635 462
868d24af 463 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
c4df9635
MD
464 }
465}
466
467/*
468 * bd_signal()
469 *
868d24af
MD
470 * This function is called whenever runningbufspace or dirtybufspace
471 * is reduced. Track threads waiting for run+dirty buffer I/O
c4df9635 472 * complete.
b1c20cfa
MD
473 *
474 * MPSAFE
c4df9635
MD
475 */
476static void
868d24af 477bd_signal(int totalspace)
c4df9635
MD
478{
479 u_int i;
480
b1c20cfa
MD
481 if (totalspace > 0) {
482 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
483 totalspace = BKVASIZE * BD_WAKE_SIZE;
484 spin_lock_wr(&needsbuffer_spin);
485 while (totalspace > 0) {
486 i = bd_wake_index++;
487 i &= BD_WAKE_MASK;
488 if (bd_wake_ary[i]) {
489 bd_wake_ary[i] = 0;
490 spin_unlock_wr(&needsbuffer_spin);
491 wakeup(&bd_wake_ary[i]);
492 spin_lock_wr(&needsbuffer_spin);
493 }
494 totalspace -= BKVASIZE;
868d24af 495 }
b1c20cfa 496 spin_unlock_wr(&needsbuffer_spin);
c4df9635 497 }
984263bc
MD
498}
499
500/*
a9a20f98
MD
501 * BIO tracking support routines.
502 *
503 * Release a ref on a bio_track. Wakeup requests are atomically released
504 * along with the last reference so bk_active will never wind up set to
505 * only 0x80000000.
506 *
507 * MPSAFE
508 */
509static
510void
511bio_track_rel(struct bio_track *track)
512{
513 int active;
514 int desired;
515
516 /*
517 * Shortcut
518 */
519 active = track->bk_active;
520 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
521 return;
522
523 /*
524 * Full-on. Note that the wait flag is only atomically released on
525 * the 1->0 count transition.
e7edae1e
MD
526 *
527 * We check for a negative count transition using bit 30 since bit 31
528 * has a different meaning.
a9a20f98
MD
529 */
530 for (;;) {
531 desired = (active & 0x7FFFFFFF) - 1;
532 if (desired)
533 desired |= active & 0x80000000;
534 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
e7edae1e 535 if (desired & 0x40000000)
a9a20f98
MD
536 panic("bio_track_rel: bad count: %p\n", track);
537 if (active & 0x80000000)
538 wakeup(track);
539 break;
540 }
541 active = track->bk_active;
542 }
543}
544
545/*
546 * Wait for the tracking count to reach 0.
547 *
548 * Use atomic ops such that the wait flag is only set atomically when
549 * bk_active is non-zero.
550 *
551 * MPSAFE
552 */
553int
554bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
555{
556 int active;
557 int desired;
558 int error;
559
560 /*
561 * Shortcut
562 */
563 if (track->bk_active == 0)
564 return(0);
565
566 /*
567 * Full-on. Note that the wait flag may only be atomically set if
568 * the active count is non-zero.
569 */
a9a20f98
MD
570 error = 0;
571 while ((active = track->bk_active) != 0) {
572 desired = active | 0x80000000;
ae8e83e6 573 tsleep_interlock(track, slp_flags);
a9a20f98
MD
574 if (active == desired ||
575 atomic_cmpset_int(&track->bk_active, active, desired)) {
d9345d3a
MD
576 error = tsleep(track, slp_flags | PINTERLOCKED,
577 "iowait", slp_timo);
a9a20f98
MD
578 if (error)
579 break;
580 }
581 }
a9a20f98
MD
582 return (error);
583}
584
585/*
3f779080
HP
586 * bufinit:
587 *
588 * Load time initialisation of the buffer cache, called from machine
589 * dependant initialization code.
590 */
984263bc
MD
591void
592bufinit(void)
593{
594 struct buf *bp;
b8bb0773 595 vm_offset_t bogus_offset;
984263bc
MD
596 int i;
597
f832287e
MD
598 spin_init(&needsbuffer_spin);
599
984263bc
MD
600 /* next, make a null set of free lists */
601 for (i = 0; i < BUFFER_QUEUES; i++)
602 TAILQ_INIT(&bufqueues[i]);
603
604 /* finally, initialize each buffer header and stick on empty q */
605 for (i = 0; i < nbuf; i++) {
606 bp = &buf[i];
607 bzero(bp, sizeof *bp);
608 bp->b_flags = B_INVAL; /* we're just an empty header */
10f3fee5 609 bp->b_cmd = BUF_CMD_DONE;
b3098c79 610 bp->b_qindex = BQUEUE_EMPTY;
81b5c339 611 initbufbio(bp);
54f51aeb 612 xio_init(&bp->b_xio);
408357d8 613 buf_dep_init(bp);
984263bc 614 BUF_LOCKINIT(bp);
b3098c79 615 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
984263bc
MD
616 }
617
618 /*
619 * maxbufspace is the absolute maximum amount of buffer space we are
620 * allowed to reserve in KVM and in real terms. The absolute maximum
621 * is nominally used by buf_daemon. hibufspace is the nominal maximum
622 * used by most other processes. The differential is required to
623 * ensure that buf_daemon is able to run when other processes might
624 * be blocked waiting for buffer space.
625 *
626 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
627 * this may result in KVM fragmentation which is not handled optimally
628 * by the system.
629 */
630 maxbufspace = nbuf * BKVASIZE;
631 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
632 lobufspace = hibufspace - MAXBSIZE;
633
634 lorunningspace = 512 * 1024;
4afeea0d 635 /* hirunningspace -- see below */
984263bc 636
868d24af
MD
637 /*
638 * Limit the amount of malloc memory since it is wired permanently
639 * into the kernel space. Even though this is accounted for in
640 * the buffer allocation, we don't want the malloced region to grow
641 * uncontrolled. The malloc scheme improves memory utilization
642 * significantly on average (small) directories.
643 */
984263bc
MD
644 maxbufmallocspace = hibufspace / 20;
645
868d24af
MD
646 /*
647 * Reduce the chance of a deadlock occuring by limiting the number
648 * of delayed-write dirty buffers we allow to stack up.
4afeea0d
MD
649 *
650 * We don't want too much actually queued to the device at once
651 * (XXX this needs to be per-mount!), because the buffers will
652 * wind up locked for a very long period of time while the I/O
653 * drains.
868d24af 654 */
4afeea0d
MD
655 hidirtybufspace = hibufspace / 2; /* dirty + running */
656 hirunningspace = hibufspace / 16; /* locked & queued to device */
657 if (hirunningspace < 1024 * 1024)
658 hirunningspace = 1024 * 1024;
659
868d24af
MD
660 dirtybufspace = 0;
661 dirtybufspacehw = 0;
984263bc 662
868d24af 663 lodirtybufspace = hidirtybufspace / 2;
984263bc 664
868d24af
MD
665 /*
666 * Maximum number of async ops initiated per buf_daemon loop. This is
667 * somewhat of a hack at the moment, we really need to limit ourselves
668 * based on the number of bytes of I/O in-transit that were initiated
669 * from buf_daemon.
670 */
984263bc 671
e4846942 672 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
c439ad8f 673 bogus_page = vm_page_alloc(&kernel_object,
e4846942
MD
674 (bogus_offset >> PAGE_SHIFT),
675 VM_ALLOC_NORMAL);
12e4aaff 676 vmstats.v_wire_count++;
984263bc
MD
677
678}
679
680/*
81b5c339
MD
681 * Initialize the embedded bio structures
682 */
683void
684initbufbio(struct buf *bp)
685{
686 bp->b_bio1.bio_buf = bp;
687 bp->b_bio1.bio_prev = NULL;
81b5c339
MD
688 bp->b_bio1.bio_offset = NOOFFSET;
689 bp->b_bio1.bio_next = &bp->b_bio2;
690 bp->b_bio1.bio_done = NULL;
ae8e83e6 691 bp->b_bio1.bio_flags = 0;
81b5c339
MD
692
693 bp->b_bio2.bio_buf = bp;
694 bp->b_bio2.bio_prev = &bp->b_bio1;
81b5c339
MD
695 bp->b_bio2.bio_offset = NOOFFSET;
696 bp->b_bio2.bio_next = NULL;
697 bp->b_bio2.bio_done = NULL;
ae8e83e6 698 bp->b_bio2.bio_flags = 0;
81b5c339
MD
699}
700
701/*
702 * Reinitialize the embedded bio structures as well as any additional
703 * translation cache layers.
704 */
705void
706reinitbufbio(struct buf *bp)
707{
708 struct bio *bio;
709
710 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
711 bio->bio_done = NULL;
81b5c339
MD
712 bio->bio_offset = NOOFFSET;
713 }
714}
715
716/*
717 * Push another BIO layer onto an existing BIO and return it. The new
718 * BIO layer may already exist, holding cached translation data.
719 */
720struct bio *
721push_bio(struct bio *bio)
722{
723 struct bio *nbio;
724
725 if ((nbio = bio->bio_next) == NULL) {
726 int index = bio - &bio->bio_buf->b_bio_array[0];
bbd44c71 727 if (index >= NBUF_BIO - 1) {
81b5c339
MD
728 panic("push_bio: too many layers bp %p\n",
729 bio->bio_buf);
730 }
731 nbio = &bio->bio_buf->b_bio_array[index + 1];
732 bio->bio_next = nbio;
733 nbio->bio_prev = bio;
734 nbio->bio_buf = bio->bio_buf;
81b5c339
MD
735 nbio->bio_offset = NOOFFSET;
736 nbio->bio_done = NULL;
737 nbio->bio_next = NULL;
738 }
739 KKASSERT(nbio->bio_done == NULL);
740 return(nbio);
741}
742
b77cfc40
MD
743/*
744 * Pop a BIO translation layer, returning the previous layer. The
745 * must have been previously pushed.
746 */
747struct bio *
81b5c339
MD
748pop_bio(struct bio *bio)
749{
b77cfc40 750 return(bio->bio_prev);
81b5c339
MD
751}
752
753void
754clearbiocache(struct bio *bio)
755{
756 while (bio) {
81b5c339
MD
757 bio->bio_offset = NOOFFSET;
758 bio = bio->bio_next;
759 }
760}
761
762/*
3f779080
HP
763 * bfreekva:
764 *
765 * Free the KVA allocation for buffer 'bp'.
984263bc 766 *
e43a034f 767 * Must be called from a critical section as this is the only locking for
984263bc
MD
768 * buffer_map.
769 *
770 * Since this call frees up buffer space, we call bufspacewakeup().
b1c20cfa
MD
771 *
772 * MPALMOSTSAFE
984263bc
MD
773 */
774static void
312dcd01 775bfreekva(struct buf *bp)
984263bc 776{
a108bf71
MD
777 int count;
778
984263bc 779 if (bp->b_kvasize) {
b1c20cfa 780 get_mplock();
984263bc 781 ++buffreekvacnt;
a108bf71 782 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
e4846942 783 vm_map_lock(&buffer_map);
984263bc 784 bufspace -= bp->b_kvasize;
e4846942 785 vm_map_delete(&buffer_map,
984263bc 786 (vm_offset_t) bp->b_kvabase,
a108bf71
MD
787 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
788 &count
984263bc 789 );
e4846942 790 vm_map_unlock(&buffer_map);
a108bf71 791 vm_map_entry_release(count);
984263bc
MD
792 bp->b_kvasize = 0;
793 bufspacewakeup();
b1c20cfa 794 rel_mplock();
984263bc
MD
795 }
796}
797
798/*
3f779080 799 * bremfree:
984263bc
MD
800 *
801 * Remove the buffer from the appropriate free list.
802 */
c3d1e862
MD
803static __inline void
804_bremfree(struct buf *bp)
984263bc 805{
b3098c79 806 if (bp->b_qindex != BQUEUE_NONE) {
77bb9400
MD
807 KASSERT(BUF_REFCNTNB(bp) == 1,
808 ("bremfree: bp %p not locked",bp));
984263bc 809 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
b3098c79 810 bp->b_qindex = BQUEUE_NONE;
984263bc 811 } else {
77bb9400 812 if (BUF_REFCNTNB(bp) <= 1)
984263bc
MD
813 panic("bremfree: removing a buffer not on a queue");
814 }
c3d1e862 815}
984263bc 816
c3d1e862
MD
817void
818bremfree(struct buf *bp)
819{
820 spin_lock_wr(&bufspin);
821 _bremfree(bp);
822 spin_unlock_wr(&bufspin);
984263bc
MD
823}
824
b1c20cfa 825static void
c3d1e862
MD
826bremfree_locked(struct buf *bp)
827{
828 _bremfree(bp);
829}
984263bc
MD
830
831/*
3f779080
HP
832 * bread:
833 *
834 * Get a buffer with the specified data. Look in the cache first. We
835 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
836 * is set, the buffer is valid and we do not have to do anything ( see
837 * getblk() ).
c3d1e862
MD
838 *
839 * MPALMOSTSAFE
984263bc
MD
840 */
841int
c8e4131d 842bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
984263bc
MD
843{
844 struct buf *bp;
845
54078292 846 bp = getblk(vp, loffset, size, 0, 0);
984263bc
MD
847 *bpp = bp;
848
849 /* if not found in cache, do some I/O */
850 if ((bp->b_flags & B_CACHE) == 0) {
c3d1e862 851 get_mplock();
ae8e83e6 852 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
10f3fee5 853 bp->b_cmd = BUF_CMD_READ;
ae8e83e6
MD
854 bp->b_bio1.bio_done = biodone_sync;
855 bp->b_bio1.bio_flags |= BIO_SYNC;
10f3fee5 856 vfs_busy_pages(vp, bp);
81b5c339 857 vn_strategy(vp, &bp->b_bio1);
c3d1e862 858 rel_mplock();
ae8e83e6 859 return (biowait(&bp->b_bio1, "biord"));
984263bc
MD
860 }
861 return (0);
862}
863
864/*
3f779080
HP
865 * breadn:
866 *
867 * Operates like bread, but also starts asynchronous I/O on
868 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
869 * to initiating I/O . If B_CACHE is set, the buffer is valid
870 * and we do not have to do anything.
b1c20cfa
MD
871 *
872 * MPALMOSTSAFE
984263bc
MD
873 */
874int
a8f169e2 875breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
c8e4131d 876 int *rabsize, int cnt, struct buf **bpp)
984263bc
MD
877{
878 struct buf *bp, *rabp;
879 int i;
880 int rv = 0, readwait = 0;
881
54078292 882 *bpp = bp = getblk(vp, loffset, size, 0, 0);
984263bc
MD
883
884 /* if not found in cache, do some I/O */
885 if ((bp->b_flags & B_CACHE) == 0) {
b1c20cfa 886 get_mplock();
ae8e83e6 887 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
10f3fee5 888 bp->b_cmd = BUF_CMD_READ;
ae8e83e6
MD
889 bp->b_bio1.bio_done = biodone_sync;
890 bp->b_bio1.bio_flags |= BIO_SYNC;
10f3fee5 891 vfs_busy_pages(vp, bp);
81b5c339 892 vn_strategy(vp, &bp->b_bio1);
984263bc 893 ++readwait;
b1c20cfa 894 rel_mplock();
984263bc
MD
895 }
896
54078292
MD
897 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
898 if (inmem(vp, *raoffset))
984263bc 899 continue;
54078292 900 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
984263bc
MD
901
902 if ((rabp->b_flags & B_CACHE) == 0) {
a22da047 903 get_mplock();
ae8e83e6 904 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
10f3fee5
MD
905 rabp->b_cmd = BUF_CMD_READ;
906 vfs_busy_pages(vp, rabp);
984263bc 907 BUF_KERNPROC(rabp);
81b5c339 908 vn_strategy(vp, &rabp->b_bio1);
b1c20cfa 909 rel_mplock();
984263bc
MD
910 } else {
911 brelse(rabp);
912 }
913 }
b1c20cfa 914 if (readwait)
ae8e83e6 915 rv = biowait(&bp->b_bio1, "biord");
984263bc
MD
916 return (rv);
917}
918
919/*
3f779080
HP
920 * bwrite:
921 *
ae8e83e6
MD
922 * Synchronous write, waits for completion.
923 *
3f779080
HP
924 * Write, release buffer on completion. (Done by iodone
925 * if async). Do not bother writing anything if the buffer
926 * is invalid.
927 *
928 * Note that we set B_CACHE here, indicating that buffer is
929 * fully valid and thus cacheable. This is true even of NFS
930 * now so we set it generally. This could be set either here
931 * or in biodone() since the I/O is synchronous. We put it
932 * here.
984263bc
MD
933 */
934int
c8e4131d 935bwrite(struct buf *bp)
984263bc 936{
ae8e83e6 937 int error;
984263bc
MD
938
939 if (bp->b_flags & B_INVAL) {
940 brelse(bp);
941 return (0);
942 }
77bb9400 943 if (BUF_REFCNTNB(bp) == 0)
984263bc 944 panic("bwrite: buffer is not busy???");
984263bc
MD
945
946 /* Mark the buffer clean */
947 bundirty(bp);
948
ae8e83e6 949 bp->b_flags &= ~(B_ERROR | B_EINTR);
6bae6177 950 bp->b_flags |= B_CACHE;
10f3fee5 951 bp->b_cmd = BUF_CMD_WRITE;
ae8e83e6
MD
952 bp->b_bio1.bio_done = biodone_sync;
953 bp->b_bio1.bio_flags |= BIO_SYNC;
10f3fee5 954 vfs_busy_pages(bp->b_vp, bp);
984263bc
MD
955
956 /*
9a71d53f
MD
957 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
958 * valid for vnode-backed buffers.
984263bc
MD
959 */
960 bp->b_runningbufspace = bp->b_bufsize;
1b30fbcc
MD
961 if (bp->b_runningbufspace) {
962 runningbufspace += bp->b_runningbufspace;
963 ++runningbufcount;
964 }
984263bc 965
81b5c339 966 vn_strategy(bp->b_vp, &bp->b_bio1);
ae8e83e6
MD
967 error = biowait(&bp->b_bio1, "biows");
968 brelse(bp);
969 return (error);
970}
984263bc 971
ae8e83e6
MD
972/*
973 * bawrite:
974 *
975 * Asynchronous write. Start output on a buffer, but do not wait for
976 * it to complete. The buffer is released when the output completes.
977 *
978 * bwrite() ( or the VOP routine anyway ) is responsible for handling
979 * B_INVAL buffers. Not us.
980 */
981void
982bawrite(struct buf *bp)
983{
984 if (bp->b_flags & B_INVAL) {
984263bc 985 brelse(bp);
ae8e83e6
MD
986 return;
987 }
988 if (BUF_REFCNTNB(bp) == 0)
989 panic("bwrite: buffer is not busy???");
990
991 /* Mark the buffer clean */
992 bundirty(bp);
993
994 bp->b_flags &= ~(B_ERROR | B_EINTR);
995 bp->b_flags |= B_CACHE;
996 bp->b_cmd = BUF_CMD_WRITE;
997 KKASSERT(bp->b_bio1.bio_done == NULL);
998 vfs_busy_pages(bp->b_vp, bp);
999
1000 /*
1001 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1002 * valid for vnode-backed buffers.
1003 */
1004 bp->b_runningbufspace = bp->b_bufsize;
1005 if (bp->b_runningbufspace) {
1006 runningbufspace += bp->b_runningbufspace;
1007 ++runningbufcount;
984263bc 1008 }
ae8e83e6
MD
1009
1010 BUF_KERNPROC(bp);
1011 vn_strategy(bp->b_vp, &bp->b_bio1);
1012}
1013
1014/*
1015 * bowrite:
1016 *
1017 * Ordered write. Start output on a buffer, and flag it so that the
1018 * device will write it in the order it was queued. The buffer is
1019 * released when the output completes. bwrite() ( or the VOP routine
1020 * anyway ) is responsible for handling B_INVAL buffers.
1021 */
1022int
1023bowrite(struct buf *bp)
1024{
1025 bp->b_flags |= B_ORDERED;
1026 bawrite(bp);
984263bc
MD
1027 return (0);
1028}
1029
984263bc 1030/*
3f779080
HP
1031 * bdwrite:
1032 *
1033 * Delayed write. (Buffer is marked dirty). Do not bother writing
1034 * anything if the buffer is marked invalid.
984263bc 1035 *
3f779080
HP
1036 * Note that since the buffer must be completely valid, we can safely
1037 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1038 * biodone() in order to prevent getblk from writing the buffer
1039 * out synchronously.
984263bc
MD
1040 */
1041void
493c516a 1042bdwrite(struct buf *bp)
984263bc 1043{
77bb9400 1044 if (BUF_REFCNTNB(bp) == 0)
984263bc
MD
1045 panic("bdwrite: buffer is not busy");
1046
1047 if (bp->b_flags & B_INVAL) {
1048 brelse(bp);
1049 return;
1050 }
1051 bdirty(bp);
1052
8c72e3d5
AH
1053 if (dsched_is_clear_buf_priv(bp))
1054 dsched_new_buf(bp);
1055
984263bc
MD
1056 /*
1057 * Set B_CACHE, indicating that the buffer is fully valid. This is
1058 * true even of NFS now.
1059 */
1060 bp->b_flags |= B_CACHE;
1061
1062 /*
1063 * This bmap keeps the system from needing to do the bmap later,
1064 * perhaps when the system is attempting to do a sync. Since it
1065 * is likely that the indirect block -- or whatever other datastructure
1066 * that the filesystem needs is still in memory now, it is a good
1067 * thing to do this. Note also, that if the pageout daemon is
1068 * requesting a sync -- there might not be enough memory to do
1069 * the bmap then... So, this is important to do.
1070 */
54078292 1071 if (bp->b_bio2.bio_offset == NOOFFSET) {
08daea96 1072 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
e92ca23a 1073 NULL, NULL, BUF_CMD_WRITE);
984263bc
MD
1074 }
1075
1076 /*
cb1cf930
MD
1077 * Because the underlying pages may still be mapped and
1078 * writable trying to set the dirty buffer (b_dirtyoff/end)
1079 * range here will be inaccurate.
1080 *
1081 * However, we must still clean the pages to satisfy the
1082 * vnode_pager and pageout daemon, so theythink the pages
1083 * have been "cleaned". What has really occured is that
1084 * they've been earmarked for later writing by the buffer
1085 * cache.
1086 *
1087 * So we get the b_dirtyoff/end update but will not actually
1088 * depend on it (NFS that is) until the pages are busied for
1089 * writing later on.
984263bc
MD
1090 */
1091 vfs_clean_pages(bp);
1092 bqrelse(bp);
1093
1094 /*
984263bc
MD
1095 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1096 * due to the softdep code.
1097 */
1098}
1099
1100/*
0a8aee15
MD
1101 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1102 * This is used by tmpfs.
1103 *
1104 * It is important for any VFS using this routine to NOT use it for
1105 * IO_SYNC or IO_ASYNC operations which occur when the system really
1106 * wants to flush VM pages to backing store.
1107 */
1108void
1109buwrite(struct buf *bp)
1110{
1111 vm_page_t m;
1112 int i;
1113
1114 /*
1115 * Only works for VMIO buffers. If the buffer is already
1116 * marked for delayed-write we can't avoid the bdwrite().
1117 */
1118 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1119 bdwrite(bp);
1120 return;
1121 }
1122
1123 /*
1124 * Set valid & dirty.
1125 */
1126 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1127 m = bp->b_xio.xio_pages[i];
1128 vfs_dirty_one_page(bp, i, m);
1129 }
1130 bqrelse(bp);
1131}
1132
1133/*
3f779080 1134 * bdirty:
984263bc 1135 *
10f3fee5
MD
1136 * Turn buffer into delayed write request by marking it B_DELWRI.
1137 * B_RELBUF and B_NOCACHE must be cleared.
984263bc 1138 *
10f3fee5
MD
1139 * We reassign the buffer to itself to properly update it in the
1140 * dirty/clean lists.
984263bc 1141 *
e43a034f 1142 * Must be called from a critical section.
b3098c79 1143 * The buffer must be on BQUEUE_NONE.
984263bc
MD
1144 */
1145void
493c516a 1146bdirty(struct buf *bp)
984263bc 1147{
b3098c79 1148 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
69f8c926 1149 if (bp->b_flags & B_NOCACHE) {
6ea70f76 1150 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
69f8c926
MD
1151 bp->b_flags &= ~B_NOCACHE;
1152 }
1153 if (bp->b_flags & B_INVAL) {
6ea70f76 1154 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
69f8c926 1155 }
10f3fee5 1156 bp->b_flags &= ~B_RELBUF;
984263bc
MD
1157
1158 if ((bp->b_flags & B_DELWRI) == 0) {
10f3fee5 1159 bp->b_flags |= B_DELWRI;
1f1ea522 1160 reassignbuf(bp);
b1c20cfa 1161 atomic_add_int(&dirtybufcount, 1);
868d24af 1162 dirtybufspace += bp->b_bufsize;
70ac7d6c 1163 if (bp->b_flags & B_HEAVY) {
b1c20cfa
MD
1164 atomic_add_int(&dirtybufcounthw, 1);
1165 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
70ac7d6c 1166 }
c4df9635 1167 bd_heatup();
984263bc
MD
1168 }
1169}
1170
1171/*
4b958e7b
MD
1172 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1173 * needs to be flushed with a different buf_daemon thread to avoid
1174 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1175 */
1176void
1177bheavy(struct buf *bp)
1178{
1179 if ((bp->b_flags & B_HEAVY) == 0) {
1180 bp->b_flags |= B_HEAVY;
70ac7d6c 1181 if (bp->b_flags & B_DELWRI) {
b1c20cfa
MD
1182 atomic_add_int(&dirtybufcounthw, 1);
1183 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
70ac7d6c 1184 }
4b958e7b
MD
1185 }
1186}
1187
1188/*
3f779080 1189 * bundirty:
984263bc
MD
1190 *
1191 * Clear B_DELWRI for buffer.
1192 *
e43a034f 1193 * Must be called from a critical section.
eaaadca0 1194 *
b3098c79 1195 * The buffer is typically on BQUEUE_NONE but there is one case in
eaaadca0
MD
1196 * brelse() that calls this function after placing the buffer on
1197 * a different queue.
b1c20cfa
MD
1198 *
1199 * MPSAFE
984263bc 1200 */
984263bc 1201void
493c516a 1202bundirty(struct buf *bp)
984263bc 1203{
984263bc
MD
1204 if (bp->b_flags & B_DELWRI) {
1205 bp->b_flags &= ~B_DELWRI;
1f1ea522 1206 reassignbuf(bp);
b1c20cfa
MD
1207 atomic_subtract_int(&dirtybufcount, 1);
1208 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
70ac7d6c 1209 if (bp->b_flags & B_HEAVY) {
b1c20cfa
MD
1210 atomic_subtract_int(&dirtybufcounthw, 1);
1211 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
70ac7d6c 1212 }
868d24af 1213 bd_signal(bp->b_bufsize);
984263bc
MD
1214 }
1215 /*
1216 * Since it is now being written, we can clear its deferred write flag.
1217 */
1218 bp->b_flags &= ~B_DEFERRED;
1219}
1220
1221/*
3f779080 1222 * brelse:
984263bc
MD
1223 *
1224 * Release a busy buffer and, if requested, free its resources. The
1225 * buffer will be stashed in the appropriate bufqueue[] allowing it
1226 * to be accessed later as a cache entity or reused for other purposes.
b1c20cfa
MD
1227 *
1228 * MPALMOSTSAFE
984263bc
MD
1229 */
1230void
c8e4131d 1231brelse(struct buf *bp)
984263bc 1232{
9188c711
MD
1233#ifdef INVARIANTS
1234 int saved_flags = bp->b_flags;
1235#endif
1236
984263bc
MD
1237 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1238
135bd6a8
MD
1239 /*
1240 * If B_NOCACHE is set we are being asked to destroy the buffer and
1241 * its backing store. Clear B_DELWRI.
1242 *
1243 * B_NOCACHE is set in two cases: (1) when the caller really wants
1244 * to destroy the buffer and backing store and (2) when the caller
1245 * wants to destroy the buffer and backing store after a write
1246 * completes.
1247 */
1248 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1249 bundirty(bp);
69f8c926
MD
1250 }
1251
78a9b77f 1252 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
984263bc 1253 /*
78a9b77f
MD
1254 * A re-dirtied buffer is only subject to destruction
1255 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
984263bc 1256 */
78a9b77f 1257 /* leave buffer intact */
10f3fee5 1258 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
78a9b77f 1259 (bp->b_bufsize <= 0)) {
984263bc 1260 /*
78a9b77f
MD
1261 * Either a failed read or we were asked to free or not
1262 * cache the buffer. This path is reached with B_DELWRI
1263 * set only if B_INVAL is already set. B_NOCACHE governs
1264 * backing store destruction.
408357d8
MD
1265 *
1266 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1267 * buffer cannot be immediately freed.
984263bc
MD
1268 */
1269 bp->b_flags |= B_INVAL;
c3d1e862 1270 if (LIST_FIRST(&bp->b_dep) != NULL) {
b1c20cfa 1271 get_mplock();
408357d8 1272 buf_deallocate(bp);
b1c20cfa 1273 rel_mplock();
c3d1e862 1274 }
984263bc 1275 if (bp->b_flags & B_DELWRI) {
b1c20cfa
MD
1276 atomic_subtract_int(&dirtybufcount, 1);
1277 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
70ac7d6c 1278 if (bp->b_flags & B_HEAVY) {
b1c20cfa
MD
1279 atomic_subtract_int(&dirtybufcounthw, 1);
1280 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
70ac7d6c 1281 }
868d24af 1282 bd_signal(bp->b_bufsize);
984263bc 1283 }
10f3fee5 1284 bp->b_flags &= ~(B_DELWRI | B_CACHE);
984263bc
MD
1285 }
1286
1287 /*
408357d8
MD
1288 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1289 * If vfs_vmio_release() is called with either bit set, the
1290 * underlying pages may wind up getting freed causing a previous
1291 * write (bdwrite()) to get 'lost' because pages associated with
1292 * a B_DELWRI bp are marked clean. Pages associated with a
1293 * B_LOCKED buffer may be mapped by the filesystem.
4b958e7b
MD
1294 *
1295 * If we want to release the buffer ourselves (rather then the
1296 * originator asking us to release it), give the originator a
1297 * chance to countermand the release by setting B_LOCKED.
984263bc
MD
1298 *
1299 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1300 * if B_DELWRI is set.
1301 *
1302 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1303 * on pages to return pages to the VM page queues.
1304 */
4b958e7b 1305 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
984263bc 1306 bp->b_flags &= ~B_RELBUF;
4b958e7b 1307 } else if (vm_page_count_severe()) {
b1c20cfa
MD
1308 if (LIST_FIRST(&bp->b_dep) != NULL) {
1309 get_mplock();
78a9b77f 1310 buf_deallocate(bp); /* can set B_LOCKED */
b1c20cfa
MD
1311 rel_mplock();
1312 }
4b958e7b
MD
1313 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1314 bp->b_flags &= ~B_RELBUF;
1315 else
1316 bp->b_flags |= B_RELBUF;
1317 }
984263bc
MD
1318
1319 /*
78a9b77f
MD
1320 * Make sure b_cmd is clear. It may have already been cleared by
1321 * biodone().
1322 *
9188c711
MD
1323 * At this point destroying the buffer is governed by the B_INVAL
1324 * or B_RELBUF flags.
1325 */
10f3fee5 1326 bp->b_cmd = BUF_CMD_DONE;
aa166ad1 1327 dsched_exit_buf(bp);
9188c711
MD
1328
1329 /*
135bd6a8
MD
1330 * VMIO buffer rundown. Make sure the VM page array is restored
1331 * after an I/O may have replaces some of the pages with bogus pages
1332 * in order to not destroy dirty pages in a fill-in read.
1333 *
1334 * Note that due to the code above, if a buffer is marked B_DELWRI
1335 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1336 * B_INVAL may still be set, however.
984263bc 1337 *
135bd6a8
MD
1338 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1339 * but not the backing store. B_NOCACHE will destroy the backing
1340 * store.
984263bc 1341 *
135bd6a8
MD
1342 * Note that dirty NFS buffers contain byte-granular write ranges
1343 * and should not be destroyed w/ B_INVAL even if the backing store
1344 * is left intact.
984263bc 1345 */
135bd6a8 1346 if (bp->b_flags & B_VMIO) {
9188c711
MD
1347 /*
1348 * Rundown for VMIO buffers which are not dirty NFS buffers.
1349 */
984263bc
MD
1350 int i, j, resid;
1351 vm_page_t m;
1352 off_t foff;
1353 vm_pindex_t poff;
1354 vm_object_t obj;
1355 struct vnode *vp;
1356
1357 vp = bp->b_vp;
1358
1359 /*
1360 * Get the base offset and length of the buffer. Note that
1361 * in the VMIO case if the buffer block size is not
1362 * page-aligned then b_data pointer may not be page-aligned.
236b2b9f 1363 * But our b_xio.xio_pages array *IS* page aligned.
984263bc
MD
1364 *
1365 * block sizes less then DEV_BSIZE (usually 512) are not
1366 * supported due to the page granularity bits (m->valid,
1367 * m->dirty, etc...).
1368 *
1369 * See man buf(9) for more information
1370 */
1371
1372 resid = bp->b_bufsize;
81b5c339 1373 foff = bp->b_loffset;
984263bc 1374
b1c20cfa 1375 get_mplock();
54f51aeb
HP
1376 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1377 m = bp->b_xio.xio_pages[i];
984263bc
MD
1378 vm_page_flag_clear(m, PG_ZERO);
1379 /*
1380 * If we hit a bogus page, fixup *all* of them
06ecca5a
MD
1381 * now. Note that we left these pages wired
1382 * when we removed them so they had better exist,
1383 * and they cannot be ripped out from under us so
e43a034f 1384 * no critical section protection is necessary.
984263bc
MD
1385 */
1386 if (m == bogus_page) {
7540ab49 1387 obj = vp->v_object;
81b5c339 1388 poff = OFF_TO_IDX(bp->b_loffset);
984263bc 1389
54f51aeb 1390 for (j = i; j < bp->b_xio.xio_npages; j++) {
984263bc
MD
1391 vm_page_t mtmp;
1392
54f51aeb 1393 mtmp = bp->b_xio.xio_pages[j];
984263bc
MD
1394 if (mtmp == bogus_page) {
1395 mtmp = vm_page_lookup(obj, poff + j);
1396 if (!mtmp) {
fc92d4aa 1397 panic("brelse: page missing");
984263bc 1398 }
54f51aeb 1399 bp->b_xio.xio_pages[j] = mtmp;
984263bc
MD
1400 }
1401 }
1402
1403 if ((bp->b_flags & B_INVAL) == 0) {
54f51aeb
HP
1404 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1405 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
984263bc 1406 }
54f51aeb 1407 m = bp->b_xio.xio_pages[i];
984263bc 1408 }
8d429613
MD
1409
1410 /*
1411 * Invalidate the backing store if B_NOCACHE is set
1412 * (e.g. used with vinvalbuf()). If this is NFS
1413 * we impose a requirement that the block size be
1414 * a multiple of PAGE_SIZE and create a temporary
1415 * hack to basically invalidate the whole page. The
1416 * problem is that NFS uses really odd buffer sizes
1417 * especially when tracking piecemeal writes and
1418 * it also vinvalbuf()'s a lot, which would result
1419 * in only partial page validation and invalidation
1420 * here. If the file page is mmap()'d, however,
1421 * all the valid bits get set so after we invalidate
1422 * here we would end up with weird m->valid values
1423 * like 0xfc. nfs_getpages() can't handle this so
1424 * we clear all the valid bits for the NFS case
1425 * instead of just some of them.
1426 *
1427 * The real bug is the VM system having to set m->valid
1428 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1429 * itself is an artifact of the whole 512-byte
1430 * granular mess that exists to support odd block
1431 * sizes and UFS meta-data block sizes (e.g. 6144).
1432 * A complete rewrite is required.
cb1cf930
MD
1433 *
1434 * XXX
8d429613 1435 */
984263bc
MD
1436 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1437 int poffset = foff & PAGE_MASK;
8d429613
MD
1438 int presid;
1439
1440 presid = PAGE_SIZE - poffset;
1441 if (bp->b_vp->v_tag == VT_NFS &&
1442 bp->b_vp->v_type == VREG) {
1443 ; /* entire page */
1444 } else if (presid > resid) {
1445 presid = resid;
1446 }
984263bc
MD
1447 KASSERT(presid >= 0, ("brelse: extra page"));
1448 vm_page_set_invalid(m, poffset, presid);
c504e38e
MD
1449
1450 /*
1451 * Also make sure any swap cache is removed
1452 * as it is now stale (HAMMER in particular
1453 * uses B_NOCACHE to deal with buffer
1454 * aliasing).
1455 */
1456 swap_pager_unswapped(m);
984263bc
MD
1457 }
1458 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1459 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1460 }
984263bc
MD
1461 if (bp->b_flags & (B_INVAL | B_RELBUF))
1462 vfs_vmio_release(bp);
b1c20cfa 1463 rel_mplock();
9188c711
MD
1464 } else {
1465 /*
1466 * Rundown for non-VMIO buffers.
1467 */
1468 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
b1c20cfa 1469 get_mplock();
9188c711
MD
1470 if (bp->b_bufsize)
1471 allocbuf(bp, 0);
f9a11477 1472 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
9188c711
MD
1473 if (bp->b_vp)
1474 brelvp(bp);
b1c20cfa 1475 rel_mplock();
9188c711 1476 }
984263bc
MD
1477 }
1478
b3098c79 1479 if (bp->b_qindex != BQUEUE_NONE)
984263bc 1480 panic("brelse: free buffer onto another queue???");
77bb9400 1481 if (BUF_REFCNTNB(bp) > 1) {
984263bc
MD
1482 /* Temporary panic to verify exclusive locking */
1483 /* This panic goes away when we allow shared refs */
1484 panic("brelse: multiple refs");
b1c20cfa 1485 /* NOT REACHED */
984263bc
MD
1486 return;
1487 }
1488
9188c711
MD
1489 /*
1490 * Figure out the correct queue to place the cleaned up buffer on.
1491 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1492 * disassociated from their vnode.
1493 */
c3d1e862 1494 spin_lock_wr(&bufspin);
408357d8
MD
1495 if (bp->b_flags & B_LOCKED) {
1496 /*
27bc0cb1
MD
1497 * Buffers that are locked are placed in the locked queue
1498 * immediately, regardless of their state.
408357d8 1499 */
27bc0cb1
MD
1500 bp->b_qindex = BQUEUE_LOCKED;
1501 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
408357d8 1502 } else if (bp->b_bufsize == 0) {
9188c711
MD
1503 /*
1504 * Buffers with no memory. Due to conditionals near the top
1505 * of brelse() such buffers should probably already be
1506 * marked B_INVAL and disassociated from their vnode.
1507 */
984263bc 1508 bp->b_flags |= B_INVAL;
54078292 1509 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1f1ea522 1510 KKASSERT((bp->b_flags & B_HASHED) == 0);
984263bc 1511 if (bp->b_kvasize) {
b3098c79 1512 bp->b_qindex = BQUEUE_EMPTYKVA;
984263bc 1513 } else {
b3098c79 1514 bp->b_qindex = BQUEUE_EMPTY;
984263bc
MD
1515 }
1516 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
78a9b77f 1517 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
9188c711
MD
1518 /*
1519 * Buffers with junk contents. Again these buffers had better
1520 * already be disassociated from their vnode.
1521 */
54078292 1522 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1f1ea522 1523 KKASSERT((bp->b_flags & B_HASHED) == 0);
984263bc 1524 bp->b_flags |= B_INVAL;
b3098c79
HP
1525 bp->b_qindex = BQUEUE_CLEAN;
1526 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
984263bc 1527 } else {
9188c711
MD
1528 /*
1529 * Remaining buffers. These buffers are still associated with
1530 * their vnode.
1531 */
b86460bf 1532 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
984263bc 1533 case B_DELWRI:
b3098c79
HP
1534 bp->b_qindex = BQUEUE_DIRTY;
1535 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
984263bc 1536 break;
4b958e7b
MD
1537 case B_DELWRI | B_HEAVY:
1538 bp->b_qindex = BQUEUE_DIRTY_HW;
1539 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1540 b_freelist);
1541 break;
984263bc 1542 default:
b86460bf
MD
1543 /*
1544 * NOTE: Buffers are always placed at the end of the
1545 * queue. If B_AGE is not set the buffer will cycle
1546 * through the queue twice.
1547 */
b3098c79
HP
1548 bp->b_qindex = BQUEUE_CLEAN;
1549 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
984263bc
MD
1550 break;
1551 }
1552 }
c3d1e862 1553 spin_unlock_wr(&bufspin);
984263bc
MD
1554
1555 /*
1556 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1557 * on the correct queue.
1558 */
1559 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1560 bundirty(bp);
1561
1562 /*
868d24af
MD
1563 * The bp is on an appropriate queue unless locked. If it is not
1564 * locked or dirty we can wakeup threads waiting for buffer space.
1565 *
984263bc
MD
1566 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1567 * if B_INVAL is set ).
1568 */
408357d8 1569 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
984263bc
MD
1570 bufcountwakeup();
1571
1572 /*
1573 * Something we can maybe free or reuse
1574 */
1575 if (bp->b_bufsize || bp->b_kvasize)
1576 bufspacewakeup();
1577
69f8c926
MD
1578 /*
1579 * Clean up temporary flags and unlock the buffer.
1580 */
ae8e83e6 1581 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
69f8c926 1582 BUF_UNLOCK(bp);
984263bc
MD
1583}
1584
1585/*
3f779080
HP
1586 * bqrelse:
1587 *
1588 * Release a buffer back to the appropriate queue but do not try to free
1589 * it. The buffer is expected to be used again soon.
984263bc 1590 *
3f779080
HP
1591 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1592 * biodone() to requeue an async I/O on completion. It is also used when
1593 * known good buffers need to be requeued but we think we may need the data
1594 * again soon.
984263bc 1595 *
3f779080 1596 * XXX we should be able to leave the B_RELBUF hint set on completion.
b1c20cfa
MD
1597 *
1598 * MPSAFE
984263bc
MD
1599 */
1600void
c8e4131d 1601bqrelse(struct buf *bp)
984263bc 1602{
984263bc
MD
1603 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1604
b3098c79 1605 if (bp->b_qindex != BQUEUE_NONE)
984263bc 1606 panic("bqrelse: free buffer onto another queue???");
77bb9400 1607 if (BUF_REFCNTNB(bp) > 1) {
984263bc
MD
1608 /* do not release to free list */
1609 panic("bqrelse: multiple refs");
984263bc
MD
1610 return;
1611 }
c3d1e862 1612
0e8bd897
MD
1613 buf_act_advance(bp);
1614
c3d1e862 1615 spin_lock_wr(&bufspin);
984263bc 1616 if (bp->b_flags & B_LOCKED) {
5e23ca53
MD
1617 /*
1618 * Locked buffers are released to the locked queue. However,
1619 * if the buffer is dirty it will first go into the dirty
1620 * queue and later on after the I/O completes successfully it
1621 * will be released to the locked queue.
1622 */
27bc0cb1
MD
1623 bp->b_qindex = BQUEUE_LOCKED;
1624 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
984263bc 1625 } else if (bp->b_flags & B_DELWRI) {
4b958e7b
MD
1626 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1627 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1628 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
984263bc
MD
1629 } else if (vm_page_count_severe()) {
1630 /*
1631 * We are too low on memory, we have to try to free the
1632 * buffer (most importantly: the wired pages making up its
1633 * backing store) *now*.
1634 */
c3d1e862 1635 spin_unlock_wr(&bufspin);
984263bc
MD
1636 brelse(bp);
1637 return;
1638 } else {
b3098c79
HP
1639 bp->b_qindex = BQUEUE_CLEAN;
1640 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
984263bc 1641 }
c3d1e862 1642 spin_unlock_wr(&bufspin);
984263bc
MD
1643
1644 if ((bp->b_flags & B_LOCKED) == 0 &&
408357d8 1645 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
984263bc
MD
1646 bufcountwakeup();
1647 }
1648
1649 /*
1650 * Something we can maybe free or reuse.
1651 */
1652 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1653 bufspacewakeup();
1654
9188c711
MD
1655 /*
1656 * Final cleanup and unlock. Clear bits that are only used while a
1657 * buffer is actively locked.
1658 */
ae8e83e6 1659 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
aa166ad1 1660 dsched_exit_buf(bp);
9188c711 1661 BUF_UNLOCK(bp);
984263bc
MD
1662}
1663
3f779080
HP
1664/*
1665 * vfs_vmio_release:
1666 *
1667 * Return backing pages held by the buffer 'bp' back to the VM system
1668 * if possible. The pages are freed if they are no longer valid or
1669 * attempt to free if it was used for direct I/O otherwise they are
1670 * sent to the page cache.
1671 *
1672 * Pages that were marked busy are left alone and skipped.
1673 *
1674 * The KVA mapping (b_data) for the underlying pages is removed by
1675 * this function.
1676 */
984263bc 1677static void
493c516a 1678vfs_vmio_release(struct buf *bp)
984263bc 1679{
e43a034f 1680 int i;
984263bc
MD
1681 vm_page_t m;
1682
e43a034f 1683 crit_enter();
54f51aeb
HP
1684 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1685 m = bp->b_xio.xio_pages[i];
1686 bp->b_xio.xio_pages[i] = NULL;
0e8bd897 1687
984263bc 1688 /*
b8a41159
MD
1689 * The VFS is telling us this is not a meta-data buffer
1690 * even if it is backed by a block device.
1691 */
1692 if (bp->b_flags & B_NOTMETA)
1693 vm_page_flag_set(m, PG_NOTMETA);
1694
1695 /*
0e8bd897
MD
1696 * This is a very important bit of code. We try to track
1697 * VM page use whether the pages are wired into the buffer
1698 * cache or not. While wired into the buffer cache the
1699 * bp tracks the act_count.
1700 *
1701 * We can choose to place unwired pages on the inactive
1702 * queue (0) or active queue (1). If we place too many
1703 * on the active queue the queue will cycle the act_count
1704 * on pages we'd like to keep, just from single-use pages
1705 * (such as when doing a tar-up or file scan).
984263bc 1706 */
0e8bd897
MD
1707 if (bp->b_act_count < vm_cycle_point)
1708 vm_page_unwire(m, 0);
1709 else
1710 vm_page_unwire(m, 1);
1711
984263bc
MD
1712 /*
1713 * We don't mess with busy pages, it is
1714 * the responsibility of the process that
1715 * busied the pages to deal with them.
1716 */
1717 if ((m->flags & PG_BUSY) || (m->busy != 0))
1718 continue;
1719
1720 if (m->wire_count == 0) {
1721 vm_page_flag_clear(m, PG_ZERO);
1722 /*
1723 * Might as well free the page if we can and it has
1724 * no valid data. We also free the page if the
1725 * buffer was used for direct I/O.
1726 */
ae8e83e6 1727#if 0
3f779080
HP
1728 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1729 m->hold_count == 0) {
984263bc
MD
1730 vm_page_busy(m);
1731 vm_page_protect(m, VM_PROT_NONE);
1732 vm_page_free(m);
ae8e83e6
MD
1733 } else
1734#endif
1735 if (bp->b_flags & B_DIRECT) {
984263bc
MD
1736 vm_page_try_to_free(m);
1737 } else if (vm_page_count_severe()) {
0e8bd897 1738 m->act_count = bp->b_act_count;
984263bc 1739 vm_page_try_to_cache(m);
0e8bd897
MD
1740 } else {
1741 m->act_count = bp->b_act_count;
984263bc
MD
1742 }
1743 }
1744 }
e43a034f 1745 crit_exit();
54f51aeb 1746 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
984263bc
MD
1747 if (bp->b_bufsize) {
1748 bufspacewakeup();
1749 bp->b_bufsize = 0;
1750 }
54f51aeb 1751 bp->b_xio.xio_npages = 0;
984263bc 1752 bp->b_flags &= ~B_VMIO;
f9a11477 1753 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
b1c20cfa
MD
1754 if (bp->b_vp) {
1755 get_mplock();
984263bc 1756 brelvp(bp);
b1c20cfa
MD
1757 rel_mplock();
1758 }
984263bc
MD
1759}
1760
1761/*
3f779080 1762 * vfs_bio_awrite:
984263bc
MD
1763 *
1764 * Implement clustered async writes for clearing out B_DELWRI buffers.
1765 * This is much better then the old way of writing only one buffer at
1766 * a time. Note that we may not be presented with the buffers in the
1767 * correct order, so we search for the cluster in both directions.
6f68d895
MD
1768 *
1769 * The buffer is locked on call.
984263bc
MD
1770 */
1771int
6f68d895 1772vfs_bio_awrite(struct buf *bp)
984263bc
MD
1773{
1774 int i;
1775 int j;
54078292 1776 off_t loffset = bp->b_loffset;
984263bc 1777 struct vnode *vp = bp->b_vp;
54078292 1778 int nbytes;
984263bc
MD
1779 struct buf *bpa;
1780 int nwritten;
1781 int size;
984263bc 1782
984263bc
MD
1783 /*
1784 * right now we support clustered writing only to regular files. If
1785 * we find a clusterable block we could be in the middle of a cluster
1786 * rather then at the beginning.
81b5c339 1787 *
54078292
MD
1788 * NOTE: b_bio1 contains the logical loffset and is aliased
1789 * to b_loffset. b_bio2 contains the translated block number.
984263bc
MD
1790 */
1791 if ((vp->v_type == VREG) &&
1792 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1793 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1794
1795 size = vp->v_mount->mnt_stat.f_iosize;
984263bc 1796
54078292 1797 for (i = size; i < MAXPHYS; i += size) {
b1c20cfa 1798 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
984263bc
MD
1799 BUF_REFCNT(bpa) == 0 &&
1800 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1801 (B_DELWRI | B_CLUSTEROK)) &&
1802 (bpa->b_bufsize == size)) {
54078292
MD
1803 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1804 (bpa->b_bio2.bio_offset !=
1805 bp->b_bio2.bio_offset + i))
984263bc
MD
1806 break;
1807 } else {
1808 break;
1809 }
1810 }
54078292 1811 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
b1c20cfa 1812 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
984263bc
MD
1813 BUF_REFCNT(bpa) == 0 &&
1814 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1815 (B_DELWRI | B_CLUSTEROK)) &&
1816 (bpa->b_bufsize == size)) {
54078292
MD
1817 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1818 (bpa->b_bio2.bio_offset !=
1819 bp->b_bio2.bio_offset - j))
984263bc
MD
1820 break;
1821 } else {
1822 break;
1823 }
1824 }
54078292
MD
1825 j -= size;
1826 nbytes = (i + j);
b1c20cfa 1827
984263bc
MD
1828 /*
1829 * this is a possible cluster write
1830 */
54078292 1831 if (nbytes != size) {
6f68d895 1832 BUF_UNLOCK(bp);
54078292
MD
1833 nwritten = cluster_wbuild(vp, size,
1834 loffset - j, nbytes);
984263bc
MD
1835 return nwritten;
1836 }
1837 }
1838
984263bc
MD
1839 /*
1840 * default (old) behavior, writing out only one block
1841 *
1842 * XXX returns b_bufsize instead of b_bcount for nwritten?
1843 */
1844 nwritten = bp->b_bufsize;
ae8e83e6
MD
1845 bremfree(bp);
1846 bawrite(bp);
984263bc
MD
1847
1848 return nwritten;
1849}
1850
1851/*
3f779080 1852 * getnewbuf:
984263bc
MD
1853 *
1854 * Find and initialize a new buffer header, freeing up existing buffers
1855 * in the bufqueues as necessary. The new buffer is returned locked.
1856 *
1857 * Important: B_INVAL is not set. If the caller wishes to throw the
1858 * buffer away, the caller must set B_INVAL prior to calling brelse().
1859 *
1860 * We block if:
1861 * We have insufficient buffer headers
1862 * We have insufficient buffer space
1863 * buffer_map is too fragmented ( space reservation fails )
1864 * If we have to flush dirty buffers ( but we try to avoid this )
1865 *
1866 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1867 * Instead we ask the buf daemon to do it for us. We attempt to
1868 * avoid piecemeal wakeups of the pageout daemon.
b1c20cfa
MD
1869 *
1870 * MPALMOSTSAFE
984263bc 1871 */
984263bc 1872static struct buf *
4b958e7b 1873getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
984263bc
MD
1874{
1875 struct buf *bp;
1876 struct buf *nbp;
1877 int defrag = 0;
1878 int nqindex;
4b958e7b 1879 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
984263bc
MD
1880 static int flushingbufs;
1881
1882 /*
1883 * We can't afford to block since we might be holding a vnode lock,
1884 * which may prevent system daemons from running. We deal with
1885 * low-memory situations by proactively returning memory and running
1886 * async I/O rather then sync I/O.
1887 */
1888
1889 ++getnewbufcalls;
1890 --getnewbufrestarts;
1891restart:
1892 ++getnewbufrestarts;
1893
1894 /*
1895 * Setup for scan. If we do not have enough free buffers,
1896 * we setup a degenerate case that immediately fails. Note
1897 * that if we are specially marked process, we are allowed to
1898 * dip into our reserves.
1899 *
1900 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1901 *
1902 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1903 * However, there are a number of cases (defragging, reusing, ...)
1904 * where we cannot backup.
1905 */
b3098c79 1906 nqindex = BQUEUE_EMPTYKVA;
c3d1e862 1907 spin_lock_wr(&bufspin);
b3098c79 1908 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
984263bc
MD
1909
1910 if (nbp == NULL) {
1911 /*
1912 * If no EMPTYKVA buffers and we are either
1913 * defragging or reusing, locate a CLEAN buffer
1914 * to free or reuse. If bufspace useage is low
1915 * skip this step so we can allocate a new buffer.
1916 */
1917 if (defrag || bufspace >= lobufspace) {
b3098c79
HP
1918 nqindex = BQUEUE_CLEAN;
1919 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
984263bc
MD
1920 }
1921
1922 /*
1923 * If we could not find or were not allowed to reuse a
1924 * CLEAN buffer, check to see if it is ok to use an EMPTY
1925 * buffer. We can only use an EMPTY buffer if allocating
1926 * its KVA would not otherwise run us out of buffer space.
1927 */
1928 if (nbp == NULL && defrag == 0 &&
1929 bufspace + maxsize < hibufspace) {
b3098c79
HP
1930 nqindex = BQUEUE_EMPTY;
1931 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
984263bc
MD
1932 }
1933 }
1934
1935 /*
1936 * Run scan, possibly freeing data and/or kva mappings on the fly
1937 * depending.
c3d1e862
MD
1938 *
1939 * WARNING! bufspin is held!
984263bc 1940 */
984263bc
MD
1941 while ((bp = nbp) != NULL) {
1942 int qindex = nqindex;
1943
b86460bf
MD
1944 nbp = TAILQ_NEXT(bp, b_freelist);
1945
1946 /*
1947 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1948 * cycles through the queue twice before being selected.
1949 */
1950 if (qindex == BQUEUE_CLEAN &&
1951 (bp->b_flags & B_AGE) == 0 && nbp) {
1952 bp->b_flags |= B_AGE;
1953 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1954 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1955 continue;
1956 }
1957
984263bc
MD
1958 /*
1959 * Calculate next bp ( we can only use it if we do not block
1960 * or do other fancy things ).
1961 */
b86460bf 1962 if (nbp == NULL) {
984263bc 1963 switch(qindex) {
b3098c79
HP
1964 case BQUEUE_EMPTY:
1965 nqindex = BQUEUE_EMPTYKVA;
1966 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
984263bc
MD
1967 break;
1968 /* fall through */
b3098c79
HP
1969 case BQUEUE_EMPTYKVA:
1970 nqindex = BQUEUE_CLEAN;
1971 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
984263bc
MD
1972 break;
1973 /* fall through */
b3098c79 1974 case BQUEUE_CLEAN:
984263bc
MD
1975 /*
1976 * nbp is NULL.
1977 */
1978 break;
1979 }
1980 }
1981
1982 /*
1983 * Sanity Checks
1984 */
3f625015 1985 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
984263bc
MD
1986
1987 /*
1988 * Note: we no longer distinguish between VMIO and non-VMIO
1989 * buffers.
1990 */
1991
1992 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1993
1994 /*
1995 * If we are defragging then we need a buffer with
1996 * b_kvasize != 0. XXX this situation should no longer
1997 * occur, if defrag is non-zero the buffer's b_kvasize
1998 * should also be non-zero at this point. XXX
1999 */
2000 if (defrag && bp->b_kvasize == 0) {
6ea70f76 2001 kprintf("Warning: defrag empty buffer %p\n", bp);
984263bc
MD
2002 continue;
2003 }
2004
2005 /*
2006 * Start freeing the bp. This is somewhat involved. nbp
b3098c79 2007 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
9188c711
MD
2008 * on the clean list must be disassociated from their
2009 * current vnode. Buffers on the empty[kva] lists have
2010 * already been disassociated.
984263bc
MD
2011 */
2012
d9dba6f6 2013 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
c3d1e862 2014 spin_unlock_wr(&bufspin);
d9dba6f6
MD
2015 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2016 goto restart;
2017 }
2018 if (bp->b_qindex != qindex) {
c3d1e862 2019 spin_unlock_wr(&bufspin);
6ea70f76 2020 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
d9dba6f6
MD
2021 BUF_UNLOCK(bp);
2022 goto restart;
2023 }
c3d1e862
MD
2024 bremfree_locked(bp);
2025 spin_unlock_wr(&bufspin);
984263bc 2026
408357d8
MD
2027 /*
2028 * Dependancies must be handled before we disassociate the
2029 * vnode.
2030 *
2031 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2032 * be immediately disassociated. HAMMER then becomes
2033 * responsible for releasing the buffer.
c3d1e862
MD
2034 *
2035 * NOTE: bufspin is UNLOCKED now.
408357d8
MD
2036 */
2037 if (LIST_FIRST(&bp->b_dep) != NULL) {
b1c20cfa 2038 get_mplock();
408357d8 2039 buf_deallocate(bp);
b1c20cfa 2040 rel_mplock();
408357d8
MD
2041 if (bp->b_flags & B_LOCKED) {
2042 bqrelse(bp);
2043 goto restart;
2044 }
4b958e7b 2045 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
408357d8
MD
2046 }
2047
b3098c79 2048 if (qindex == BQUEUE_CLEAN) {
b1c20cfa 2049 get_mplock();
984263bc 2050 if (bp->b_flags & B_VMIO) {
b1c20cfa 2051 get_mplock();
984263bc 2052 vfs_vmio_release(bp);
b1c20cfa 2053 rel_mplock();
984263bc
MD
2054 }
2055 if (bp->b_vp)
2056 brelvp(bp);
b1c20cfa 2057 rel_mplock();
984263bc
MD
2058 }
2059
2060 /*
2061 * NOTE: nbp is now entirely invalid. We can only restart
2062 * the scan from this point on.
2063 *
2064 * Get the rest of the buffer freed up. b_kva* is still
2065 * valid after this operation.
2066 */
2067
54078292 2068 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1f1ea522 2069 KKASSERT((bp->b_flags & B_HASHED) == 0);
984263bc 2070
06ecca5a 2071 /*
e43a034f
MD
2072 * critical section protection is not required when
2073 * scrapping a buffer's contents because it is already
2074 * wired.
06ecca5a 2075 */
b1c20cfa
MD
2076 if (bp->b_bufsize) {
2077 get_mplock();
984263bc 2078 allocbuf(bp, 0);
b1c20cfa
MD
2079 rel_mplock();
2080 }
984263bc 2081
4414f2c9 2082 bp->b_flags = B_BNOCLIP;
10f3fee5 2083 bp->b_cmd = BUF_CMD_DONE;
984263bc 2084 bp->b_vp = NULL;
984263bc
MD
2085 bp->b_error = 0;
2086 bp->b_resid = 0;
2087 bp->b_bcount = 0;
54f51aeb 2088 bp->b_xio.xio_npages = 0;
984263bc 2089 bp->b_dirtyoff = bp->b_dirtyend = 0;
0e8bd897 2090 bp->b_act_count = ACT_INIT;
81b5c339 2091 reinitbufbio(bp);
b86460bf 2092 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
408357d8 2093 buf_dep_init(bp);
4b958e7b
MD
2094 if (blkflags & GETBLK_BHEAVY)
2095 bp->b_flags |= B_HEAVY;
984263bc
MD
2096
2097 /*
2098 * If we are defragging then free the buffer.
2099 */
2100 if (defrag) {
2101 bp->b_flags |= B_INVAL;
2102 bfreekva(bp);
2103 brelse(bp);
2104 defrag = 0;
2105 goto restart;
2106 }
2107
2108 /*
2109 * If we are overcomitted then recover the buffer and its
2110 * KVM space. This occurs in rare situations when multiple
2111 * processes are blocked in getnewbuf() or allocbuf().
2112 */
2113 if (bufspace >= hibufspace)
2114 flushingbufs = 1;
2115 if (flushingbufs && bp->b_kvasize != 0) {
2116 bp->b_flags |= B_INVAL;
2117 bfreekva(bp);
2118 brelse(bp);
2119 goto restart;
2120 }
2121 if (bufspace < lobufspace)
2122 flushingbufs = 0;
2123 break;
c3d1e862 2124 /* NOT REACHED, bufspin not held */
984263bc
MD
2125 }
2126
2127 /*
2128 * If we exhausted our list, sleep as appropriate. We may have to
2129 * wakeup various daemons and write out some dirty buffers.
2130 *
2131 * Generally we are sleeping due to insufficient buffer space.
c3d1e862
MD
2132 *
2133 * NOTE: bufspin is held if bp is NULL, else it is not held.
984263bc 2134 */
984263bc
MD
2135 if (bp == NULL) {
2136 int flags;
2137 char *waitmsg;
2138
c3d1e862 2139 spin_unlock_wr(&bufspin);
984263bc
MD
2140 if (defrag) {
2141 flags = VFS_BIO_NEED_BUFSPACE;
2142 waitmsg = "nbufkv";
2143 } else if (bufspace >= hibufspace) {
2144 waitmsg = "nbufbs";
2145 flags = VFS_BIO_NEED_BUFSPACE;
2146 } else {
2147 waitmsg = "newbuf";
2148 flags = VFS_BIO_NEED_ANY;
2149 }
2150
984263bc 2151 needsbuffer |= flags;
4b958e7b 2152 bd_speedup(); /* heeeelp */
984263bc 2153 while (needsbuffer & flags) {
4b958e7b 2154 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
984263bc
MD
2155 return (NULL);
2156 }
2157 } else {
2158 /*
2159 * We finally have a valid bp. We aren't quite out of the
2160 * woods, we still have to reserve kva space. In order
2161 * to keep fragmentation sane we only allocate kva in
2162 * BKVASIZE chunks.
c3d1e862
MD
2163 *
2164 * (bufspin is not held)
984263bc
MD
2165 */
2166 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2167
2168 if (maxsize != bp->b_kvasize) {
2169 vm_offset_t addr = 0;
a108bf71 2170 int count;
984263bc
MD
2171
2172 bfreekva(bp);
2173
b1c20cfa 2174 get_mplock();
a108bf71 2175 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
e4846942 2176 vm_map_lock(&buffer_map);
984263bc 2177
e4846942
MD
2178 if (vm_map_findspace(&buffer_map,
2179 vm_map_min(&buffer_map), maxsize,
c809941b 2180 maxsize, 0, &addr)) {
984263bc 2181 /*
3f779080 2182 * Uh oh. Buffer map is too fragmented. We
984263bc
MD
2183 * must defragment the map.
2184 */
e4846942 2185 vm_map_unlock(&buffer_map);
a108bf71 2186 vm_map_entry_release(count);
984263bc
MD
2187 ++bufdefragcnt;
2188 defrag = 1;
2189 bp->b_flags |= B_INVAL;
b1c20cfa 2190 rel_mplock();
984263bc
MD
2191 brelse(bp);
2192 goto restart;
2193 }
2194 if (addr) {
e4846942 2195 vm_map_insert(&buffer_map, &count,
a108bf71 2196 NULL, 0,
984263bc 2197 addr, addr + maxsize,
1b874851
MD
2198 VM_MAPTYPE_NORMAL,
2199 VM_PROT_ALL, VM_PROT_ALL,
2200 MAP_NOFAULT);
984263bc
MD
2201
2202 bp->b_kvabase = (caddr_t) addr;
2203 bp->b_kvasize = maxsize;
2204 bufspace += bp->b_kvasize;
2205 ++bufreusecnt;
2206 }
e4846942 2207 vm_map_unlock(&buffer_map);
a108bf71 2208 vm_map_entry_release(count);
b1c20cfa 2209 rel_mplock();
984263bc
MD
2210 }
2211 bp->b_data = bp->b_kvabase;
2212 }
2213 return(bp);
2214}
2215
2216/*
4ecf7cc9
MD
2217 * This routine is called in an emergency to recover VM pages from the
2218 * buffer cache by cashing in clean buffers. The idea is to recover
2219 * enough pages to be able to satisfy a stuck bio_page_alloc().
2220 */
2221static int
2222recoverbufpages(void)
2223{
2224 struct buf *bp;
2225 int bytes = 0;
2226
2227 ++recoverbufcalls;
2228
c3d1e862 2229 spin_lock_wr(&bufspin);
4ecf7cc9
MD
2230 while (bytes < MAXBSIZE) {
2231 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2232 if (bp == NULL)
2233 break;
2234
2235 /*
2236 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2237 * cycles through the queue twice before being selected.
2238 */
2239 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2240 bp->b_flags |= B_AGE;
2241 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2242 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2243 bp, b_freelist);
2244 continue;
2245 }
2246
2247 /*
2248 * Sanity Checks
2249 */
2250 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2251 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2252
2253 /*
2254 * Start freeing the bp. This is somewhat involved.
2255 *
2256 * Buffers on the clean list must be disassociated from
2257 * their current vnode
2258 */
2259
2260 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2261 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2262 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2263 continue;
2264 }
2265 if (bp->b_qindex != BQUEUE_CLEAN) {
2266 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2267 BUF_UNLOCK(bp);
2268 continue;
2269 }
c3d1e862
MD
2270 bremfree_locked(bp);
2271 spin_unlock_wr(&bufspin);
4ecf7cc9
MD
2272
2273 /*
2274 * Dependancies must be handled before we disassociate the
2275 * vnode.
2276 *
2277 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2278 * be immediately disassociated. HAMMER then becomes
2279 * responsible for releasing the buffer.
2280 */
2281 if (LIST_FIRST(&bp->b_dep) != NULL) {
2282 buf_deallocate(bp);
2283 if (bp->b_flags & B_LOCKED) {
2284 bqrelse(bp);
c3d1e862 2285 spin_lock_wr(&bufspin);
4ecf7cc9
MD
2286 continue;
2287 }
2288 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2289 }
2290
2291 bytes += bp->b_bufsize;
2292
b1c20cfa 2293 get_mplock();
4ecf7cc9 2294 if (bp->b_flags & B_VMIO) {
4ecf7cc9
MD
2295 bp->b_flags |= B_DIRECT; /* try to free pages */
2296 vfs_vmio_release(bp);
2297 }
2298 if (bp->b_vp)
2299 brelvp(bp);
2300
2301 KKASSERT(bp->b_vp == NULL);
2302 KKASSERT((bp->b_flags & B_HASHED) == 0);
2303
2304 /*
2305 * critical section protection is not required when
2306 * scrapping a buffer's contents because it is already
2307 * wired.
2308 */
2309 if (bp->b_bufsize)
2310 allocbuf(bp, 0);
b1c20cfa 2311 rel_mplock();
4ecf7cc9
MD
2312
2313 bp->b_flags = B_BNOCLIP;
2314 bp->b_cmd = BUF_CMD_DONE;
2315 bp->b_vp = NULL;
2316 bp->b_error = 0;
2317 bp->b_resid = 0;
2318 bp->b_bcount = 0;
2319 bp->b_xio.xio_npages = 0;
2320 bp->b_dirtyoff = bp->b_dirtyend = 0;
2321 reinitbufbio(bp);
2322 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2323 buf_dep_init(bp);
2324 bp->b_flags |= B_INVAL;
2325 /* bfreekva(bp); */
2326 brelse(bp);
c3d1e862 2327 spin_lock_wr(&bufspin);
4ecf7cc9 2328 }
c3d1e862 2329 spin_unlock_wr(&bufspin);
4ecf7cc9
MD
2330 return(bytes);
2331}
2332
2333/*
3f779080 2334 * buf_daemon:
984263bc 2335 *
3f779080 2336 * Buffer flushing daemon. Buffers are normally flushed by the
984263bc
MD
2337 * update daemon but if it cannot keep up this process starts to
2338 * take the load in an attempt to prevent getnewbuf() from blocking.
4b958e7b
MD
2339 *
2340 * Once a flush is initiated it does not stop until the number
2341 * of buffers falls below lodirtybuffers, but we will wake up anyone
2342 * waiting at the mid-point.
984263bc
MD
2343 */
2344
984263bc
MD
2345static struct kproc_desc buf_kp = {
2346 "bufdaemon",
2347 buf_daemon,
4b958e7b
MD
2348 &bufdaemon_td
2349};
2350SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2351 kproc_start, &buf_kp)
2352
2353static struct kproc_desc bufhw_kp = {
2354 "bufdaemon_hw",
2355 buf_daemon_hw,
2356 &bufdaemonhw_td
984263bc 2357};
4b958e7b
MD
2358SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2359 kproc_start, &bufhw_kp)
984263bc
MD
2360
2361static void
c972a82f 2362buf_daemon(void)
984263bc 2363{
cd083340
MD
2364 int limit;
2365
984263bc
MD
2366 /*
2367 * This process needs to be suspended prior to shutdown sync.
2368 */
bc6dffab 2369 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
4b958e7b 2370 bufdaemon_td, SHUTDOWN_PRI_LAST);
4ecf7cc9 2371 curthread->td_flags |= TDF_SYSTHREAD;
984263bc
MD
2372
2373 /*
2374 * This process is allowed to take the buffer cache to the limit
2375 */
e43a034f 2376 crit_enter();
984263bc
MD
2377
2378 for (;;) {
0cfcada1 2379 kproc_suspend_loop();
984263bc
MD
2380
2381 /*
4afeea0d
MD
2382 * Do the flush as long as the number of dirty buffers
2383 * (including those running) exceeds lodirtybufspace.
2384 *
2385 * When flushing limit running I/O to hirunningspace
984263bc
MD
2386 * Do the flush. Limit the amount of in-transit I/O we
2387 * allow to build up, otherwise we would completely saturate
2388 * the I/O system. Wakeup any waiting processes before we
2389 * normally would so they can run in parallel with our drain.
cd083340
MD
2390 *
2391 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2392 * but because we split the operation into two threads we
2393 * have to cut it in half for each thread.
984263bc 2394 */
4afeea0d 2395 waitrunningbufspace();
cd083340 2396 limit = lodirtybufspace / 2;
70ac7d6c
MD
2397 while (runningbufspace + dirtybufspace > limit ||
2398 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
4b958e7b 2399 if (flushbufqueues(BQUEUE_DIRTY) == 0)
984263bc 2400 break;
4afeea0d
MD
2401 if (runningbufspace < hirunningspace)
2402 continue;
2403 waitrunningbufspace();
1b30fbcc 2404 }
984263bc
MD
2405
2406 /*
cd083340
MD
2407 * We reached our low water mark, reset the
2408 * request and sleep until we are needed again.
2409 * The sleep is just so the suspend code works.
984263bc 2410 */
cd083340
MD
2411 spin_lock_wr(&needsbuffer_spin);
2412 if (bd_request == 0) {
e590ee86 2413 ssleep(&bd_request, &needsbuffer_spin, 0,
4b958e7b 2414 "psleep", hz);
984263bc 2415 }
cd083340
MD
2416 bd_request = 0;
2417 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
2418 }
2419}
2420
4b958e7b
MD
2421static void
2422buf_daemon_hw(void)
2423{
cd083340
MD
2424 int limit;
2425
4b958e7b
MD
2426 /*
2427 * This process needs to be suspended prior to shutdown sync.
2428 */
2429 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2430 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
4ecf7cc9 2431 curthread->td_flags |= TDF_SYSTHREAD;
4b958e7b
MD
2432
2433 /*
2434 * This process is allowed to take the buffer cache to the limit
2435 */
2436 crit_enter();
2437
2438 for (;;) {
2439 kproc_suspend_loop();
2440
2441 /*
2442 * Do the flush. Limit the amount of in-transit I/O we
2443 * allow to build up, otherwise we would completely saturate
2444 * the I/O system. Wakeup any waiting processes before we
2445 * normally would so they can run in parallel with our drain.
cd083340 2446 *
4afeea0d
MD
2447 * Once we decide to flush push the queued I/O up to
2448 * hirunningspace in order to trigger bursting by the bioq
2449 * subsystem.
2450 *
cd083340
MD
2451 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2452 * but because we split the operation into two threads we
2453 * have to cut it in half for each thread.
4b958e7b 2454 */
4afeea0d 2455 waitrunningbufspace();
cd083340 2456 limit = lodirtybufspace / 2;
70ac7d6c
MD
2457 while (runningbufspace + dirtybufspacehw > limit ||
2458 dirtybufcounthw >= nbuf / 2) {
4b958e7b
MD
2459 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2460 break;
4afeea0d
MD
2461 if (runningbufspace < hirunningspace)
2462 continue;
2463 waitrunningbufspace();
1b30fbcc 2464 }
4b958e7b
MD
2465
2466 /*
cd083340
MD
2467 * We reached our low water mark, reset the
2468 * request and sleep until we are needed again.
2469 * The sleep is just so the suspend code works.
4b958e7b 2470 */
cd083340
MD
2471 spin_lock_wr(&needsbuffer_spin);
2472 if (bd_request_hw == 0) {
e590ee86 2473 ssleep(&bd_request_hw, &needsbuffer_spin, 0,
4b958e7b 2474 "psleep", hz);
4b958e7b 2475 }
cd083340
MD
2476 bd_request_hw = 0;
2477 spin_unlock_wr(&needsbuffer_spin);
4b958e7b
MD
2478 }
2479}
2480
984263bc 2481/*
3f779080 2482 * flushbufqueues:
984263bc
MD
2483 *
2484 * Try to flush a buffer in the dirty queue. We must be careful to
2485 * free up B_INVAL buffers instead of write them, which NFS is
2486 * particularly sensitive to.
b86460bf
MD
2487 *
2488 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2489 * that we really want to try to get the buffer out and reuse it
2490 * due to the write load on the machine.
984263bc 2491 */
984263bc 2492static int
4b958e7b 2493flushbufqueues(bufq_type_t q)
984263bc
MD
2494{
2495 struct buf *bp;
2496 int r = 0;
c3d1e862
MD
2497 int spun;
2498
2499 spin_lock_wr(&bufspin);
2500 spun = 1;
984263bc 2501
4b958e7b 2502 bp = TAILQ_FIRST(&bufqueues[q]);
984263bc 2503 while (bp) {
408357d8
MD
2504 KASSERT((bp->b_flags & B_DELWRI),
2505 ("unexpected clean buffer %p", bp));
b86460bf 2506
70371608 2507 if (bp->b_flags & B_DELWRI) {
984263bc 2508 if (bp->b_flags & B_INVAL) {
c3d1e862
MD
2509 spin_unlock_wr(&bufspin);
2510 spun = 0;
984263bc
MD
2511 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2512 panic("flushbufqueues: locked buf");
2513 bremfree(bp);
2514 brelse(bp);
2515 ++r;
2516 break;
2517 }
2518 if (LIST_FIRST(&bp->b_dep) != NULL &&
984263bc 2519 (bp->b_flags & B_DEFERRED) == 0 &&
408357d8 2520 buf_countdeps(bp, 0)) {
4b958e7b
MD
2521 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2522 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2523 b_freelist);
984263bc 2524 bp->b_flags |= B_DEFERRED;
4b958e7b 2525 bp = TAILQ_FIRST(&bufqueues[q]);
984263bc
MD
2526 continue;
2527 }
6f68d895
MD
2528
2529 /*
2530 * Only write it out if we can successfully lock
27bc0cb1 2531 * it. If the buffer has a dependancy,
78a9b77f
MD
2532 * buf_checkwrite must also return 0 for us to
2533 * be able to initate the write.
2534 *
2535 * If the buffer is flagged B_ERROR it may be
2536 * requeued over and over again, we try to
2537 * avoid a live lock.
6f68d895
MD
2538 */
2539 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
c3d1e862
MD
2540 spin_unlock_wr(&bufspin);
2541 spun = 0;
27bc0cb1
MD
2542 if (LIST_FIRST(&bp->b_dep) != NULL &&
2543 buf_checkwrite(bp)) {
2544 bremfree(bp);
2545 brelse(bp);
78a9b77f
MD
2546 } else if (bp->b_flags & B_ERROR) {
2547 tsleep(bp, 0, "bioer", 1);
2548 bp->b_flags &= ~B_AGE;
2549 vfs_bio_awrite(bp);
27bc0cb1 2550 } else {
b86460bf 2551 bp->b_flags |= B_AGE;
27bc0cb1
MD
2552 vfs_bio_awrite(bp);
2553 }
6f68d895
MD
2554 ++r;
2555 break;
2556 }
984263bc
MD
2557 }
2558 bp = TAILQ_NEXT(bp, b_freelist);
2559 }
c3d1e862
MD
2560 if (spun)
2561 spin_unlock_wr(&bufspin);
984263bc
MD
2562 return (r);
2563}
2564
2565/*
3f779080
HP
2566 * inmem:
2567 *
2568 * Returns true if no I/O is needed to access the associated VM object.
1f1ea522 2569 * This is like findblk except it also hunts around in the VM system for
3f779080 2570 * the data.
06ecca5a 2571 *
3f779080
HP
2572 * Note that we ignore vm_page_free() races from interrupts against our
2573 * lookup, since if the caller is not protected our return value will not
2574 * be any more valid then otherwise once we exit the critical section.
984263bc 2575 */
984263bc 2576int
54078292 2577inmem(struct vnode *vp, off_t loffset)
984263bc
MD
2578{
2579 vm_object_t obj;
2580 vm_offset_t toff, tinc, size;
2581 vm_page_t m;
984263bc 2582
b1c20cfa 2583 if (findblk(vp, loffset, FINDBLK_TEST))
984263bc
MD
2584 return 1;
2585 if (vp->v_mount == NULL)
2586 return 0;
7540ab49
MD
2587 if ((obj = vp->v_object) == NULL)
2588 return 0;
984263bc
MD
2589
2590 size = PAGE_SIZE;
2591 if (size > vp->v_mount->mnt_stat.f_iosize)
2592 size = vp->v_mount->mnt_stat.f_iosize;
984263bc
MD
2593
2594 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
54078292
MD
2595 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2596 if (m == NULL)
984263bc
MD
2597 return 0;
2598 tinc = size;
54078292
MD
2599 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2600 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
984263bc 2601 if (vm_page_is_valid(m,
54078292 2602 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
984263bc
MD
2603 return 0;
2604 }
2605 return 1;
2606}
2607
2608/*
1f1ea522
MD
2609 * findblk:
2610 *
b1c20cfa
MD
2611 * Locate and return the specified buffer. Unless flagged otherwise,
2612 * a locked buffer will be returned if it exists or NULL if it does not.
0202303b 2613 *
ae8e83e6
MD
2614 * findblk()'d buffers are still on the bufqueues and if you intend
2615 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2616 * and possibly do other stuff to it.
2617 *
b1c20cfa
MD
2618 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2619 * for locking the buffer and ensuring that it remains
2620 * the desired buffer after locking.
0202303b 2621 *
b1c20cfa
MD
2622 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2623 * to acquire the lock we return NULL, even if the
2624 * buffer exists.
2625 *
c0885fab
MD
2626 * (0) - Lock the buffer blocking.
2627 *
b1c20cfa 2628 * MPSAFE
1f1ea522
MD
2629 */
2630struct buf *
b1c20cfa 2631findblk(struct vnode *vp, off_t loffset, int flags)
1f1ea522
MD
2632{
2633 struct buf *bp;
b1c20cfa
MD
2634 int lkflags;
2635
2636 lkflags = LK_EXCLUSIVE;
2637 if (flags & FINDBLK_NBLOCK)
2638 lkflags |= LK_NOWAIT;
1f1ea522 2639
b1c20cfa 2640 for (;;) {
3b998fa9 2641 lwkt_gettoken(&vp->v_token);
b1c20cfa 2642 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
3b998fa9 2643 lwkt_reltoken(&vp->v_token);
b1c20cfa
MD
2644 if (bp == NULL || (flags & FINDBLK_TEST))
2645 break;
c0885fab 2646 if (BUF_LOCK(bp, lkflags)) {
b1c20cfa
MD
2647 bp = NULL;
2648 break;
2649 }
2650 if (bp->b_vp == vp && bp->b_loffset == loffset)
2651 break;
2652 BUF_UNLOCK(bp);
2653 }
1f1ea522
MD
2654 return(bp);
2655}
2656
2657/*
c0885fab
MD
2658 * getcacheblk:
2659 *
2660 * Similar to getblk() except only returns the buffer if it is
2661 * B_CACHE and requires no other manipulation. Otherwise NULL
2662 * is returned.
2663 *
2664 * If B_RAM is set the buffer might be just fine, but we return
2665 * NULL anyway because we want the code to fall through to the
2666 * cluster read. Otherwise read-ahead breaks.
2667 */
2668struct buf *
2669getcacheblk(struct vnode *vp, off_t loffset)
2670{
2671 struct buf *bp;
2672
2673 bp = findblk(vp, loffset, 0);
2674 if (bp) {
2675 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2676 bp->b_flags &= ~B_AGE;
2677 bremfree(bp);
2678 } else {
2679 BUF_UNLOCK(bp);
2680 bp = NULL;
2681 }
2682 }
2683 return (bp);
2684}
2685
2686/*
3f779080 2687 * getblk:
984263bc
MD
2688 *
2689 * Get a block given a specified block and offset into a file/device.
10f3fee5
MD
2690 * B_INVAL may or may not be set on return. The caller should clear
2691 * B_INVAL prior to initiating a READ.
984263bc 2692 *
77bb9400
MD
2693 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2694 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2695 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2696 * without doing any of those things the system will likely believe
2697 * the buffer to be valid (especially if it is not B_VMIO), and the
2698 * next getblk() will return the buffer with B_CACHE set.
2699 *
984263bc
MD
2700 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2701 * an existing buffer.
2702 *
2703 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2704 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2705 * and then cleared based on the backing VM. If the previous buffer is
2706 * non-0-sized but invalid, B_CACHE will be cleared.
2707 *
2708 * If getblk() must create a new buffer, the new buffer is returned with
2709 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2710 * case it is returned with B_INVAL clear and B_CACHE set based on the
2711 * backing VM.
2712 *
62cfda27 2713 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
984263bc
MD
2714 * B_CACHE bit is clear.
2715 *
2716 * What this means, basically, is that the caller should use B_CACHE to
2717 * determine whether the buffer is fully valid or not and should clear
2718 * B_INVAL prior to issuing a read. If the caller intends to validate
2719 * the buffer by loading its data area with something, the caller needs
2720 * to clear B_INVAL. If the caller does this without issuing an I/O,
2721 * the caller should set B_CACHE ( as an optimization ), else the caller
2722 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2723 * a write attempt or if it was a successfull read. If the caller
2724 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2725 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
4b958e7b
MD
2726 *
2727 * getblk flags:
2728 *
2729 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2730 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
b1c20cfa
MD
2731 *
2732 * MPALMOSTSAFE
984263bc
MD
2733 */
2734struct buf *
4b958e7b 2735getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
984263bc
MD
2736{
2737 struct buf *bp;
4b958e7b 2738 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
e92ca23a 2739 int error;
b1c20cfa 2740 int lkflags;
984263bc
MD
2741
2742 if (size > MAXBSIZE)
fc92d4aa 2743 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
7540ab49
MD
2744 if (vp->v_object == NULL)
2745 panic("getblk: vnode %p has no object!", vp);
984263bc 2746
984263bc 2747loop:
b1c20cfa 2748 if ((bp = findblk(vp, loffset, FINDBLK_TEST)) != NULL) {
984263bc 2749 /*
a0da602d
MD
2750 * The buffer was found in the cache, but we need to lock it.
2751 * Even with LK_NOWAIT the lockmgr may break our critical
2752 * section, so double-check the validity of the buffer
2753 * once the lock has been obtained.
984263bc 2754 */
984263bc 2755 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
b1c20cfa 2756 if (blkflags & GETBLK_NOWAIT)
b77cfc40 2757 return(NULL);
b1c20cfa 2758 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
4b958e7b 2759 if (blkflags & GETBLK_PCATCH)
f2770c70 2760 lkflags |= LK_PCATCH;
e92ca23a
MD
2761 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2762 if (error) {
2763 if (error == ENOLCK)
2764 goto loop;
e92ca23a 2765 return (NULL);
f2770c70 2766 }
b1c20cfa 2767 /* buffer may have changed on us */
984263bc
MD
2768 }
2769
2770 /*
a0da602d
MD
2771 * Once the buffer has been locked, make sure we didn't race
2772 * a buffer recyclement. Buffers that are no longer hashed
2773 * will have b_vp == NULL, so this takes care of that check
2774 * as well.
2775 */
54078292 2776 if (bp->b_vp != vp || bp->b_loffset != loffset) {
973c11b9
MD
2777 kprintf("Warning buffer %p (vp %p loffset %lld) "
2778 "was recycled\n",
2779 bp, vp, (long long)loffset);
a9518ecf 2780 BUF_UNLOCK(bp);
a0da602d
MD
2781 goto loop;
2782 }
2783
2784 /*
b77cfc40
MD
2785 * If SZMATCH any pre-existing buffer must be of the requested
2786 * size or NULL is returned. The caller absolutely does not
2787 * want getblk() to bwrite() the buffer on a size mismatch.
2788 */
2789 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2790 BUF_UNLOCK(bp);
b77cfc40
MD
2791 return(NULL);
2792 }
2793
2794 /*
4baec531
MD
2795 * All vnode-based buffers must be backed by a VM object.
2796 */
2797 KKASSERT(bp->b_flags & B_VMIO);
10f3fee5 2798 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
b86460bf 2799 bp->b_flags &= ~B_AGE;
4baec531
MD
2800
2801 /*
a0da602d
MD
2802 * Make sure that B_INVAL buffers do not have a cached
2803 * block number translation.
2804 */
54078292 2805 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
973c11b9
MD
2806 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2807 " did not have cleared bio_offset cache\n",
2808 bp, vp, (long long)loffset);
81b5c339 2809 clearbiocache(&bp->b_bio2);
a0da602d
MD
2810 }
2811
2812 /*
984263bc 2813 * The buffer is locked. B_CACHE is cleared if the buffer is
4baec531 2814 * invalid.
984263bc
MD
2815 */
2816 if (bp->b_flags & B_INVAL)
2817 bp->b_flags &= ~B_CACHE;
984263bc
MD
2818 bremfree(bp);
2819
2820 /*
4baec531
MD
2821 * Any size inconsistancy with a dirty buffer or a buffer
2822 * with a softupdates dependancy must be resolved. Resizing
2823 * the buffer in such circumstances can lead to problems.
cb1cf930
MD
2824 *
2825 * Dirty or dependant buffers are written synchronously.
2826 * Other types of buffers are simply released and
2827 * reconstituted as they may be backed by valid, dirty VM
2828 * pages (but not marked B_DELWRI).
2829 *
2830 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2831 * and may be left over from a prior truncation (and thus
2832 * no longer represent the actual EOF point), so we
2833 * definitely do not want to B_NOCACHE the backing store.
984263bc 2834 */
4baec531 2835 if (size != bp->b_bcount) {
b1c20cfa 2836 get_mplock();
4baec531 2837 if (bp->b_flags & B_DELWRI) {
cb1cf930 2838 bp->b_flags |= B_RELBUF;
62cfda27 2839 bwrite(bp);
4baec531 2840 } else if (LIST_FIRST(&bp->b_dep)) {
cb1cf930 2841 bp->b_flags |= B_RELBUF;
62cfda27 2842 bwrite(bp);
4baec531
MD
2843 } else {
2844 bp->b_flags |= B_RELBUF;
2845 brelse(bp);
984263bc 2846 }
b1c20cfa 2847 rel_mplock();
4baec531 2848 goto loop;
984263bc 2849 }
4baec531 2850 KKASSERT(size <= bp->b_kvasize);
81b5c339
MD
2851 KASSERT(bp->b_loffset != NOOFFSET,
2852 ("getblk: no buffer offset"));
984263bc
MD
2853
2854 /*
2855 * A buffer with B_DELWRI set and B_CACHE clear must
2856 * be committed before we can return the buffer in
2857 * order to prevent the caller from issuing a read
2858 * ( due to B_CACHE not being set ) and overwriting
2859 * it.
2860 *
2861 * Most callers, including NFS and FFS, need this to
2862 * operate properly either because they assume they
2863 * can issue a read if B_CACHE is not set, or because
2864 * ( for example ) an uncached B_DELWRI might loop due
2865 * to softupdates re-dirtying the buffer. In the latter
2866 * case, B_CACHE is set after the first write completes,
2867 * preventing further loops.
2868 *
2869 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2870 * above while extending the buffer, we cannot allow the
2871 * buffer to remain with B_CACHE set after the write
2872 * completes or it will represent a corrupt state. To
2873 * deal with this we set B_NOCACHE to scrap the buffer
2874 * after the write.
2875 *
cb1cf930
MD
2876 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2877 * I'm not even sure this state is still possible
2878 * now that getblk() writes out any dirty buffers
2879 * on size changes.
2880 *
984263bc
MD
2881 * We might be able to do something fancy, like setting
2882 * B_CACHE in bwrite() except if B_DELWRI is already set,
2883 * so the below call doesn't set B_CACHE, but that gets real
2884 * confusing. This is much easier.
2885 */
2886
2887 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
b1c20cfa 2888 get_mplock();
cb1cf930
MD
2889 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2890 "and CACHE clear, b_flags %08x\n",
2891 bp, (intmax_t)bp->b_loffset, bp->b_flags);
984263bc 2892 bp->b_flags |= B_NOCACHE;
62cfda27 2893 bwrite(bp);
b1c20cfa 2894 rel_mplock();
984263bc
MD
2895 goto loop;
2896 }
984263bc
MD
2897 } else {
2898 /*
2899 * Buffer is not in-core, create new buffer. The buffer
2900 * returned by getnewbuf() is locked. Note that the returned
2901 * buffer is also considered valid (not marked B_INVAL).
21ab32bd
MD
2902 *
2903 * Calculating the offset for the I/O requires figuring out
2904 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2905 * the mount's f_iosize otherwise. If the vnode does not
2906 * have an associated mount we assume that the passed size is
2907 * the block size.
2908 *
2909 * Note that vn_isdisk() cannot be used here since it may
2910 * return a failure for numerous reasons. Note that the
2911 * buffer size may be larger then the block size (the caller
2912 * will use block numbers with the proper multiple). Beware
2913 * of using any v_* fields which are part of unions. In
2914 * particular, in DragonFly the mount point overloading
1d505369
MD
2915 * mechanism uses the namecache only and the underlying
2916 * directory vnode is not a special case.
984263bc 2917 */
7540ab49 2918 int bsize, maxsize;
984263bc 2919
21ab32bd 2920 if (vp->v_type == VBLK || vp->v_type == VCHR)
984263bc 2921 bsize = DEV_BSIZE;
984263bc
MD
2922 else if (vp->v_mount)
2923 bsize = vp->v_mount->mnt_stat.f_iosize;
2924 else
2925 bsize = size;
2926
7540ab49 2927 maxsize = size + (loffset & PAGE_MASK);
984263bc
MD
2928 maxsize = imax(maxsize, bsize);
2929
b1c20cfa
MD
2930 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2931 if (bp == NULL) {
2932 if (slpflags || slptimeo)
984263bc 2933 return NULL;
984263bc
MD
2934 goto loop;
2935 }
2936
2937 /*
b1c20cfa
MD
2938 * Atomically insert the buffer into the hash, so that it can
2939 * be found by findblk().
2940 *
2941 * If bgetvp() returns non-zero a collision occured, and the
2942 * bp will not be associated with the vnode.
1f1ea522
MD
2943 *
2944 * Make sure the translation layer has been cleared.
984263bc 2945 */
54078292
MD
2946 bp->b_loffset = loffset;
2947 bp->b_bio2.bio_offset = NOOFFSET;
1f1ea522 2948 /* bp->b_bio2.bio_next = NULL; */
984263bc 2949
b1c20cfa
MD
2950 if (bgetvp(vp, bp)) {
2951 bp->b_flags |= B_INVAL;
2952 brelse(bp);
2953 goto loop;
2954 }
984263bc
MD
2955
2956 /*
4baec531 2957 * All vnode-based buffers must be backed by a VM object.
984263bc 2958 */
4baec531
MD
2959 KKASSERT(vp->v_object != NULL);
2960 bp->b_flags |= B_VMIO;
10f3fee5 2961 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
984263bc 2962
b1c20cfa 2963 get_mplock();
984263bc 2964 allocbuf(bp, size);
b1c20cfa 2965 rel_mplock();
984263bc 2966 }
8c72e3d5 2967 KKASSERT(dsched_is_clear_buf_priv(bp));
984263bc
MD
2968 return (bp);
2969}
2970
2971/*
5e23ca53
MD
2972 * regetblk(bp)
2973 *
27bc0cb1
MD
2974 * Reacquire a buffer that was previously released to the locked queue,
2975 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2976 * set B_LOCKED (which handles the acquisition race).
5e23ca53 2977 *
27bc0cb1
MD
2978 * To this end, either B_LOCKED must be set or the dependancy list must be
2979 * non-empty.
b1c20cfa
MD
2980 *
2981 * MPSAFE
5e23ca53
MD
2982 */
2983void
2984regetblk(struct buf *bp)
2985{
27bc0cb1 2986 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
5e23ca53 2987 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
5e23ca53 2988 bremfree(bp);
5e23ca53
MD
2989}
2990
2991/*
3f779080
HP
2992 * geteblk:
2993 *
2994 * Get an empty, disassociated buffer of given size. The buffer is
2995 * initially set to B_INVAL.
06ecca5a 2996 *
3f779080
HP
2997 * critical section protection is not required for the allocbuf()
2998 * call because races are impossible here.
b1c20cfa
MD
2999 *
3000 * MPALMOSTSAFE
984263bc
MD
3001 */
3002struct buf *
3003geteblk(int size)
3004{
3005 struct buf *bp;
984263bc
MD
3006 int maxsize;
3007
3008 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3009
e43a034f
MD
3010 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
3011 ;
b1c20cfa 3012 get_mplock();
984263bc 3013 allocbuf(bp, size);
b1c20cfa 3014 rel_mplock();
984263bc 3015 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
8c72e3d5 3016 KKASSERT(dsched_is_clear_buf_priv(bp));
984263bc
MD
3017 return (bp);
3018}
3019
3020
3021/*
3f779080 3022 * allocbuf:
984263bc 3023 *
3f779080
HP
3024 * This code constitutes the buffer memory from either anonymous system
3025 * memory (in the case of non-VMIO operations) or from an associated
3026 * VM object (in the case of VMIO operations). This code is able to
3027 * resize a buffer up or down.
984263bc 3028 *
3f779080
HP
3029 * Note that this code is tricky, and has many complications to resolve
3030 * deadlock or inconsistant data situations. Tread lightly!!!
3031 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3032 * the caller. Calling this code willy nilly can result in the loss of data.
06ecca5a 3033 *
3f779080
HP
3034 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3035 * B_CACHE for the non-VMIO case.
3036 *
3037 * This routine does not need to be called from a critical section but you
3038 * must own the buffer.
b1c20cfa
MD
3039 *
3040 * NOTMPSAFE
984263bc 3041 */
984263bc
MD
3042int
3043allocbuf(struct buf *bp, int size)
3044{
3045 int newbsize, mbsize;
3046 int i;
3047
3048 if (BUF_REFCNT(bp) == 0)
3049 panic("allocbuf: buffer not busy");
3050
3051 if (bp->b_kvasize < size)
3052 panic("allocbuf: buffer too small");
3053
3054 if ((bp->b_flags & B_VMIO) == 0) {
3055 caddr_t origbuf;
3056 int origbufsize;
3057 /*
3058 * Just get anonymous memory from the kernel. Don't
3059 * mess with B_CACHE.
3060 */
3061 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
984263bc
MD
3062 if (bp->b_flags & B_MALLOC)
3063 newbsize = mbsize;
3064 else
984263bc
MD
3065 newbsize = round_page(size);
3066
3067 if (newbsize < bp->b_bufsize) {
984263bc 3068 /*
312dcd01 3069 * Malloced buffers are not shrunk
984263bc
MD
3070 */
3071 if (bp->b_flags & B_MALLOC) {
3072 if (newbsize) {
3073 bp->b_bcount = size;
3074 } else {
efda3bd0 3075 kfree(bp->b_data, M_BIOBUF);
984263bc
MD
3076 if (bp->b_bufsize) {
3077 bufmallocspace -= bp->b_bufsize;
3078 bufspacewakeup();
3079 bp->b_bufsize = 0;
3080 }
3081 bp->b_data = bp->b_kvabase;
3082 bp->b_bcount = 0;
3083 bp->b_flags &= ~B_MALLOC;
3084 }
3085 return 1;
3086 }
984263bc
MD
3087 vm_hold_free_pages(
3088 bp,
3089 (vm_offset_t) bp->b_data + newbsize,
3090 (vm_offset_t) bp->b_data + bp->b_bufsize);
3091 } else if (newbsize > bp->b_bufsize) {
984263bc
MD
3092 /*
3093 * We only use malloced memory on the first allocation.
3094 * and revert to page-allocated memory when the buffer
3095 * grows.
3096 */
4baec531 3097 if ((bufmallocspace < maxbufmallocspace) &&
984263bc
MD
3098 (bp->b_bufsize == 0) &&
3099 (mbsize <= PAGE_SIZE/2)) {
3100
efda3bd0 3101 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
984263bc
MD
3102 bp->b_bufsize = mbsize;
3103 bp->b_bcount = size;
3104 bp->b_flags |= B_MALLOC;
3105 bufmallocspace += mbsize;
3106 return 1;
3107 }
984263bc
MD
3108 origbuf = NULL;
3109 origbufsize = 0;
984263bc 3110 /*
4baec531
MD
3111 * If the buffer is growing on its other-than-first
3112 * allocation, then we revert to the page-allocation
3113 * scheme.
984263bc
MD
3114 */
3115 if (bp->b_flags & B_MALLOC) {
3116 origbuf = bp->b_data;
3117 origbufsize = bp->b_bufsize;
3118 bp->b_data = bp->b_kvabase;
3119 if (bp->b_bufsize) {
3120 bufmallocspace -= bp->b_bufsize;
3121 bufspacewakeup();
3122 bp->b_bufsize = 0;
3123 }
3124 bp->b_flags &= ~B_MALLOC;
3125 newbsize = round_page(newbsize);
3126 }
984263bc
MD
3127 vm_hold_load_pages(
3128 bp,
3129 (vm_offset_t) bp->b_data + bp->b_bufsize,
3130 (vm_offset_t) bp->b_data + newbsize);
984263bc
MD
3131 if (origbuf) {
3132 bcopy(origbuf, bp->b_data, origbufsize);
efda3bd0 3133 kfree(origbuf, M_BIOBUF);
984263bc 3134 }
984263bc
MD
3135 }
3136 } else {
3137 vm_page_t m;
3138 int desiredpages;
3139
3140 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
4baec531
MD
3141 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3142 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3143 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
984263bc 3144
984263bc
MD
3145 if (bp->b_flags & B_MALLOC)
3146 panic("allocbuf: VMIO buffer can't be malloced");
984263bc
MD
3147 /*
3148 * Set B_CACHE initially if buffer is 0 length or will become
3149 * 0-length.
3150 */
3151 if (size == 0 || bp->b_bufsize == 0)
3152 bp->b_flags |= B_CACHE;
3153
3154 if (newbsize < bp->b_bufsize) {
3155 /*
3156 * DEV_BSIZE aligned new buffer size is less then the
3157 * DEV_BSIZE aligned existing buffer size. Figure out
3158 * if we have to remove any pages.
3159 */
54f51aeb
HP
3160 if (desiredpages < bp->b_xio.xio_npages) {
3161 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
984263bc
MD
3162 /*
3163 * the page is not freed here -- it
3164 * is the responsibility of
3165 * vnode_pager_setsize
3166 */
54f51aeb 3167 m = bp->b_xio.xio_pages[i];
984263bc
MD
3168 KASSERT(m != bogus_page,
3169 ("allocbuf: bogus page found"));
3170 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3171 ;
3172
54f51aeb 3173 bp->b_xio.xio_pages[i] = NULL;
984263bc
MD
3174 vm_page_unwire(m, 0);
3175 }
3176 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
54f51aeb
HP
3177 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3178 bp->b_xio.xio_npages = desiredpages;
984263bc
MD
3179 }
3180 } else if (size > bp->b_bcount) {
3181 /*
3182 * We are growing the buffer, possibly in a
3183 * byte-granular fashion.
3184 */
3185 struct vnode *vp;
3186 vm_object_t obj;
3187 vm_offset_t toff;
3188 vm_offset_t tinc;
3189
3190 /*
3191 * Step 1, bring in the VM pages from the object,
3192 * allocating them if necessary. We must clear
3193 * B_CACHE if these pages are not valid for the
3194 * range covered by the buffer.
06ecca5a 3195 *
e43a034f
MD
3196 * critical section protection is required to protect
3197 * against interrupts unbusying and freeing pages
3198 * between our vm_page_lookup() and our
3199 * busycheck/wiring call.
984263bc 3200 */
984263bc 3201 vp = bp->b_vp;
7540ab49 3202 obj = vp->v_object;
984263bc 3203
654a39f0 3204 crit_enter();
54f51aeb 3205 while (bp->b_xio.xio_npages < desiredpages) {
984263bc
MD
3206 vm_page_t m;
3207 vm_pindex_t pi;
3208
81b5c339 3209 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
984263bc
MD
3210 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3211 /*
3212 * note: must allocate system pages
3213 * since blocking here could intefere
3214 * with paging I/O, no matter which
3215 * process we are.
3216 */
4ecf7cc9
MD
3217 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3218 if (m) {
984263bc
MD
3219 vm_page_wire(m);
3220 vm_page_wakeup(m);
6362a262 3221 vm_page_flag_clear(m, PG_ZERO);
984263bc 3222 bp->b_flags &= ~B_CACHE;
54f51aeb
HP
3223 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3224 ++bp->b_xio.xio_npages;
984263bc
MD
3225 }
3226 continue;
3227 }
3228
3229 /*
3230 * We found a page. If we have to sleep on it,
3231 * retry because it might have gotten freed out
3232 * from under us.
3233 *
3234 * We can only test PG_BUSY here. Blocking on
3235 * m->busy might lead to a deadlock:
3236 *
3237 * vm_fault->getpages->cluster_read->allocbuf
3238 *
3239 */
3240
3241 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3242 continue;
984263bc
MD
3243 vm_page_flag_clear(m, PG_ZERO);
3244 vm_page_wire(m);
54f51aeb
HP
3245 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3246 ++bp->b_xio.xio_npages;
0e8bd897
MD
3247 if (bp->b_act_count < m->act_count)
3248 bp->b_act_count = m->act_count;
984263bc 3249 }
654a39f0 3250 crit_exit();
984263bc
MD
3251
3252 /*
3253 * Step 2. We've loaded the pages into the buffer,
3254 * we have to figure out if we can still have B_CACHE
3255 * set. Note that B_CACHE is set according to the
3f779080 3256 * byte-granular range ( bcount and size ), not the
984263bc
MD
3257 * aligned range ( newbsize ).
3258 *
3259 * The VM test is against m->valid, which is DEV_BSIZE
3260 * aligned. Needless to say, the validity of the data
3261 * needs to also be DEV_BSIZE aligned. Note that this
3262 * fails with NFS if the server or some other client
3263 * extends the file's EOF. If our buffer is resized,
3264 * B_CACHE may remain set! XXX
3265 */
3266
3267 toff = bp->b_bcount;
81b5c339 3268 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
984263bc
MD
3269
3270 while ((bp->b_flags & B_CACHE) && toff < size) {
3271 vm_pindex_t pi;
3272
3273 if (tinc > (size - toff))
3274 tinc = size - toff;
3275
81b5c339 3276 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
984263bc
MD
3277 PAGE_SHIFT;
3278
3279 vfs_buf_test_cache(
3280 bp,
81b5c339 3281 bp->b_loffset,
984263bc
MD
3282 toff,
3283 tinc,
54f51aeb 3284 bp->b_xio.xio_pages[pi]
984263bc
MD
3285 );
3286 toff += tinc;
3287 tinc = PAGE_SIZE;
3288 }
3289
3290 /*
3291 * Step 3, fixup the KVM pmap. Remember that
81b5c339
MD
3292 * bp->b_data is relative to bp->b_loffset, but
3293 * bp->b_loffset may be offset into the first page.
984263bc
MD
3294 */
3295
3296 bp->b_data = (caddr_t)
3297 trunc_page((vm_offset_t)bp->b_data);
3298 pmap_qenter(
3299 (vm_offset_t)bp->b_data,
54f51aeb
HP
3300 bp->b_xio.xio_pages,
3301 bp->b_xio.xio_npages
984263bc
MD
3302 );
3303 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
81b5c339 3304 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
984263bc
MD
3305 }
3306 }
70ac7d6c
MD
3307
3308 /* adjust space use on already-dirty buffer */
868d24af
MD
3309 if (bp->b_flags & B_DELWRI) {
3310 dirtybufspace += newbsize - bp->b_bufsize;
3311 if (bp->b_flags & B_HEAVY)
3312 dirtybufspacehw += newbsize - bp->b_bufsize;
3313 }
984263bc
MD
3314 if (newbsize < bp->b_bufsize)
3315 bufspacewakeup();
3316 bp->b_bufsize = newbsize; /* actual buffer allocation */
3317 bp->b_bcount = size; /* requested buffer size */
3318 return 1;
3319}
3320
3321/*
3f779080 3322 * biowait:
984263bc 3323 *
ae8e83e6
MD
3324 * Wait for buffer I/O completion, returning error status. B_EINTR
3325 * is converted into an EINTR error but not cleared (since a chain
3326 * of biowait() calls may occur).
10f3fee5 3327 *
ae8e83e6
MD
3328 * On return bpdone() will have been called but the buffer will remain
3329 * locked and will not have been brelse()'d.
3330 *
3331 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3332 * likely still in progress on return.
3333 *
3334 * NOTE! This operation is on a BIO, not a BUF.
3335 *
3336 * NOTE! BIO_DONE is cleared by vn_strategy()
343b537b
MD
3337 *
3338 * MPSAFE
984263bc 3339 */
ae8e83e6
MD
3340static __inline int
3341_biowait(struct bio *bio, const char *wmesg, int to)
984263bc 3342{
ae8e83e6
MD
3343 struct buf *bp = bio->bio_buf;
3344 u_int32_t flags;
3345 u_int32_t nflags;
3346 int error;
3347
3348 KKASSERT(bio == &bp->b_bio1);
3349 for (;;) {
3350 flags = bio->bio_flags;
3351 if (flags & BIO_DONE)
3352 break;
3353 tsleep_interlock(bio, 0);
3354 nflags = flags | BIO_WANT;
3355 tsleep_interlock(bio, 0);
3356 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3357 if (wmesg)
3358 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3359 else if (bp->b_cmd == BUF_CMD_READ)
3360 error = tsleep(bio, PINTERLOCKED, "biord", to);
343b537b 3361 else
ae8e83e6
MD
3362 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3363 if (error) {
3364 kprintf("tsleep error biowait %d\n", error);
3365 return (error);
3366 }
343b537b 3367 }
984263bc 3368 }
ae8e83e6
MD
3369
3370 /*
3371 * Finish up.
3372 */
3373 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3374 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3375 if (bp->b_flags & B_EINTR)
984263bc 3376 return (EINTR);
ae8e83e6 3377 if (bp->b_flags & B_ERROR)
984263bc 3378 return (bp->b_error ? bp->b_error : EIO);
ae8e83e6
MD
3379 return (0);
3380}
3381
3382int
3383biowait(struct bio *bio, const char *wmesg)
3384{
3385 return(_biowait(bio, wmesg, 0));
3386}
3387
3388int
3389biowait_timeout(struct bio *bio, const char *wmesg, int to)
3390{
3391 return(_biowait(bio, wmesg, to));
984263bc
MD
3392}
3393
3394/*
81b5c339
MD
3395 * This associates a tracking count with an I/O. vn_strategy() and
3396 * dev_dstrategy() do this automatically but there are a few cases
3397 * where a vnode or device layer is bypassed when a block translation
3398 * is cached. In such cases bio_start_transaction() may be called on
3399 * the bypassed layers so the system gets an I/O in progress indication