This is a major revamping of the pageout and low-memory handling code.
[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>
3020e3be 47#include <sys/proc.h>
984263bc
MD
48#include <vm/vm.h>
49#include <vm/vm_param.h>
50#include <vm/vm_kern.h>
51#include <vm/vm_pageout.h>
52#include <vm/vm_page.h>
53#include <vm/vm_object.h>
54#include <vm/vm_extern.h>
55#include <vm/vm_map.h>
654a39f0 56
3020e3be 57#include <sys/buf2.h>
654a39f0 58#include <sys/thread2.h>
f832287e 59#include <sys/spinlock2.h>
12e4aaff 60#include <vm/vm_page2.h>
984263bc 61
135bd6a8
MD
62#include "opt_ddb.h"
63#ifdef DDB
64#include <ddb/ddb.h>
65#endif
66
b3098c79
HP
67/*
68 * Buffer queues.
69 */
b3098c79
HP
70enum bufq_type {
71 BQUEUE_NONE, /* not on any queue */
72 BQUEUE_LOCKED, /* locked buffers */
73 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
74 BQUEUE_DIRTY, /* B_DELWRI buffers */
4b958e7b 75 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
b3098c79 76 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
4b958e7b
MD
77 BQUEUE_EMPTY, /* empty buffer headers */
78
79 BUFFER_QUEUES /* number of buffer queues */
b3098c79 80};
4b958e7b
MD
81
82typedef enum bufq_type bufq_type_t;
83
c4df9635
MD
84#define BD_WAKE_SIZE 128
85#define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
86
b3098c79
HP
87TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
88
984263bc
MD
89static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
90
984263bc 91struct buf *buf; /* buffer header pool */
984263bc 92
984263bc
MD
93static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
94 int pageno, vm_page_t m);
c8e4131d 95static void vfs_clean_pages(struct buf *bp);
984263bc
MD
96static void vfs_setdirty(struct buf *bp);
97static void vfs_vmio_release(struct buf *bp);
4b958e7b 98static int flushbufqueues(bufq_type_t q);
4ecf7cc9 99static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
984263bc 100
868d24af 101static void bd_signal(int totalspace);
4b958e7b
MD
102static void buf_daemon(void);
103static void buf_daemon_hw(void);
c4df9635 104
984263bc
MD
105/*
106 * bogus page -- for I/O to/from partially complete buffers
107 * this is a temporary solution to the problem, but it is not
108 * really that bad. it would be better to split the buffer
109 * for input in the case of buffers partially already in memory,
110 * but the code is intricate enough already.
111 */
112vm_page_t bogus_page;
a0c36a34 113
460426e6
MD
114/*
115 * These are all static, but make the ones we export globals so we do
116 * not need to use compiler magic.
117 */
118int bufspace, maxbufspace,
984263bc
MD
119 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
120static int bufreusecnt, bufdefragcnt, buffreekvacnt;
984263bc 121static int lorunningspace, hirunningspace, runningbufreq;
868d24af 122int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
70ac7d6c 123int dirtybufcount, dirtybufcounthw;
1b30fbcc 124int runningbufspace, runningbufcount;
984263bc
MD
125static int getnewbufcalls;
126static int getnewbufrestarts;
4ecf7cc9 127static int recoverbufcalls;
f832287e
MD
128static int needsbuffer; /* locked by needsbuffer_spin */
129static int bd_request; /* locked by needsbuffer_spin */
4b958e7b 130static int bd_request_hw; /* locked by needsbuffer_spin */
c4df9635
MD
131static u_int bd_wake_ary[BD_WAKE_SIZE];
132static u_int bd_wake_index;
f832287e
MD
133static struct spinlock needsbuffer_spin;
134
4ecf7cc9
MD
135static struct thread *bufdaemon_td;
136static struct thread *bufdaemonhw_td;
137
138
3f779080
HP
139/*
140 * Sysctls for operational control of the buffer cache.
141 */
868d24af 142SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
3f779080 143 "Number of dirty buffers to flush before bufdaemon becomes inactive");
868d24af 144SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
bb606263 145 "High watermark used to trigger explicit flushing of dirty buffers");
3f779080
HP
146SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
147 "Minimum amount of buffer space required for active I/O");
148SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
149 "Maximum amount of buffer space to usable for active I/O");
3f779080
HP
150/*
151 * Sysctls determining current state of the buffer cache.
152 */
17cde63e
MD
153SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
154 "Total number of buffers in buffer cache");
868d24af 155SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
70ac7d6c 156 "Pending bytes of dirty buffers (all)");
868d24af 157SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
70ac7d6c
MD
158 "Pending bytes of dirty buffers (heavy weight)");
159SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
160 "Pending number of dirty buffers");
161SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
4b958e7b 162 "Pending number of dirty buffers (heavy weight)");
3f779080 163SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
bb606263 164 "I/O bytes currently in progress due to asynchronous writes");
1b30fbcc
MD
165SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
166 "I/O buffers currently in progress due to asynchronous writes");
3f779080
HP
167SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
168 "Hard limit on maximum amount of memory usable for buffer space");
169SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
170 "Soft limit on maximum amount of memory usable for buffer space");
171SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
172 "Minimum amount of memory to reserve for system buffer space");
173SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
174 "Amount of memory available for buffers");
175SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
bb606263 176 0, "Maximum amount of memory reserved for buffers using malloc");
3f779080
HP
177SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
178 "Amount of memory left for buffers using malloc-scheme");
179SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
180 "New buffer header acquisition requests");
181SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
182 0, "New buffer header acquisition restarts");
4ecf7cc9
MD
183SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
184 "Recover VM space in an emergency");
3f779080 185SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
bb606263 186 "Buffer acquisition restarts due to fragmented buffer map");
3f779080
HP
187SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
188 "Amount of time KVA space was deallocated in an arbitrary buffer");
189SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
190 "Amount of time buffer re-use operations were successful");
306ab3cb
HP
191SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
192 "sizeof(struct buf)");
984263bc 193
984263bc
MD
194char *buf_wmesg = BUF_WMESG;
195
196extern int vm_swap_size;
197
198#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
c4df9635 199#define VFS_BIO_NEED_UNUSED02 0x02
868d24af 200#define VFS_BIO_NEED_UNUSED04 0x04
984263bc
MD
201#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
202
203/*
3f779080 204 * bufspacewakeup:
984263bc
MD
205 *
206 * Called when buffer space is potentially available for recovery.
207 * getnewbuf() will block on this flag when it is unable to free
208 * sufficient buffer space. Buffer space becomes recoverable when
209 * bp's get placed back in the queues.
210 */
211
212static __inline void
213bufspacewakeup(void)
214{
215 /*
216 * If someone is waiting for BUF space, wake them up. Even
217 * though we haven't freed the kva space yet, the waiting
218 * process will be able to now.
219 */
220 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
f832287e 221 spin_lock_wr(&needsbuffer_spin);
984263bc 222 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
f832287e 223 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
224 wakeup(&needsbuffer);
225 }
226}
227
228/*
3f779080
HP
229 * runningbufwakeup:
230 *
231 * Accounting for I/O in progress.
984263bc
MD
232 *
233 */
234static __inline void
235runningbufwakeup(struct buf *bp)
236{
868d24af
MD
237 int totalspace;
238
239 if ((totalspace = bp->b_runningbufspace) != 0) {
240 runningbufspace -= totalspace;
1b30fbcc 241 --runningbufcount;
984263bc
MD
242 bp->b_runningbufspace = 0;
243 if (runningbufreq && runningbufspace <= lorunningspace) {
244 runningbufreq = 0;
245 wakeup(&runningbufreq);
246 }
868d24af 247 bd_signal(totalspace);
984263bc
MD
248 }
249}
250
251/*
3f779080 252 * bufcountwakeup:
984263bc
MD
253 *
254 * Called when a buffer has been added to one of the free queues to
255 * account for the buffer and to wakeup anyone waiting for free buffers.
256 * This typically occurs when large amounts of metadata are being handled
257 * by the buffer cache ( else buffer space runs out first, usually ).
258 */
259
260static __inline void
261bufcountwakeup(void)
262{
984263bc 263 if (needsbuffer) {
f832287e 264 spin_lock_wr(&needsbuffer_spin);
984263bc 265 needsbuffer &= ~VFS_BIO_NEED_ANY;
f832287e 266 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
267 wakeup(&needsbuffer);
268 }
269}
270
271/*
3f779080 272 * waitrunningbufspace()
984263bc 273 *
cd083340 274 * Wait for the amount of running I/O to drop to a reasonable level.
984263bc 275 *
cd083340
MD
276 * The caller may be using this function to block in a tight loop, we
277 * must block of runningbufspace is greater then the passed limit.
278 * And even with that it may not be enough, due to the presence of
279 * B_LOCKED dirty buffers, so also wait for at least one running buffer
280 * to complete.
984263bc
MD
281 */
282static __inline void
cd083340 283waitrunningbufspace(int limit)
984263bc 284{
cd083340
MD
285 int lorun;
286
287 if (lorunningspace < limit)
288 lorun = lorunningspace;
289 else
290 lorun = limit;
291
292 crit_enter();
293 if (runningbufspace > lorun) {
294 while (runningbufspace > lorun) {
e43a034f
MD
295 ++runningbufreq;
296 tsleep(&runningbufreq, 0, "wdrain", 0);
297 }
cd083340
MD
298 } else if (runningbufspace) {
299 ++runningbufreq;
300 tsleep(&runningbufreq, 0, "wdrain2", 1);
984263bc 301 }
cd083340 302 crit_exit();
984263bc
MD
303}
304
305/*
3f779080 306 * vfs_buf_test_cache:
984263bc
MD
307 *
308 * Called when a buffer is extended. This function clears the B_CACHE
309 * bit if the newly extended portion of the buffer does not contain
310 * valid data.
311 */
312static __inline__
313void
314vfs_buf_test_cache(struct buf *bp,
315 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
316 vm_page_t m)
317{
318 if (bp->b_flags & B_CACHE) {
319 int base = (foff + off) & PAGE_MASK;
320 if (vm_page_is_valid(m, base, size) == 0)
321 bp->b_flags &= ~B_CACHE;
322 }
323}
324
3f779080 325/*
cd083340 326 * bd_speedup()
4b958e7b 327 *
cd083340
MD
328 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
329 * low water mark.
3f779080 330 */
984263bc
MD
331static __inline__
332void
c4df9635 333bd_speedup(void)
984263bc 334{
70ac7d6c 335 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
cd083340
MD
336 return;
337
338 if (bd_request == 0 &&
70ac7d6c
MD
339 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
340 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
f832287e 341 spin_lock_wr(&needsbuffer_spin);
984263bc 342 bd_request = 1;
f832287e 343 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
344 wakeup(&bd_request);
345 }
cd083340 346 if (bd_request_hw == 0 &&
70ac7d6c
MD
347 (dirtybufspacehw > lodirtybufspace / 2 ||
348 dirtybufcounthw >= nbuf / 2)) {
4b958e7b
MD
349 spin_lock_wr(&needsbuffer_spin);
350 bd_request_hw = 1;
351 spin_unlock_wr(&needsbuffer_spin);
352 wakeup(&bd_request_hw);
353 }
984263bc
MD
354}
355
356/*
c4df9635 357 * bd_heatup()
3f779080 358 *
c4df9635
MD
359 * Get the buf_daemon heated up when the number of running and dirty
360 * buffers exceeds the mid-point.
984263bc 361 */
c4df9635
MD
362int
363bd_heatup(void)
364{
365 int mid1;
366 int mid2;
868d24af 367 int totalspace;
984263bc 368
868d24af 369 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
c4df9635 370
868d24af 371 totalspace = runningbufspace + dirtybufspace;
70ac7d6c 372 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
c4df9635 373 bd_speedup();
868d24af
MD
374 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
375 if (totalspace >= mid2)
376 return(totalspace - mid2);
c4df9635
MD
377 }
378 return(0);
379}
380
381/*
382 * bd_wait()
383 *
868d24af
MD
384 * Wait for the buffer cache to flush (totalspace) bytes worth of
385 * buffers, then return.
c4df9635
MD
386 *
387 * Regardless this function blocks while the number of dirty buffers
868d24af 388 * exceeds hidirtybufspace.
c4df9635 389 */
984263bc 390void
868d24af 391bd_wait(int totalspace)
984263bc 392{
c4df9635 393 u_int i;
868d24af 394 int count;
c4df9635 395
4ecf7cc9
MD
396 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
397 return;
398
868d24af 399 while (totalspace > 0) {
c4df9635
MD
400 bd_heatup();
401 crit_enter();
868d24af
MD
402 if (totalspace > runningbufspace + dirtybufspace)
403 totalspace = runningbufspace + dirtybufspace;
404 count = totalspace / BKVASIZE;
c4df9635
MD
405 if (count >= BD_WAKE_SIZE)
406 count = BD_WAKE_SIZE - 1;
407 i = (bd_wake_index + count) & BD_WAKE_MASK;
408 ++bd_wake_ary[i];
409 tsleep(&bd_wake_ary[i], 0, "flstik", hz);
410 crit_exit();
411
868d24af 412 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
c4df9635
MD
413 }
414}
415
416/*
417 * bd_signal()
418 *
868d24af
MD
419 * This function is called whenever runningbufspace or dirtybufspace
420 * is reduced. Track threads waiting for run+dirty buffer I/O
c4df9635
MD
421 * complete.
422 */
423static void
868d24af 424bd_signal(int totalspace)
c4df9635
MD
425{
426 u_int i;
427
868d24af
MD
428 while (totalspace > 0) {
429 i = atomic_fetchadd_int(&bd_wake_index, 1);
430 i &= BD_WAKE_MASK;
431 if (bd_wake_ary[i]) {
432 bd_wake_ary[i] = 0;
433 wakeup(&bd_wake_ary[i]);
434 }
435 totalspace -= BKVASIZE;
c4df9635 436 }
984263bc
MD
437}
438
439/*
3f779080
HP
440 * bufinit:
441 *
442 * Load time initialisation of the buffer cache, called from machine
443 * dependant initialization code.
444 */
984263bc
MD
445void
446bufinit(void)
447{
448 struct buf *bp;
b8bb0773 449 vm_offset_t bogus_offset;
984263bc
MD
450 int i;
451
f832287e
MD
452 spin_init(&needsbuffer_spin);
453
984263bc
MD
454 /* next, make a null set of free lists */
455 for (i = 0; i < BUFFER_QUEUES; i++)
456 TAILQ_INIT(&bufqueues[i]);
457
458 /* finally, initialize each buffer header and stick on empty q */
459 for (i = 0; i < nbuf; i++) {
460 bp = &buf[i];
461 bzero(bp, sizeof *bp);
462 bp->b_flags = B_INVAL; /* we're just an empty header */
10f3fee5 463 bp->b_cmd = BUF_CMD_DONE;
b3098c79 464 bp->b_qindex = BQUEUE_EMPTY;
81b5c339 465 initbufbio(bp);
54f51aeb 466 xio_init(&bp->b_xio);
408357d8 467 buf_dep_init(bp);
984263bc 468 BUF_LOCKINIT(bp);
b3098c79 469 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
984263bc
MD
470 }
471
472 /*
473 * maxbufspace is the absolute maximum amount of buffer space we are
474 * allowed to reserve in KVM and in real terms. The absolute maximum
475 * is nominally used by buf_daemon. hibufspace is the nominal maximum
476 * used by most other processes. The differential is required to
477 * ensure that buf_daemon is able to run when other processes might
478 * be blocked waiting for buffer space.
479 *
480 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
481 * this may result in KVM fragmentation which is not handled optimally
482 * by the system.
483 */
484 maxbufspace = nbuf * BKVASIZE;
485 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
486 lobufspace = hibufspace - MAXBSIZE;
487
488 lorunningspace = 512 * 1024;
489 hirunningspace = 1024 * 1024;
490
868d24af
MD
491 /*
492 * Limit the amount of malloc memory since it is wired permanently
493 * into the kernel space. Even though this is accounted for in
494 * the buffer allocation, we don't want the malloced region to grow
495 * uncontrolled. The malloc scheme improves memory utilization
496 * significantly on average (small) directories.
497 */
984263bc
MD
498 maxbufmallocspace = hibufspace / 20;
499
868d24af
MD
500 /*
501 * Reduce the chance of a deadlock occuring by limiting the number
502 * of delayed-write dirty buffers we allow to stack up.
503 */
504 hidirtybufspace = hibufspace / 2;
505 dirtybufspace = 0;
506 dirtybufspacehw = 0;
984263bc 507
868d24af 508 lodirtybufspace = hidirtybufspace / 2;
984263bc 509
868d24af
MD
510 /*
511 * Maximum number of async ops initiated per buf_daemon loop. This is
512 * somewhat of a hack at the moment, we really need to limit ourselves
513 * based on the number of bytes of I/O in-transit that were initiated
514 * from buf_daemon.
515 */
984263bc 516
e4846942 517 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
c439ad8f 518 bogus_page = vm_page_alloc(&kernel_object,
e4846942
MD
519 (bogus_offset >> PAGE_SHIFT),
520 VM_ALLOC_NORMAL);
12e4aaff 521 vmstats.v_wire_count++;
984263bc
MD
522
523}
524
525/*
81b5c339
MD
526 * Initialize the embedded bio structures
527 */
528void
529initbufbio(struct buf *bp)
530{
531 bp->b_bio1.bio_buf = bp;
532 bp->b_bio1.bio_prev = NULL;
81b5c339
MD
533 bp->b_bio1.bio_offset = NOOFFSET;
534 bp->b_bio1.bio_next = &bp->b_bio2;
535 bp->b_bio1.bio_done = NULL;
536
537 bp->b_bio2.bio_buf = bp;
538 bp->b_bio2.bio_prev = &bp->b_bio1;
81b5c339
MD
539 bp->b_bio2.bio_offset = NOOFFSET;
540 bp->b_bio2.bio_next = NULL;
541 bp->b_bio2.bio_done = NULL;
542}
543
544/*
545 * Reinitialize the embedded bio structures as well as any additional
546 * translation cache layers.
547 */
548void
549reinitbufbio(struct buf *bp)
550{
551 struct bio *bio;
552
553 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
554 bio->bio_done = NULL;
81b5c339
MD
555 bio->bio_offset = NOOFFSET;
556 }
557}
558
559/*
560 * Push another BIO layer onto an existing BIO and return it. The new
561 * BIO layer may already exist, holding cached translation data.
562 */
563struct bio *
564push_bio(struct bio *bio)
565{
566 struct bio *nbio;
567
568 if ((nbio = bio->bio_next) == NULL) {
569 int index = bio - &bio->bio_buf->b_bio_array[0];
bbd44c71 570 if (index >= NBUF_BIO - 1) {
81b5c339
MD
571 panic("push_bio: too many layers bp %p\n",
572 bio->bio_buf);
573 }
574 nbio = &bio->bio_buf->b_bio_array[index + 1];
575 bio->bio_next = nbio;
576 nbio->bio_prev = bio;
577 nbio->bio_buf = bio->bio_buf;
81b5c339
MD
578 nbio->bio_offset = NOOFFSET;
579 nbio->bio_done = NULL;
580 nbio->bio_next = NULL;
581 }
582 KKASSERT(nbio->bio_done == NULL);
583 return(nbio);
584}
585
b77cfc40
MD
586/*
587 * Pop a BIO translation layer, returning the previous layer. The
588 * must have been previously pushed.
589 */
590struct bio *
81b5c339
MD
591pop_bio(struct bio *bio)
592{
b77cfc40 593 return(bio->bio_prev);
81b5c339
MD
594}
595
596void
597clearbiocache(struct bio *bio)
598{
599 while (bio) {
81b5c339
MD
600 bio->bio_offset = NOOFFSET;
601 bio = bio->bio_next;
602 }
603}
604
605/*
3f779080
HP
606 * bfreekva:
607 *
608 * Free the KVA allocation for buffer 'bp'.
984263bc 609 *
e43a034f 610 * Must be called from a critical section as this is the only locking for
984263bc
MD
611 * buffer_map.
612 *
613 * Since this call frees up buffer space, we call bufspacewakeup().
614 */
615static void
312dcd01 616bfreekva(struct buf *bp)
984263bc 617{
a108bf71
MD
618 int count;
619
984263bc
MD
620 if (bp->b_kvasize) {
621 ++buffreekvacnt;
a108bf71 622 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
e4846942 623 vm_map_lock(&buffer_map);
984263bc 624 bufspace -= bp->b_kvasize;
e4846942 625 vm_map_delete(&buffer_map,
984263bc 626 (vm_offset_t) bp->b_kvabase,
a108bf71
MD
627 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
628 &count
984263bc 629 );
e4846942 630 vm_map_unlock(&buffer_map);
a108bf71 631 vm_map_entry_release(count);
984263bc
MD
632 bp->b_kvasize = 0;
633 bufspacewakeup();
634 }
635}
636
637/*
3f779080 638 * bremfree:
984263bc
MD
639 *
640 * Remove the buffer from the appropriate free list.
641 */
642void
c8e4131d 643bremfree(struct buf *bp)
984263bc 644{
e43a034f 645 crit_enter();
984263bc 646
b3098c79 647 if (bp->b_qindex != BQUEUE_NONE) {
77bb9400
MD
648 KASSERT(BUF_REFCNTNB(bp) == 1,
649 ("bremfree: bp %p not locked",bp));
984263bc 650 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
b3098c79 651 bp->b_qindex = BQUEUE_NONE;
984263bc 652 } else {
77bb9400 653 if (BUF_REFCNTNB(bp) <= 1)
984263bc
MD
654 panic("bremfree: removing a buffer not on a queue");
655 }
656
e43a034f 657 crit_exit();
984263bc
MD
658}
659
660
661/*
3f779080
HP
662 * bread:
663 *
664 * Get a buffer with the specified data. Look in the cache first. We
665 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
666 * is set, the buffer is valid and we do not have to do anything ( see
667 * getblk() ).
984263bc
MD
668 */
669int
c8e4131d 670bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
984263bc
MD
671{
672 struct buf *bp;
673
54078292 674 bp = getblk(vp, loffset, size, 0, 0);
984263bc
MD
675 *bpp = bp;
676
677 /* if not found in cache, do some I/O */
678 if ((bp->b_flags & B_CACHE) == 0) {
78a9b77f
MD
679 KASSERT(!(bp->b_flags & B_ASYNC),
680 ("bread: illegal async bp %p", bp));
984263bc 681 bp->b_flags &= ~(B_ERROR | B_INVAL);
10f3fee5
MD
682 bp->b_cmd = BUF_CMD_READ;
683 vfs_busy_pages(vp, bp);
81b5c339 684 vn_strategy(vp, &bp->b_bio1);
984263bc
MD
685 return (biowait(bp));
686 }
687 return (0);
688}
689
690/*
3f779080
HP
691 * breadn:
692 *
693 * Operates like bread, but also starts asynchronous I/O on
694 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
695 * to initiating I/O . If B_CACHE is set, the buffer is valid
696 * and we do not have to do anything.
984263bc
MD
697 */
698int
a8f169e2 699breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
c8e4131d 700 int *rabsize, int cnt, struct buf **bpp)
984263bc
MD
701{
702 struct buf *bp, *rabp;
703 int i;
704 int rv = 0, readwait = 0;
705
54078292 706 *bpp = bp = getblk(vp, loffset, size, 0, 0);
984263bc
MD
707
708 /* if not found in cache, do some I/O */
709 if ((bp->b_flags & B_CACHE) == 0) {
984263bc 710 bp->b_flags &= ~(B_ERROR | B_INVAL);
10f3fee5
MD
711 bp->b_cmd = BUF_CMD_READ;
712 vfs_busy_pages(vp, bp);
81b5c339 713 vn_strategy(vp, &bp->b_bio1);
984263bc
MD
714 ++readwait;
715 }
716
54078292
MD
717 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
718 if (inmem(vp, *raoffset))
984263bc 719 continue;
54078292 720 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
984263bc
MD
721
722 if ((rabp->b_flags & B_CACHE) == 0) {
10f3fee5 723 rabp->b_flags |= B_ASYNC;
984263bc 724 rabp->b_flags &= ~(B_ERROR | B_INVAL);
10f3fee5
MD
725 rabp->b_cmd = BUF_CMD_READ;
726 vfs_busy_pages(vp, rabp);
984263bc 727 BUF_KERNPROC(rabp);
81b5c339 728 vn_strategy(vp, &rabp->b_bio1);
984263bc
MD
729 } else {
730 brelse(rabp);
731 }
732 }
733
734 if (readwait) {
735 rv = biowait(bp);
736 }
737 return (rv);
738}
739
740/*
3f779080
HP
741 * bwrite:
742 *
743 * Write, release buffer on completion. (Done by iodone
744 * if async). Do not bother writing anything if the buffer
745 * is invalid.
746 *
747 * Note that we set B_CACHE here, indicating that buffer is
748 * fully valid and thus cacheable. This is true even of NFS
749 * now so we set it generally. This could be set either here
750 * or in biodone() since the I/O is synchronous. We put it
751 * here.
984263bc
MD
752 */
753int
c8e4131d 754bwrite(struct buf *bp)
984263bc 755{
e43a034f 756 int oldflags;
984263bc
MD
757
758 if (bp->b_flags & B_INVAL) {
759 brelse(bp);
760 return (0);
761 }
762
763 oldflags = bp->b_flags;
764
77bb9400 765 if (BUF_REFCNTNB(bp) == 0)
984263bc 766 panic("bwrite: buffer is not busy???");
e43a034f 767 crit_enter();
984263bc
MD
768
769 /* Mark the buffer clean */
770 bundirty(bp);
771
10f3fee5 772 bp->b_flags &= ~B_ERROR;
6bae6177 773 bp->b_flags |= B_CACHE;
10f3fee5
MD
774 bp->b_cmd = BUF_CMD_WRITE;
775 vfs_busy_pages(bp->b_vp, bp);
984263bc
MD
776
777 /*
9a71d53f
MD
778 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
779 * valid for vnode-backed buffers.
984263bc
MD
780 */
781 bp->b_runningbufspace = bp->b_bufsize;
1b30fbcc
MD
782 if (bp->b_runningbufspace) {
783 runningbufspace += bp->b_runningbufspace;
784 ++runningbufcount;
785 }
984263bc 786
e43a034f 787 crit_exit();
984263bc
MD
788 if (oldflags & B_ASYNC)
789 BUF_KERNPROC(bp);
81b5c339 790 vn_strategy(bp->b_vp, &bp->b_bio1);
984263bc
MD
791
792 if ((oldflags & B_ASYNC) == 0) {
793 int rtval = biowait(bp);
794 brelse(bp);
795 return (rtval);
984263bc 796 }
984263bc
MD
797 return (0);
798}
799
984263bc 800/*
3f779080
HP
801 * bdwrite:
802 *
803 * Delayed write. (Buffer is marked dirty). Do not bother writing
804 * anything if the buffer is marked invalid.
984263bc 805 *
3f779080
HP
806 * Note that since the buffer must be completely valid, we can safely
807 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
808 * biodone() in order to prevent getblk from writing the buffer
809 * out synchronously.
984263bc
MD
810 */
811void
493c516a 812bdwrite(struct buf *bp)
984263bc 813{
77bb9400 814 if (BUF_REFCNTNB(bp) == 0)
984263bc
MD
815 panic("bdwrite: buffer is not busy");
816
817 if (bp->b_flags & B_INVAL) {
818 brelse(bp);
819 return;
820 }
821 bdirty(bp);
822
823 /*
824 * Set B_CACHE, indicating that the buffer is fully valid. This is
825 * true even of NFS now.
826 */
827 bp->b_flags |= B_CACHE;
828
829 /*
830 * This bmap keeps the system from needing to do the bmap later,
831 * perhaps when the system is attempting to do a sync. Since it
832 * is likely that the indirect block -- or whatever other datastructure
833 * that the filesystem needs is still in memory now, it is a good
834 * thing to do this. Note also, that if the pageout daemon is
835 * requesting a sync -- there might not be enough memory to do
836 * the bmap then... So, this is important to do.
837 */
54078292 838 if (bp->b_bio2.bio_offset == NOOFFSET) {
08daea96 839 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
e92ca23a 840 NULL, NULL, BUF_CMD_WRITE);
984263bc
MD
841 }
842
843 /*
844 * Set the *dirty* buffer range based upon the VM system dirty pages.
845 */
846 vfs_setdirty(bp);
847
848 /*
849 * We need to do this here to satisfy the vnode_pager and the
850 * pageout daemon, so that it thinks that the pages have been
851 * "cleaned". Note that since the pages are in a delayed write
852 * buffer -- the VFS layer "will" see that the pages get written
853 * out on the next sync, or perhaps the cluster will be completed.
854 */
855 vfs_clean_pages(bp);
856 bqrelse(bp);
857
858 /*
984263bc
MD
859 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
860 * due to the softdep code.
861 */
862}
863
864/*
3f779080 865 * bdirty:
984263bc 866 *
10f3fee5
MD
867 * Turn buffer into delayed write request by marking it B_DELWRI.
868 * B_RELBUF and B_NOCACHE must be cleared.
984263bc 869 *
10f3fee5
MD
870 * We reassign the buffer to itself to properly update it in the
871 * dirty/clean lists.
984263bc 872 *
e43a034f 873 * Must be called from a critical section.
b3098c79 874 * The buffer must be on BQUEUE_NONE.
984263bc
MD
875 */
876void
493c516a 877bdirty(struct buf *bp)
984263bc 878{
b3098c79 879 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
69f8c926 880 if (bp->b_flags & B_NOCACHE) {
6ea70f76 881 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
69f8c926
MD
882 bp->b_flags &= ~B_NOCACHE;
883 }
884 if (bp->b_flags & B_INVAL) {
6ea70f76 885 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
69f8c926 886 }
10f3fee5 887 bp->b_flags &= ~B_RELBUF;
984263bc
MD
888
889 if ((bp->b_flags & B_DELWRI) == 0) {
10f3fee5 890 bp->b_flags |= B_DELWRI;
1f1ea522 891 reassignbuf(bp);
70ac7d6c 892 ++dirtybufcount;
868d24af 893 dirtybufspace += bp->b_bufsize;
70ac7d6c
MD
894 if (bp->b_flags & B_HEAVY) {
895 ++dirtybufcounthw;
868d24af 896 dirtybufspacehw += bp->b_bufsize;
70ac7d6c 897 }
c4df9635 898 bd_heatup();
984263bc
MD
899 }
900}
901
902/*
4b958e7b
MD
903 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
904 * needs to be flushed with a different buf_daemon thread to avoid
905 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
906 */
907void
908bheavy(struct buf *bp)
909{
910 if ((bp->b_flags & B_HEAVY) == 0) {
911 bp->b_flags |= B_HEAVY;
70ac7d6c
MD
912 if (bp->b_flags & B_DELWRI) {
913 ++dirtybufcounthw;
868d24af 914 dirtybufspacehw += bp->b_bufsize;
70ac7d6c 915 }
4b958e7b
MD
916 }
917}
918
919/*
3f779080 920 * bundirty:
984263bc
MD
921 *
922 * Clear B_DELWRI for buffer.
923 *
e43a034f 924 * Must be called from a critical section.
eaaadca0 925 *
b3098c79 926 * The buffer is typically on BQUEUE_NONE but there is one case in
eaaadca0
MD
927 * brelse() that calls this function after placing the buffer on
928 * a different queue.
984263bc
MD
929 */
930
931void
493c516a 932bundirty(struct buf *bp)
984263bc 933{
984263bc
MD
934 if (bp->b_flags & B_DELWRI) {
935 bp->b_flags &= ~B_DELWRI;
1f1ea522 936 reassignbuf(bp);
70ac7d6c 937 --dirtybufcount;
868d24af 938 dirtybufspace -= bp->b_bufsize;
70ac7d6c
MD
939 if (bp->b_flags & B_HEAVY) {
940 --dirtybufcounthw;
868d24af 941 dirtybufspacehw -= bp->b_bufsize;
70ac7d6c 942 }
868d24af 943 bd_signal(bp->b_bufsize);
984263bc
MD
944 }
945 /*
946 * Since it is now being written, we can clear its deferred write flag.
947 */
948 bp->b_flags &= ~B_DEFERRED;
949}
950
951/*
3f779080 952 * bawrite:
984263bc
MD
953 *
954 * Asynchronous write. Start output on a buffer, but do not wait for
955 * it to complete. The buffer is released when the output completes.
956 *
957 * bwrite() ( or the VOP routine anyway ) is responsible for handling
958 * B_INVAL buffers. Not us.
959 */
960void
c8e4131d 961bawrite(struct buf *bp)
984263bc
MD
962{
963 bp->b_flags |= B_ASYNC;
62cfda27 964 bwrite(bp);
984263bc
MD
965}
966
967/*
3f779080 968 * bowrite:
984263bc
MD
969 *
970 * Ordered write. Start output on a buffer, and flag it so that the
971 * device will write it in the order it was queued. The buffer is
972 * released when the output completes. bwrite() ( or the VOP routine
973 * anyway ) is responsible for handling B_INVAL buffers.
974 */
975int
c8e4131d 976bowrite(struct buf *bp)
984263bc
MD
977{
978 bp->b_flags |= B_ORDERED | B_ASYNC;
62cfda27 979 return (bwrite(bp));
984263bc
MD
980}
981
982/*
3f779080
HP
983 * buf_dirty_count_severe:
984 *
985 * Return true if we have too many dirty buffers.
984263bc
MD
986 */
987int
988buf_dirty_count_severe(void)
989{
70ac7d6c
MD
990 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
991 dirtybufcount >= nbuf / 2);
984263bc
MD
992}
993
994/*
3f779080 995 * brelse:
984263bc
MD
996 *
997 * Release a busy buffer and, if requested, free its resources. The
998 * buffer will be stashed in the appropriate bufqueue[] allowing it
999 * to be accessed later as a cache entity or reused for other purposes.
1000 */
1001void
c8e4131d 1002brelse(struct buf *bp)
984263bc 1003{
9188c711
MD
1004#ifdef INVARIANTS
1005 int saved_flags = bp->b_flags;
1006#endif
1007
984263bc
MD
1008 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1009
e43a034f 1010 crit_enter();
984263bc 1011
135bd6a8
MD
1012 /*
1013 * If B_NOCACHE is set we are being asked to destroy the buffer and
1014 * its backing store. Clear B_DELWRI.
1015 *
1016 * B_NOCACHE is set in two cases: (1) when the caller really wants
1017 * to destroy the buffer and backing store and (2) when the caller
1018 * wants to destroy the buffer and backing store after a write
1019 * completes.
1020 */
1021 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1022 bundirty(bp);
69f8c926
MD
1023 }
1024
78a9b77f 1025 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
984263bc 1026 /*
78a9b77f
MD
1027 * A re-dirtied buffer is only subject to destruction
1028 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
984263bc 1029 */
78a9b77f 1030 /* leave buffer intact */
10f3fee5 1031 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
78a9b77f 1032 (bp->b_bufsize <= 0)) {
984263bc 1033 /*
78a9b77f
MD
1034 * Either a failed read or we were asked to free or not
1035 * cache the buffer. This path is reached with B_DELWRI
1036 * set only if B_INVAL is already set. B_NOCACHE governs
1037 * backing store destruction.
408357d8
MD
1038 *
1039 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1040 * buffer cannot be immediately freed.
984263bc
MD
1041 */
1042 bp->b_flags |= B_INVAL;
408357d8
MD
1043 if (LIST_FIRST(&bp->b_dep) != NULL)
1044 buf_deallocate(bp);
984263bc 1045 if (bp->b_flags & B_DELWRI) {
70ac7d6c 1046 --dirtybufcount;
868d24af 1047 dirtybufspace -= bp->b_bufsize;
70ac7d6c
MD
1048 if (bp->b_flags & B_HEAVY) {
1049 --dirtybufcounthw;
868d24af 1050 dirtybufspacehw -= bp->b_bufsize;
70ac7d6c 1051 }
868d24af 1052 bd_signal(bp->b_bufsize);
984263bc 1053 }
10f3fee5 1054 bp->b_flags &= ~(B_DELWRI | B_CACHE);
984263bc
MD
1055 }
1056
1057 /*
408357d8
MD
1058 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1059 * If vfs_vmio_release() is called with either bit set, the
1060 * underlying pages may wind up getting freed causing a previous
1061 * write (bdwrite()) to get 'lost' because pages associated with
1062 * a B_DELWRI bp are marked clean. Pages associated with a
1063 * B_LOCKED buffer may be mapped by the filesystem.
4b958e7b
MD
1064 *
1065 * If we want to release the buffer ourselves (rather then the
1066 * originator asking us to release it), give the originator a
1067 * chance to countermand the release by setting B_LOCKED.
984263bc
MD
1068 *
1069 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1070 * if B_DELWRI is set.
1071 *
1072 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1073 * on pages to return pages to the VM page queues.
1074 */
4b958e7b 1075 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
984263bc 1076 bp->b_flags &= ~B_RELBUF;
4b958e7b 1077 } else if (vm_page_count_severe()) {
389ee6be 1078 if (LIST_FIRST(&bp->b_dep) != NULL)
78a9b77f 1079 buf_deallocate(bp); /* can set B_LOCKED */
4b958e7b
MD
1080 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1081 bp->b_flags &= ~B_RELBUF;
1082 else
1083 bp->b_flags |= B_RELBUF;
1084 }
984263bc
MD
1085
1086 /*
78a9b77f
MD
1087 * Make sure b_cmd is clear. It may have already been cleared by
1088 * biodone().
1089 *
9188c711
MD
1090 * At this point destroying the buffer is governed by the B_INVAL
1091 * or B_RELBUF flags.
1092 */
10f3fee5 1093 bp->b_cmd = BUF_CMD_DONE;
9188c711
MD
1094
1095 /*
135bd6a8
MD
1096 * VMIO buffer rundown. Make sure the VM page array is restored
1097 * after an I/O may have replaces some of the pages with bogus pages
1098 * in order to not destroy dirty pages in a fill-in read.
1099 *
1100 * Note that due to the code above, if a buffer is marked B_DELWRI
1101 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1102 * B_INVAL may still be set, however.
984263bc 1103 *
135bd6a8
MD
1104 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1105 * but not the backing store. B_NOCACHE will destroy the backing
1106 * store.
984263bc 1107 *
135bd6a8
MD
1108 * Note that dirty NFS buffers contain byte-granular write ranges
1109 * and should not be destroyed w/ B_INVAL even if the backing store
1110 * is left intact.
984263bc 1111 */
135bd6a8 1112 if (bp->b_flags & B_VMIO) {
9188c711
MD
1113 /*
1114 * Rundown for VMIO buffers which are not dirty NFS buffers.
1115 */
984263bc
MD
1116 int i, j, resid;
1117 vm_page_t m;
1118 off_t foff;
1119 vm_pindex_t poff;
1120 vm_object_t obj;
1121 struct vnode *vp;
1122
1123 vp = bp->b_vp;
1124
1125 /*
1126 * Get the base offset and length of the buffer. Note that
1127 * in the VMIO case if the buffer block size is not
1128 * page-aligned then b_data pointer may not be page-aligned.
236b2b9f 1129 * But our b_xio.xio_pages array *IS* page aligned.
984263bc
MD
1130 *
1131 * block sizes less then DEV_BSIZE (usually 512) are not
1132 * supported due to the page granularity bits (m->valid,
1133 * m->dirty, etc...).
1134 *
1135 * See man buf(9) for more information
1136 */
1137
1138 resid = bp->b_bufsize;
81b5c339 1139 foff = bp->b_loffset;
984263bc 1140
54f51aeb
HP
1141 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1142 m = bp->b_xio.xio_pages[i];
984263bc
MD
1143 vm_page_flag_clear(m, PG_ZERO);
1144 /*
1145 * If we hit a bogus page, fixup *all* of them
06ecca5a
MD
1146 * now. Note that we left these pages wired
1147 * when we removed them so they had better exist,
1148 * and they cannot be ripped out from under us so
e43a034f 1149 * no critical section protection is necessary.
984263bc
MD
1150 */
1151 if (m == bogus_page) {
7540ab49 1152 obj = vp->v_object;
81b5c339 1153 poff = OFF_TO_IDX(bp->b_loffset);
984263bc 1154
54f51aeb 1155 for (j = i; j < bp->b_xio.xio_npages; j++) {
984263bc
MD
1156 vm_page_t mtmp;
1157
54f51aeb 1158 mtmp = bp->b_xio.xio_pages[j];
984263bc
MD
1159 if (mtmp == bogus_page) {
1160 mtmp = vm_page_lookup(obj, poff + j);
1161 if (!mtmp) {
fc92d4aa 1162 panic("brelse: page missing");
984263bc 1163 }
54f51aeb 1164 bp->b_xio.xio_pages[j] = mtmp;
984263bc
MD
1165 }
1166 }
1167
1168 if ((bp->b_flags & B_INVAL) == 0) {
54f51aeb
HP
1169 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1170 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
984263bc 1171 }
54f51aeb 1172 m = bp->b_xio.xio_pages[i];
984263bc 1173 }
8d429613
MD
1174
1175 /*
1176 * Invalidate the backing store if B_NOCACHE is set
1177 * (e.g. used with vinvalbuf()). If this is NFS
1178 * we impose a requirement that the block size be
1179 * a multiple of PAGE_SIZE and create a temporary
1180 * hack to basically invalidate the whole page. The
1181 * problem is that NFS uses really odd buffer sizes
1182 * especially when tracking piecemeal writes and
1183 * it also vinvalbuf()'s a lot, which would result
1184 * in only partial page validation and invalidation
1185 * here. If the file page is mmap()'d, however,
1186 * all the valid bits get set so after we invalidate
1187 * here we would end up with weird m->valid values
1188 * like 0xfc. nfs_getpages() can't handle this so
1189 * we clear all the valid bits for the NFS case
1190 * instead of just some of them.
1191 *
1192 * The real bug is the VM system having to set m->valid
1193 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1194 * itself is an artifact of the whole 512-byte
1195 * granular mess that exists to support odd block
1196 * sizes and UFS meta-data block sizes (e.g. 6144).
1197 * A complete rewrite is required.
1198 */
984263bc
MD
1199 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1200 int poffset = foff & PAGE_MASK;
8d429613
MD
1201 int presid;
1202
1203 presid = PAGE_SIZE - poffset;
1204 if (bp->b_vp->v_tag == VT_NFS &&
1205 bp->b_vp->v_type == VREG) {
1206 ; /* entire page */
1207 } else if (presid > resid) {
1208 presid = resid;
1209 }
984263bc
MD
1210 KASSERT(presid >= 0, ("brelse: extra page"));
1211 vm_page_set_invalid(m, poffset, presid);
1212 }
1213 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1214 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1215 }
984263bc
MD
1216 if (bp->b_flags & (B_INVAL | B_RELBUF))
1217 vfs_vmio_release(bp);
9188c711
MD
1218 } else {
1219 /*
1220 * Rundown for non-VMIO buffers.
1221 */
1222 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1223#if 0
1224 if (bp->b_vp)
6ea70f76 1225 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
9188c711
MD
1226#endif
1227 if (bp->b_bufsize)
1228 allocbuf(bp, 0);
f9a11477 1229 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
9188c711
MD
1230 if (bp->b_vp)
1231 brelvp(bp);
1232 }
984263bc
MD
1233 }
1234
b3098c79 1235 if (bp->b_qindex != BQUEUE_NONE)
984263bc 1236 panic("brelse: free buffer onto another queue???");
77bb9400 1237 if (BUF_REFCNTNB(bp) > 1) {
984263bc
MD
1238 /* Temporary panic to verify exclusive locking */
1239 /* This panic goes away when we allow shared refs */
1240 panic("brelse: multiple refs");
1241 /* do not release to free list */
1242 BUF_UNLOCK(bp);
e43a034f 1243 crit_exit();
984263bc
MD
1244 return;
1245 }
1246
9188c711
MD
1247 /*
1248 * Figure out the correct queue to place the cleaned up buffer on.
1249 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1250 * disassociated from their vnode.
1251 */
408357d8
MD
1252 if (bp->b_flags & B_LOCKED) {
1253 /*
27bc0cb1
MD
1254 * Buffers that are locked are placed in the locked queue
1255 * immediately, regardless of their state.
408357d8 1256 */
27bc0cb1
MD
1257 bp->b_qindex = BQUEUE_LOCKED;
1258 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
408357d8 1259 } else if (bp->b_bufsize == 0) {
9188c711
MD
1260 /*
1261 * Buffers with no memory. Due to conditionals near the top
1262 * of brelse() such buffers should probably already be
1263 * marked B_INVAL and disassociated from their vnode.
1264 */
984263bc 1265 bp->b_flags |= B_INVAL;
54078292 1266 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 1267 KKASSERT((bp->b_flags & B_HASHED) == 0);
984263bc 1268 if (bp->b_kvasize) {
b3098c79 1269 bp->b_qindex = BQUEUE_EMPTYKVA;
984263bc 1270 } else {
b3098c79 1271 bp->b_qindex = BQUEUE_EMPTY;
984263bc
MD
1272 }
1273 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
78a9b77f 1274 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
9188c711
MD
1275 /*
1276 * Buffers with junk contents. Again these buffers had better
1277 * already be disassociated from their vnode.
1278 */
54078292 1279 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 1280 KKASSERT((bp->b_flags & B_HASHED) == 0);
984263bc 1281 bp->b_flags |= B_INVAL;
b3098c79
HP
1282 bp->b_qindex = BQUEUE_CLEAN;
1283 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
984263bc 1284 } else {
9188c711
MD
1285 /*
1286 * Remaining buffers. These buffers are still associated with
1287 * their vnode.
1288 */
b86460bf 1289 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
984263bc 1290 case B_DELWRI:
b3098c79
HP
1291 bp->b_qindex = BQUEUE_DIRTY;
1292 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
984263bc 1293 break;
4b958e7b
MD
1294 case B_DELWRI | B_HEAVY:
1295 bp->b_qindex = BQUEUE_DIRTY_HW;
1296 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1297 b_freelist);
1298 break;
984263bc 1299 default:
b86460bf
MD
1300 /*
1301 * NOTE: Buffers are always placed at the end of the
1302 * queue. If B_AGE is not set the buffer will cycle
1303 * through the queue twice.
1304 */
b3098c79
HP
1305 bp->b_qindex = BQUEUE_CLEAN;
1306 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
984263bc
MD
1307 break;
1308 }
1309 }
1310
1311 /*
1312 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1313 * on the correct queue.
1314 */
1315 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1316 bundirty(bp);
1317
1318 /*
868d24af
MD
1319 * The bp is on an appropriate queue unless locked. If it is not
1320 * locked or dirty we can wakeup threads waiting for buffer space.
1321 *
984263bc
MD
1322 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1323 * if B_INVAL is set ).
1324 */
408357d8 1325 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
984263bc
MD
1326 bufcountwakeup();
1327
1328 /*
1329 * Something we can maybe free or reuse
1330 */
1331 if (bp->b_bufsize || bp->b_kvasize)
1332 bufspacewakeup();
1333
69f8c926
MD
1334 /*
1335 * Clean up temporary flags and unlock the buffer.
1336 */
b86460bf 1337 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
69f8c926 1338 BUF_UNLOCK(bp);
e43a034f 1339 crit_exit();
984263bc
MD
1340}
1341
1342/*
3f779080
HP
1343 * bqrelse:
1344 *
1345 * Release a buffer back to the appropriate queue but do not try to free
1346 * it. The buffer is expected to be used again soon.
984263bc 1347 *
3f779080
HP
1348 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1349 * biodone() to requeue an async I/O on completion. It is also used when
1350 * known good buffers need to be requeued but we think we may need the data
1351 * again soon.
984263bc 1352 *
3f779080 1353 * XXX we should be able to leave the B_RELBUF hint set on completion.
984263bc
MD
1354 */
1355void
c8e4131d 1356bqrelse(struct buf *bp)
984263bc 1357{
e43a034f 1358 crit_enter();
984263bc
MD
1359
1360 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1361
b3098c79 1362 if (bp->b_qindex != BQUEUE_NONE)
984263bc 1363 panic("bqrelse: free buffer onto another queue???");
77bb9400 1364 if (BUF_REFCNTNB(bp) > 1) {
984263bc
MD
1365 /* do not release to free list */
1366 panic("bqrelse: multiple refs");
1367 BUF_UNLOCK(bp);
e43a034f 1368 crit_exit();
984263bc
MD
1369 return;
1370 }
1371 if (bp->b_flags & B_LOCKED) {
5e23ca53
MD
1372 /*
1373 * Locked buffers are released to the locked queue. However,
1374 * if the buffer is dirty it will first go into the dirty
1375 * queue and later on after the I/O completes successfully it
1376 * will be released to the locked queue.
1377 */
27bc0cb1
MD
1378 bp->b_qindex = BQUEUE_LOCKED;
1379 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
984263bc 1380 } else if (bp->b_flags & B_DELWRI) {
4b958e7b
MD
1381 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1382 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1383 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
984263bc
MD
1384 } else if (vm_page_count_severe()) {
1385 /*
1386 * We are too low on memory, we have to try to free the
1387 * buffer (most importantly: the wired pages making up its
1388 * backing store) *now*.
1389 */
e43a034f 1390 crit_exit();
984263bc
MD
1391 brelse(bp);
1392 return;
1393 } else {
b3098c79
HP
1394 bp->b_qindex = BQUEUE_CLEAN;
1395 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
984263bc
MD
1396 }
1397
1398 if ((bp->b_flags & B_LOCKED) == 0 &&
408357d8 1399 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
984263bc
MD
1400 bufcountwakeup();
1401 }
1402
1403 /*
1404 * Something we can maybe free or reuse.
1405 */
1406 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1407 bufspacewakeup();
1408
9188c711
MD
1409 /*
1410 * Final cleanup and unlock. Clear bits that are only used while a
1411 * buffer is actively locked.
1412 */
b86460bf 1413 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF);
9188c711 1414 BUF_UNLOCK(bp);
e43a034f 1415 crit_exit();
984263bc
MD
1416}
1417
3f779080
HP
1418/*
1419 * vfs_vmio_release:
1420 *
1421 * Return backing pages held by the buffer 'bp' back to the VM system
1422 * if possible. The pages are freed if they are no longer valid or
1423 * attempt to free if it was used for direct I/O otherwise they are
1424 * sent to the page cache.
1425 *
1426 * Pages that were marked busy are left alone and skipped.
1427 *
1428 * The KVA mapping (b_data) for the underlying pages is removed by
1429 * this function.
1430 */
984263bc 1431static void
493c516a 1432vfs_vmio_release(struct buf *bp)
984263bc 1433{
e43a034f 1434 int i;
984263bc
MD
1435 vm_page_t m;
1436
e43a034f 1437 crit_enter();
54f51aeb
HP
1438 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1439 m = bp->b_xio.xio_pages[i];
1440 bp->b_xio.xio_pages[i] = NULL;
984263bc
MD
1441 /*
1442 * In order to keep page LRU ordering consistent, put
1443 * everything on the inactive queue.
1444 */
1445 vm_page_unwire(m, 0);
1446 /*
1447 * We don't mess with busy pages, it is
1448 * the responsibility of the process that
1449 * busied the pages to deal with them.
1450 */
1451 if ((m->flags & PG_BUSY) || (m->busy != 0))
1452 continue;
1453
1454 if (m->wire_count == 0) {
1455 vm_page_flag_clear(m, PG_ZERO);
1456 /*
1457 * Might as well free the page if we can and it has
1458 * no valid data. We also free the page if the
1459 * buffer was used for direct I/O.
1460 */
3f779080
HP
1461 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1462 m->hold_count == 0) {
984263bc
MD
1463 vm_page_busy(m);
1464 vm_page_protect(m, VM_PROT_NONE);
1465 vm_page_free(m);
1466 } else if (bp->b_flags & B_DIRECT) {
1467 vm_page_try_to_free(m);
1468 } else if (vm_page_count_severe()) {
1469 vm_page_try_to_cache(m);
1470 }
1471 }
1472 }
e43a034f 1473 crit_exit();
54f51aeb 1474 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
984263bc
MD
1475 if (bp->b_bufsize) {
1476 bufspacewakeup();
1477 bp->b_bufsize = 0;
1478 }
54f51aeb 1479 bp->b_xio.xio_npages = 0;
984263bc 1480 bp->b_flags &= ~B_VMIO;
f9a11477 1481 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
984263bc
MD
1482 if (bp->b_vp)
1483 brelvp(bp);
1484}
1485
1486/*
3f779080 1487 * vfs_bio_awrite:
984263bc
MD
1488 *
1489 * Implement clustered async writes for clearing out B_DELWRI buffers.
1490 * This is much better then the old way of writing only one buffer at
1491 * a time. Note that we may not be presented with the buffers in the
1492 * correct order, so we search for the cluster in both directions.
6f68d895
MD
1493 *
1494 * The buffer is locked on call.
984263bc
MD
1495 */
1496int
6f68d895 1497vfs_bio_awrite(struct buf *bp)
984263bc
MD
1498{
1499 int i;
1500 int j;
54078292 1501 off_t loffset = bp->b_loffset;
984263bc 1502 struct vnode *vp = bp->b_vp;
54078292 1503 int nbytes;
984263bc
MD
1504 struct buf *bpa;
1505 int nwritten;
1506 int size;
984263bc 1507
e43a034f 1508 crit_enter();
984263bc
MD
1509 /*
1510 * right now we support clustered writing only to regular files. If
1511 * we find a clusterable block we could be in the middle of a cluster
1512 * rather then at the beginning.
81b5c339 1513 *
54078292
MD
1514 * NOTE: b_bio1 contains the logical loffset and is aliased
1515 * to b_loffset. b_bio2 contains the translated block number.
984263bc
MD
1516 */
1517 if ((vp->v_type == VREG) &&
1518 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1519 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1520
1521 size = vp->v_mount->mnt_stat.f_iosize;
984263bc 1522
54078292
MD
1523 for (i = size; i < MAXPHYS; i += size) {
1524 if ((bpa = findblk(vp, loffset + i)) &&
984263bc
MD
1525 BUF_REFCNT(bpa) == 0 &&
1526 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1527 (B_DELWRI | B_CLUSTEROK)) &&
1528 (bpa->b_bufsize == size)) {
54078292
MD
1529 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1530 (bpa->b_bio2.bio_offset !=
1531 bp->b_bio2.bio_offset + i))
984263bc
MD
1532 break;
1533 } else {
1534 break;
1535 }
1536 }
54078292
MD
1537 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1538 if ((bpa = findblk(vp, loffset - j)) &&
984263bc
MD
1539 BUF_REFCNT(bpa) == 0 &&
1540 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1541 (B_DELWRI | B_CLUSTEROK)) &&
1542 (bpa->b_bufsize == size)) {
54078292
MD
1543 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1544 (bpa->b_bio2.bio_offset !=
1545 bp->b_bio2.bio_offset - j))
984263bc
MD
1546 break;
1547 } else {
1548 break;
1549 }
1550 }
54078292
MD
1551 j -= size;
1552 nbytes = (i + j);
984263bc
MD
1553 /*
1554 * this is a possible cluster write
1555 */
54078292 1556 if (nbytes != size) {
6f68d895 1557 BUF_UNLOCK(bp);
54078292
MD
1558 nwritten = cluster_wbuild(vp, size,
1559 loffset - j, nbytes);
e43a034f 1560 crit_exit();
984263bc
MD
1561 return nwritten;
1562 }
1563 }
1564
984263bc
MD
1565 bremfree(bp);
1566 bp->b_flags |= B_ASYNC;
1567
e43a034f 1568 crit_exit();
984263bc
MD
1569 /*
1570 * default (old) behavior, writing out only one block
1571 *
1572 * XXX returns b_bufsize instead of b_bcount for nwritten?
1573 */
1574 nwritten = bp->b_bufsize;
62cfda27 1575 bwrite(bp);
984263bc
MD
1576
1577 return nwritten;
1578}
1579
1580/*
3f779080 1581 * getnewbuf:
984263bc
MD
1582 *
1583 * Find and initialize a new buffer header, freeing up existing buffers
1584 * in the bufqueues as necessary. The new buffer is returned locked.
1585 *
1586 * Important: B_INVAL is not set. If the caller wishes to throw the
1587 * buffer away, the caller must set B_INVAL prior to calling brelse().
1588 *
1589 * We block if:
1590 * We have insufficient buffer headers
1591 * We have insufficient buffer space
1592 * buffer_map is too fragmented ( space reservation fails )
1593 * If we have to flush dirty buffers ( but we try to avoid this )
1594 *
1595 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1596 * Instead we ask the buf daemon to do it for us. We attempt to
1597 * avoid piecemeal wakeups of the pageout daemon.
1598 */
1599
1600static struct buf *
4b958e7b 1601getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
984263bc
MD
1602{
1603 struct buf *bp;
1604 struct buf *nbp;
1605 int defrag = 0;
1606 int nqindex;
4b958e7b 1607 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
984263bc
MD
1608 static int flushingbufs;
1609
1610 /*
1611 * We can't afford to block since we might be holding a vnode lock,
1612 * which may prevent system daemons from running. We deal with
1613 * low-memory situations by proactively returning memory and running
1614 * async I/O rather then sync I/O.
1615 */
1616
1617 ++getnewbufcalls;
1618 --getnewbufrestarts;
1619restart:
1620 ++getnewbufrestarts;
1621
1622 /*
1623 * Setup for scan. If we do not have enough free buffers,
1624 * we setup a degenerate case that immediately fails. Note
1625 * that if we are specially marked process, we are allowed to
1626 * dip into our reserves.
1627 *
1628 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1629 *
1630 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1631 * However, there are a number of cases (defragging, reusing, ...)
1632 * where we cannot backup.
1633 */
b3098c79
HP
1634 nqindex = BQUEUE_EMPTYKVA;
1635 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
984263bc
MD
1636
1637 if (nbp == NULL) {
1638 /*
1639 * If no EMPTYKVA buffers and we are either
1640 * defragging or reusing, locate a CLEAN buffer
1641 * to free or reuse. If bufspace useage is low
1642 * skip this step so we can allocate a new buffer.
1643 */
1644 if (defrag || bufspace >= lobufspace) {
b3098c79
HP
1645 nqindex = BQUEUE_CLEAN;
1646 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
984263bc
MD
1647 }
1648
1649 /*
1650 * If we could not find or were not allowed to reuse a
1651 * CLEAN buffer, check to see if it is ok to use an EMPTY
1652 * buffer. We can only use an EMPTY buffer if allocating
1653 * its KVA would not otherwise run us out of buffer space.
1654 */
1655 if (nbp == NULL && defrag == 0 &&
1656 bufspace + maxsize < hibufspace) {
b3098c79
HP
1657 nqindex = BQUEUE_EMPTY;
1658 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
984263bc
MD
1659 }
1660 }
1661
1662 /*
1663 * Run scan, possibly freeing data and/or kva mappings on the fly
1664 * depending.
1665 */
1666
1667 while ((bp = nbp) != NULL) {
1668 int qindex = nqindex;
1669
b86460bf
MD
1670 nbp = TAILQ_NEXT(bp, b_freelist);
1671
1672 /*
1673 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1674 * cycles through the queue twice before being selected.
1675 */
1676 if (qindex == BQUEUE_CLEAN &&
1677 (bp->b_flags & B_AGE) == 0 && nbp) {
1678 bp->b_flags |= B_AGE;
1679 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1680 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1681 continue;
1682 }
1683
984263bc
MD
1684 /*
1685 * Calculate next bp ( we can only use it if we do not block
1686 * or do other fancy things ).
1687 */
b86460bf 1688 if (nbp == NULL) {
984263bc 1689 switch(qindex) {
b3098c79
HP
1690 case BQUEUE_EMPTY:
1691 nqindex = BQUEUE_EMPTYKVA;
1692 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
984263bc
MD
1693 break;
1694 /* fall through */
b3098c79
HP
1695 case BQUEUE_EMPTYKVA:
1696 nqindex = BQUEUE_CLEAN;
1697 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
984263bc
MD
1698 break;
1699 /* fall through */
b3098c79 1700 case BQUEUE_CLEAN:
984263bc
MD
1701 /*
1702 * nbp is NULL.
1703 */
1704 break;
1705 }
1706 }
1707
1708 /*
1709 * Sanity Checks
1710 */
3f625015 1711 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
984263bc
MD
1712
1713 /*
1714 * Note: we no longer distinguish between VMIO and non-VMIO
1715 * buffers.
1716 */
1717
1718 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1719
1720 /*
1721 * If we are defragging then we need a buffer with
1722 * b_kvasize != 0. XXX this situation should no longer
1723 * occur, if defrag is non-zero the buffer's b_kvasize
1724 * should also be non-zero at this point. XXX
1725 */
1726 if (defrag && bp->b_kvasize == 0) {
6ea70f76 1727 kprintf("Warning: defrag empty buffer %p\n", bp);
984263bc
MD
1728 continue;
1729 }
1730
1731 /*
1732 * Start freeing the bp. This is somewhat involved. nbp
b3098c79 1733 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
9188c711
MD
1734 * on the clean list must be disassociated from their
1735 * current vnode. Buffers on the empty[kva] lists have
1736 * already been disassociated.
984263bc
MD
1737 */
1738
d9dba6f6 1739 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
6ea70f76 1740 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
d9dba6f6
MD
1741 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1742 goto restart;
1743 }
1744 if (bp->b_qindex != qindex) {
6ea70f76 1745 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
d9dba6f6
MD
1746 BUF_UNLOCK(bp);
1747 goto restart;
1748 }
984263bc
MD
1749 bremfree(bp);
1750
408357d8
MD
1751 /*
1752 * Dependancies must be handled before we disassociate the
1753 * vnode.
1754 *
1755 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1756 * be immediately disassociated. HAMMER then becomes
1757 * responsible for releasing the buffer.
1758 */
1759 if (LIST_FIRST(&bp->b_dep) != NULL) {
1760 buf_deallocate(bp);
1761 if (bp->b_flags & B_LOCKED) {
1762 bqrelse(bp);
1763 goto restart;
1764 }
4b958e7b 1765 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
408357d8
MD
1766 }
1767
b3098c79 1768 if (qindex == BQUEUE_CLEAN) {
984263bc
MD
1769 if (bp->b_flags & B_VMIO) {
1770 bp->b_flags &= ~B_ASYNC;
1771 vfs_vmio_release(bp);
1772 }
1773 if (bp->b_vp)
1774 brelvp(bp);
1775 }
1776
1777 /*
1778 * NOTE: nbp is now entirely invalid. We can only restart
1779 * the scan from this point on.
1780 *
1781 * Get the rest of the buffer freed up. b_kva* is still
1782 * valid after this operation.
1783 */
1784
54078292 1785 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 1786 KKASSERT((bp->b_flags & B_HASHED) == 0);
984263bc 1787
06ecca5a 1788 /*
e43a034f
MD
1789 * critical section protection is not required when
1790 * scrapping a buffer's contents because it is already
1791 * wired.
06ecca5a 1792 */
984263bc
MD
1793 if (bp->b_bufsize)
1794 allocbuf(bp, 0);
1795
4414f2c9 1796 bp->b_flags = B_BNOCLIP;
10f3fee5 1797 bp->b_cmd = BUF_CMD_DONE;
984263bc 1798 bp->b_vp = NULL;
984263bc
MD
1799 bp->b_error = 0;
1800 bp->b_resid = 0;
1801 bp->b_bcount = 0;
54f51aeb 1802 bp->b_xio.xio_npages = 0;
984263bc 1803 bp->b_dirtyoff = bp->b_dirtyend = 0;
81b5c339 1804 reinitbufbio(bp);
b86460bf 1805 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
408357d8 1806 buf_dep_init(bp);
4b958e7b
MD
1807 if (blkflags & GETBLK_BHEAVY)
1808 bp->b_flags |= B_HEAVY;
984263bc
MD
1809
1810 /*
1811 * If we are defragging then free the buffer.
1812 */
1813 if (defrag) {
1814 bp->b_flags |= B_INVAL;
1815 bfreekva(bp);
1816 brelse(bp);
1817 defrag = 0;
1818 goto restart;
1819 }
1820
1821 /*
1822 * If we are overcomitted then recover the buffer and its
1823 * KVM space. This occurs in rare situations when multiple
1824 * processes are blocked in getnewbuf() or allocbuf().
1825 */
1826 if (bufspace >= hibufspace)
1827 flushingbufs = 1;
1828 if (flushingbufs && bp->b_kvasize != 0) {
1829 bp->b_flags |= B_INVAL;
1830 bfreekva(bp);
1831 brelse(bp);
1832 goto restart;
1833 }
1834 if (bufspace < lobufspace)
1835 flushingbufs = 0;
1836 break;
1837 }
1838
1839 /*
1840 * If we exhausted our list, sleep as appropriate. We may have to
1841 * wakeup various daemons and write out some dirty buffers.
1842 *
1843 * Generally we are sleeping due to insufficient buffer space.
1844 */
1845
1846 if (bp == NULL) {
1847 int flags;
1848 char *waitmsg;
1849
1850 if (defrag) {
1851 flags = VFS_BIO_NEED_BUFSPACE;
1852 waitmsg = "nbufkv";
1853 } else if (bufspace >= hibufspace) {
1854 waitmsg = "nbufbs";
1855 flags = VFS_BIO_NEED_BUFSPACE;
1856 } else {
1857 waitmsg = "newbuf";
1858 flags = VFS_BIO_NEED_ANY;
1859 }
1860
984263bc 1861 needsbuffer |= flags;
4b958e7b 1862 bd_speedup(); /* heeeelp */
984263bc 1863 while (needsbuffer & flags) {
4b958e7b 1864 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
984263bc
MD
1865 return (NULL);
1866 }
1867 } else {
1868 /*
1869 * We finally have a valid bp. We aren't quite out of the
1870 * woods, we still have to reserve kva space. In order
1871 * to keep fragmentation sane we only allocate kva in
1872 * BKVASIZE chunks.
1873 */
1874 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1875
1876 if (maxsize != bp->b_kvasize) {
1877 vm_offset_t addr = 0;
a108bf71 1878 int count;
984263bc
MD
1879
1880 bfreekva(bp);
1881
a108bf71 1882 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
e4846942 1883 vm_map_lock(&buffer_map);
984263bc 1884
e4846942
MD
1885 if (vm_map_findspace(&buffer_map,
1886 vm_map_min(&buffer_map), maxsize,
e9bb90e8 1887 maxsize, &addr)) {
984263bc 1888 /*
3f779080 1889 * Uh oh. Buffer map is too fragmented. We
984263bc
MD
1890 * must defragment the map.
1891 */
e4846942 1892 vm_map_unlock(&buffer_map);
a108bf71 1893 vm_map_entry_release(count);
984263bc
MD
1894 ++bufdefragcnt;
1895 defrag = 1;
1896 bp->b_flags |= B_INVAL;
1897 brelse(bp);
1898 goto restart;
1899 }
1900 if (addr) {
e4846942 1901 vm_map_insert(&buffer_map, &count,
a108bf71 1902 NULL, 0,
984263bc 1903 addr, addr + maxsize,
1b874851
MD
1904 VM_MAPTYPE_NORMAL,
1905 VM_PROT_ALL, VM_PROT_ALL,
1906 MAP_NOFAULT);
984263bc
MD
1907
1908 bp->b_kvabase = (caddr_t) addr;
1909 bp->b_kvasize = maxsize;
1910 bufspace += bp->b_kvasize;
1911 ++bufreusecnt;
1912 }
e4846942 1913 vm_map_unlock(&buffer_map);
a108bf71 1914 vm_map_entry_release(count);
984263bc
MD
1915 }
1916 bp->b_data = bp->b_kvabase;
1917 }
1918 return(bp);
1919}
1920
1921/*
4ecf7cc9
MD
1922 * This routine is called in an emergency to recover VM pages from the
1923 * buffer cache by cashing in clean buffers. The idea is to recover
1924 * enough pages to be able to satisfy a stuck bio_page_alloc().
1925 */
1926static int
1927recoverbufpages(void)
1928{
1929 struct buf *bp;
1930 int bytes = 0;
1931
1932 ++recoverbufcalls;
1933
1934 while (bytes < MAXBSIZE) {
1935 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1936 if (bp == NULL)
1937 break;
1938
1939 /*
1940 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1941 * cycles through the queue twice before being selected.
1942 */
1943 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
1944 bp->b_flags |= B_AGE;
1945 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1946 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
1947 bp, b_freelist);
1948 continue;
1949 }
1950
1951 /*
1952 * Sanity Checks
1953 */
1954 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
1955 KKASSERT((bp->b_flags & B_DELWRI) == 0);
1956
1957 /*
1958 * Start freeing the bp. This is somewhat involved.
1959 *
1960 * Buffers on the clean list must be disassociated from
1961 * their current vnode
1962 */
1963
1964 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1965 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
1966 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1967 continue;
1968 }
1969 if (bp->b_qindex != BQUEUE_CLEAN) {
1970 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
1971 BUF_UNLOCK(bp);
1972 continue;
1973 }
1974 bremfree(bp);
1975
1976 /*
1977 * Dependancies must be handled before we disassociate the
1978 * vnode.
1979 *
1980 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1981 * be immediately disassociated. HAMMER then becomes
1982 * responsible for releasing the buffer.
1983 */
1984 if (LIST_FIRST(&bp->b_dep) != NULL) {
1985 buf_deallocate(bp);
1986 if (bp->b_flags & B_LOCKED) {
1987 bqrelse(bp);
1988 continue;
1989 }
1990 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1991 }
1992
1993 bytes += bp->b_bufsize;
1994
1995 if (bp->b_flags & B_VMIO) {
1996 bp->b_flags &= ~B_ASYNC;
1997 bp->b_flags |= B_DIRECT; /* try to free pages */
1998 vfs_vmio_release(bp);
1999 }
2000 if (bp->b_vp)
2001 brelvp(bp);
2002
2003 KKASSERT(bp->b_vp == NULL);
2004 KKASSERT((bp->b_flags & B_HASHED) == 0);
2005
2006 /*
2007 * critical section protection is not required when
2008 * scrapping a buffer's contents because it is already
2009 * wired.
2010 */
2011 if (bp->b_bufsize)
2012 allocbuf(bp, 0);
2013
2014 bp->b_flags = B_BNOCLIP;
2015 bp->b_cmd = BUF_CMD_DONE;
2016 bp->b_vp = NULL;
2017 bp->b_error = 0;
2018 bp->b_resid = 0;
2019 bp->b_bcount = 0;
2020 bp->b_xio.xio_npages = 0;
2021 bp->b_dirtyoff = bp->b_dirtyend = 0;
2022 reinitbufbio(bp);
2023 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2024 buf_dep_init(bp);
2025 bp->b_flags |= B_INVAL;
2026 /* bfreekva(bp); */
2027 brelse(bp);
2028 }
2029 return(bytes);
2030}
2031
2032/*
3f779080 2033 * buf_daemon:
984263bc 2034 *
3f779080 2035 * Buffer flushing daemon. Buffers are normally flushed by the
984263bc
MD
2036 * update daemon but if it cannot keep up this process starts to
2037 * take the load in an attempt to prevent getnewbuf() from blocking.
4b958e7b
MD
2038 *
2039 * Once a flush is initiated it does not stop until the number
2040 * of buffers falls below lodirtybuffers, but we will wake up anyone
2041 * waiting at the mid-point.
984263bc
MD
2042 */
2043
984263bc
MD
2044static struct kproc_desc buf_kp = {
2045 "bufdaemon",
2046 buf_daemon,
4b958e7b
MD
2047 &bufdaemon_td
2048};
2049SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2050 kproc_start, &buf_kp)
2051
2052static struct kproc_desc bufhw_kp = {
2053 "bufdaemon_hw",
2054 buf_daemon_hw,
2055 &bufdaemonhw_td
984263bc 2056};
4b958e7b
MD
2057SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2058 kproc_start, &bufhw_kp)
984263bc
MD
2059
2060static void
c972a82f 2061buf_daemon(void)
984263bc 2062{
cd083340
MD
2063 int limit;
2064
984263bc
MD
2065 /*
2066 * This process needs to be suspended prior to shutdown sync.
2067 */
bc6dffab 2068 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
4b958e7b 2069 bufdaemon_td, SHUTDOWN_PRI_LAST);
4ecf7cc9 2070 curthread->td_flags |= TDF_SYSTHREAD;
984263bc
MD
2071
2072 /*
2073 * This process is allowed to take the buffer cache to the limit
2074 */
e43a034f 2075 crit_enter();
984263bc
MD
2076
2077 for (;;) {
0cfcada1 2078 kproc_suspend_loop();
984263bc
MD
2079
2080 /*
2081 * Do the flush. Limit the amount of in-transit I/O we
2082 * allow to build up, otherwise we would completely saturate
2083 * the I/O system. Wakeup any waiting processes before we
2084 * normally would so they can run in parallel with our drain.
cd083340
MD
2085 *
2086 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2087 * but because we split the operation into two threads we
2088 * have to cut it in half for each thread.
984263bc 2089 */
cd083340
MD
2090 limit = lodirtybufspace / 2;
2091 waitrunningbufspace(limit);
70ac7d6c
MD
2092 while (runningbufspace + dirtybufspace > limit ||
2093 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
4b958e7b 2094 if (flushbufqueues(BQUEUE_DIRTY) == 0)
984263bc 2095 break;
cd083340 2096 waitrunningbufspace(limit);
1b30fbcc 2097 }
984263bc
MD
2098
2099 /*
cd083340
MD
2100 * We reached our low water mark, reset the
2101 * request and sleep until we are needed again.
2102 * The sleep is just so the suspend code works.
984263bc 2103 */
cd083340
MD
2104 spin_lock_wr(&needsbuffer_spin);
2105 if (bd_request == 0) {
4b958e7b
MD
2106 msleep(&bd_request, &needsbuffer_spin, 0,
2107 "psleep", hz);
984263bc 2108 }
cd083340
MD
2109 bd_request = 0;
2110 spin_unlock_wr(&needsbuffer_spin);
984263bc
MD
2111 }
2112}
2113
4b958e7b
MD
2114static void
2115buf_daemon_hw(void)
2116{
cd083340
MD
2117 int limit;
2118
4b958e7b
MD
2119 /*
2120 * This process needs to be suspended prior to shutdown sync.
2121 */
2122 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2123 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
4ecf7cc9 2124 curthread->td_flags |= TDF_SYSTHREAD;
4b958e7b
MD
2125
2126 /*
2127 * This process is allowed to take the buffer cache to the limit
2128 */
2129 crit_enter();
2130
2131 for (;;) {
2132 kproc_suspend_loop();
2133
2134 /*
2135 * Do the flush. Limit the amount of in-transit I/O we
2136 * allow to build up, otherwise we would completely saturate
2137 * the I/O system. Wakeup any waiting processes before we
2138 * normally would so they can run in parallel with our drain.
cd083340
MD
2139 *
2140 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2141 * but because we split the operation into two threads we
2142 * have to cut it in half for each thread.
4b958e7b 2143 */
cd083340
MD
2144 limit = lodirtybufspace / 2;
2145 waitrunningbufspace(limit);
70ac7d6c
MD
2146 while (runningbufspace + dirtybufspacehw > limit ||
2147 dirtybufcounthw >= nbuf / 2) {
4b958e7b
MD
2148 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2149 break;
cd083340 2150 waitrunningbufspace(limit);
1b30fbcc 2151 }
4b958e7b
MD
2152
2153 /*
cd083340
MD
2154 * We reached our low water mark, reset the
2155 * request and sleep until we are needed again.
2156 * The sleep is just so the suspend code works.
4b958e7b 2157 */
cd083340
MD
2158 spin_lock_wr(&needsbuffer_spin);
2159 if (bd_request_hw == 0) {
4b958e7b
MD
2160 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2161 "psleep", hz);
4b958e7b 2162 }
cd083340
MD
2163 bd_request_hw = 0;
2164 spin_unlock_wr(&needsbuffer_spin);
4b958e7b
MD
2165 }
2166}
2167
984263bc 2168/*
3f779080 2169 * flushbufqueues:
984263bc
MD
2170 *
2171 * Try to flush a buffer in the dirty queue. We must be careful to
2172 * free up B_INVAL buffers instead of write them, which NFS is
2173 * particularly sensitive to.
b86460bf
MD
2174 *
2175 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2176 * that we really want to try to get the buffer out and reuse it
2177 * due to the write load on the machine.
984263bc
MD
2178 */
2179
2180static int
4b958e7b 2181flushbufqueues(bufq_type_t q)
984263bc
MD
2182{
2183 struct buf *bp;
2184 int r = 0;
2185
4b958e7b 2186 bp = TAILQ_FIRST(&bufqueues[q]);
984263bc 2187 while (bp) {
408357d8
MD
2188 KASSERT((bp->b_flags & B_DELWRI),
2189 ("unexpected clean buffer %p", bp));
b86460bf 2190
70371608 2191 if (bp->b_flags & B_DELWRI) {
984263bc
MD
2192 if (bp->b_flags & B_INVAL) {
2193 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2194 panic("flushbufqueues: locked buf");
2195 bremfree(bp);
2196 brelse(bp);
2197 ++r;
2198 break;
2199 }
2200 if (LIST_FIRST(&bp->b_dep) != NULL &&
984263bc 2201 (bp->b_flags & B_DEFERRED) == 0 &&
408357d8 2202 buf_countdeps(bp, 0)) {
4b958e7b
MD
2203 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2204 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2205 b_freelist);
984263bc 2206 bp->b_flags |= B_DEFERRED;
4b958e7b 2207 bp = TAILQ_FIRST(&bufqueues[q]);
984263bc
MD
2208 continue;
2209 }
6f68d895
MD
2210
2211 /*
2212 * Only write it out if we can successfully lock
27bc0cb1 2213 * it. If the buffer has a dependancy,
78a9b77f
MD
2214 * buf_checkwrite must also return 0 for us to
2215 * be able to initate the write.
2216 *
2217 * If the buffer is flagged B_ERROR it may be
2218 * requeued over and over again, we try to
2219 * avoid a live lock.
6f68d895
MD
2220 */
2221 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
27bc0cb1
MD
2222 if (LIST_FIRST(&bp->b_dep) != NULL &&
2223 buf_checkwrite(bp)) {
2224 bremfree(bp);
2225 brelse(bp);
78a9b77f
MD
2226 } else if (bp->b_flags & B_ERROR) {
2227 tsleep(bp, 0, "bioer", 1);
2228 bp->b_flags &= ~B_AGE;
2229 vfs_bio_awrite(bp);
27bc0cb1 2230 } else {
b86460bf 2231 bp->b_flags |= B_AGE;
27bc0cb1
MD
2232 vfs_bio_awrite(bp);
2233 }
6f68d895
MD
2234 ++r;
2235 break;
2236 }
984263bc
MD
2237 }
2238 bp = TAILQ_NEXT(bp, b_freelist);
2239 }
2240 return (r);
2241}
2242
2243/*
3f779080
HP
2244 * inmem:
2245 *
2246 * Returns true if no I/O is needed to access the associated VM object.
1f1ea522 2247 * This is like findblk except it also hunts around in the VM system for
3f779080 2248 * the data.
06ecca5a 2249 *
3f779080
HP
2250 * Note that we ignore vm_page_free() races from interrupts against our
2251 * lookup, since if the caller is not protected our return value will not
2252 * be any more valid then otherwise once we exit the critical section.
984263bc 2253 */
984263bc 2254int
54078292 2255inmem(struct vnode *vp, off_t loffset)
984263bc
MD
2256{
2257 vm_object_t obj;
2258 vm_offset_t toff, tinc, size;
2259 vm_page_t m;
984263bc 2260
54078292 2261 if (findblk(vp, loffset))
984263bc
MD
2262 return 1;
2263 if (vp->v_mount == NULL)
2264 return 0;
7540ab49
MD
2265 if ((obj = vp->v_object) == NULL)
2266 return 0;
984263bc
MD
2267
2268 size = PAGE_SIZE;
2269 if (size > vp->v_mount->mnt_stat.f_iosize)
2270 size = vp->v_mount->mnt_stat.f_iosize;
984263bc
MD
2271
2272 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
54078292
MD
2273 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2274 if (m == NULL)
984263bc
MD
2275 return 0;
2276 tinc = size;
54078292
MD
2277 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2278 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
984263bc 2279 if (vm_page_is_valid(m,
54078292 2280 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
984263bc
MD
2281 return 0;
2282 }
2283 return 1;
2284}
2285
2286/*
3f779080 2287 * vfs_setdirty:
984263bc
MD
2288 *
2289 * Sets the dirty range for a buffer based on the status of the dirty
2290 * bits in the pages comprising the buffer.
2291 *
2292 * The range is limited to the size of the buffer.
2293 *
2294 * This routine is primarily used by NFS, but is generalized for the
2295 * B_VMIO case.
2296 */
2297static void
2298vfs_setdirty(struct buf *bp)
2299{
2300 int i;
2301 vm_object_t object;
2302
2303 /*
2304 * Degenerate case - empty buffer
2305 */
2306
2307 if (bp->b_bufsize == 0)
2308 return;
2309
2310 /*
2311 * We qualify the scan for modified pages on whether the
2312 * object has been flushed yet. The OBJ_WRITEABLE flag
2313 * is not cleared simply by protecting pages off.
2314 */
2315
2316 if ((bp->b_flags & B_VMIO) == 0)
2317 return;
2318
54f51aeb 2319 object = bp->b_xio.xio_pages[0]->object;
984263bc
MD
2320
2321 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
6ea70f76 2322 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
984263bc 2323 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
6ea70f76 2324 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
984263bc
MD
2325
2326 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2327 vm_offset_t boffset;
2328 vm_offset_t eoffset;
2329
2330 /*
2331 * test the pages to see if they have been modified directly
2332 * by users through the VM system.
2333 */
54f51aeb
HP
2334 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2335 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2336 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
984263bc
MD
2337 }
2338
2339 /*
2340 * Calculate the encompassing dirty range, boffset and eoffset,
2341 * (eoffset - boffset) bytes.
2342 */
2343
54f51aeb
HP
2344 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2345 if (bp->b_xio.xio_pages[i]->dirty)
984263bc
MD
2346 break;
2347 }
81b5c339 2348 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
984263bc 2349
54f51aeb
HP
2350 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2351 if (bp->b_xio.xio_pages[i]->dirty) {
984263bc
MD
2352 break;
2353 }
2354 }
81b5c339 2355 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
984263bc
MD
2356
2357 /*
2358 * Fit it to the buffer.
2359 */
2360
2361 if (eoffset > bp->b_bcount)
2362 eoffset = bp->b_bcount;
2363
2364 /*
2365 * If we have a good dirty range, merge with the existing
2366 * dirty range.
2367 */
2368
2369 if (boffset < eoffset) {
2370 if (bp->b_dirtyoff > boffset)
2371 bp->b_dirtyoff = boffset;
2372 if (bp->b_dirtyend < eoffset)
2373 bp->b_dirtyend = eoffset;
2374 }
2375 }
2376}
2377
2378/*
1f1ea522
MD
2379 * findblk:
2380 *
2381 * Locate and return the specified buffer, or NULL if the buffer does
2382 * not exist. Do not attempt to lock the buffer or manipulate it in
2383 * any way. The caller must validate that the correct buffer has been
2384 * obtain after locking it.
2385 */
2386struct buf *
54078292 2387findblk(struct vnode *vp, off_t loffset)
1f1ea522
MD
2388{
2389 struct buf *bp;
2390
2391 crit_enter();
54078292 2392 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
1f1ea522
MD
2393 crit_exit();
2394 return(bp);
2395}
2396
2397/*
3f779080 2398 * getblk:
984263bc
MD
2399 *
2400 * Get a block given a specified block and offset into a file/device.
10f3fee5
MD
2401 * B_INVAL may or may not be set on return. The caller should clear
2402 * B_INVAL prior to initiating a READ.
984263bc 2403 *
77bb9400
MD
2404 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2405 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2406 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2407 * without doing any of those things the system will likely believe
2408 * the buffer to be valid (especially if it is not B_VMIO), and the
2409 * next getblk() will return the buffer with B_CACHE set.
2410 *
984263bc
MD
2411 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2412 * an existing buffer.
2413 *
2414 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2415 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2416 * and then cleared based on the backing VM. If the previous buffer is
2417 * non-0-sized but invalid, B_CACHE will be cleared.
2418 *
2419 * If getblk() must create a new buffer, the new buffer is returned with
2420 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2421 * case it is returned with B_INVAL clear and B_CACHE set based on the
2422 * backing VM.
2423 *
62cfda27 2424 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
984263bc
MD
2425 * B_CACHE bit is clear.
2426 *
2427 * What this means, basically, is that the caller should use B_CACHE to
2428 * determine whether the buffer is fully valid or not and should clear
2429 * B_INVAL prior to issuing a read. If the caller intends to validate
2430 * the buffer by loading its data area with something, the caller needs
2431 * to clear B_INVAL. If the caller does this without issuing an I/O,
2432 * the caller should set B_CACHE ( as an optimization ), else the caller
2433 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2434 * a write attempt or if it was a successfull read. If the caller
2435 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2436 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
4b958e7b
MD
2437 *
2438 * getblk flags:
2439 *
2440 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2441 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
984263bc
MD
2442 */
2443struct buf *
4b958e7b 2444getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
984263bc
MD
2445{
2446 struct buf *bp;
4b958e7b 2447 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
e92ca23a 2448 int error;
984263bc
MD
2449
2450 if (size > MAXBSIZE)
fc92d4aa 2451 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
7540ab49
MD
2452 if (vp->v_object == NULL)
2453 panic("getblk: vnode %p has no object!", vp);
984263bc 2454
e43a034f 2455 crit_enter();
984263bc 2456loop:
54078292 2457 if ((bp = findblk(vp, loffset))) {
984263bc 2458 /*
a0da602d
MD
2459 * The buffer was found in the cache, but we need to lock it.
2460 * Even with LK_NOWAIT the lockmgr may break our critical
2461 * section, so double-check the validity of the buffer
2462 * once the lock has been obtained.
984263bc 2463 */
984263bc 2464 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
b77cfc40
MD
2465 if (blkflags & GETBLK_NOWAIT) {
2466 crit_exit();
2467 return(NULL);
2468 }
f2770c70 2469 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
4b958e7b 2470 if (blkflags & GETBLK_PCATCH)
f2770c70 2471 lkflags |= LK_PCATCH;
e92ca23a
MD
2472 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2473 if (error) {
2474 if (error == ENOLCK)
2475 goto loop;
2476 crit_exit();
2477 return (NULL);
f2770c70 2478 }
984263bc
MD
2479 }
2480
2481 /*
a0da602d
MD
2482 * Once the buffer has been locked, make sure we didn't race
2483 * a buffer recyclement. Buffers that are no longer hashed
2484 * will have b_vp == NULL, so this takes care of that check
2485 * as well.
2486 */
54078292 2487 if (bp->b_vp != vp || bp->b_loffset != loffset) {
6ea70f76 2488 kprintf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset);
a9518ecf 2489 BUF_UNLOCK(bp);
a0da602d
MD
2490 goto loop;
2491 }
2492
2493 /*
b77cfc40
MD
2494 * If SZMATCH any pre-existing buffer must be of the requested
2495 * size or NULL is returned. The caller absolutely does not
2496 * want getblk() to bwrite() the buffer on a size mismatch.
2497 */
2498 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2499 BUF_UNLOCK(bp);
2500 crit_exit();
2501 return(NULL);
2502 }
2503
2504 /*
4baec531
MD
2505 * All vnode-based buffers must be backed by a VM object.
2506 */
2507 KKASSERT(bp->b_flags & B_VMIO);
10f3fee5 2508 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
b86460bf 2509 bp->b_flags &= ~B_AGE;
4baec531
MD
2510
2511 /*
a0da602d
MD
2512 * Make sure that B_INVAL buffers do not have a cached
2513 * block number translation.
2514 */
54078292 2515 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
6ea70f76 2516 kprintf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset);
81b5c339 2517 clearbiocache(&bp->b_bio2);
a0da602d
MD
2518 }
2519
2520 /*
984263bc 2521 * The buffer is locked. B_CACHE is cleared if the buffer is
4baec531 2522 * invalid.
984263bc
MD
2523 */
2524 if (bp->b_flags & B_INVAL)
2525 bp->b_flags &= ~B_CACHE;
984263bc
MD
2526 bremfree(bp);
2527
2528 /*
4baec531
MD
2529 * Any size inconsistancy with a dirty buffer or a buffer
2530 * with a softupdates dependancy must be resolved. Resizing
2531 * the buffer in such circumstances can lead to problems.
984263bc 2532 */
4baec531
MD
2533 if (size != bp->b_bcount) {
2534 if (bp->b_flags & B_DELWRI) {
2535 bp->b_flags |= B_NOCACHE;
62cfda27 2536 bwrite(bp);
4baec531
MD
2537 } else if (LIST_FIRST(&bp->b_dep)) {
2538 bp->b_flags |= B_NOCACHE;
62cfda27 2539 bwrite(bp);
4baec531
MD
2540 } else {
2541 bp->b_flags |= B_RELBUF;
2542 brelse(bp);
984263bc 2543 }
4baec531 2544 goto loop;
984263bc 2545 }
4baec531 2546 KKASSERT(size <= bp->b_kvasize);
81b5c339
MD
2547 KASSERT(bp->b_loffset != NOOFFSET,
2548 ("getblk: no buffer offset"));
984263bc
MD
2549
2550 /*
2551 * A buffer with B_DELWRI set and B_CACHE clear must
2552 * be committed before we can return the buffer in
2553 * order to prevent the caller from issuing a read
2554 * ( due to B_CACHE not being set ) and overwriting
2555 * it.
2556 *
2557 * Most callers, including NFS and FFS, need this to
2558 * operate properly either because they assume they
2559 * can issue a read if B_CACHE is not set, or because
2560 * ( for example ) an uncached B_DELWRI might loop due
2561 * to softupdates re-dirtying the buffer. In the latter
2562 * case, B_CACHE is set after the first write completes,
2563 * preventing further loops.
2564 *
2565 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2566 * above while extending the buffer, we cannot allow the
2567 * buffer to remain with B_CACHE set after the write
2568 * completes or it will represent a corrupt state. To
2569 * deal with this we set B_NOCACHE to scrap the buffer
2570 * after the write.
2571 *
2572 * We might be able to do something fancy, like setting
2573 * B_CACHE in bwrite() except if B_DELWRI is already set,
2574 * so the below call doesn't set B_CACHE, but that gets real
2575 * confusing. This is much easier.
2576 */
2577
2578 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2579 bp->b_flags |= B_NOCACHE;
62cfda27 2580 bwrite(bp);
984263bc
MD
2581 goto loop;
2582 }
e43a034f 2583 crit_exit();
984263bc
MD
2584 } else {
2585 /*
2586 * Buffer is not in-core, create new buffer. The buffer
2587 * returned by getnewbuf() is locked. Note that the returned
2588 * buffer is also considered valid (not marked B_INVAL).
21ab32bd
MD
2589 *
2590 * Calculating the offset for the I/O requires figuring out
2591 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2592 * the mount's f_iosize otherwise. If the vnode does not
2593 * have an associated mount we assume that the passed size is
2594 * the block size.
2595 *
2596 * Note that vn_isdisk() cannot be used here since it may
2597 * return a failure for numerous reasons. Note that the
2598 * buffer size may be larger then the block size (the caller
2599 * will use block numbers with the proper multiple). Beware
2600 * of using any v_* fields which are part of unions. In
2601 * particular, in DragonFly the mount point overloading
1d505369
MD
2602 * mechanism uses the namecache only and the underlying
2603 * directory vnode is not a special case.
984263bc 2604 */
7540ab49 2605 int bsize, maxsize;
984263bc 2606
21ab32bd 2607 if (vp->v_type == VBLK || vp->v_type == VCHR)
984263bc 2608 bsize = DEV_BSIZE;
984263bc
MD
2609 else if (vp->v_mount)
2610 bsize = vp->v_mount->mnt_stat.f_iosize;
2611 else
2612 bsize = size;
2613
7540ab49 2614 maxsize = size + (loffset & PAGE_MASK);
984263bc
MD
2615 maxsize = imax(maxsize, bsize);
2616
4b958e7b
MD
2617 if ((bp = getnewbuf(blkflags, slptimeo, size, maxsize)) == NULL) {
2618 if (slpflags || slptimeo) {
e43a034f 2619 crit_exit();
984263bc
MD
2620 return NULL;
2621 }
2622 goto loop;
2623 }
2624
2625 /*
2626 * This code is used to make sure that a buffer is not
2627 * created while the getnewbuf routine is blocked.
2628 * This can be a problem whether the vnode is locked or not.
2629 * If the buffer is created out from under us, we have to
179024a5 2630 * throw away the one we just created. There is no window
e43a034f
MD
2631 * race because we are safely running in a critical section
2632 * from the point of the duplicate buffer creation through
2633 * to here, and we've locked the buffer.
984263bc 2634 */
54078292 2635 if (findblk(vp, loffset)) {
984263bc
MD
2636 bp->b_flags |= B_INVAL;
2637 brelse(bp);
2638 goto loop;
2639 }
2640
2641 /*
2642 * Insert the buffer into the hash, so that it can
1f1ea522
MD
2643 * be found by findblk().
2644 *
2645 * Make sure the translation layer has been cleared.
984263bc 2646 */
54078292
MD
2647 bp->b_loffset = loffset;
2648 bp->b_bio2.bio_offset = NOOFFSET;
1f1ea522 2649 /* bp->b_bio2.bio_next = NULL; */
984263bc
MD
2650
2651 bgetvp(vp, bp);
984263bc
MD
2652
2653 /*
4baec531 2654 * All vnode-based buffers must be backed by a VM object.
984263bc 2655 */
4baec531
MD
2656 KKASSERT(vp->v_object != NULL);
2657 bp->b_flags |= B_VMIO;
10f3fee5 2658 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
984263bc
MD
2659
2660 allocbuf(bp, size);
2661
e43a034f 2662 crit_exit();
984263bc
MD
2663 }
2664 return (bp);
2665}
2666
2667/*
5e23ca53
MD
2668 * regetblk(bp)
2669 *
27bc0cb1
MD
2670 * Reacquire a buffer that was previously released to the locked queue,
2671 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2672 * set B_LOCKED (which handles the acquisition race).
5e23ca53 2673 *
27bc0cb1
MD
2674 * To this end, either B_LOCKED must be set or the dependancy list must be
2675 * non-empty.
5e23ca53
MD
2676 */
2677void
2678regetblk(struct buf *bp)
2679{
27bc0cb1 2680 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
5e23ca53
MD
2681 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2682 crit_enter();
2683 bremfree(bp);
2684 crit_exit();
2685}
2686
2687/*
3f779080
HP
2688 * geteblk:
2689 *
2690 * Get an empty, disassociated buffer of given size. The buffer is
2691 * initially set to B_INVAL.
06ecca5a 2692 *
3f779080
HP
2693 * critical section protection is not required for the allocbuf()
2694 * call because races are impossible here.
984263bc
MD
2695 */
2696struct buf *
2697geteblk(int size)
2698{
2699 struct buf *bp;
984263bc
MD
2700 int maxsize;
2701
2702 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2703
e43a034f
MD
2704 crit_enter();
2705 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2706 ;
2707 crit_exit();
984263bc
MD
2708 allocbuf(bp, size);
2709 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2710 return (bp);
2711}
2712
2713
2714/*
3f779080 2715 * allocbuf:
984263bc 2716 *
3f779080
HP
2717 * This code constitutes the buffer memory from either anonymous system
2718 * memory (in the case of non-VMIO operations) or from an associated
2719 * VM object (in the case of VMIO operations). This code is able to
2720 * resize a buffer up or down.
984263bc 2721 *
3f779080
HP
2722 * Note that this code is tricky, and has many complications to resolve
2723 * deadlock or inconsistant data situations. Tread lightly!!!
2724 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2725 * the caller. Calling this code willy nilly can result in the loss of data.
06ecca5a 2726 *
3f779080
HP
2727 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2728 * B_CACHE for the non-VMIO case.
2729 *
2730 * This routine does not need to be called from a critical section but you
2731 * must own the buffer.
984263bc 2732 */
984263bc
MD
2733int
2734allocbuf(struct buf *bp, int size)
2735{
2736 int newbsize, mbsize;
2737 int i;
2738
2739 if (BUF_REFCNT(bp) == 0)
2740 panic("allocbuf: buffer not busy");
2741
2742 if (bp->b_kvasize < size)
2743 panic("allocbuf: buffer too small");
2744
2745 if ((bp->b_flags & B_VMIO) == 0) {
2746 caddr_t origbuf;
2747 int origbufsize;
2748 /*
2749 * Just get anonymous memory from the kernel. Don't
2750 * mess with B_CACHE.
2751 */
2752 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
984263bc
MD
2753 if (bp->b_flags & B_MALLOC)
2754 newbsize = mbsize;
2755 else
984263bc
MD
2756 newbsize = round_page(size);
2757
2758 if (newbsize < bp->b_bufsize) {
984263bc 2759 /*
312dcd01 2760 * Malloced buffers are not shrunk
984263bc
MD
2761 */
2762 if (bp->b_flags & B_MALLOC) {
2763 if (newbsize) {
2764 bp->b_bcount = size;
2765 } else {
efda3bd0 2766 kfree(bp->b_data, M_BIOBUF);
984263bc
MD
2767 if (bp->b_bufsize) {
2768 bufmallocspace -= bp->b_bufsize;
2769 bufspacewakeup();
2770 bp->b_bufsize = 0;
2771 }
2772 bp->b_data = bp->b_kvabase;
2773 bp->b_bcount = 0;
2774 bp->b_flags &= ~B_MALLOC;
2775 }
2776 return 1;
2777 }
984263bc
MD
2778 vm_hold_free_pages(
2779 bp,
2780 (vm_offset_t) bp->b_data + newbsize,
2781 (vm_offset_t) bp->b_data + bp->b_bufsize);
2782 } else if (newbsize > bp->b_bufsize) {
984263bc
MD
2783 /*
2784 * We only use malloced memory on the first allocation.
2785 * and revert to page-allocated memory when the buffer
2786 * grows.
2787 */
4baec531 2788 if ((bufmallocspace < maxbufmallocspace) &&
984263bc
MD
2789 (bp->b_bufsize == 0) &&
2790 (mbsize <= PAGE_SIZE/2)) {
2791
efda3bd0 2792 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
984263bc
MD
2793 bp->b_bufsize = mbsize;
2794 bp->b_bcount = size;
2795 bp->b_flags |= B_MALLOC;
2796 bufmallocspace += mbsize;
2797 return 1;
2798 }
984263bc
MD
2799 origbuf = NULL;
2800 origbufsize = 0;
984263bc 2801 /*
4baec531
MD
2802 * If the buffer is growing on its other-than-first
2803 * allocation, then we revert to the page-allocation
2804 * scheme.
984263bc
MD
2805 */
2806 if (bp->b_flags & B_MALLOC) {
2807 origbuf = bp->b_data;
2808 origbufsize = bp->b_bufsize;
2809 bp->b_data = bp->b_kvabase;
2810 if (bp->b_bufsize) {
2811 bufmallocspace -= bp->b_bufsize;
2812 bufspacewakeup();
2813 bp->b_bufsize = 0;
2814 }
2815 bp->b_flags &= ~B_MALLOC;
2816 newbsize = round_page(newbsize);
2817 }
984263bc
MD
2818 vm_hold_load_pages(
2819 bp,
2820 (vm_offset_t) bp->b_data + bp->b_bufsize,
2821 (vm_offset_t) bp->b_data + newbsize);
984263bc
MD
2822 if (origbuf) {
2823 bcopy(origbuf, bp->b_data, origbufsize);
efda3bd0 2824 kfree(origbuf, M_BIOBUF);
984263bc 2825 }
984263bc
MD
2826 }
2827 } else {
2828 vm_page_t m;
2829 int desiredpages;
2830
2831 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
4baec531
MD
2832 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2833 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2834 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
984263bc 2835
984263bc
MD
2836 if (bp->b_flags & B_MALLOC)
2837 panic("allocbuf: VMIO buffer can't be malloced");
984263bc
MD
2838 /*
2839 * Set B_CACHE initially if buffer is 0 length or will become
2840 * 0-length.
2841 */
2842 if (size == 0 || bp->b_bufsize == 0)
2843 bp->b_flags |= B_CACHE;
2844
2845 if (newbsize < bp->b_bufsize) {
2846 /*
2847 * DEV_BSIZE aligned new buffer size is less then the
2848 * DEV_BSIZE aligned existing buffer size. Figure out
2849 * if we have to remove any pages.
2850 */
54f51aeb
HP
2851 if (desiredpages < bp->b_xio.xio_npages) {
2852 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
984263bc
MD
2853 /*
2854 * the page is not freed here -- it
2855 * is the responsibility of
2856 * vnode_pager_setsize
2857 */
54f51aeb 2858 m = bp->b_xio.xio_pages[i];
984263bc
MD
2859 KASSERT(m != bogus_page,
2860 ("allocbuf: bogus page found"));
2861 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2862 ;
2863
54f51aeb 2864 bp->b_xio.xio_pages[i] = NULL;
984263bc
MD
2865 vm_page_unwire(m, 0);
2866 }
2867 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
54f51aeb
HP
2868 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2869 bp->b_xio.xio_npages = desiredpages;
984263bc
MD
2870 }
2871 } else if (size > bp->b_bcount) {
2872 /*
2873 * We are growing the buffer, possibly in a
2874 * byte-granular fashion.
2875 */
2876 struct vnode *vp;
2877 vm_object_t obj;
2878 vm_offset_t toff;
2879 vm_offset_t tinc;
2880
2881 /*
2882 * Step 1, bring in the VM pages from the object,
2883 * allocating them if necessary. We must clear
2884 * B_CACHE if these pages are not valid for the
2885 * range covered by the buffer.
06ecca5a 2886 *
e43a034f
MD
2887 * critical section protection is required to protect
2888 * against interrupts unbusying and freeing pages
2889 * between our vm_page_lookup() and our
2890 * busycheck/wiring call.
984263bc 2891 */
984263bc 2892 vp = bp->b_vp;
7540ab49 2893 obj = vp->v_object;
984263bc 2894
654a39f0 2895 crit_enter();
54f51aeb 2896 while (bp->b_xio.xio_npages < desiredpages) {
984263bc
MD
2897 vm_page_t m;
2898 vm_pindex_t pi;
2899
81b5c339 2900 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
984263bc
MD
2901 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2902 /*
2903 * note: must allocate system pages
2904 * since blocking here could intefere
2905 * with paging I/O, no matter which
2906 * process we are.
2907 */
4ecf7cc9
MD
2908 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
2909 if (m) {
984263bc
MD
2910 vm_page_wire(m);
2911 vm_page_wakeup(m);
2912 bp->b_flags &= ~B_CACHE;
54f51aeb
HP
2913 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2914 ++bp->b_xio.xio_npages;
984263bc
MD
2915 }
2916 continue;
2917 }
2918
2919 /*
2920 * We found a page. If we have to sleep on it,
2921 * retry because it might have gotten freed out
2922 * from under us.
2923 *
2924 * We can only test PG_BUSY here. Blocking on
2925 * m->busy might lead to a deadlock:
2926 *
2927 * vm_fault->getpages->cluster_read->allocbuf
2928 *
2929 */
2930
2931 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2932 continue;
984263bc
MD
2933 vm_page_flag_clear(m, PG_ZERO);
2934 vm_page_wire(m);
54f51aeb
HP
2935 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2936 ++bp->b_xio.xio_npages;
984263bc 2937 }
654a39f0 2938 crit_exit();
984263bc
MD
2939
2940 /*
2941 * Step 2. We've loaded the pages into the buffer,
2942 * we have to figure out if we can still have B_CACHE
2943 * set. Note that B_CACHE is set according to the
3f779080 2944 * byte-granular range ( bcount and size ), not the
984263bc
MD
2945 * aligned range ( newbsize ).
2946 *
2947 * The VM test is against m->valid, which is DEV_BSIZE
2948 * aligned. Needless to say, the validity of the data
2949 * needs to also be DEV_BSIZE aligned. Note that this
2950 * fails with NFS if the server or some other client
2951 * extends the file's EOF. If our buffer is resized,
2952 * B_CACHE may remain set! XXX
2953 */
2954
2955 toff = bp->b_bcount;
81b5c339 2956 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
984263bc
MD
2957
2958 while ((bp->b_flags & B_CACHE) && toff < size) {
2959 vm_pindex_t pi;
2960
2961 if (tinc > (size - toff))
2962 tinc = size - toff;
2963
81b5c339 2964 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
984263bc
MD
2965 PAGE_SHIFT;
2966
2967 vfs_buf_test_cache(
2968 bp,
81b5c339 2969 bp->b_loffset,
984263bc
MD
2970 toff,
2971 tinc,
54f51aeb 2972 bp->b_xio.xio_pages[pi]
984263bc
MD
2973 );
2974 toff += tinc;
2975 tinc = PAGE_SIZE;
2976 }
2977
2978 /*
2979 * Step 3, fixup the KVM pmap. Remember that
81b5c339
MD
2980 * bp->b_data is relative to bp->b_loffset, but
2981 * bp->b_loffset may be offset into the first page.
984263bc
MD
2982 */
2983
2984 bp->b_data = (caddr_t)
2985 trunc_page((vm_offset_t)bp->b_data);
2986 pmap_qenter(
2987 (vm_offset_t)bp->b_data,
54f51aeb
HP
2988 bp->b_xio.xio_pages,
2989 bp->b_xio.xio_npages
984263bc
MD
2990 );
2991 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
81b5c339 2992 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
984263bc
MD
2993 }
2994 }
70ac7d6c
MD
2995
2996 /* adjust space use on already-dirty buffer */
868d24af
MD
2997 if (bp->b_flags & B_DELWRI) {
2998 dirtybufspace += newbsize - bp->b_bufsize;
2999 if (bp->b_flags & B_HEAVY)
3000 dirtybufspacehw += newbsize - bp->b_bufsize;
3001 }
984263bc
MD
3002 if (newbsize < bp->b_bufsize)
3003 bufspacewakeup();
3004 bp->b_bufsize = newbsize; /* actual buffer allocation */
3005 bp->b_bcount = size; /* requested buffer size */
3006 return 1;
3007}
3008
3009/*
3f779080 3010 * biowait:
984263bc
MD
3011 *
3012 * Wait for buffer I/O completion, returning error status. The buffer
10f3fee5
MD
3013 * is left locked on return. B_EINTR is converted into an EINTR error
3014 * and cleared.
3015 *
3016 * NOTE! The original b_cmd is lost on return, since b_cmd will be
3017 * set to BUF_CMD_DONE.
984263bc
MD
3018 */
3019int
c8e4131d 3020biowait(struct buf *bp)
984263bc 3021{
e43a034f 3022 crit_enter();
10f3fee5
MD
3023 while (bp->b_cmd != BUF_CMD_DONE) {
3024 if (bp->b_cmd == BUF_CMD_READ)
377d4740 3025 tsleep(bp, 0, "biord", 0);
984263bc 3026 else
377d4740 3027 tsleep(bp, 0, "biowr", 0);
984263bc 3028 }
e43a034f 3029 crit_exit();
984263bc
MD
3030 if (bp->b_flags & B_EINTR) {
3031 bp->b_flags &= ~B_EINTR;
3032 return (EINTR);
3033 }
3034 if (bp->b_flags & B_ERROR) {
3035 return (bp->b_error ? bp->b_error : EIO);
3036 } else {
3037 return (0);
3038 }
3039}
3040
3041/*
81b5c339
MD
3042 * This associates a tracking count with an I/O. vn_strategy() and
3043 * dev_dstrategy() do this automatically but there are a few cases
3044 * where a vnode or device layer is bypassed when a block translation
3045 * is cached. In such cases bio_start_transaction() may be called on
3046 * the bypassed layers so the system gets an I/O in progress indication
3047 * for those higher layers.
3048 */
3049void
3050bio_start_transaction(struct bio *bio, struct bio_track *track)
3051{
3052 bio->bio_track = track;
3053 atomic_add_int(&track->bk_active, 1);
3054}
3055
3056/*
3057 * Initiate I/O on a vnode.
3058 */
3059void
3060vn_strategy(struct vnode *vp, struct bio *bio)
3061{
3062 struct bio_track *track;
3063
10f3fee5
MD
3064 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3065 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
81b5c339
MD
3066 track = &vp->v_track_read;
3067 else
3068 track = &vp->v_track_write;
3069 bio->bio_track = track;
3070 atomic_add_int(&track->bk_active, 1);
3071 vop_strategy(*vp->v_ops, vp, bio);
3072}
3073
3074
3075/*
3f779080 3076 * biodone:
984263bc
MD
3077 *
3078 * Finish I/O on a buffer, optionally calling a completion function.
3079 * This is usually called from an interrupt so process blocking is
3080 * not allowed.
3081 *
3082 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3083 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3084 * assuming B_INVAL is clear.
3085 *
3086 * For the VMIO case, we set B_CACHE if the op was a read and no
3087 * read error occured, or if the op was a write. B_CACHE is never
3088 * set if the buffer is invalid or otherwise uncacheable.
3089 *
3090 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3091 * initiator to leave B_INVAL set to brelse the buffer out of existance
3092 * in the biodone routine.
3093 */
3094void
81b5c339 3095biodone(struct bio *bio)
984263bc 3096{
81b5c339 3097 struct buf *bp = bio->bio_buf;
10f3fee5 3098 buf_cmd_t cmd;
984263bc 3099
e43a034f 3100 crit_enter();
984263bc 3101
81b5c339
MD
3102 KASSERT(BUF_REFCNTNB(bp) > 0,
3103 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
10f3fee5
MD
3104 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3105 ("biodone: bp %p already done!", bp));
984263bc 3106
984263bc
MD
3107 runningbufwakeup(bp);
3108
81b5c339 3109 /*
10f3fee5 3110 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
81b5c339
MD
3111 */
3112 while (bio) {
3113 biodone_t *done_func;
3114 struct bio_track *track;
984263bc 3115
81b5c339
MD
3116 /*
3117 * BIO tracking. Most but not all BIOs are tracked.
3118 */
3119 if ((track = bio->bio_track) != NULL) {
3120 atomic_subtract_int(&track->bk_active, 1);
3121 if (track->bk_active < 0) {
3122 panic("biodone: bad active count bio %p\n",
3123 bio);
3124 }
3125 if (track->bk_waitflag) {
3126 track->bk_waitflag = 0;
3127 wakeup(track);
3128 }
3129 bio->bio_track = NULL;
3130 }
3131
3132 /*
3133 * A bio_done function terminates the loop. The function
3134 * will be responsible for any further chaining and/or
3135 * buffer management.
10f3fee5
MD
3136 *
3137 * WARNING! The done function can deallocate the buffer!
81b5c339
MD
3138 */
3139 if ((done_func = bio->bio_done) != NULL) {
3140 bio->bio_done = NULL;
3141 done_func(bio);
3142 crit_exit();
3143 return;
3144 }
3145 bio = bio->bio_prev;
984263bc
MD
3146 }
3147
10f3fee5
MD
3148 cmd = bp->b_cmd;
3149 bp->b_cmd = BUF_CMD_DONE;
3150
81b5c339 3151 /*
10f3fee5 3152 * Only reads and writes are processed past this point.
81b5c339 3153 */
10f3fee5 3154 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
78a9b77f
MD
3155 if (cmd == BUF_CMD_FREEBLKS)
3156 bp->b_flags |= B_NOCACHE;
81b5c339 3157 brelse(bp);
e43a034f 3158 crit_exit();
984263bc
MD
3159 return;
3160 }
81b5c339 3161
69f8c926 3162 /*
78a9b77f
MD
3163 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3164 * a lot worse. XXX - move this above the clearing of b_cmd
69f8c926 3165 */
408357d8
MD
3166 if (LIST_FIRST(&bp->b_dep) != NULL)
3167 buf_complete(bp);
984263bc 3168
78a9b77f
MD
3169 /*
3170 * A failed write must re-dirty the buffer unless B_INVAL
3171 * was set.
3172 */
3173 if (cmd == BUF_CMD_WRITE &&
3174 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3175 bp->b_flags &= ~B_NOCACHE;
3176 bdirty(bp);
3177 }
3178
3179
984263bc
MD
3180 if (bp->b_flags & B_VMIO) {
3181 int i;
3182 vm_ooffset_t foff;
3183 vm_page_t m;
3184 vm_object_t obj;
3185 int iosize;
3186 struct vnode *vp = bp->b_vp;
3187
7540ab49 3188 obj = vp->v_object;
984263bc
MD
3189
3190#if defined(VFS_BIO_DEBUG)
3c37c940 3191 if (vp->v_auxrefs == 0)
8a6722ed 3192 panic("biodone: zero vnode hold count");
7540ab49 3193 if ((vp->v_flag & VOBJBUF) == 0)
984263bc 3194 panic("biodone: vnode is not setup for merged cache");
984263bc
MD
3195#endif
3196
81b5c339
MD
3197 foff = bp->b_loffset;
3198 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
7540ab49 3199 KASSERT(obj != NULL, ("biodone: missing VM object"));
984263bc 3200
984263bc 3201#if defined(VFS_BIO_DEBUG)
54f51aeb 3202 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
6ea70f76 3203 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
54f51aeb 3204 obj->paging_in_progress, bp->b_xio.xio_npages);
984263bc
MD
3205 }
3206#endif
3207
3208 /*
3209 * Set B_CACHE if the op was a normal read and no error
3210 * occured. B_CACHE is set for writes in the b*write()
3211 * routines.
3212 */
3213 iosize = bp->b_bcount - bp->b_resid;
10f3fee5 3214 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
984263bc
MD
3215 bp->b_flags |= B_CACHE;
3216 }
3217
54f51aeb 3218 for (i = 0; i < bp->b_xio.xio_npages; i++) {
984263bc
MD
3219 int bogusflag = 0;
3220 int resid;
3221
3222 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3223 if (resid > iosize)
3224 resid = iosize;
3225
3226 /*
06ecca5a
MD
3227 * cleanup bogus pages, restoring the originals. Since
3228 * the originals should still be wired, we don't have
3229 * to worry about interrupt/freeing races destroying
3230 * the VM object association.
984263bc 3231 */
54f51aeb 3232 m = bp->b_xio.xio_pages[i];
984263bc
MD
3233 if (m == bogus_page) {
3234 bogusflag = 1;
3235 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3236 if (m == NULL)
3237 panic("biodone: page disappeared");
54f51aeb
HP
3238 bp->b_xio.xio_pages[i] = m;
3239 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3240 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
984263bc
MD
3241 }
3242#if defined(VFS_BIO_DEBUG)
3243 if (OFF_TO_IDX(foff) != m->pindex) {
6ea70f76 3244 kprintf(
984263bc
MD
3245"biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3246 (unsigned long)foff, m->pindex);
3247 }
3248#endif
3249
3250 /*
3251 * In the write case, the valid and clean bits are
3252 * already changed correctly ( see bdwrite() ), so we
3253 * only need to do this here in the read case.
3254 */
10f3fee5 3255 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
984263bc
MD
3256 vfs_page_set_valid(bp, foff, i, m);
3257 }
3258 vm_page_flag_clear(m, PG_ZERO);
3259
3260 /*
3261 * when debugging new filesystems or buffer I/O methods, this
3262 * is the most common error that pops up. if you see this, you
3263 * have not set the page busy flag correctly!!!
3264 */
3265 if (m->busy == 0) {
6ea70f76 3266 kprintf("biodone: page busy < 0, "
984263bc
MD
3267 "pindex: %d, foff: 0x(%x,%x), "
3268 "resid: %d, index: %d\n",
3269 (int) m->pindex, (int)(foff >> 32),
3270 (int) foff & 0xffffffff, resid, i);
3271 if (!vn_isdisk(vp, NULL))
6ea70f76 3272 kprintf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
984263bc 3273 bp->b_vp->v_mount->mnt_stat.f_iosize,
54078292 3274 bp->b_loffset,
54f51aeb 3275 bp->b_flags, bp->b_xio.xio_npages);
984263bc 3276 else
6ea70f76 3277 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
54078292 3278 bp->b_loffset,
54f51aeb 3279 bp->b_flags, bp->b_xio.xio_npages);
6ea70f76 3280 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
984263bc 3281 m->valid, m->dirty, m->wire_count);
fc92d4aa 3282 panic("biodone: page busy < 0");
984263bc
MD
3283 }
3284 vm_page_io_finish(m);
3285 vm_object_pip_subtract(obj, 1);
3286 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3287 iosize -= resid;
3288 }
3289 if (obj)
3290 vm_object_pip_wakeupn(obj, 0);
3291 }
3292
3293 /*
3294 * For asynchronous completions, release the buffer now. The brelse
3295 * will do a wakeup there if necessary - so no need to do a wakeup
3296 * here in the async case. The sync case always needs to do a wakeup.
3297 */
3298
3299 if (bp->b_flags & B_ASYNC) {
3300 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3301 brelse(bp);
3302 else
3303 bqrelse(bp);
3304 } else {
3305 wakeup(bp);
3306 }
e43a034f 3307 crit_exit();
984263bc
MD
3308}
3309
3310/*
3f779080
HP
3311 * vfs_unbusy_pages:
3312 *
3313 * This routine is called in lieu of iodone in the case of
3314 * incomplete I/O. This keeps the busy status for pages
3315 * consistant.
984263bc
MD
3316 */
3317void
493c516a 3318vfs_unbusy_pages(struct buf *bp)
984263bc
MD
3319{
3320 int i;
3321
3322 runningbufwakeup(bp);
3323 if (bp->b_flags & B_VMIO) {
3324 struct vnode *vp = bp->b_vp;
3325 vm_object_t obj;
3326
7540ab49 3327 obj = vp->v_object;
984263bc 3328
54f51aeb
HP
3329 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3330 vm_page_t m = bp->b_xio.xio_pages[i];
984263bc 3331
06ecca5a
MD
3332 /*
3333 * When restoring bogus changes the original pages
3334 * should still be wired, so we are in no danger of
3335 * losing the object association and do not need
e43a034f 3336 * critical section protection particularly.
06ecca5a 3337 */
984263bc 3338 if (m == bogus_page) {
81b5c339 3339 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
984263bc 3340 if (!m) {
fc92d4aa 3341 panic("vfs_unbusy_pages: page missing");
984263bc 3342 }
54f51aeb
HP
3343 bp->b_xio.xio_pages[i] = m;
3344 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3345 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
984263bc
MD
3346 }
3347 vm_object_pip_subtract(obj, 1);
3348 vm_page_flag_clear(m, PG_ZERO);
3349 vm_page_io_finish(m);
3350 }
3351 vm_object_pip_wakeupn(obj, 0);
3352 }
3353}
3354
3355/*
3356 * vfs_page_set_valid:
3357 *
3358 * Set the valid bits in a page based on the supplied offset. The
3359 * range is restricted to the buffer's size.
3360 *
3361 * This routine is typically called after a read completes.
3362 */
3363static void
3364vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3365{
3366 vm_ooffset_t soff, eoff;
3367
3368 /*
3369 * Start and end offsets in buffer. eoff - soff may not cross a
3370 * page boundry or cross the end of the buffer. The end of the
3371 * buffer, in this case, is our file EOF, not the allocation size
3372 * of the buffer.
3373 */
3374 soff = off;
3375 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
81b5c339
MD
3376 if (eoff > bp->b_loffset + bp->b_bcount)
3377 eoff = bp->b_loffset + bp->b_bcount;
984263bc
MD
3378
3379 /*
3380 * Set valid range. This is typically the entire buffer and thus the
3381 * entire page.
3382 */
3383 if (eoff > soff) {
3384 vm_page_set_validclean(
3385 m,
3386 (vm_offset_t) (soff & PAGE_MASK),
3387 (vm_offset_t) (eoff - soff)
3388 );
3389 }
3390}
3391
3392/*
3f779080
HP
3393 * vfs_busy_pages:
3394 *
3395 * This routine is called before a device strategy routine.
3396 * It is used to tell the VM system that paging I/O is in
3397 * progress, and treat the pages associated with the buffer
3398 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3399 * flag is handled to make sure that the object doesn't become
3400 * inconsistant.
3401 *
3402 * Since I/O has not been initiated yet, certain buffer flags
3403 * such as B_ERROR or B_INVAL may be in an inconsistant state
3404 * and should be ignored.
984263bc
MD
3405 */
3406void
10f3fee5 3407vfs_busy_pages(struct vnode *vp, struct buf *bp)
984263bc
MD
3408{
3409 int i, bogus;
fde7ac71 3410 struct lwp *lp = curthread->td_lwp;
984263bc 3411
a8f169e2 3412 /*
10f3fee5
MD
3413 * The buffer's I/O command must already be set. If reading,
3414 * B_CACHE must be 0 (double check against callers only doing
3415 * I/O when B_CACHE is 0).
a8f169e2 3416 */
10f3fee5
MD
3417 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3418 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
a8f169e2 3419
984263bc 3420 if (bp->b_flags & B_VMIO) {
984263bc
MD
3421 vm_object_t obj;
3422 vm_ooffset_t foff;
3423
7540ab49 3424 obj = vp->v_object;
81b5c339
MD
3425 foff = bp->b_loffset;
3426 KASSERT(bp->b_loffset != NOOFFSET,
3427 ("vfs_busy_pages: no buffer offset"));
984263bc
MD
3428 vfs_setdirty(bp);
3429
17cde63e
MD
3430 /*
3431 * Loop until none of the pages are busy.
3432 */
984263bc 3433retry:
54f51aeb
HP
3434 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3435 vm_page_t m = bp->b_xio.xio_pages[i];
17cde63e 3436
984263bc
MD
3437 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3438 goto retry;
3439 }
3440
17cde63e
MD
3441 /*
3442 * Setup for I/O, soft-busy the page right now because
3443 * the next loop may block.
3444 */
54f51aeb
HP
3445 for (i = 0; i < bp-