FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_bio.c
1 /*
2 * Copyright (c) 1994,1997 John S. Dyson
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
12 * John S. Dyson.
13 *
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15 */
16
17 /*
18 * this file contains a new buffer I/O scheme implementing a coherent
19 * VM object and buffer cache scheme. Pains have been taken to make
20 * sure that the performance degradation associated with schemes such
21 * as this is not realized.
22 *
23 * Author: John S. Dyson
24 * Significant help during the development and debugging phases
25 * had been provided by David Greenman, also of the FreeBSD core team.
26 *
27 * see man buf(9) for more info.
28 */
29
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/buf.h>
33 #include <sys/conf.h>
34 #include <sys/devicestat.h>
35 #include <sys/eventhandler.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/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>
47 #include <sys/dsched.h>
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>
56 #include <vm/vm_pager.h>
57 #include <vm/swap_pager.h>
58
59 #include <sys/buf2.h>
60 #include <sys/thread2.h>
61 #include <sys/spinlock2.h>
62 #include <sys/mplock2.h>
63 #include <vm/vm_page2.h>
64
65 #include "opt_ddb.h"
66 #ifdef DDB
67 #include <ddb/ddb.h>
68 #endif
69
70 /*
71 * Buffer queues.
72 */
73 enum bufq_type {
74 BQUEUE_NONE, /* not on any queue */
75 BQUEUE_LOCKED, /* locked buffers */
76 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
77 BQUEUE_DIRTY, /* B_DELWRI buffers */
78 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
79 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
80 BQUEUE_EMPTY, /* empty buffer headers */
81
82 BUFFER_QUEUES /* number of buffer queues */
83 };
84
85 typedef enum bufq_type bufq_type_t;
86
87 #define BD_WAKE_SIZE 16384
88 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
89
90 TAILQ_HEAD(bqueues, buf);
91
92 struct bufpcpu {
93 struct spinlock spin;
94 struct bqueues bufqueues[BUFFER_QUEUES];
95 } __cachealign;
96
97 struct bufpcpu bufpcpu[MAXCPU];
98
99 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
100
101 struct buf *buf; /* buffer header pool */
102
103 static void vfs_clean_pages(struct buf *bp);
104 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
105 #if 0
106 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
107 #endif
108 static void vfs_vmio_release(struct buf *bp);
109 static int flushbufqueues(struct buf *marker, bufq_type_t q);
110 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj,
111 vm_pindex_t pg, int deficit);
112
113 static void bd_signal(long totalspace);
114 static void buf_daemon(void);
115 static void buf_daemon_hw(void);
116
117 /*
118 * bogus page -- for I/O to/from partially complete buffers
119 * this is a temporary solution to the problem, but it is not
120 * really that bad. it would be better to split the buffer
121 * for input in the case of buffers partially already in memory,
122 * but the code is intricate enough already.
123 */
124 vm_page_t bogus_page;
125
126 /*
127 * These are all static, but make the ones we export globals so we do
128 * not need to use compiler magic.
129 */
130 long bufspace; /* locked by buffer_map */
131 long maxbufspace;
132 static long bufmallocspace; /* atomic ops */
133 long maxbufmallocspace, lobufspace, hibufspace;
134 static long bufreusecnt, bufdefragcnt, buffreekvacnt;
135 static long lorunningspace;
136 static long hirunningspace;
137 static long dirtykvaspace; /* atomic */
138 static long dirtybufspace; /* atomic */
139 static long dirtybufcount; /* atomic */
140 static long dirtybufspacehw; /* atomic */
141 static long dirtybufcounthw; /* atomic */
142 static long runningbufspace; /* atomic */
143 static long runningbufcount; /* atomic */
144 long lodirtybufspace;
145 long hidirtybufspace;
146 static int getnewbufcalls;
147 static int getnewbufrestarts;
148 static int recoverbufcalls;
149 static int needsbuffer; /* atomic */
150 static int runningbufreq; /* atomic */
151 static int bd_request; /* atomic */
152 static int bd_request_hw; /* atomic */
153 static u_int bd_wake_ary[BD_WAKE_SIZE];
154 static u_int bd_wake_index;
155 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
156 static int debug_commit;
157 static int debug_bufbio;
158
159 static struct thread *bufdaemon_td;
160 static struct thread *bufdaemonhw_td;
161 static u_int lowmempgallocs;
162 static u_int lowmempgfails;
163
164 /*
165 * Sysctls for operational control of the buffer cache.
166 */
167 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
168 "Number of dirty buffers to flush before bufdaemon becomes inactive");
169 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
170 "High watermark used to trigger explicit flushing of dirty buffers");
171 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
172 "Minimum amount of buffer space required for active I/O");
173 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
174 "Maximum amount of buffer space to usable for active I/O");
175 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
176 "Page allocations done during periods of very low free memory");
177 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0,
178 "Page allocations which failed during periods of very low free memory");
179 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
180 "Recycle pages to active or inactive queue transition pt 0-64");
181 /*
182 * Sysctls determining current state of the buffer cache.
183 */
184 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
185 "Total number of buffers in buffer cache");
186 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0,
187 "KVA reserved by dirty buffers (all)");
188 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
189 "Pending bytes of dirty buffers (all)");
190 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
191 "Pending bytes of dirty buffers (heavy weight)");
192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
193 "Pending number of dirty buffers");
194 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
195 "Pending number of dirty buffers (heavy weight)");
196 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
197 "I/O bytes currently in progress due to asynchronous writes");
198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
199 "I/O buffers currently in progress due to asynchronous writes");
200 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
201 "Hard limit on maximum amount of memory usable for buffer space");
202 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
203 "Soft limit on maximum amount of memory usable for buffer space");
204 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
205 "Minimum amount of memory to reserve for system buffer space");
206 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
207 "Amount of memory available for buffers");
208 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
209 0, "Maximum amount of memory reserved for buffers using malloc");
210 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
211 "Amount of memory left for buffers using malloc-scheme");
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
213 "New buffer header acquisition requests");
214 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
215 0, "New buffer header acquisition restarts");
216 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
217 "Recover VM space in an emergency");
218 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
219 "Buffer acquisition restarts due to fragmented buffer map");
220 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
221 "Amount of time KVA space was deallocated in an arbitrary buffer");
222 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
223 "Amount of time buffer re-use operations were successful");
224 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
225 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, "");
226 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
227 "sizeof(struct buf)");
228
229 char *buf_wmesg = BUF_WMESG;
230
231 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
232 #define VFS_BIO_NEED_UNUSED02 0x02
233 #define VFS_BIO_NEED_UNUSED04 0x04
234 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
235
236 /*
237 * bufspacewakeup:
238 *
239 * Called when buffer space is potentially available for recovery.
240 * getnewbuf() will block on this flag when it is unable to free
241 * sufficient buffer space. Buffer space becomes recoverable when
242 * bp's get placed back in the queues.
243 */
244 static __inline void
245 bufspacewakeup(void)
246 {
247 /*
248 * If someone is waiting for BUF space, wake them up. Even
249 * though we haven't freed the kva space yet, the waiting
250 * process will be able to now.
251 */
252 for (;;) {
253 int flags = needsbuffer;
254 cpu_ccfence();
255 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0)
256 break;
257 if (atomic_cmpset_int(&needsbuffer, flags,
258 flags & ~VFS_BIO_NEED_BUFSPACE)) {
259 wakeup(&needsbuffer);
260 break;
261 }
262 /* retry */
263 }
264 }
265
266 /*
267 * runningbufwakeup:
268 *
269 * Accounting for I/O in progress.
270 *
271 */
272 static __inline void
273 runningbufwakeup(struct buf *bp)
274 {
275 long totalspace;
276 long limit;
277 long flags;
278
279 if ((totalspace = bp->b_runningbufspace) != 0) {
280 atomic_add_long(&runningbufspace, -totalspace);
281 atomic_add_long(&runningbufcount, -1);
282 bp->b_runningbufspace = 0;
283
284 /*
285 * see waitrunningbufspace() for limit test.
286 */
287 limit = hirunningspace * 3 / 6;
288 for (;;) {
289 flags = runningbufreq;
290 cpu_ccfence();
291 if (flags == 0)
292 break;
293 if (atomic_cmpset_int(&runningbufreq, flags, 0)) {
294 wakeup(&runningbufreq);
295 break;
296 }
297 /* retry */
298 }
299 bd_signal(totalspace);
300 }
301 }
302
303 /*
304 * bufcountwakeup:
305 *
306 * Called when a buffer has been added to one of the free queues to
307 * account for the buffer and to wakeup anyone waiting for free buffers.
308 * This typically occurs when large amounts of metadata are being handled
309 * by the buffer cache ( else buffer space runs out first, usually ).
310 */
311 static __inline void
312 bufcountwakeup(void)
313 {
314 long flags;
315
316 for (;;) {
317 flags = needsbuffer;
318 if (flags == 0)
319 break;
320 if (atomic_cmpset_int(&needsbuffer, flags,
321 (flags & ~VFS_BIO_NEED_ANY))) {
322 wakeup(&needsbuffer);
323 break;
324 }
325 /* retry */
326 }
327 }
328
329 /*
330 * waitrunningbufspace()
331 *
332 * If runningbufspace exceeds 4/6 hirunningspace we block until
333 * runningbufspace drops to 3/6 hirunningspace. We also block if another
334 * thread blocked here in order to be fair, even if runningbufspace
335 * is now lower than the limit.
336 *
337 * The caller may be using this function to block in a tight loop, we
338 * must block while runningbufspace is greater than at least
339 * hirunningspace * 3 / 6.
340 */
341 void
342 waitrunningbufspace(void)
343 {
344 long limit = hirunningspace * 4 / 6;
345 long flags;
346
347 while (runningbufspace > limit || runningbufreq) {
348 tsleep_interlock(&runningbufreq, 0);
349 flags = atomic_fetchadd_int(&runningbufreq, 1);
350 if (runningbufspace > limit || flags)
351 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz);
352 }
353 }
354
355 /*
356 * buf_dirty_count_severe:
357 *
358 * Return true if we have too many dirty buffers.
359 */
360 int
361 buf_dirty_count_severe(void)
362 {
363 return (runningbufspace + dirtykvaspace >= hidirtybufspace ||
364 dirtybufcount >= nbuf / 2);
365 }
366
367 /*
368 * Return true if the amount of running I/O is severe and BIOQ should
369 * start bursting.
370 */
371 int
372 buf_runningbufspace_severe(void)
373 {
374 return (runningbufspace >= hirunningspace * 4 / 6);
375 }
376
377 /*
378 * vfs_buf_test_cache:
379 *
380 * Called when a buffer is extended. This function clears the B_CACHE
381 * bit if the newly extended portion of the buffer does not contain
382 * valid data.
383 *
384 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
385 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
386 * them while a clean buffer was present.
387 */
388 static __inline__
389 void
390 vfs_buf_test_cache(struct buf *bp,
391 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
392 vm_page_t m)
393 {
394 if (bp->b_flags & B_CACHE) {
395 int base = (foff + off) & PAGE_MASK;
396 if (vm_page_is_valid(m, base, size) == 0)
397 bp->b_flags &= ~B_CACHE;
398 }
399 }
400
401 /*
402 * bd_speedup()
403 *
404 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
405 * low water mark.
406 */
407 static __inline__
408 void
409 bd_speedup(void)
410 {
411 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2)
412 return;
413
414 if (bd_request == 0 &&
415 (dirtykvaspace > lodirtybufspace / 2 ||
416 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
417 if (atomic_fetchadd_int(&bd_request, 1) == 0)
418 wakeup(&bd_request);
419 }
420 if (bd_request_hw == 0 &&
421 (dirtykvaspace > lodirtybufspace / 2 ||
422 dirtybufcounthw >= nbuf / 2)) {
423 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0)
424 wakeup(&bd_request_hw);
425 }
426 }
427
428 /*
429 * bd_heatup()
430 *
431 * Get the buf_daemon heated up when the number of running and dirty
432 * buffers exceeds the mid-point.
433 *
434 * Return the total number of dirty bytes past the second mid point
435 * as a measure of how much excess dirty data there is in the system.
436 */
437 long
438 bd_heatup(void)
439 {
440 long mid1;
441 long mid2;
442 long totalspace;
443
444 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
445
446 totalspace = runningbufspace + dirtykvaspace;
447 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
448 bd_speedup();
449 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
450 if (totalspace >= mid2)
451 return(totalspace - mid2);
452 }
453 return(0);
454 }
455
456 /*
457 * bd_wait()
458 *
459 * Wait for the buffer cache to flush (totalspace) bytes worth of
460 * buffers, then return.
461 *
462 * Regardless this function blocks while the number of dirty buffers
463 * exceeds hidirtybufspace.
464 */
465 void
466 bd_wait(long totalspace)
467 {
468 u_int i;
469 u_int j;
470 u_int mi;
471 int count;
472
473 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
474 return;
475
476 while (totalspace > 0) {
477 bd_heatup();
478
479 /*
480 * Order is important. Suppliers adjust bd_wake_index after
481 * updating runningbufspace/dirtykvaspace. We want to fetch
482 * bd_wake_index before accessing. Any error should thus
483 * be in our favor.
484 */
485 i = atomic_fetchadd_int(&bd_wake_index, 0);
486 if (totalspace > runningbufspace + dirtykvaspace)
487 totalspace = runningbufspace + dirtykvaspace;
488 count = totalspace / BKVASIZE;
489 if (count >= BD_WAKE_SIZE / 2)
490 count = BD_WAKE_SIZE / 2;
491 i = i + count;
492 mi = i & BD_WAKE_MASK;
493
494 /*
495 * This is not a strict interlock, so we play a bit loose
496 * with locking access to dirtybufspace*. We have to re-check
497 * bd_wake_index to ensure that it hasn't passed us.
498 */
499 tsleep_interlock(&bd_wake_ary[mi], 0);
500 atomic_add_int(&bd_wake_ary[mi], 1);
501 j = atomic_fetchadd_int(&bd_wake_index, 0);
502 if ((int)(i - j) >= 0)
503 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz);
504
505 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace;
506 }
507 }
508
509 /*
510 * bd_signal()
511 *
512 * This function is called whenever runningbufspace or dirtykvaspace
513 * is reduced. Track threads waiting for run+dirty buffer I/O
514 * complete.
515 */
516 static void
517 bd_signal(long totalspace)
518 {
519 u_int i;
520
521 if (totalspace > 0) {
522 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
523 totalspace = BKVASIZE * BD_WAKE_SIZE;
524 while (totalspace > 0) {
525 i = atomic_fetchadd_int(&bd_wake_index, 1);
526 i &= BD_WAKE_MASK;
527 if (atomic_readandclear_int(&bd_wake_ary[i]))
528 wakeup(&bd_wake_ary[i]);
529 totalspace -= BKVASIZE;
530 }
531 }
532 }
533
534 /*
535 * BIO tracking support routines.
536 *
537 * Release a ref on a bio_track. Wakeup requests are atomically released
538 * along with the last reference so bk_active will never wind up set to
539 * only 0x80000000.
540 */
541 static
542 void
543 bio_track_rel(struct bio_track *track)
544 {
545 int active;
546 int desired;
547
548 /*
549 * Shortcut
550 */
551 active = track->bk_active;
552 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
553 return;
554
555 /*
556 * Full-on. Note that the wait flag is only atomically released on
557 * the 1->0 count transition.
558 *
559 * We check for a negative count transition using bit 30 since bit 31
560 * has a different meaning.
561 */
562 for (;;) {
563 desired = (active & 0x7FFFFFFF) - 1;
564 if (desired)
565 desired |= active & 0x80000000;
566 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
567 if (desired & 0x40000000)
568 panic("bio_track_rel: bad count: %p", track);
569 if (active & 0x80000000)
570 wakeup(track);
571 break;
572 }
573 active = track->bk_active;
574 }
575 }
576
577 /*
578 * Wait for the tracking count to reach 0.
579 *
580 * Use atomic ops such that the wait flag is only set atomically when
581 * bk_active is non-zero.
582 */
583 int
584 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
585 {
586 int active;
587 int desired;
588 int error;
589
590 /*
591 * Shortcut
592 */
593 if (track->bk_active == 0)
594 return(0);
595
596 /*
597 * Full-on. Note that the wait flag may only be atomically set if
598 * the active count is non-zero.
599 *
600 * NOTE: We cannot optimize active == desired since a wakeup could
601 * clear active prior to our tsleep_interlock().
602 */
603 error = 0;
604 while ((active = track->bk_active) != 0) {
605 cpu_ccfence();
606 desired = active | 0x80000000;
607 tsleep_interlock(track, slp_flags);
608 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
609 error = tsleep(track, slp_flags | PINTERLOCKED,
610 "trwait", slp_timo);
611 if (error)
612 break;
613 }
614 }
615 return (error);
616 }
617
618 /*
619 * bufinit:
620 *
621 * Load time initialisation of the buffer cache, called from machine
622 * dependant initialization code.
623 */
624 static
625 void
626 bufinit(void *dummy __unused)
627 {
628 struct bufpcpu *pcpu;
629 struct buf *bp;
630 vm_offset_t bogus_offset;
631 int i;
632 int j;
633 long n;
634
635 /* next, make a null set of free lists */
636 for (i = 0; i < ncpus; ++i) {
637 pcpu = &bufpcpu[i];
638 spin_init(&pcpu->spin);
639 for (j = 0; j < BUFFER_QUEUES; j++)
640 TAILQ_INIT(&pcpu->bufqueues[j]);
641 }
642
643 /* finally, initialize each buffer header and stick on empty q */
644 i = 0;
645 pcpu = &bufpcpu[i];
646
647 for (n = 0; n < nbuf; n++) {
648 bp = &buf[n];
649 bzero(bp, sizeof *bp);
650 bp->b_flags = B_INVAL; /* we're just an empty header */
651 bp->b_cmd = BUF_CMD_DONE;
652 bp->b_qindex = BQUEUE_EMPTY;
653 bp->b_qcpu = i;
654 initbufbio(bp);
655 xio_init(&bp->b_xio);
656 buf_dep_init(bp);
657 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
658 bp, b_freelist);
659
660 i = (i + 1) % ncpus;
661 pcpu = &bufpcpu[i];
662 }
663
664 /*
665 * maxbufspace is the absolute maximum amount of buffer space we are
666 * allowed to reserve in KVM and in real terms. The absolute maximum
667 * is nominally used by buf_daemon. hibufspace is the nominal maximum
668 * used by most other processes. The differential is required to
669 * ensure that buf_daemon is able to run when other processes might
670 * be blocked waiting for buffer space.
671 *
672 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
673 * this may result in KVM fragmentation which is not handled optimally
674 * by the system.
675 */
676 maxbufspace = nbuf * BKVASIZE;
677 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
678 lobufspace = hibufspace - MAXBSIZE;
679
680 lorunningspace = 512 * 1024;
681 /* hirunningspace -- see below */
682
683 /*
684 * Limit the amount of malloc memory since it is wired permanently
685 * into the kernel space. Even though this is accounted for in
686 * the buffer allocation, we don't want the malloced region to grow
687 * uncontrolled. The malloc scheme improves memory utilization
688 * significantly on average (small) directories.
689 */
690 maxbufmallocspace = hibufspace / 20;
691
692 /*
693 * Reduce the chance of a deadlock occuring by limiting the number
694 * of delayed-write dirty buffers we allow to stack up.
695 *
696 * We don't want too much actually queued to the device at once
697 * (XXX this needs to be per-mount!), because the buffers will
698 * wind up locked for a very long period of time while the I/O
699 * drains.
700 */
701 hidirtybufspace = hibufspace / 2; /* dirty + running */
702 hirunningspace = hibufspace / 16; /* locked & queued to device */
703 if (hirunningspace < 1024 * 1024)
704 hirunningspace = 1024 * 1024;
705
706 dirtykvaspace = 0;
707 dirtybufspace = 0;
708 dirtybufspacehw = 0;
709
710 lodirtybufspace = hidirtybufspace / 2;
711
712 /*
713 * Maximum number of async ops initiated per buf_daemon loop. This is
714 * somewhat of a hack at the moment, we really need to limit ourselves
715 * based on the number of bytes of I/O in-transit that were initiated
716 * from buf_daemon.
717 */
718
719 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
720 vm_object_hold(&kernel_object);
721 bogus_page = vm_page_alloc(&kernel_object,
722 (bogus_offset >> PAGE_SHIFT),
723 VM_ALLOC_NORMAL);
724 vm_object_drop(&kernel_object);
725 vmstats.v_wire_count++;
726
727 }
728
729 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL);
730
731 /*
732 * Initialize the embedded bio structures, typically used by
733 * deprecated code which tries to allocate its own struct bufs.
734 */
735 void
736 initbufbio(struct buf *bp)
737 {
738 bp->b_bio1.bio_buf = bp;
739 bp->b_bio1.bio_prev = NULL;
740 bp->b_bio1.bio_offset = NOOFFSET;
741 bp->b_bio1.bio_next = &bp->b_bio2;
742 bp->b_bio1.bio_done = NULL;
743 bp->b_bio1.bio_flags = 0;
744
745 bp->b_bio2.bio_buf = bp;
746 bp->b_bio2.bio_prev = &bp->b_bio1;
747 bp->b_bio2.bio_offset = NOOFFSET;
748 bp->b_bio2.bio_next = NULL;
749 bp->b_bio2.bio_done = NULL;
750 bp->b_bio2.bio_flags = 0;
751
752 BUF_LOCKINIT(bp);
753 }
754
755 /*
756 * Reinitialize the embedded bio structures as well as any additional
757 * translation cache layers.
758 */
759 void
760 reinitbufbio(struct buf *bp)
761 {
762 struct bio *bio;
763
764 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
765 bio->bio_done = NULL;
766 bio->bio_offset = NOOFFSET;
767 }
768 }
769
770 /*
771 * Undo the effects of an initbufbio().
772 */
773 void
774 uninitbufbio(struct buf *bp)
775 {
776 dsched_exit_buf(bp);
777 BUF_LOCKFREE(bp);
778 }
779
780 /*
781 * Push another BIO layer onto an existing BIO and return it. The new
782 * BIO layer may already exist, holding cached translation data.
783 */
784 struct bio *
785 push_bio(struct bio *bio)
786 {
787 struct bio *nbio;
788
789 if ((nbio = bio->bio_next) == NULL) {
790 int index = bio - &bio->bio_buf->b_bio_array[0];
791 if (index >= NBUF_BIO - 1) {
792 panic("push_bio: too many layers bp %p",
793 bio->bio_buf);
794 }
795 nbio = &bio->bio_buf->b_bio_array[index + 1];
796 bio->bio_next = nbio;
797 nbio->bio_prev = bio;
798 nbio->bio_buf = bio->bio_buf;
799 nbio->bio_offset = NOOFFSET;
800 nbio->bio_done = NULL;
801 nbio->bio_next = NULL;
802 }
803 KKASSERT(nbio->bio_done == NULL);
804 return(nbio);
805 }
806
807 /*
808 * Pop a BIO translation layer, returning the previous layer. The
809 * must have been previously pushed.
810 */
811 struct bio *
812 pop_bio(struct bio *bio)
813 {
814 return(bio->bio_prev);
815 }
816
817 void
818 clearbiocache(struct bio *bio)
819 {
820 while (bio) {
821 bio->bio_offset = NOOFFSET;
822 bio = bio->bio_next;
823 }
824 }
825
826 /*
827 * bfreekva:
828 *
829 * Free the KVA allocation for buffer 'bp'.
830 *
831 * Must be called from a critical section as this is the only locking for
832 * buffer_map.
833 *
834 * Since this call frees up buffer space, we call bufspacewakeup().
835 */
836 static void
837 bfreekva(struct buf *bp)
838 {
839 int count;
840
841 if (bp->b_kvasize) {
842 ++buffreekvacnt;
843 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
844 vm_map_lock(&buffer_map);
845 bufspace -= bp->b_kvasize;
846 vm_map_delete(&buffer_map,
847 (vm_offset_t) bp->b_kvabase,
848 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
849 &count
850 );
851 vm_map_unlock(&buffer_map);
852 vm_map_entry_release(count);
853 bp->b_kvasize = 0;
854 bp->b_kvabase = NULL;
855 bufspacewakeup();
856 }
857 }
858
859 /*
860 * Remove the buffer from the appropriate free list.
861 * (caller must be locked)
862 */
863 static __inline void
864 _bremfree(struct buf *bp)
865 {
866 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
867
868 if (bp->b_qindex != BQUEUE_NONE) {
869 KASSERT(BUF_REFCNTNB(bp) == 1,
870 ("bremfree: bp %p not locked",bp));
871 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist);
872 bp->b_qindex = BQUEUE_NONE;
873 } else {
874 if (BUF_REFCNTNB(bp) <= 1)
875 panic("bremfree: removing a buffer not on a queue");
876 }
877 }
878
879 /*
880 * bremfree() - must be called with a locked buffer
881 */
882 void
883 bremfree(struct buf *bp)
884 {
885 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
886
887 spin_lock(&pcpu->spin);
888 _bremfree(bp);
889 spin_unlock(&pcpu->spin);
890 }
891
892 /*
893 * bremfree_locked - must be called with pcpu->spin locked
894 */
895 static void
896 bremfree_locked(struct buf *bp)
897 {
898 _bremfree(bp);
899 }
900
901 /*
902 * This version of bread issues any required I/O asyncnronously and
903 * makes a callback on completion.
904 *
905 * The callback must check whether BIO_DONE is set in the bio and issue
906 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
907 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
908 */
909 void
910 breadcb(struct vnode *vp, off_t loffset, int size,
911 void (*func)(struct bio *), void *arg)
912 {
913 struct buf *bp;
914
915 bp = getblk(vp, loffset, size, 0, 0);
916
917 /* if not found in cache, do some I/O */
918 if ((bp->b_flags & B_CACHE) == 0) {
919 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
920 bp->b_cmd = BUF_CMD_READ;
921 bp->b_bio1.bio_done = func;
922 bp->b_bio1.bio_caller_info1.ptr = arg;
923 vfs_busy_pages(vp, bp);
924 BUF_KERNPROC(bp);
925 vn_strategy(vp, &bp->b_bio1);
926 } else if (func) {
927 /*
928 * Since we are issuing the callback synchronously it cannot
929 * race the BIO_DONE, so no need for atomic ops here.
930 */
931 /*bp->b_bio1.bio_done = func;*/
932 bp->b_bio1.bio_caller_info1.ptr = arg;
933 bp->b_bio1.bio_flags |= BIO_DONE;
934 func(&bp->b_bio1);
935 } else {
936 bqrelse(bp);
937 }
938 }
939
940 /*
941 * breadnx() - Terminal function for bread() and breadn().
942 *
943 * This function will start asynchronous I/O on read-ahead blocks as well
944 * as satisfy the primary request.
945 *
946 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is
947 * set, the buffer is valid and we do not have to do anything.
948 */
949 int
950 breadnx(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
951 int *rabsize, int cnt, struct buf **bpp)
952 {
953 struct buf *bp, *rabp;
954 int i;
955 int rv = 0, readwait = 0;
956
957 if (*bpp)
958 bp = *bpp;
959 else
960 *bpp = bp = getblk(vp, loffset, size, 0, 0);
961
962 /* if not found in cache, do some I/O */
963 if ((bp->b_flags & B_CACHE) == 0) {
964 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
965 bp->b_cmd = BUF_CMD_READ;
966 bp->b_bio1.bio_done = biodone_sync;
967 bp->b_bio1.bio_flags |= BIO_SYNC;
968 vfs_busy_pages(vp, bp);
969 vn_strategy(vp, &bp->b_bio1);
970 ++readwait;
971 }
972
973 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
974 if (inmem(vp, *raoffset))
975 continue;
976 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
977
978 if ((rabp->b_flags & B_CACHE) == 0) {
979 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
980 rabp->b_cmd = BUF_CMD_READ;
981 vfs_busy_pages(vp, rabp);
982 BUF_KERNPROC(rabp);
983 vn_strategy(vp, &rabp->b_bio1);
984 } else {
985 brelse(rabp);
986 }
987 }
988 if (readwait)
989 rv = biowait(&bp->b_bio1, "biord");
990 return (rv);
991 }
992
993 /*
994 * bwrite:
995 *
996 * Synchronous write, waits for completion.
997 *
998 * Write, release buffer on completion. (Done by iodone
999 * if async). Do not bother writing anything if the buffer
1000 * is invalid.
1001 *
1002 * Note that we set B_CACHE here, indicating that buffer is
1003 * fully valid and thus cacheable. This is true even of NFS
1004 * now so we set it generally. This could be set either here
1005 * or in biodone() since the I/O is synchronous. We put it
1006 * here.
1007 */
1008 int
1009 bwrite(struct buf *bp)
1010 {
1011 int error;
1012
1013 if (bp->b_flags & B_INVAL) {
1014 brelse(bp);
1015 return (0);
1016 }
1017 if (BUF_REFCNTNB(bp) == 0)
1018 panic("bwrite: buffer is not busy???");
1019
1020 /* Mark the buffer clean */
1021 bundirty(bp);
1022
1023 bp->b_flags &= ~(B_ERROR | B_EINTR);
1024 bp->b_flags |= B_CACHE;
1025 bp->b_cmd = BUF_CMD_WRITE;
1026 bp->b_bio1.bio_done = biodone_sync;
1027 bp->b_bio1.bio_flags |= BIO_SYNC;
1028 vfs_busy_pages(bp->b_vp, bp);
1029
1030 /*
1031 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1032 * valid for vnode-backed buffers.
1033 */
1034 bsetrunningbufspace(bp, bp->b_bufsize);
1035 vn_strategy(bp->b_vp, &bp->b_bio1);
1036 error = biowait(&bp->b_bio1, "biows");
1037 brelse(bp);
1038
1039 return (error);
1040 }
1041
1042 /*
1043 * bawrite:
1044 *
1045 * Asynchronous write. Start output on a buffer, but do not wait for
1046 * it to complete. The buffer is released when the output completes.
1047 *
1048 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1049 * B_INVAL buffers. Not us.
1050 */
1051 void
1052 bawrite(struct buf *bp)
1053 {
1054 if (bp->b_flags & B_INVAL) {
1055 brelse(bp);
1056 return;
1057 }
1058 if (BUF_REFCNTNB(bp) == 0)
1059 panic("bwrite: buffer is not busy???");
1060
1061 /* Mark the buffer clean */
1062 bundirty(bp);
1063
1064 bp->b_flags &= ~(B_ERROR | B_EINTR);
1065 bp->b_flags |= B_CACHE;
1066 bp->b_cmd = BUF_CMD_WRITE;
1067 KKASSERT(bp->b_bio1.bio_done == NULL);
1068 vfs_busy_pages(bp->b_vp, bp);
1069
1070 /*
1071 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1072 * valid for vnode-backed buffers.
1073 */
1074 bsetrunningbufspace(bp, bp->b_bufsize);
1075 BUF_KERNPROC(bp);
1076 vn_strategy(bp->b_vp, &bp->b_bio1);
1077 }
1078
1079 /*
1080 * bowrite:
1081 *
1082 * Ordered write. Start output on a buffer, and flag it so that the
1083 * device will write it in the order it was queued. The buffer is
1084 * released when the output completes. bwrite() ( or the VOP routine
1085 * anyway ) is responsible for handling B_INVAL buffers.
1086 */
1087 int
1088 bowrite(struct buf *bp)
1089 {
1090 bp->b_flags |= B_ORDERED;
1091 bawrite(bp);
1092 return (0);
1093 }
1094
1095 /*
1096 * bdwrite:
1097 *
1098 * Delayed write. (Buffer is marked dirty). Do not bother writing
1099 * anything if the buffer is marked invalid.
1100 *
1101 * Note that since the buffer must be completely valid, we can safely
1102 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1103 * biodone() in order to prevent getblk from writing the buffer
1104 * out synchronously.
1105 */
1106 void
1107 bdwrite(struct buf *bp)
1108 {
1109 if (BUF_REFCNTNB(bp) == 0)
1110 panic("bdwrite: buffer is not busy");
1111
1112 if (bp->b_flags & B_INVAL) {
1113 brelse(bp);
1114 return;
1115 }
1116 bdirty(bp);
1117
1118 if (dsched_is_clear_buf_priv(bp))
1119 dsched_new_buf(bp);
1120
1121 /*
1122 * Set B_CACHE, indicating that the buffer is fully valid. This is
1123 * true even of NFS now.
1124 */
1125 bp->b_flags |= B_CACHE;
1126
1127 /*
1128 * This bmap keeps the system from needing to do the bmap later,
1129 * perhaps when the system is attempting to do a sync. Since it
1130 * is likely that the indirect block -- or whatever other datastructure
1131 * that the filesystem needs is still in memory now, it is a good
1132 * thing to do this. Note also, that if the pageout daemon is
1133 * requesting a sync -- there might not be enough memory to do
1134 * the bmap then... So, this is important to do.
1135 */
1136 if (bp->b_bio2.bio_offset == NOOFFSET) {
1137 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1138 NULL, NULL, BUF_CMD_WRITE);
1139 }
1140
1141 /*
1142 * Because the underlying pages may still be mapped and
1143 * writable trying to set the dirty buffer (b_dirtyoff/end)
1144 * range here will be inaccurate.
1145 *
1146 * However, we must still clean the pages to satisfy the
1147 * vnode_pager and pageout daemon, so theythink the pages
1148 * have been "cleaned". What has really occured is that
1149 * they've been earmarked for later writing by the buffer
1150 * cache.
1151 *
1152 * So we get the b_dirtyoff/end update but will not actually
1153 * depend on it (NFS that is) until the pages are busied for
1154 * writing later on.
1155 */
1156 vfs_clean_pages(bp);
1157 bqrelse(bp);
1158
1159 /*
1160 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1161 * due to the softdep code.
1162 */
1163 }
1164
1165 /*
1166 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1167 * This is used by tmpfs.
1168 *
1169 * It is important for any VFS using this routine to NOT use it for
1170 * IO_SYNC or IO_ASYNC operations which occur when the system really
1171 * wants to flush VM pages to backing store.
1172 */
1173 void
1174 buwrite(struct buf *bp)
1175 {
1176 vm_page_t m;
1177 int i;
1178
1179 /*
1180 * Only works for VMIO buffers. If the buffer is already
1181 * marked for delayed-write we can't avoid the bdwrite().
1182 */
1183 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1184 bdwrite(bp);
1185 return;
1186 }
1187
1188 /*
1189 * Mark as needing a commit.
1190 */
1191 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1192 m = bp->b_xio.xio_pages[i];
1193 vm_page_need_commit(m);
1194 }
1195 bqrelse(bp);
1196 }
1197
1198 /*
1199 * bdirty:
1200 *
1201 * Turn buffer into delayed write request by marking it B_DELWRI.
1202 * B_RELBUF and B_NOCACHE must be cleared.
1203 *
1204 * We reassign the buffer to itself to properly update it in the
1205 * dirty/clean lists.
1206 *
1207 * Must be called from a critical section.
1208 * The buffer must be on BQUEUE_NONE.
1209 */
1210 void
1211 bdirty(struct buf *bp)
1212 {
1213 KASSERT(bp->b_qindex == BQUEUE_NONE,
1214 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1215 if (bp->b_flags & B_NOCACHE) {
1216 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1217 bp->b_flags &= ~B_NOCACHE;
1218 }
1219 if (bp->b_flags & B_INVAL) {
1220 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1221 }
1222 bp->b_flags &= ~B_RELBUF;
1223
1224 if ((bp->b_flags & B_DELWRI) == 0) {
1225 lwkt_gettoken(&bp->b_vp->v_token);
1226 bp->b_flags |= B_DELWRI;
1227 reassignbuf(bp);
1228 lwkt_reltoken(&bp->b_vp->v_token);
1229
1230 atomic_add_long(&dirtybufcount, 1);
1231 atomic_add_long(&dirtykvaspace, bp->b_kvasize);
1232 atomic_add_long(&dirtybufspace, bp->b_bufsize);
1233 if (bp->b_flags & B_HEAVY) {
1234 atomic_add_long(&dirtybufcounthw, 1);
1235 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1236 }
1237 bd_heatup();
1238 }
1239 }
1240
1241 /*
1242 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1243 * needs to be flushed with a different buf_daemon thread to avoid
1244 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1245 */
1246 void
1247 bheavy(struct buf *bp)
1248 {
1249 if ((bp->b_flags & B_HEAVY) == 0) {
1250 bp->b_flags |= B_HEAVY;
1251 if (bp->b_flags & B_DELWRI) {
1252 atomic_add_long(&dirtybufcounthw, 1);
1253 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1254 }
1255 }
1256 }
1257
1258 /*
1259 * bundirty:
1260 *
1261 * Clear B_DELWRI for buffer.
1262 *
1263 * Must be called from a critical section.
1264 *
1265 * The buffer is typically on BQUEUE_NONE but there is one case in
1266 * brelse() that calls this function after placing the buffer on
1267 * a different queue.
1268 */
1269 void
1270 bundirty(struct buf *bp)
1271 {
1272 if (bp->b_flags & B_DELWRI) {
1273 lwkt_gettoken(&bp->b_vp->v_token);
1274 bp->b_flags &= ~B_DELWRI;
1275 reassignbuf(bp);
1276 lwkt_reltoken(&bp->b_vp->v_token);
1277
1278 atomic_add_long(&dirtybufcount, -1);
1279 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1280 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1281 if (bp->b_flags & B_HEAVY) {
1282 atomic_add_long(&dirtybufcounthw, -1);
1283 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize);
1284 }
1285 bd_signal(bp->b_bufsize);
1286 }
1287 /*
1288 * Since it is now being written, we can clear its deferred write flag.
1289 */
1290 bp->b_flags &= ~B_DEFERRED;
1291 }
1292
1293 /*
1294 * Set the b_runningbufspace field, used to track how much I/O is
1295 * in progress at any given moment.
1296 */
1297 void
1298 bsetrunningbufspace(struct buf *bp, int bytes)
1299 {
1300 bp->b_runningbufspace = bytes;
1301 if (bytes) {
1302 atomic_add_long(&runningbufspace, bytes);
1303 atomic_add_long(&runningbufcount, 1);
1304 }
1305 }
1306
1307 /*
1308 * brelse:
1309 *
1310 * Release a busy buffer and, if requested, free its resources. The
1311 * buffer will be stashed in the appropriate bufqueue[] allowing it
1312 * to be accessed later as a cache entity or reused for other purposes.
1313 */
1314 void
1315 brelse(struct buf *bp)
1316 {
1317 struct bufpcpu *pcpu;
1318 #ifdef INVARIANTS
1319 int saved_flags = bp->b_flags;
1320 #endif
1321
1322 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1323 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1324
1325 /*
1326 * If B_NOCACHE is set we are being asked to destroy the buffer and
1327 * its backing store. Clear B_DELWRI.
1328 *
1329 * B_NOCACHE is set in two cases: (1) when the caller really wants
1330 * to destroy the buffer and backing store and (2) when the caller
1331 * wants to destroy the buffer and backing store after a write
1332 * completes.
1333 */
1334 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1335 bundirty(bp);
1336 }
1337
1338 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1339 /*
1340 * A re-dirtied buffer is only subject to destruction
1341 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1342 */
1343 /* leave buffer intact */
1344 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1345 (bp->b_bufsize <= 0)) {
1346 /*
1347 * Either a failed read or we were asked to free or not
1348 * cache the buffer. This path is reached with B_DELWRI
1349 * set only if B_INVAL is already set. B_NOCACHE governs
1350 * backing store destruction.
1351 *
1352 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1353 * buffer cannot be immediately freed.
1354 */
1355 bp->b_flags |= B_INVAL;
1356 if (LIST_FIRST(&bp->b_dep) != NULL)
1357 buf_deallocate(bp);
1358 if (bp->b_flags & B_DELWRI) {
1359 atomic_add_long(&dirtybufcount, -1);
1360 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1361 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1362 if (bp->b_flags & B_HEAVY) {
1363 atomic_add_long(&dirtybufcounthw, -1);
1364 atomic_add_long(&dirtybufspacehw,
1365 -bp->b_bufsize);
1366 }
1367 bd_signal(bp->b_bufsize);
1368 }
1369 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1370 }
1371
1372 /*
1373 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1374 * or if b_refs is non-zero.
1375 *
1376 * If vfs_vmio_release() is called with either bit set, the
1377 * underlying pages may wind up getting freed causing a previous
1378 * write (bdwrite()) to get 'lost' because pages associated with
1379 * a B_DELWRI bp are marked clean. Pages associated with a
1380 * B_LOCKED buffer may be mapped by the filesystem.
1381 *
1382 * If we want to release the buffer ourselves (rather then the
1383 * originator asking us to release it), give the originator a
1384 * chance to countermand the release by setting B_LOCKED.
1385 *
1386 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1387 * if B_DELWRI is set.
1388 *
1389 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1390 * on pages to return pages to the VM page queues.
1391 */
1392 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1393 bp->b_flags &= ~B_RELBUF;
1394 } else if (vm_page_count_min(0)) {
1395 if (LIST_FIRST(&bp->b_dep) != NULL)
1396 buf_deallocate(bp); /* can set B_LOCKED */
1397 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1398 bp->b_flags &= ~B_RELBUF;
1399 else
1400 bp->b_flags |= B_RELBUF;
1401 }
1402
1403 /*
1404 * Make sure b_cmd is clear. It may have already been cleared by
1405 * biodone().
1406 *
1407 * At this point destroying the buffer is governed by the B_INVAL
1408 * or B_RELBUF flags.
1409 */
1410 bp->b_cmd = BUF_CMD_DONE;
1411 dsched_exit_buf(bp);
1412
1413 /*
1414 * VMIO buffer rundown. Make sure the VM page array is restored
1415 * after an I/O may have replaces some of the pages with bogus pages
1416 * in order to not destroy dirty pages in a fill-in read.
1417 *
1418 * Note that due to the code above, if a buffer is marked B_DELWRI
1419 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1420 * B_INVAL may still be set, however.
1421 *
1422 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1423 * but not the backing store. B_NOCACHE will destroy the backing
1424 * store.
1425 *
1426 * Note that dirty NFS buffers contain byte-granular write ranges
1427 * and should not be destroyed w/ B_INVAL even if the backing store
1428 * is left intact.
1429 */
1430 if (bp->b_flags & B_VMIO) {
1431 /*
1432 * Rundown for VMIO buffers which are not dirty NFS buffers.
1433 */
1434 int i, j, resid;
1435 vm_page_t m;
1436 off_t foff;
1437 vm_pindex_t poff;
1438 vm_object_t obj;
1439 struct vnode *vp;
1440
1441 vp = bp->b_vp;
1442
1443 /*
1444 * Get the base offset and length of the buffer. Note that
1445 * in the VMIO case if the buffer block size is not
1446 * page-aligned then b_data pointer may not be page-aligned.
1447 * But our b_xio.xio_pages array *IS* page aligned.
1448 *
1449 * block sizes less then DEV_BSIZE (usually 512) are not
1450 * supported due to the page granularity bits (m->valid,
1451 * m->dirty, etc...).
1452 *
1453 * See man buf(9) for more information
1454 */
1455
1456 resid = bp->b_bufsize;
1457 foff = bp->b_loffset;
1458
1459 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1460 m = bp->b_xio.xio_pages[i];
1461 vm_page_flag_clear(m, PG_ZERO);
1462 /*
1463 * If we hit a bogus page, fixup *all* of them
1464 * now. Note that we left these pages wired
1465 * when we removed them so they had better exist,
1466 * and they cannot be ripped out from under us so
1467 * no critical section protection is necessary.
1468 */
1469 if (m == bogus_page) {
1470 obj = vp->v_object;
1471 poff = OFF_TO_IDX(bp->b_loffset);
1472
1473 vm_object_hold(obj);
1474 for (j = i; j < bp->b_xio.xio_npages; j++) {
1475 vm_page_t mtmp;
1476
1477 mtmp = bp->b_xio.xio_pages[j];
1478 if (mtmp == bogus_page) {
1479 mtmp = vm_page_lookup(obj, poff + j);
1480 if (!mtmp) {
1481 panic("brelse: page missing");
1482 }
1483 bp->b_xio.xio_pages[j] = mtmp;
1484 }
1485 }
1486 bp->b_flags &= ~B_HASBOGUS;
1487 vm_object_drop(obj);
1488
1489 if ((bp->b_flags & B_INVAL) == 0) {
1490 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1491 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1492 }
1493 m = bp->b_xio.xio_pages[i];
1494 }
1495
1496 /*
1497 * Invalidate the backing store if B_NOCACHE is set
1498 * (e.g. used with vinvalbuf()). If this is NFS
1499 * we impose a requirement that the block size be
1500 * a multiple of PAGE_SIZE and create a temporary
1501 * hack to basically invalidate the whole page. The
1502 * problem is that NFS uses really odd buffer sizes
1503 * especially when tracking piecemeal writes and
1504 * it also vinvalbuf()'s a lot, which would result
1505 * in only partial page validation and invalidation
1506 * here. If the file page is mmap()'d, however,
1507 * all the valid bits get set so after we invalidate
1508 * here we would end up with weird m->valid values
1509 * like 0xfc. nfs_getpages() can't handle this so
1510 * we clear all the valid bits for the NFS case
1511 * instead of just some of them.
1512 *
1513 * The real bug is the VM system having to set m->valid
1514 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1515 * itself is an artifact of the whole 512-byte
1516 * granular mess that exists to support odd block
1517 * sizes and UFS meta-data block sizes (e.g. 6144).
1518 * A complete rewrite is required.
1519 *
1520 * XXX
1521 */
1522 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1523 int poffset = foff & PAGE_MASK;
1524 int presid;
1525
1526 presid = PAGE_SIZE - poffset;
1527 if (bp->b_vp->v_tag == VT_NFS &&
1528 bp->b_vp->v_type == VREG) {
1529 ; /* entire page */
1530 } else if (presid > resid) {
1531 presid = resid;
1532 }
1533 KASSERT(presid >= 0, ("brelse: extra page"));
1534 vm_page_set_invalid(m, poffset, presid);
1535
1536 /*
1537 * Also make sure any swap cache is removed
1538 * as it is now stale (HAMMER in particular
1539 * uses B_NOCACHE to deal with buffer
1540 * aliasing).
1541 */
1542 swap_pager_unswapped(m);
1543 }
1544 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1545 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1546 }
1547 if (bp->b_flags & (B_INVAL | B_RELBUF))
1548 vfs_vmio_release(bp);
1549 } else {
1550 /*
1551 * Rundown for non-VMIO buffers.
1552 */
1553 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1554 if (bp->b_bufsize)
1555 allocbuf(bp, 0);
1556 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1557 if (bp->b_vp)
1558 brelvp(bp);
1559 }
1560 }
1561
1562 if (bp->b_qindex != BQUEUE_NONE)
1563 panic("brelse: free buffer onto another queue???");
1564 if (BUF_REFCNTNB(bp) > 1) {
1565 /* Temporary panic to verify exclusive locking */
1566 /* This panic goes away when we allow shared refs */
1567 panic("brelse: multiple refs");
1568 /* NOT REACHED */
1569 return;
1570 }
1571
1572 /*
1573 * Figure out the correct queue to place the cleaned up buffer on.
1574 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1575 * disassociated from their vnode.
1576 *
1577 * Return the buffer to its original pcpu area
1578 */
1579 pcpu = &bufpcpu[bp->b_qcpu];
1580 spin_lock(&pcpu->spin);
1581
1582 if (bp->b_flags & B_LOCKED) {
1583 /*
1584 * Buffers that are locked are placed in the locked queue
1585 * immediately, regardless of their state.
1586 */
1587 bp->b_qindex = BQUEUE_LOCKED;
1588 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1589 bp, b_freelist);
1590 } else if (bp->b_bufsize == 0) {
1591 /*
1592 * Buffers with no memory. Due to conditionals near the top
1593 * of brelse() such buffers should probably already be
1594 * marked B_INVAL and disassociated from their vnode.
1595 */
1596 bp->b_flags |= B_INVAL;
1597 KASSERT(bp->b_vp == NULL,
1598 ("bp1 %p flags %08x/%08x vnode %p "
1599 "unexpectededly still associated!",
1600 bp, saved_flags, bp->b_flags, bp->b_vp));
1601 KKASSERT((bp->b_flags & B_HASHED) == 0);
1602 if (bp->b_kvasize) {
1603 bp->b_qindex = BQUEUE_EMPTYKVA;
1604 } else {
1605 bp->b_qindex = BQUEUE_EMPTY;
1606 }
1607 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1608 bp, b_freelist);
1609 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1610 /*
1611 * Buffers with junk contents. Again these buffers had better
1612 * already be disassociated from their vnode.
1613 */
1614 KASSERT(bp->b_vp == NULL,
1615 ("bp2 %p flags %08x/%08x vnode %p unexpectededly "
1616 "still associated!",
1617 bp, saved_flags, bp->b_flags, bp->b_vp));
1618 KKASSERT((bp->b_flags & B_HASHED) == 0);
1619 bp->b_flags |= B_INVAL;
1620 bp->b_qindex = BQUEUE_CLEAN;
1621 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1622 bp, b_freelist);
1623 } else {
1624 /*
1625 * Remaining buffers. These buffers are still associated with
1626 * their vnode.
1627 */
1628 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1629 case B_DELWRI:
1630 bp->b_qindex = BQUEUE_DIRTY;
1631 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1632 bp, b_freelist);
1633 break;
1634 case B_DELWRI | B_HEAVY:
1635 bp->b_qindex = BQUEUE_DIRTY_HW;
1636 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1637 bp, b_freelist);
1638 break;
1639 default:
1640 /*
1641 * NOTE: Buffers are always placed at the end of the
1642 * queue. If B_AGE is not set the buffer will cycle
1643 * through the queue twice.
1644 */
1645 bp->b_qindex = BQUEUE_CLEAN;
1646 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1647 bp, b_freelist);
1648 break;
1649 }
1650 }
1651 spin_unlock(&pcpu->spin);
1652
1653 /*
1654 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1655 * on the correct queue but we have not yet unlocked it.
1656 */
1657 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1658 bundirty(bp);
1659
1660 /*
1661 * The bp is on an appropriate queue unless locked. If it is not
1662 * locked or dirty we can wakeup threads waiting for buffer space.
1663 *
1664 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1665 * if B_INVAL is set ).
1666 */
1667 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1668 bufcountwakeup();
1669
1670 /*
1671 * Something we can maybe free or reuse
1672 */
1673 if (bp->b_bufsize || bp->b_kvasize)
1674 bufspacewakeup();
1675
1676 /*
1677 * Clean up temporary flags and unlock the buffer.
1678 */
1679 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1680 BUF_UNLOCK(bp);
1681 }
1682
1683 /*
1684 * bqrelse:
1685 *
1686 * Release a buffer back to the appropriate queue but do not try to free
1687 * it. The buffer is expected to be used again soon.
1688 *
1689 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1690 * biodone() to requeue an async I/O on completion. It is also used when
1691 * known good buffers need to be requeued but we think we may need the data
1692 * again soon.
1693 *
1694 * XXX we should be able to leave the B_RELBUF hint set on completion.
1695 */
1696 void
1697 bqrelse(struct buf *bp)
1698 {
1699 struct bufpcpu *pcpu;
1700
1701 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1702 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1703
1704 if (bp->b_qindex != BQUEUE_NONE)
1705 panic("bqrelse: free buffer onto another queue???");
1706 if (BUF_REFCNTNB(bp) > 1) {
1707 /* do not release to free list */
1708 panic("bqrelse: multiple refs");
1709 return;
1710 }
1711
1712 buf_act_advance(bp);
1713
1714 pcpu = &bufpcpu[bp->b_qcpu];
1715 spin_lock(&pcpu->spin);
1716
1717 if (bp->b_flags & B_LOCKED) {
1718 /*
1719 * Locked buffers are released to the locked queue. However,
1720 * if the buffer is dirty it will first go into the dirty
1721 * queue and later on after the I/O completes successfully it
1722 * will be released to the locked queue.
1723 */
1724 bp->b_qindex = BQUEUE_LOCKED;
1725 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1726 bp, b_freelist);
1727 } else if (bp->b_flags & B_DELWRI) {
1728 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1729 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1730 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1731 bp, b_freelist);
1732 } else if (vm_page_count_min(0)) {
1733 /*
1734 * We are too low on memory, we have to try to free the
1735 * buffer (most importantly: the wired pages making up its
1736 * backing store) *now*.
1737 */
1738 spin_unlock(&pcpu->spin);
1739 brelse(bp);
1740 return;
1741 } else {
1742 bp->b_qindex = BQUEUE_CLEAN;
1743 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1744 bp, b_freelist);
1745 }
1746 spin_unlock(&pcpu->spin);
1747
1748 /*
1749 * We have now placed the buffer on the proper queue, but have yet
1750 * to unlock it.
1751 */
1752 if ((bp->b_flags & B_LOCKED) == 0 &&
1753 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1754 bufcountwakeup();
1755 }
1756
1757 /*
1758 * Something we can maybe free or reuse.
1759 */
1760 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1761 bufspacewakeup();
1762
1763 /*
1764 * Final cleanup and unlock. Clear bits that are only used while a
1765 * buffer is actively locked.
1766 */
1767 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1768 dsched_exit_buf(bp);
1769 BUF_UNLOCK(bp);
1770 }
1771
1772 /*
1773 * Hold a buffer, preventing it from being reused. This will prevent
1774 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1775 * operations. If a B_INVAL operation occurs the buffer will remain held
1776 * but the underlying pages may get ripped out.
1777 *
1778 * These functions are typically used in VOP_READ/VOP_WRITE functions
1779 * to hold a buffer during a copyin or copyout, preventing deadlocks
1780 * or recursive lock panics when read()/write() is used over mmap()'d
1781 * space.
1782 *
1783 * NOTE: bqhold() requires that the buffer be locked at the time of the
1784 * hold. bqdrop() has no requirements other than the buffer having
1785 * previously been held.
1786 */
1787 void
1788 bqhold(struct buf *bp)
1789 {
1790 atomic_add_int(&bp->b_refs, 1);
1791 }
1792
1793 void
1794 bqdrop(struct buf *bp)
1795 {
1796 KKASSERT(bp->b_refs > 0);
1797 atomic_add_int(&bp->b_refs, -1);
1798 }
1799
1800 /*
1801 * Return backing pages held by the buffer 'bp' back to the VM system.
1802 * This routine is called when the bp is invalidated, released, or
1803 * reused.
1804 *
1805 * The KVA mapping (b_data) for the underlying pages is removed by
1806 * this function.
1807 *
1808 * WARNING! This routine is integral to the low memory critical path
1809 * when a buffer is B_RELBUF'd. If the system has a severe page
1810 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE
1811 * queues so they can be reused in the current pageout daemon
1812 * pass.
1813 */
1814 static void
1815 vfs_vmio_release(struct buf *bp)
1816 {
1817 int i;
1818 vm_page_t m;
1819
1820 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1821 m = bp->b_xio.xio_pages[i];
1822 bp->b_xio.xio_pages[i] = NULL;
1823
1824 /*
1825 * We need to own the page in order to safely unwire it.
1826 */
1827 vm_page_busy_wait(m, FALSE, "vmiopg");
1828
1829 /*
1830 * The VFS is telling us this is not a meta-data buffer
1831 * even if it is backed by a block device.
1832 */
1833 if (bp->b_flags & B_NOTMETA)
1834 vm_page_flag_set(m, PG_NOTMETA);
1835
1836 /*
1837 * This is a very important bit of code. We try to track
1838 * VM page use whether the pages are wired into the buffer
1839 * cache or not. While wired into the buffer cache the
1840 * bp tracks the act_count.
1841 *
1842 * We can choose to place unwired pages on the inactive
1843 * queue (0) or active queue (1). If we place too many
1844 * on the active queue the queue will cycle the act_count
1845 * on pages we'd like to keep, just from single-use pages
1846 * (such as when doing a tar-up or file scan).
1847 */
1848 if (bp->b_act_count < vm_cycle_point)
1849 vm_page_unwire(m, 0);
1850 else
1851 vm_page_unwire(m, 1);
1852
1853 /*
1854 * If the wire_count has dropped to 0 we may need to take
1855 * further action before unbusying the page.
1856 *
1857 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us.
1858 */
1859 if (m->wire_count == 0) {
1860 vm_page_flag_clear(m, PG_ZERO);
1861
1862 if (bp->b_flags & B_DIRECT) {
1863 /*
1864 * Attempt to free the page if B_DIRECT is
1865 * set, the caller does not desire the page
1866 * to be cached.
1867 */
1868 vm_page_wakeup(m);
1869 vm_page_try_to_free(m);
1870 } else if ((bp->b_flags & B_NOTMETA) ||
1871 vm_page_count_min(0)) {
1872 /*
1873 * Attempt to move the page to PQ_CACHE
1874 * if B_NOTMETA is set. This flag is set
1875 * by HAMMER to remove one of the two pages
1876 * present when double buffering is enabled.
1877 *
1878 * Attempt to move the page to PQ_CACHE
1879 * If we have a severe page deficit. This
1880 * will cause buffer cache operations related
1881 * to pageouts to recycle the related pages
1882 * in order to avoid a low memory deadlock.
1883 */
1884 m->act_count = bp->b_act_count;
1885 vm_page_wakeup(m);
1886 vm_page_try_to_cache(m);
1887 } else {
1888 /*
1889 * Nominal case, leave the page on the
1890 * queue the original unwiring placed it on
1891 * (active or inactive).
1892 */
1893 m->act_count = bp->b_act_count;
1894 vm_page_wakeup(m);
1895 }
1896 } else {
1897 vm_page_wakeup(m);
1898 }
1899 }
1900
1901 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
1902 bp->b_xio.xio_npages);
1903 if (bp->b_bufsize) {
1904 bufspacewakeup();
1905 bp->b_bufsize = 0;
1906 }
1907 bp->b_xio.xio_npages = 0;
1908 bp->b_flags &= ~B_VMIO;
1909 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1910 if (bp->b_vp)
1911 brelvp(bp);
1912 }
1913
1914 /*
1915 * Find and initialize a new buffer header, freeing up existing buffers
1916 * in the bufqueues as necessary. The new buffer is returned locked.
1917 *
1918 * Important: B_INVAL is not set. If the caller wishes to throw the
1919 * buffer away, the caller must set B_INVAL prior to calling brelse().
1920 *
1921 * We block if:
1922 * We have insufficient buffer headers
1923 * We have insufficient buffer space
1924 * buffer_map is too fragmented ( space reservation fails )
1925 * If we have to flush dirty buffers ( but we try to avoid this )
1926 *
1927 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1928 * Instead we ask the buf daemon to do it for us. We attempt to
1929 * avoid piecemeal wakeups of the pageout daemon.
1930 */
1931 struct buf *
1932 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1933 {
1934 struct bufpcpu *pcpu;
1935 struct buf *bp;
1936 struct buf *nbp;
1937 int defrag = 0;
1938 int nqindex;
1939 int nqcpu;
1940 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1941 int maxloops = 200000;
1942 int restart_reason = 0;
1943 struct buf *restart_bp = NULL;
1944 static int flushingbufs;
1945
1946 /*
1947 * We can't afford to block since we might be holding a vnode lock,
1948 * which may prevent system daemons from running. We deal with
1949 * low-memory situations by proactively returning memory and running
1950 * async I/O rather then sync I/O.
1951 */
1952
1953 ++getnewbufcalls;
1954 --getnewbufrestarts;
1955 nqcpu = mycpu->gd_cpuid;
1956 restart:
1957 ++getnewbufrestarts;
1958
1959 if (debug_bufbio && --maxloops == 0)
1960 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)",
1961 mycpu->gd_cpuid, restart_reason, restart_bp);
1962
1963 /*
1964 * Setup for scan. If we do not have enough free buffers,
1965 * we setup a degenerate case that immediately fails. Note
1966 * that if we are specially marked process, we are allowed to
1967 * dip into our reserves.
1968 *
1969 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1970 *
1971 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1972 * However, there are a number of cases (defragging, reusing, ...)
1973 * where we cannot backup.
1974 */
1975 pcpu = &bufpcpu[nqcpu];
1976 nqindex = BQUEUE_EMPTYKVA;
1977 spin_lock(&pcpu->spin);
1978
1979 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]);
1980
1981 if (nbp == NULL) {
1982 /*
1983 * If no EMPTYKVA buffers and we are either
1984 * defragging or reusing, locate a CLEAN buffer
1985 * to free or reuse. If bufspace useage is low
1986 * skip this step so we can allocate a new buffer.
1987 */
1988 if (defrag || bufspace >= lobufspace) {
1989 nqindex = BQUEUE_CLEAN;
1990 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]);
1991 }
1992
1993 /*
1994 * If we could not find or were not allowed to reuse a
1995 * CLEAN buffer, check to see if it is ok to use an EMPTY
1996 * buffer. We can only use an EMPTY buffer if allocating
1997 * its KVA would not otherwise run us out of buffer space.
1998 */
1999 if (nbp == NULL && defrag == 0 &&
2000 bufspace + maxsize < hibufspace) {
2001 nqindex = BQUEUE_EMPTY;
2002 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]);
2003 }
2004 }
2005
2006 /*
2007 * Run scan, possibly freeing data and/or kva mappings on the fly
2008 * depending.
2009 *
2010 * WARNING! spin is held!
2011 */
2012 while ((bp = nbp) != NULL) {
2013 int qindex = nqindex;
2014
2015 nbp = TAILQ_NEXT(bp, b_freelist);
2016
2017 /*
2018 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2019 * cycles through the queue twice before being selected.
2020 */
2021 if (qindex == BQUEUE_CLEAN &&
2022 (bp->b_flags & B_AGE) == 0 && nbp) {
2023 bp->b_flags |= B_AGE;
2024 TAILQ_REMOVE(&pcpu->bufqueues[qindex],
2025 bp, b_freelist);
2026 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex],
2027 bp, b_freelist);
2028 continue;
2029 }
2030
2031 /*
2032 * Calculate next bp ( we can only use it if we do not block
2033 * or do other fancy things ).
2034 */
2035 if (nbp == NULL) {
2036 switch(qindex) {
2037 case BQUEUE_EMPTY:
2038 nqindex = BQUEUE_EMPTYKVA;
2039 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA])))
2040 break;
2041 /* fall through */
2042 case BQUEUE_EMPTYKVA:
2043 nqindex = BQUEUE_CLEAN;
2044 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN])))
2045 break;
2046 /* fall through */
2047 case BQUEUE_CLEAN:
2048 /*
2049 * nbp is NULL.
2050 */
2051 break;
2052 }
2053 }
2054
2055 /*
2056 * Sanity Checks
2057 */
2058 KASSERT(bp->b_qindex == qindex,
2059 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2060
2061 /*
2062 * Note: we no longer distinguish between VMIO and non-VMIO
2063 * buffers.
2064 */
2065 KASSERT((bp->b_flags & B_DELWRI) == 0,
2066 ("delwri buffer %p found in queue %d", bp, qindex));
2067
2068 /*
2069 * Do not try to reuse a buffer with a non-zero b_refs.
2070 * This is an unsynchronized test. A synchronized test
2071 * is also performed after we lock the buffer.
2072 */
2073 if (bp->b_refs)
2074 continue;
2075
2076 /*
2077 * If we are defragging then we need a buffer with
2078 * b_kvasize != 0. XXX this situation should no longer
2079 * occur, if defrag is non-zero the buffer's b_kvasize
2080 * should also be non-zero at this point. XXX
2081 */
2082 if (defrag && bp->b_kvasize == 0) {
2083 kprintf("Warning: defrag empty buffer %p\n", bp);
2084 continue;
2085 }
2086
2087 /*
2088 * Start freeing the bp. This is somewhat involved. nbp
2089 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2090 * on the clean list must be disassociated from their
2091 * current vnode. Buffers on the empty[kva] lists have
2092 * already been disassociated.
2093 *
2094 * b_refs is checked after locking along with queue changes.
2095 * We must check here to deal with zero->nonzero transitions
2096 * made by the owner of the buffer lock, which is used by
2097 * VFS's to hold the buffer while issuing an unlocked
2098 * uiomove()s. We cannot invalidate the buffer's pages
2099 * for this case. Once we successfully lock a buffer the
2100 * only 0->1 transitions of b_refs will occur via findblk().
2101 *
2102 * We must also check for queue changes after successful
2103 * locking as the current lock holder may dispose of the
2104 * buffer and change its queue.
2105 */
2106 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2107 spin_unlock(&pcpu->spin);
2108 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2109 restart_reason = 1;
2110 restart_bp = bp;
2111 goto restart;
2112 }
2113 if (bp->b_qindex != qindex || bp->b_refs) {
2114 spin_unlock(&pcpu->spin);
2115 BUF_UNLOCK(bp);
2116 restart_reason = 2;
2117 restart_bp = bp;
2118 goto restart;
2119 }
2120 bremfree_locked(bp);
2121 spin_unlock(&pcpu->spin);
2122
2123 /*
2124 * Dependancies must be handled before we disassociate the
2125 * vnode.
2126 *
2127 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2128 * be immediately disassociated. HAMMER then becomes
2129 * responsible for releasing the buffer.
2130 *
2131 * NOTE: spin is UNLOCKED now.
2132 */
2133 if (LIST_FIRST(&bp->b_dep) != NULL) {
2134 buf_deallocate(bp);
2135 if (bp->b_flags & B_LOCKED) {
2136 bqrelse(bp);
2137 restart_reason = 3;
2138 restart_bp = bp;
2139 goto restart;
2140 }
2141 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2142 }
2143
2144 if (qindex == BQUEUE_CLEAN) {
2145 if (bp->b_flags & B_VMIO)
2146 vfs_vmio_release(bp);
2147 if (bp->b_vp)
2148 brelvp(bp);
2149 }
2150
2151 /*
2152 * NOTE: nbp is now entirely invalid. We can only restart
2153 * the scan from this point on.
2154 *
2155 * Get the rest of the buffer freed up. b_kva* is still
2156 * valid after this operation.
2157 */
2158 KASSERT(bp->b_vp == NULL,
2159 ("bp3 %p flags %08x vnode %p qindex %d "
2160 "unexpectededly still associated!",
2161 bp, bp->b_flags, bp->b_vp, qindex));
2162 KKASSERT((bp->b_flags & B_HASHED) == 0);
2163
2164 /*
2165 * critical section protection is not required when
2166 * scrapping a buffer's contents because it is already
2167 * wired.
2168 */
2169 if (bp->b_bufsize)
2170 allocbuf(bp, 0);
2171
2172 bp->b_flags = B_BNOCLIP;
2173 bp->b_cmd = BUF_CMD_DONE;
2174 bp->b_vp = NULL;
2175 bp->b_error = 0;
2176 bp->b_resid = 0;
2177 bp->b_bcount = 0;
2178 bp->b_xio.xio_npages = 0;
2179 bp->b_dirtyoff = bp->b_dirtyend = 0;
2180 bp->b_act_count = ACT_INIT;
2181 reinitbufbio(bp);
2182 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2183 buf_dep_init(bp);
2184 if (blkflags & GETBLK_BHEAVY)
2185 bp->b_flags |= B_HEAVY;
2186
2187 /*
2188 * If we are defragging then free the buffer.
2189 */
2190 if (defrag) {
2191 bp->b_flags |= B_INVAL;
2192 bfreekva(bp);
2193 brelse(bp);
2194 defrag = 0;
2195 restart_reason = 4;
2196 restart_bp = bp;
2197 goto restart;
2198 }
2199
2200 /*
2201 * If we are overcomitted then recover the buffer and its
2202 * KVM space. This occurs in rare situations when multiple
2203 * processes are blocked in getnewbuf() or allocbuf().
2204 *
2205 * On 64-bit systems BKVASIZE == MAXBSIZE and overcommit
2206 * should not be possible.
2207 */
2208 if (bufspace >= hibufspace)
2209 flushingbufs = 1;
2210 if (BKVASIZE != MAXBSIZE) {
2211 if (flushingbufs && bp->b_kvasize != 0) {
2212 bp->b_flags |= B_INVAL;
2213 bfreekva(bp);
2214 brelse(bp);
2215 restart_reason = 5;
2216 restart_bp = bp;
2217 goto restart;
2218 }
2219 }
2220 if (bufspace < lobufspace)
2221 flushingbufs = 0;
2222
2223 /*
2224 * b_refs can transition to a non-zero value while we hold
2225 * the buffer locked due to a findblk(). Our brelvp() above
2226 * interlocked any future possible transitions due to
2227 * findblk()s.
2228 *
2229 * If we find b_refs to be non-zero we can destroy the
2230 * buffer's contents but we cannot yet reuse the buffer.
2231 */
2232 if (bp->b_refs) {
2233 bp->b_flags |= B_INVAL;
2234 if (BKVASIZE != MAXBSIZE)
2235 bfreekva(bp);
2236 brelse(bp);
2237 restart_reason = 6;
2238 restart_bp = bp;
2239 goto restart;
2240 }
2241 break;
2242 /* NOT REACHED, spin not held */
2243 }
2244
2245 /*
2246 * If we exhausted our list, iterate other cpus. If that fails,
2247 * sleep as appropriate. We may have to wakeup various daemons
2248 * and write out some dirty buffers.
2249 *
2250 * Generally we are sleeping due to insufficient buffer space.
2251 *
2252 * NOTE: spin is held if bp is NULL, else it is not held.
2253 */
2254 if (bp == NULL) {
2255 int flags;
2256 char *waitmsg;
2257
2258 spin_unlock(&pcpu->spin);
2259
2260 nqcpu = (nqcpu + 1) % ncpus;
2261 if (nqcpu != mycpu->gd_cpuid) {
2262 restart_reason = 7;
2263 restart_bp = bp;
2264 goto restart;
2265 }
2266
2267 if (defrag) {
2268 flags = VFS_BIO_NEED_BUFSPACE;
2269 waitmsg = "nbufkv";
2270 } else if (bufspace >= hibufspace) {
2271 waitmsg = "nbufbs";
2272 flags = VFS_BIO_NEED_BUFSPACE;
2273 } else {
2274 waitmsg = "newbuf";
2275 flags = VFS_BIO_NEED_ANY;
2276 }
2277
2278 bd_speedup(); /* heeeelp */
2279 atomic_set_int(&needsbuffer, flags);
2280 while (needsbuffer & flags) {
2281 int value;
2282
2283 tsleep_interlock(&needsbuffer, 0);
2284 value = atomic_fetchadd_int(&needsbuffer, 0);
2285 if (value & flags) {
2286 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags,
2287 waitmsg, slptimeo)) {
2288 return (NULL);
2289 }
2290 }
2291 }
2292 } else {
2293 /*
2294 * We finally have a valid bp. We aren't quite out of the
2295 * woods, we still have to reserve kva space. In order
2296 * to keep fragmentation sane we only allocate kva in
2297 * BKVASIZE chunks.
2298 *
2299 * (spin is not held)
2300 */
2301 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2302
2303 if (maxsize != bp->b_kvasize) {
2304 vm_offset_t addr = 0;
2305 int count;
2306
2307 bfreekva(bp);
2308
2309 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2310 vm_map_lock(&buffer_map);
2311
2312 if (vm_map_findspace(&buffer_map,
2313 vm_map_min(&buffer_map), maxsize,
2314 maxsize, 0, &addr)) {
2315 /*
2316 * Uh oh. Buffer map is too fragmented. We
2317 * must defragment the map.
2318 */
2319 vm_map_unlock(&buffer_map);
2320 vm_map_entry_release(count);
2321 ++bufdefragcnt;
2322 defrag = 1;
2323 bp->b_flags |= B_INVAL;
2324 brelse(bp);
2325 restart_reason = 8;
2326 restart_bp = bp;
2327 goto restart;
2328 }
2329 if (addr) {
2330 vm_map_insert(&buffer_map, &count,
2331 NULL, 0,
2332 addr, addr + maxsize,
2333 VM_MAPTYPE_NORMAL,
2334 VM_PROT_ALL, VM_PROT_ALL,
2335 MAP_NOFAULT);
2336
2337 bp->b_kvabase = (caddr_t) addr;
2338 bp->b_kvasize = maxsize;
2339 bufspace += bp->b_kvasize;
2340 ++bufreusecnt;
2341 }
2342 vm_map_unlock(&buffer_map);
2343 vm_map_entry_release(count);
2344 }
2345 bp->b_data = bp->b_kvabase;
2346 }
2347 return(bp);
2348 }
2349
2350 /*
2351 * buf_daemon:
2352 *
2353 * Buffer flushing daemon. Buffers are normally flushed by the
2354 * update daemon but if it cannot keep up this process starts to
2355 * take the load in an attempt to prevent getnewbuf() from blocking.
2356 *
2357 * Once a flush is initiated it does not stop until the number
2358 * of buffers falls below lodirtybuffers, but we will wake up anyone
2359 * waiting at the mid-point.
2360 */
2361 static struct kproc_desc buf_kp = {
2362 "bufdaemon",
2363 buf_daemon,
2364 &bufdaemon_td
2365 };
2366 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2367 kproc_start, &buf_kp)
2368
2369 static struct kproc_desc bufhw_kp = {
2370 "bufdaemon_hw",
2371 buf_daemon_hw,
2372 &bufdaemonhw_td
2373 };
2374 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2375 kproc_start, &bufhw_kp)
2376
2377 static void
2378 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long),
2379 int *bd_req)
2380 {
2381 long limit;
2382 struct buf *marker;
2383
2384 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO);
2385 marker->b_flags |= B_MARKER;
2386 marker->b_qindex = BQUEUE_NONE;
2387 marker->b_qcpu = 0;
2388
2389 /*
2390 * This process needs to be suspended prior to shutdown sync.
2391 */
2392 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2393 td, SHUTDOWN_PRI_LAST);
2394 curthread->td_flags |= TDF_SYSTHREAD;
2395
2396 /*
2397 * This process is allowed to take the buffer cache to the limit
2398 */
2399 for (;;) {
2400 kproc_suspend_loop();
2401
2402 /*
2403 * Do the flush as long as the number of dirty buffers
2404 * (including those running) exceeds lodirtybufspace.
2405 *
2406 * When flushing limit running I/O to hirunningspace
2407 * Do the flush. Limit the amount of in-transit I/O we
2408 * allow to build up, otherwise we would completely saturate
2409 * the I/O system. Wakeup any waiting processes before we
2410 * normally would so they can run in parallel with our drain.
2411 *
2412 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2413 * but because we split the operation into two threads we
2414 * have to cut it in half for each thread.
2415 */
2416 waitrunningbufspace();
2417 limit = lodirtybufspace / 2;
2418 while (buf_limit_fn(limit)) {
2419 if (flushbufqueues(marker, queue) == 0)
2420 break;
2421 if (runningbufspace < hirunningspace)
2422 continue;
2423 waitrunningbufspace();
2424 }
2425
2426 /*
2427 * We reached our low water mark, reset the
2428 * request and sleep until we are needed again.
2429 * The sleep is just so the suspend code works.
2430 */
2431 tsleep_interlock(bd_req, 0);
2432 if (atomic_swap_int(bd_req, 0) == 0)
2433 tsleep(bd_req, PINTERLOCKED, "psleep", hz);
2434 }
2435 /* NOT REACHED */
2436 /*kfree(marker, M_BIOBUF);*/
2437 }
2438
2439 static int
2440 buf_daemon_limit(long limit)
2441 {
2442 return (runningbufspace + dirtykvaspace > limit ||
2443 dirtybufcount - dirtybufcounthw >= nbuf / 2);
2444 }
2445
2446 static int
2447 buf_daemon_hw_limit(long limit)
2448 {
2449 return (runningbufspace + dirtykvaspace > limit ||
2450 dirtybufcounthw >= nbuf / 2);
2451 }
2452
2453 static void
2454 buf_daemon(void)
2455 {
2456 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit,
2457 &bd_request);
2458 }
2459
2460 static void
2461 buf_daemon_hw(void)
2462 {
2463 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit,
2464 &bd_request_hw);
2465 }
2466
2467 /*
2468 * flushbufqueues:
2469 *
2470 * Try to flush a buffer in the dirty queue. We must be careful to
2471 * free up B_INVAL buffers instead of write them, which NFS is
2472 * particularly sensitive to.
2473 *
2474 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2475 * that we really want to try to get the buffer out and reuse it
2476 * due to the write load on the machine.
2477 *
2478 * We must lock the buffer in order to check its validity before we
2479 * can mess with its contents. spin isn't enough.
2480 */
2481 static int
2482 flushbufqueues(struct buf *marker, bufq_type_t q)
2483 {
2484 struct bufpcpu *pcpu;
2485 struct buf *bp;
2486 int r = 0;
2487 int lcpu = marker->b_qcpu;
2488
2489 KKASSERT(marker->b_qindex == BQUEUE_NONE);
2490 KKASSERT(marker->b_flags & B_MARKER);
2491
2492 again:
2493 /*
2494 * Spinlock needed to perform operations on the queue and may be
2495 * held through a non-blocking BUF_LOCK(), but cannot be held when
2496 * BUF_UNLOCK()ing or through any other major operation.
2497 */
2498 pcpu = &bufpcpu[marker->b_qcpu];
2499 spin_lock(&pcpu->spin);
2500 marker->b_qindex = q;
2501 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist);
2502 bp = marker;
2503
2504 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) {
2505 /*
2506 * NOTE: spinlock is always held at the top of the loop
2507 */
2508 if (bp->b_flags & B_MARKER)
2509 continue;
2510 if ((bp->b_flags & B_DELWRI) == 0) {
2511 kprintf("Unexpected clean buffer %p\n", bp);
2512 continue;
2513 }
2514 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
2515 continue;
2516 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q);
2517
2518 /*
2519 * Once the buffer is locked we will have no choice but to
2520 * unlock the spinlock around a later BUF_UNLOCK and re-set
2521 * bp = marker when looping. Move the marker now to make
2522 * things easier.
2523 */
2524 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2525 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist);
2526
2527 /*
2528 * Must recheck B_DELWRI after successfully locking
2529 * the buffer.
2530 */
2531 if ((bp->b_flags & B_DELWRI) == 0) {
2532 spin_unlock(&pcpu->spin);
2533 BUF_UNLOCK(bp);
2534 spin_lock(&pcpu->spin);
2535 bp = marker;
2536 continue;
2537 }
2538
2539 /*
2540 * Remove the buffer from its queue. We still own the
2541 * spinlock here.
2542 */
2543 _bremfree(bp);
2544
2545 /*
2546 * Disposing of an invalid buffer counts as a flush op
2547 */
2548 if (bp->b_flags & B_INVAL) {
2549 spin_unlock(&pcpu->spin);
2550 brelse(bp);
2551 spin_lock(&pcpu->spin);
2552 ++r;
2553 break;
2554 }
2555
2556 /*
2557 * Release the spinlock for the more complex ops we
2558 * are now going to do.
2559 */
2560 spin_unlock(&pcpu->spin);
2561 lwkt_yield();
2562
2563 /*
2564 * This is a bit messy
2565 */
2566 if (LIST_FIRST(&bp->b_dep) != NULL &&
2567 (bp->b_flags & B_DEFERRED) == 0 &&
2568 buf_countdeps(bp, 0)) {
2569 spin_lock(&pcpu->spin);
2570 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist);
2571 bp->b_qindex = q;
2572 bp->b_flags |= B_DEFERRED;
2573 spin_unlock(&pcpu->spin);
2574 BUF_UNLOCK(bp);
2575 spin_lock(&pcpu->spin);
2576 bp = marker;
2577 continue;
2578 }
2579
2580 /*
2581 * spinlock not held here.
2582 *
2583 * If the buffer has a dependancy, buf_checkwrite() must
2584 * also return 0 for us to be able to initate the write.
2585 *
2586 * If the buffer is flagged B_ERROR it may be requeued
2587 * over and over again, we try to avoid a live lock.
2588 */
2589 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2590 brelse(bp);
2591 } else if (bp->b_flags & B_ERROR) {
2592 tsleep(bp, 0, "bioer", 1);
2593 bp->b_flags &= ~B_AGE;
2594 cluster_awrite(bp);
2595 } else {
2596 bp->b_flags |= B_AGE;
2597 cluster_awrite(bp);
2598 }
2599 spin_lock(&pcpu->spin);
2600 ++r;
2601 break;
2602 }
2603
2604 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2605 marker->b_qindex = BQUEUE_NONE;
2606 spin_unlock(&pcpu->spin);
2607
2608 /*
2609 * Advance the marker to be fair.
2610 */
2611 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus;
2612 if (bp == NULL) {
2613 if (marker->b_qcpu != lcpu)
2614 goto again;
2615 }
2616
2617 return (r);
2618 }
2619
2620 /*
2621 * inmem:
2622 *
2623 * Returns true if no I/O is needed to access the associated VM object.
2624 * This is like findblk except it also hunts around in the VM system for
2625 * the data.
2626 *
2627 * Note that we ignore vm_page_free() races from interrupts against our
2628 * lookup, since if the caller is not protected our return value will not
2629 * be any more valid then otherwise once we exit the critical section.
2630 */
2631 int
2632 inmem(struct vnode *vp, off_t loffset)
2633 {
2634 vm_object_t obj;
2635 vm_offset_t toff, tinc, size;
2636 vm_page_t m;
2637 int res = 1;
2638
2639 if (findblk(vp, loffset, FINDBLK_TEST))
2640 return 1;
2641 if (vp->v_mount == NULL)
2642 return 0;
2643 if ((obj = vp->v_object) == NULL)
2644 return 0;
2645
2646 size = PAGE_SIZE;
2647 if (size > vp->v_mount->mnt_stat.f_iosize)
2648 size = vp->v_mount->mnt_stat.f_iosize;
2649
2650 vm_object_hold(obj);
2651 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2652 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2653 if (m == NULL) {
2654 res = 0;
2655 break;
2656 }
2657 tinc = size;
2658 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2659 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2660 if (vm_page_is_valid(m,
2661 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2662 res = 0;
2663 break;
2664 }
2665 }
2666 vm_object_drop(obj);
2667 return (res);
2668 }
2669
2670 /*
2671 * findblk:
2672 *
2673 * Locate and return the specified buffer. Unless flagged otherwise,
2674 * a locked buffer will be returned if it exists or NULL if it does not.
2675 *
2676 * findblk()'d buffers are still on the bufqueues and if you intend
2677 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2678 * and possibly do other stuff to it.
2679 *
2680 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2681 * for locking the buffer and ensuring that it remains
2682 * the desired buffer after locking.
2683 *
2684 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2685 * to acquire the lock we return NULL, even if the
2686 * buffer exists.
2687 *
2688 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2689 * reuse by getnewbuf() but does not prevent
2690 * disassociation (B_INVAL). Used to avoid deadlocks
2691 * against random (vp,loffset)s due to reassignment.
2692 *
2693 * (0) - Lock the buffer blocking.
2694 */
2695 struct buf *
2696 findblk(struct vnode *vp, off_t loffset, int flags)
2697 {
2698 struct buf *bp;
2699 int lkflags;
2700
2701 lkflags = LK_EXCLUSIVE;
2702 if (flags & FINDBLK_NBLOCK)
2703 lkflags |= LK_NOWAIT;
2704
2705 for (;;) {
2706 /*
2707 * Lookup. Ref the buf while holding v_token to prevent
2708 * reuse (but does not prevent diassociation).
2709 */
2710 lwkt_gettoken_shared(&vp->v_token);
2711 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2712 if (bp == NULL) {
2713 lwkt_reltoken(&vp->v_token);
2714 return(NULL);
2715 }
2716 bqhold(bp);
2717 lwkt_reltoken(&vp->v_token);
2718
2719 /*
2720 * If testing only break and return bp, do not lock.
2721 */
2722 if (flags & FINDBLK_TEST)
2723 break;
2724
2725 /*
2726 * Lock the buffer, return an error if the lock fails.
2727 * (only FINDBLK_NBLOCK can cause the lock to fail).
2728 */
2729 if (BUF_LOCK(bp, lkflags)) {
2730 atomic_subtract_int(&bp->b_refs, 1);
2731 /* bp = NULL; not needed */
2732 return(NULL);
2733 }
2734
2735 /*
2736 * Revalidate the locked buf before allowing it to be
2737 * returned.
2738 */
2739 if (bp->b_vp == vp && bp->b_loffset == loffset)
2740 break;
2741 atomic_subtract_int(&bp->b_refs, 1);
2742 BUF_UNLOCK(bp);
2743 }
2744
2745 /*
2746 * Success
2747 */
2748 if ((flags & FINDBLK_REF) == 0)
2749 atomic_subtract_int(&bp->b_refs, 1);
2750 return(bp);
2751 }
2752
2753 /*
2754 * getcacheblk:
2755 *
2756 * Similar to getblk() except only returns the buffer if it is
2757 * B_CACHE and requires no other manipulation. Otherwise NULL
2758 * is returned.
2759 *
2760 * If B_RAM is set the buffer might be just fine, but we return
2761 * NULL anyway because we want the code to fall through to the
2762 * cluster read. Otherwise read-ahead breaks.
2763 *
2764 * If blksize is 0 the buffer cache buffer must already be fully
2765 * cached.
2766 *
2767 * If blksize is non-zero getblk() will be used, allowing a buffer
2768 * to be reinstantiated from its VM backing store. The buffer must
2769 * still be fully cached after reinstantiation to be returned.
2770 */
2771 struct buf *
2772 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags)
2773 {
2774 struct buf *bp;
2775 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0;
2776
2777 if (blksize) {
2778 bp = getblk(vp, loffset, blksize, blkflags, 0);
2779 if (bp) {
2780 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2781 B_CACHE) {
2782 bp->b_flags &= ~B_AGE;
2783 } else {
2784 brelse(bp);
2785 bp = NULL;
2786 }
2787 }
2788 } else {
2789 bp = findblk(vp, loffset, fndflags);
2790 if (bp) {
2791 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2792 B_CACHE) {
2793 bp->b_flags &= ~B_AGE;
2794 bremfree(bp);
2795 } else {
2796 BUF_UNLOCK(bp);
2797 bp = NULL;
2798 }
2799 }
2800 }
2801 return (bp);
2802 }
2803
2804 /*
2805 * getblk:
2806 *
2807 * Get a block given a specified block and offset into a file/device.
2808 * B_INVAL may or may not be set on return. The caller should clear
2809 * B_INVAL prior to initiating a READ.
2810 *
2811 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2812 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2813 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2814 * without doing any of those things the system will likely believe
2815 * the buffer to be valid (especially if it is not B_VMIO), and the
2816 * next getblk() will return the buffer with B_CACHE set.
2817 *
2818 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2819 * an existing buffer.
2820 *
2821 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2822 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2823 * and then cleared based on the backing VM. If the previous buffer is
2824 * non-0-sized but invalid, B_CACHE will be cleared.
2825 *
2826 * If getblk() must create a new buffer, the new buffer is returned with
2827 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2828 * case it is returned with B_INVAL clear and B_CACHE set based on the
2829 * backing VM.
2830 *
2831 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2832 * B_CACHE bit is clear.
2833 *
2834 * What this means, basically, is that the caller should use B_CACHE to
2835 * determine whether the buffer is fully valid or not and should clear
2836 * B_INVAL prior to issuing a read. If the caller intends to validate
2837 * the buffer by loading its data area with something, the caller needs
2838 * to clear B_INVAL. If the caller does this without issuing an I/O,
2839 * the caller should set B_CACHE ( as an optimization ), else the caller
2840 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2841 * a write attempt or if it was a successfull read. If the caller
2842 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2843 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2844 *
2845 * getblk flags:
2846 *
2847 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2848 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2849 */
2850 struct buf *
2851 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2852 {
2853 struct buf *bp;
2854 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2855 int error;
2856 int lkflags;
2857
2858 if (size > MAXBSIZE)
2859 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2860 if (vp->v_object == NULL)
2861 panic("getblk: vnode %p has no object!", vp);
2862
2863 loop:
2864 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2865 /*
2866 * The buffer was found in the cache, but we need to lock it.
2867 * We must acquire a ref on the bp to prevent reuse, but
2868 * this will not prevent disassociation (brelvp()) so we
2869 * must recheck (vp,loffset) after acquiring the lock.
2870 *
2871 * Without the ref the buffer could potentially be reused
2872 * before we acquire the lock and create a deadlock
2873 * situation between the thread trying to reuse the buffer
2874 * and us due to the fact that we would wind up blocking
2875 * on a random (vp,loffset).
2876 */
2877 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2878 if (blkflags & GETBLK_NOWAIT) {
2879 bqdrop(bp);
2880 return(NULL);
2881 }
2882 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2883 if (blkflags & GETBLK_PCATCH)
2884 lkflags |= LK_PCATCH;
2885 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2886 if (error) {
2887 bqdrop(bp);
2888 if (error == ENOLCK)
2889 goto loop;
2890 return (NULL);
2891 }
2892 /* buffer may have changed on us */
2893 }
2894 bqdrop(bp);
2895
2896 /*
2897 * Once the buffer has been locked, make sure we didn't race
2898 * a buffer recyclement. Buffers that are no longer hashed
2899 * will have b_vp == NULL, so this takes care of that check
2900 * as well.
2901 */
2902 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2903 kprintf("Warning buffer %p (vp %p loffset %lld) "
2904 "was recycled\n",
2905 bp, vp, (long long)loffset);
2906 BUF_UNLOCK(bp);
2907 goto loop;
2908 }
2909
2910 /*
2911 * If SZMATCH any pre-existing buffer must be of the requested
2912 * size or NULL is returned. The caller absolutely does not
2913 * want getblk() to bwrite() the buffer on a size mismatch.
2914 */
2915 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2916 BUF_UNLOCK(bp);
2917 return(NULL);
2918 }
2919
2920 /*
2921 * All vnode-based buffers must be backed by a VM object.
2922 */
2923 KKASSERT(bp->b_flags & B_VMIO);
2924 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2925 bp->b_flags &= ~B_AGE;
2926
2927 /*
2928 * Make sure that B_INVAL buffers do not have a cached
2929 * block number translation.
2930 */
2931 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2932 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2933 " did not have cleared bio_offset cache\n",
2934 bp, vp, (long long)loffset);
2935 clearbiocache(&bp->b_bio2);
2936 }
2937
2938 /*
2939 * The buffer is locked. B_CACHE is cleared if the buffer is
2940 * invalid.
2941 */
2942 if (bp->b_flags & B_INVAL)
2943 bp->b_flags &= ~B_CACHE;
2944 bremfree(bp);
2945
2946 /*
2947 * Any size inconsistancy with a dirty buffer or a buffer
2948 * with a softupdates dependancy must be resolved. Resizing
2949 * the buffer in such circumstances can lead to problems.
2950 *
2951 * Dirty or dependant buffers are written synchronously.
2952 * Other types of buffers are simply released and
2953 * reconstituted as they may be backed by valid, dirty VM
2954 * pages (but not marked B_DELWRI).
2955 *
2956 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2957 * and may be left over from a prior truncation (and thus
2958 * no longer represent the actual EOF point), so we
2959 * definitely do not want to B_NOCACHE the backing store.
2960 */
2961 if (size != bp->b_bcount) {
2962 if (bp->b_flags & B_DELWRI) {
2963 bp->b_flags |= B_RELBUF;
2964 bwrite(bp);
2965 } else if (LIST_FIRST(&bp->b_dep)) {
2966 bp->b_flags |= B_RELBUF;
2967 bwrite(bp);
2968 } else {
2969 bp->b_flags |= B_RELBUF;
2970 brelse(bp);
2971 }
2972 goto loop;
2973 }
2974 KKASSERT(size <= bp->b_kvasize);
2975 KASSERT(bp->b_loffset != NOOFFSET,
2976 ("getblk: no buffer offset"));
2977
2978 /*
2979 * A buffer with B_DELWRI set and B_CACHE clear must
2980 * be committed before we can return the buffer in
2981 * order to prevent the caller from issuing a read
2982 * ( due to B_CACHE not being set ) and overwriting
2983 * it.
2984 *
2985 * Most callers, including NFS and FFS, need this to
2986 * operate properly either because they assume they
2987 * can issue a read if B_CACHE is not set, or because
2988 * ( for example ) an uncached B_DELWRI might loop due
2989 * to softupdates re-dirtying the buffer. In the latter
2990 * case, B_CACHE is set after the first write completes,
2991 * preventing further loops.
2992 *
2993 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2994 * above while extending the buffer, we cannot allow the
2995 * buffer to remain with B_CACHE set after the write
2996 * completes or it will represent a corrupt state. To
2997 * deal with this we set B_NOCACHE to scrap the buffer
2998 * after the write.
2999 *
3000 * XXX Should this be B_RELBUF instead of B_NOCACHE?
3001 * I'm not even sure this state is still possible
3002 * now that getblk() writes out any dirty buffers
3003 * on size changes.
3004 *
3005 * We might be able to do something fancy, like setting
3006 * B_CACHE in bwrite() except if B_DELWRI is already set,
3007 * so the below call doesn't set B_CACHE, but that gets real
3008 * confusing. This is much easier.
3009 */
3010
3011 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3012 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
3013 "and CACHE clear, b_flags %08x\n",
3014 bp, (uintmax_t)bp->b_loffset, bp->b_flags);
3015 bp->b_flags |= B_NOCACHE;
3016 bwrite(bp);
3017 goto loop;
3018 }
3019 } else {
3020 /*
3021 * Buffer is not in-core, create new buffer. The buffer
3022 * returned by getnewbuf() is locked. Note that the returned
3023 * buffer is also considered valid (not marked B_INVAL).
3024 *
3025 * Calculating the offset for the I/O requires figuring out
3026 * the block size. We use DEV_BSIZE for VBLK or VCHR and
3027 * the mount's f_iosize otherwise. If the vnode does not
3028 * have an associated mount we assume that the passed size is
3029 * the block size.
3030 *
3031 * Note that vn_isdisk() cannot be used here since it may
3032 * return a failure for numerous reasons. Note that the
3033 * buffer size may be larger then the block size (the caller
3034 * will use block numbers with the proper multiple). Beware
3035 * of using any v_* fields which are part of unions. In
3036 * particular, in DragonFly the mount point overloading
3037 * mechanism uses the namecache only and the underlying
3038 * directory vnode is not a special case.
3039 */
3040 int bsize, maxsize;
3041
3042 if (vp->v_type == VBLK || vp->v_type == VCHR)
3043 bsize = DEV_BSIZE;
3044 else if (vp->v_mount)
3045 bsize = vp->v_mount->mnt_stat.f_iosize;
3046 else
3047 bsize = size;
3048
3049 maxsize = size + (loffset & PAGE_MASK);
3050 maxsize = imax(maxsize, bsize);
3051
3052 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
3053 if (bp == NULL) {
3054 if (slpflags || slptimeo)
3055 return NULL;
3056 goto loop;
3057 }
3058
3059 /*
3060 * Atomically insert the buffer into the hash, so that it can
3061 * be found by findblk().
3062 *
3063 * If bgetvp() returns non-zero a collision occured, and the
3064 * bp will not be associated with the vnode.
3065 *
3066 * Make sure the translation layer has been cleared.
3067 */
3068 bp->b_loffset = loffset;
3069 bp->b_bio2.bio_offset = NOOFFSET;
3070 /* bp->b_bio2.bio_next = NULL; */
3071
3072 if (bgetvp(vp, bp, size)) {
3073 bp->b_flags |= B_INVAL;
3074 brelse(bp);
3075 goto loop;
3076 }
3077
3078 /*
3079 * All vnode-based buffers must be backed by a VM object.
3080 */
3081 KKASSERT(vp->v_object != NULL);
3082 bp->b_flags |= B_VMIO;
3083 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3084
3085 allocbuf(bp, size);
3086 }
3087 KKASSERT(dsched_is_clear_buf_priv(bp));
3088 return (bp);
3089 }
3090
3091 /*
3092 * regetblk(bp)
3093 *
3094 * Reacquire a buffer that was previously released to the locked queue,
3095 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3096 * set B_LOCKED (which handles the acquisition race).
3097 *
3098 * To this end, either B_LOCKED must be set or the dependancy list must be
3099 * non-empty.
3100 */
3101 void
3102 regetblk(struct buf *bp)
3103 {
3104 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3105 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3106 bremfree(bp);
3107 }
3108
3109 /*
3110 * geteblk:
3111 *
3112 * Get an empty, disassociated buffer of given size. The buffer is
3113 * initially set to B_INVAL.
3114 *
3115 * critical section protection is not required for the allocbuf()
3116 * call because races are impossible here.
3117 */
3118 struct buf *
3119 geteblk(int size)
3120 {
3121 struct buf *bp;
3122 int maxsize;
3123
3124 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3125
3126 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL)
3127 ;
3128 allocbuf(bp, size);
3129 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3130 KKASSERT(dsched_is_clear_buf_priv(bp));
3131 return (bp);
3132 }
3133
3134
3135 /*
3136 * allocbuf:
3137 *
3138 * This code constitutes the buffer memory from either anonymous system
3139 * memory (in the case of non-VMIO operations) or from an associated
3140 * VM object (in the case of VMIO operations). This code is able to
3141 * resize a buffer up or down.
3142 *
3143 * Note that this code is tricky, and has many complications to resolve
3144 * deadlock or inconsistant data situations. Tread lightly!!!
3145 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3146 * the caller. Calling this code willy nilly can result in the loss of
3147 * data.
3148 *
3149 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3150 * B_CACHE for the non-VMIO case.
3151 *
3152 * This routine does not need to be called from a critical section but you
3153 * must own the buffer.
3154 */
3155 int
3156 allocbuf(struct buf *bp, int size)
3157 {
3158 int newbsize, mbsize;
3159 int i;
3160
3161 if (BUF_REFCNT(bp) == 0)
3162 panic("allocbuf: buffer not busy");
3163
3164 if (bp->b_kvasize < size)
3165 panic("allocbuf: buffer too small");
3166
3167 if ((bp->b_flags & B_VMIO) == 0) {
3168 caddr_t origbuf;
3169 int origbufsize;
3170 /*
3171 * Just get anonymous memory from the kernel. Don't
3172 * mess with B_CACHE.
3173 */
3174 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3175 if (bp->b_flags & B_MALLOC)
3176 newbsize = mbsize;
3177 else
3178 newbsize = round_page(size);
3179
3180 if (newbsize < bp->b_bufsize) {
3181 /*
3182 * Malloced buffers are not shrunk
3183 */
3184 if (bp->b_flags & B_MALLOC) {
3185 if (newbsize) {
3186 bp->b_bcount = size;
3187 } else {
3188 kfree(bp->b_data, M_BIOBUF);
3189 if (bp->b_bufsize) {
3190 atomic_subtract_long(&bufmallocspace, bp->b_bufsize);
3191 bufspacewakeup();
3192 bp->b_bufsize = 0;
3193 }
3194 bp->b_data = bp->b_kvabase;
3195 bp->b_bcount = 0;
3196 bp->b_flags &= ~B_MALLOC;
3197 }
3198 return 1;
3199 }
3200 vm_hold_free_pages(
3201 bp,
3202 (vm_offset_t) bp->b_data + newbsize,
3203 (vm_offset_t) bp->b_data + bp->b_bufsize);
3204 } else if (newbsize > bp->b_bufsize) {
3205 /*
3206 * We only use malloced memory on the first allocation.
3207 * and revert to page-allocated memory when the buffer
3208 * grows.
3209 */
3210 if ((bufmallocspace < maxbufmallocspace) &&
3211 (bp->b_bufsize == 0) &&
3212 (mbsize <= PAGE_SIZE/2)) {
3213
3214 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3215 bp->b_bufsize = mbsize;
3216 bp->b_bcount = size;
3217 bp->b_flags |= B_MALLOC;
3218 atomic_add_long(&bufmallocspace, mbsize);
3219 return 1;
3220 }
3221 origbuf = NULL;
3222 origbufsize = 0;
3223 /*
3224 * If the buffer is growing on its other-than-first
3225 * allocation, then we revert to the page-allocation
3226 * scheme.
3227 */
3228 if (bp->b_flags & B_MALLOC) {
3229 origbuf = bp->b_data;
3230 origbufsize = bp->b_bufsize;
3231 bp->b_data = bp->b_kvabase;
3232 if (bp->b_bufsize) {
3233 atomic_subtract_long(&bufmallocspace,
3234 bp->b_bufsize);
3235 bufspacewakeup();
3236 bp->b_bufsize = 0;
3237 }
3238 bp->b_flags &= ~B_MALLOC;
3239 newbsize = round_page(newbsize);
3240 }
3241 vm_hold_load_pages(
3242 bp,
3243 (vm_offset_t) bp->b_data + bp->b_bufsize,
3244 (vm_offset_t) bp->b_data + newbsize);
3245 if (origbuf) {
3246 bcopy(origbuf, bp->b_data, origbufsize);
3247 kfree(origbuf, M_BIOBUF);
3248 }
3249 }
3250 } else {
3251 vm_page_t m;
3252 int desiredpages;
3253
3254 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3255 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3256 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3257 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3258
3259 if (bp->b_flags & B_MALLOC)
3260 panic("allocbuf: VMIO buffer can't be malloced");
3261 /*
3262 * Set B_CACHE initially if buffer is 0 length or will become
3263 * 0-length.
3264 */
3265 if (size == 0 || bp->b_bufsize == 0)
3266 bp->b_flags |= B_CACHE;
3267
3268 if (newbsize < bp->b_bufsize) {
3269 /*
3270 * DEV_BSIZE aligned new buffer size is less then the
3271 * DEV_BSIZE aligned existing buffer size. Figure out
3272 * if we have to remove any pages.
3273 */
3274 if (desiredpages < bp->b_xio.xio_npages) {
3275 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3276 /*
3277 * the page is not freed here -- it
3278 * is the responsibility of
3279 * vnode_pager_setsize
3280 */
3281 m = bp->b_xio.xio_pages[i];
3282 KASSERT(m != bogus_page,
3283 ("allocbuf: bogus page found"));
3284 vm_page_busy_wait(m, TRUE, "biodep");
3285 bp->b_xio.xio_pages[i] = NULL;
3286 vm_page_unwire(m, 0);
3287 vm_page_wakeup(m);
3288 }
3289 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3290 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3291 bp->b_xio.xio_npages = desiredpages;
3292 }
3293 } else if (size > bp->b_bcount) {
3294 /*
3295 * We are growing the buffer, possibly in a
3296 * byte-granular fashion.
3297 */
3298 struct vnode *vp;
3299 vm_object_t obj;
3300 vm_offset_t toff;
3301 vm_offset_t tinc;
3302
3303 /*
3304 * Step 1, bring in the VM pages from the object,
3305 * allocating them if necessary. We must clear
3306 * B_CACHE if these pages are not valid for the
3307 * range covered by the buffer.
3308 *
3309 * critical section protection is required to protect
3310 * against interrupts unbusying and freeing pages
3311 * between our vm_page_lookup() and our
3312 * busycheck/wiring call.
3313 */
3314 vp = bp->b_vp;
3315 obj = vp->v_object;
3316
3317 vm_object_hold(obj);
3318 while (bp->b_xio.xio_npages < desiredpages) {
3319 vm_page_t m;
3320 vm_pindex_t pi;
3321 int error;
3322
3323 pi = OFF_TO_IDX(bp->b_loffset) +
3324 bp->b_xio.xio_npages;
3325
3326 /*
3327 * Blocking on m->busy might lead to a
3328 * deadlock:
3329 *
3330 * vm_fault->getpages->cluster_read->allocbuf
3331 */
3332 m = vm_page_lookup_busy_try(obj, pi, FALSE,
3333 &error);
3334 if (error) {
3335 vm_page_sleep_busy(m, FALSE, "pgtblk");
3336 continue;
3337 }
3338 if (m == NULL) {
3339 /*
3340 * note: must allocate system pages
3341 * since blocking here could intefere
3342 * with paging I/O, no matter which
3343 * process we are.
3344 */
3345 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages);
3346 if (m) {
3347 vm_page_wire(m);
3348 vm_page_flag_clear(m, PG_ZERO);
3349 vm_page_wakeup(m);
3350 bp->b_flags &= ~B_CACHE;
3351 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3352 ++bp->b_xio.xio_npages;
3353 }
3354 continue;
3355 }
3356
3357 /*
3358 * We found a page and were able to busy it.
3359 */
3360 vm_page_flag_clear(m, PG_ZERO);
3361 vm_page_wire(m);
3362 vm_page_wakeup(m);
3363 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3364 ++bp->b_xio.xio_npages;
3365 if (bp->b_act_count < m->act_count)
3366 bp->b_act_count = m->act_count;
3367 }
3368 vm_object_drop(obj);
3369
3370 /*
3371 * Step 2. We've loaded the pages into the buffer,
3372 * we have to figure out if we can still have B_CACHE
3373 * set. Note that B_CACHE is set according to the
3374 * byte-granular range ( bcount and size ), not the
3375 * aligned range ( newbsize ).
3376 *
3377 * The VM test is against m->valid, which is DEV_BSIZE
3378 * aligned. Needless to say, the validity of the data
3379 * needs to also be DEV_BSIZE aligned. Note that this
3380 * fails with NFS if the server or some other client
3381 * extends the file's EOF. If our buffer is resized,
3382 * B_CACHE may remain set! XXX
3383 */
3384
3385 toff = bp->b_bcount;
3386 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3387
3388 while ((bp->b_flags & B_CACHE) && toff < size) {
3389 vm_pindex_t pi;
3390
3391 if (tinc > (size - toff))
3392 tinc = size - toff;
3393
3394 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3395 PAGE_SHIFT;
3396
3397 vfs_buf_test_cache(
3398 bp,
3399 bp->b_loffset,
3400 toff,
3401 tinc,
3402 bp->b_xio.xio_pages[pi]
3403 );
3404 toff += tinc;
3405 tinc = PAGE_SIZE;
3406 }
3407
3408 /*
3409 * Step 3, fixup the KVM pmap. Remember that
3410 * bp->b_data is relative to bp->b_loffset, but
3411 * bp->b_loffset may be offset into the first page.
3412 */
3413
3414 bp->b_data = (caddr_t)
3415 trunc_page((vm_offset_t)bp->b_data);
3416 pmap_qenter(
3417 (vm_offset_t)bp->b_data,
3418 bp->b_xio.xio_pages,
3419 bp->b_xio.xio_npages
3420 );
3421 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3422 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3423 }
3424 }
3425
3426 /* adjust space use on already-dirty buffer */
3427 if (bp->b_flags & B_DELWRI) {
3428 /* dirtykvaspace unchanged */
3429 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize);
3430 if (bp->b_flags & B_HEAVY) {
3431 atomic_add_long(&dirtybufspacehw,
3432 newbsize - bp->b_bufsize);
3433 }
3434 }
3435 if (newbsize < bp->b_bufsize)
3436 bufspacewakeup();
3437 bp->b_bufsize = newbsize; /* actual buffer allocation */
3438 bp->b_bcount = size; /* requested buffer size */
3439 return 1;
3440 }
3441
3442 /*
3443 * biowait:
3444 *
3445 * Wait for buffer I/O completion, returning error status. B_EINTR
3446 * is converted into an EINTR error but not cleared (since a chain
3447 * of biowait() calls may occur).
3448 *
3449 * On return bpdone() will have been called but the buffer will remain
3450 * locked and will not have been brelse()'d.
3451 *
3452 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3453 * likely still in progress on return.
3454 *
3455 * NOTE! This operation is on a BIO, not a BUF.
3456 *
3457 * NOTE! BIO_DONE is cleared by vn_strategy()
3458 */
3459 static __inline int
3460 _biowait(struct bio *bio, const char *wmesg, int to)
3461 {
3462 struct buf *bp = bio->bio_buf;
3463 u_int32_t flags;
3464 u_int32_t nflags;
3465 int error;
3466
3467 KKASSERT(bio == &bp->b_bio1);
3468 for (;;) {
3469 flags = bio->bio_flags;
3470 if (flags & BIO_DONE)
3471 break;
3472 nflags = flags | BIO_WANT;
3473 tsleep_interlock(bio, 0);
3474 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3475 if (wmesg)
3476 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3477 else if (bp->b_cmd == BUF_CMD_READ)
3478 error = tsleep(bio, PINTERLOCKED, "biord", to);
3479 else
3480 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3481 if (error) {
3482 kprintf("tsleep error biowait %d\n", error);
3483 return (error);
3484 }
3485 }
3486 }
3487
3488 /*
3489 * Finish up.
3490 */
3491 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3492 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3493 if (bp->b_flags & B_EINTR)
3494 return (EINTR);
3495 if (bp->b_flags & B_ERROR)
3496 return (bp->b_error ? bp->b_error : EIO);
3497 return (0);
3498 }
3499
3500 int
3501 biowait(struct bio *bio, const char *wmesg)
3502 {
3503 return(_biowait(bio, wmesg, 0));
3504 }
3505
3506 int
3507 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3508 {
3509 return(_biowait(bio, wmesg, to));
3510 }
3511
3512 /*
3513 * This associates a tracking count with an I/O. vn_strategy() and
3514 * dev_dstrategy() do this automatically but there are a few cases
3515 * where a vnode or device layer is bypassed when a block translation
3516 * is cached. In such cases bio_start_transaction() may be called on
3517 * the bypassed layers so the system gets an I/O in progress indication
3518 * for those higher layers.
3519 */
3520 void
3521 bio_start_transaction(struct bio *bio, struct bio_track *track)
3522 {
3523 bio->bio_track = track;
3524 if (dsched_is_clear_buf_priv(bio->bio_buf))
3525 dsched_new_buf(bio->bio_buf);
3526 bio_track_ref(track);
3527 }
3528
3529 /*
3530 * Initiate I/O on a vnode.
3531 *
3532 * SWAPCACHE OPERATION:
3533 *
3534 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3535 * devfs also uses b_vp for fake buffers so we also have to check
3536 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3537 * underlying block device. The swap assignments are related to the
3538 * buffer cache buffer's b_vp, not the passed vp.
3539 *
3540 * The passed vp == bp->b_vp only in the case where the strategy call
3541 * is made on the vp itself for its own buffers (a regular file or
3542 * block device vp). The filesystem usually then re-calls vn_strategy()
3543 * after translating the request to an underlying device.
3544 *
3545 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3546 * underlying buffer cache buffers.
3547 *
3548 * We can only deal with page-aligned buffers at the moment, because
3549 * we can't tell what the real dirty state for pages straddling a buffer
3550 * are.
3551 *
3552 * In order to call swap_pager_strategy() we must provide the VM object
3553 * and base offset for the underlying buffer cache pages so it can find
3554 * the swap blocks.
3555 */
3556 void
3557 vn_strategy(struct vnode *vp, struct bio *bio)
3558 {
3559 struct bio_track *track;
3560 struct buf *bp = bio->bio_buf;
3561
3562 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3563
3564 /*
3565 * Set when an I/O is issued on the bp. Cleared by consumers
3566 * (aka HAMMER), allowing the consumer to determine if I/O had
3567 * actually occurred.
3568 */
3569 bp->b_flags |= B_IODEBUG;
3570
3571 /*
3572 * Handle the swap cache intercept.
3573 */
3574 if (vn_cache_strategy(vp, bio))
3575 return;
3576
3577 /*
3578 * Otherwise do the operation through the filesystem
3579 */
3580 if (bp->b_cmd == BUF_CMD_READ)
3581 track = &vp->v_track_read;
3582 else
3583 track = &vp->v_track_write;
3584 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3585 bio->bio_track = track;
3586 if (dsched_is_clear_buf_priv(bio->bio_buf))
3587 dsched_new_buf(bio->bio_buf);
3588 bio_track_ref(track);
3589 vop_strategy(*vp->v_ops, vp, bio);
3590 }
3591
3592 static void vn_cache_strategy_callback(struct bio *bio);
3593
3594 int
3595 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3596 {
3597 struct buf *bp = bio->bio_buf;
3598 struct bio *nbio;
3599 vm_object_t object;
3600 vm_page_t m;
3601 int i;
3602
3603 /*
3604 * Is this buffer cache buffer suitable for reading from
3605 * the swap cache?
3606 */
3607 if (vm_swapcache_read_enable == 0 ||
3608 bp->b_cmd != BUF_CMD_READ ||
3609 ((bp->b_flags & B_CLUSTER) == 0 &&
3610 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3611 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3612 (bp->b_bcount & PAGE_MASK) != 0) {
3613 return(0);
3614 }
3615
3616 /*
3617 * Figure out the original VM object (it will match the underlying
3618 * VM pages). Note that swap cached data uses page indices relative
3619 * to that object, not relative to bio->bio_offset.
3620 */
3621 if (bp->b_flags & B_CLUSTER)
3622 object = vp->v_object;
3623 else
3624 object = bp->b_vp->v_object;
3625
3626 /*
3627 * In order to be able to use the swap cache all underlying VM
3628 * pages must be marked as such, and we can't have any bogus pages.
3629 */
3630 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3631 m = bp->b_xio.xio_pages[i];
3632 if ((m->flags & PG_SWAPPED) == 0)
3633 break;
3634 if (m == bogus_page)
3635 break;
3636 }
3637
3638 /*
3639 * If we are good then issue the I/O using swap_pager_strategy().
3640 *
3641 * We can only do this if the buffer actually supports object-backed
3642 * I/O. If it doesn't npages will be 0.
3643 */
3644 if (i && i == bp->b_xio.xio_npages) {
3645 m = bp->b_xio.xio_pages[0];
3646 nbio = push_bio(bio);
3647 nbio->bio_done = vn_cache_strategy_callback;
3648 nbio->bio_offset = ptoa(m->pindex);
3649 KKASSERT(m->object == object);
3650 swap_pager_strategy(object, nbio);
3651 return(1);
3652 }
3653 return(0);
3654 }
3655
3656 /*
3657 * This is a bit of a hack but since the vn_cache_strategy() function can
3658 * override a VFS's strategy function we must make sure that the bio, which
3659 * is probably bio2, doesn't leak an unexpected offset value back to the
3660 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3661 * bio went through its own file strategy function and the the bio2 offset
3662 * is a cached disk offset when, in fact, it isn't.
3663 */
3664 static void
3665 vn_cache_strategy_callback(struct bio *bio)
3666 {
3667 bio->bio_offset = NOOFFSET;
3668 biodone(pop_bio(bio));
3669 }
3670
3671 /*
3672 * bpdone:
3673 *
3674 * Finish I/O on a buffer after all BIOs have been processed.
3675 * Called when the bio chain is exhausted or by biowait. If called
3676 * by biowait, elseit is typically 0.
3677 *
3678 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3679 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3680 * assuming B_INVAL is clear.
3681 *
3682 * For the VMIO case, we set B_CACHE if the op was a read and no
3683 * read error occured, or if the op was a write. B_CACHE is never
3684 * set if the buffer is invalid or otherwise uncacheable.
3685 *
3686 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3687 * initiator to leave B_INVAL set to brelse the buffer out of existance
3688 * in the biodone routine.
3689 */
3690 void
3691 bpdone(struct buf *bp, int elseit)
3692 {
3693 buf_cmd_t cmd;
3694
3695 KASSERT(BUF_REFCNTNB(bp) > 0,
3696 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3697 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3698 ("biodone: bp %p already done!", bp));
3699
3700 /*
3701 * No more BIOs are left. All completion functions have been dealt
3702 * with, now we clean up the buffer.
3703 */
3704 cmd = bp->b_cmd;
3705 bp->b_cmd = BUF_CMD_DONE;
3706
3707 /*
3708 * Only reads and writes are processed past this point.
3709 */
3710 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3711 if (cmd == BUF_CMD_FREEBLKS)
3712 bp->b_flags |= B_NOCACHE;
3713 if (elseit)
3714 brelse(bp);
3715 return;
3716 }
3717
3718 /*
3719 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3720 * a lot worse. XXX - move this above the clearing of b_cmd
3721 */
3722 if (LIST_FIRST(&bp->b_dep) != NULL)
3723 buf_complete(bp);
3724
3725 /*
3726 * A failed write must re-dirty the buffer unless B_INVAL
3727 * was set. Only applicable to normal buffers (with VPs).
3728 * vinum buffers may not have a vp.
3729 */
3730 if (cmd == BUF_CMD_WRITE &&
3731 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3732 bp->b_flags &= ~B_NOCACHE;
3733 if (bp->b_vp)
3734 bdirty(bp);
3735 }
3736
3737 if (bp->b_flags & B_VMIO) {
3738 int i;
3739 vm_ooffset_t foff;
3740 vm_page_t m;
3741 vm_object_t obj;
3742 int iosize;
3743 struct vnode *vp = bp->b_vp;
3744
3745 obj = vp->v_object;
3746
3747 #if defined(VFS_BIO_DEBUG)
3748 if (vp->v_auxrefs == 0)
3749 panic("biodone: zero vnode hold count");
3750 if ((vp->v_flag & VOBJBUF) == 0)
3751 panic("biodone: vnode is not setup for merged cache");
3752 #endif
3753
3754 foff = bp->b_loffset;
3755 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3756 KASSERT(obj != NULL, ("biodone: missing VM object"));
3757
3758 #if defined(VFS_BIO_DEBUG)
3759 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3760 kprintf("biodone: paging in progress(%d) < "
3761 "bp->b_xio.xio_npages(%d)\n",
3762 obj->paging_in_progress,
3763 bp->b_xio.xio_npages);
3764 }
3765 #endif
3766
3767 /*
3768 * Set B_CACHE if the op was a normal read and no error
3769 * occured. B_CACHE is set for writes in the b*write()
3770 * routines.
3771 */
3772 iosize = bp->b_bcount - bp->b_resid;
3773 if (cmd == BUF_CMD_READ &&
3774 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3775 bp->b_flags |= B_CACHE;
3776 }
3777
3778 vm_object_hold(obj);
3779 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3780 int bogusflag = 0;
3781 int resid;
3782
3783 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3784 if (resid > iosize)
3785 resid = iosize;
3786
3787 /*
3788 * cleanup bogus pages, restoring the originals. Since
3789 * the originals should still be wired, we don't have
3790 * to worry about interrupt/freeing races destroying
3791 * the VM object association.
3792 */
3793 m = bp->b_xio.xio_pages[i];
3794 if (m == bogus_page) {
3795 bogusflag = 1;
3796 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3797 if (m == NULL)
3798 panic("biodone: page disappeared");
3799 bp->b_xio.xio_pages[i] = m;
3800 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3801 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3802 }
3803 #if defined(VFS_BIO_DEBUG)
3804 if (OFF_TO_IDX(foff) != m->pindex) {
3805 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3806 "mismatch\n",
3807 (unsigned long)foff, (long)m->pindex);
3808 }
3809 #endif
3810
3811 /*
3812 * In the write case, the valid and clean bits are
3813 * already changed correctly (see bdwrite()), so we
3814 * only need to do this here in the read case.
3815 */
3816 vm_page_busy_wait(m, FALSE, "bpdpgw");
3817 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3818 vfs_clean_one_page(bp, i, m);
3819 }
3820 vm_page_flag_clear(m, PG_ZERO);
3821
3822 /*
3823 * when debugging new filesystems or buffer I/O
3824 * methods, this is the most common error that pops
3825 * up. if you see this, you have not set the page
3826 * busy flag correctly!!!
3827 */
3828 if (m->busy == 0) {
3829 kprintf("biodone: page busy < 0, "
3830 "pindex: %d, foff: 0x(%x,%x), "
3831 "resid: %d, index: %d\n",
3832 (int) m->pindex, (int)(foff >> 32),
3833 (int) foff & 0xffffffff, resid, i);
3834 if (!vn_isdisk(vp, NULL))
3835 kprintf(" iosize: %ld, loffset: %lld, "
3836 "flags: 0x%08x, npages: %d\n",
3837 bp->b_vp->v_mount->mnt_stat.f_iosize,
3838 (long long)bp->b_loffset,
3839 bp->b_flags, bp->b_xio.xio_npages);
3840 else
3841 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3842 (long long)bp->b_loffset,
3843 bp->b_flags, bp->b_xio.xio_npages);
3844 kprintf(" valid: 0x%x, dirty: 0x%x, "
3845 "wired: %d\n",
3846 m->valid, m->dirty,
3847 m->wire_count);
3848 panic("biodone: page busy < 0");
3849 }
3850 vm_page_io_finish(m);
3851 vm_page_wakeup(m);
3852 vm_object_pip_wakeup(obj);
3853 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3854 iosize -= resid;
3855 }
3856 bp->b_flags &= ~B_HASBOGUS;
3857 vm_object_drop(obj);
3858 }
3859
3860 /*
3861 * Finish up by releasing the buffer. There are no more synchronous
3862 * or asynchronous completions, those were handled by bio_done
3863 * callbacks.
3864 */
3865 if (elseit) {
3866 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3867 brelse(bp);
3868 else
3869 bqrelse(bp);
3870 }
3871 }
3872
3873 /*
3874 * Normal biodone.
3875 */
3876 void
3877 biodone(struct bio *bio)
3878 {
3879 struct buf *bp = bio->bio_buf;
3880
3881 runningbufwakeup(bp);
3882
3883 /*
3884 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3885 */
3886 while (bio) {
3887 biodone_t *done_func;
3888 struct bio_track *track;
3889
3890 /*
3891 * BIO tracking. Most but not all BIOs are tracked.
3892 */
3893 if ((track = bio->bio_track) != NULL) {
3894 bio_track_rel(track);
3895 bio->bio_track = NULL;
3896 }
3897
3898 /*
3899 * A bio_done function terminates the loop. The function
3900 * will be responsible for any further chaining and/or
3901 * buffer management.
3902 *
3903 * WARNING! The done function can deallocate the buffer!
3904 */
3905 if ((done_func = bio->bio_done) != NULL) {
3906 bio->bio_done = NULL;
3907 done_func(bio);
3908 return;
3909 }
3910 bio = bio->bio_prev;
3911 }
3912
3913 /*
3914 * If we've run out of bio's do normal [a]synchronous completion.
3915 */
3916 bpdone(bp, 1);
3917 }
3918
3919 /*
3920 * Synchronous biodone - this terminates a synchronous BIO.
3921 *
3922 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3923 * but still locked. The caller must brelse() the buffer after waiting
3924 * for completion.
3925 */
3926 void
3927 biodone_sync(struct bio *bio)
3928 {
3929 struct buf *bp = bio->bio_buf;
3930 int flags;
3931 int nflags;
3932
3933 KKASSERT(bio == &bp->b_bio1);
3934 bpdone(bp, 0);
3935
3936 for (;;) {
3937 flags = bio->bio_flags;
3938 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3939
3940 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3941 if (flags & BIO_WANT)
3942 wakeup(bio);
3943 break;
3944 }
3945 }
3946 }
3947
3948 /*
3949 * vfs_unbusy_pages:
3950 *
3951 * This routine is called in lieu of iodone in the case of
3952 * incomplete I/O. This keeps the busy status for pages
3953 * consistant.
3954 */
3955 void
3956 vfs_unbusy_pages(struct buf *bp)
3957 {
3958 int i;
3959
3960 runningbufwakeup(bp);
3961
3962 if (bp->b_flags & B_VMIO) {
3963 struct vnode *vp = bp->b_vp;
3964 vm_object_t obj;
3965
3966 obj = vp->v_object;
3967 vm_object_hold(obj);
3968
3969 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3970 vm_page_t m = bp->b_xio.xio_pages[i];
3971
3972 /*
3973 * When restoring bogus changes the original pages
3974 * should still be wired, so we are in no danger of
3975 * losing the object association and do not need
3976 * critical section protection particularly.
3977 */
3978 if (m == bogus_page) {
3979 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3980 if (!m) {
3981 panic("vfs_unbusy_pages: page missing");
3982 }
3983 bp->b_xio.xio_pages[i] = m;
3984 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3985 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3986 }
3987 vm_page_busy_wait(m, FALSE, "bpdpgw");
3988 vm_page_flag_clear(m, PG_ZERO);
3989 vm_page_io_finish(m);
3990 vm_page_wakeup(m);
3991 vm_object_pip_wakeup(obj);
3992 }
3993 bp->b_flags &= ~B_HASBOGUS;
3994 vm_object_drop(obj);
3995 }
3996 }
3997
3998 /*
3999 * vfs_busy_pages:
4000 *
4001 * This routine is called before a device strategy routine.
4002 * It is used to tell the VM system that paging I/O is in
4003 * progress, and treat the pages associated with the buffer
4004 * almost as being PG_BUSY. Also the object 'paging_in_progress'
4005 * flag is handled to make sure that the object doesn't become
4006 * inconsistant.
4007 *
4008 * Since I/O has not been initiated yet, certain buffer flags
4009 * such as B_ERROR or B_INVAL may be in an inconsistant state
4010 * and should be ignored.
4011 */
4012 void
4013 vfs_busy_pages(struct vnode *vp, struct buf *bp)
4014 {
4015 int i, bogus;
4016 struct lwp *lp = curthread->td_lwp;
4017
4018 /*
4019 * The buffer's I/O command must already be set. If reading,
4020 * B_CACHE must be 0 (double check against callers only doing
4021 * I/O when B_CACHE is 0).
4022 */
4023 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4024 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
4025
4026 if (bp->b_flags & B_VMIO) {
4027 vm_object_t obj;
4028
4029 obj = vp->v_object;
4030 KASSERT(bp->b_loffset != NOOFFSET,
4031 ("vfs_busy_pages: no buffer offset"));
4032
4033 /*
4034 * Busy all the pages. We have to busy them all at once
4035 * to avoid deadlocks.
4036 */
4037 retry:
4038 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4039 vm_page_t m = bp->b_xio.xio_pages[i];
4040
4041 if (vm_page_busy_try(m, FALSE)) {
4042 vm_page_sleep_busy(m, FALSE, "vbpage");
4043 while (--i >= 0)
4044 vm_page_wakeup(bp->b_xio.xio_pages[i]);
4045 goto retry;
4046 }
4047 }
4048
4049 /*
4050 * Setup for I/O, soft-busy the page right now because
4051 * the next loop may block.
4052 */
4053 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4054 vm_page_t m = bp->b_xio.xio_pages[i];
4055
4056 vm_page_flag_clear(m, PG_ZERO);
4057 if ((bp->b_flags & B_CLUSTER) == 0) {
4058 vm_object_pip_add(obj, 1);
4059 vm_page_io_start(m);
4060 }
4061 }
4062
4063 /*
4064 * Adjust protections for I/O and do bogus-page mapping.
4065 * Assume that vm_page_protect() can block (it can block
4066 * if VM_PROT_NONE, don't take any chances regardless).
4067 *
4068 * In particular note that for writes we must incorporate
4069 * page dirtyness from the VM system into the buffer's
4070 * dirty range.
4071 *
4072 * For reads we theoretically must incorporate page dirtyness
4073 * from the VM system to determine if the page needs bogus
4074 * replacement, but we shortcut the test by simply checking
4075 * that all m->valid bits are set, indicating that the page
4076 * is fully valid and does not need to be re-read. For any
4077 * VM system dirtyness the page will also be fully valid
4078 * since it was mapped at one point.
4079 */
4080 bogus = 0;
4081 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4082 vm_page_t m = bp->b_xio.xio_pages[i];
4083
4084 vm_page_flag_clear(m, PG_ZERO); /* XXX */
4085 if (bp->b_cmd == BUF_CMD_WRITE) {
4086 /*
4087 * When readying a vnode-backed buffer for
4088 * a write we must zero-fill any invalid
4089 * portions of the backing VM pages, mark
4090 * it valid and clear related dirty bits.
4091 *
4092 * vfs_clean_one_page() incorporates any
4093 * VM dirtyness and updates the b_dirtyoff
4094 * range (after we've made the page RO).
4095 *
4096 * It is also expected that the pmap modified
4097 * bit has already been cleared by the
4098 * vm_page_protect(). We may not be able
4099 * to clear all dirty bits for a page if it
4100 * was also memory mapped (NFS).
4101 *
4102 * Finally be sure to unassign any swap-cache
4103 * backing store as it is now stale.
4104 */
4105 vm_page_protect(m, VM_PROT_READ);
4106 vfs_clean_one_page(bp, i, m);
4107 swap_pager_unswapped(m);
4108 } else if (m->valid == VM_PAGE_BITS_ALL) {
4109 /*
4110 * When readying a vnode-backed buffer for
4111 * read we must replace any dirty pages with
4112 * a bogus page so dirty data is not destroyed
4113 * when filling gaps.
4114 *
4115 * To avoid testing whether the page is
4116 * dirty we instead test that the page was
4117 * at some point mapped (m->valid fully
4118 * valid) with the understanding that
4119 * this also covers the dirty case.
4120 */
4121 bp->b_xio.xio_pages[i] = bogus_page;
4122 bp->b_flags |= B_HASBOGUS;
4123 bogus++;
4124 } else if (m->valid & m->dirty) {
4125 /*
4126 * This case should not occur as partial
4127 * dirtyment can only happen if the buffer
4128 * is B_CACHE, and this code is not entered
4129 * if the buffer is B_CACHE.
4130 */
4131 kprintf("Warning: vfs_busy_pages - page not "
4132 "fully valid! loff=%jx bpf=%08x "
4133 "idx=%d val=%02x dir=%02x\n",
4134 (uintmax_t)bp->b_loffset, bp->b_flags,
4135 i, m->valid, m->dirty);
4136 vm_page_protect(m, VM_PROT_NONE);
4137 } else {
4138 /*
4139 * The page is not valid and can be made
4140 * part of the read.
4141 */
4142 vm_page_protect(m, VM_PROT_NONE);
4143 }
4144 vm_page_wakeup(m);
4145 }
4146 if (bogus) {
4147 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4148 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4149 }
4150 }
4151
4152 /*
4153 * This is the easiest place to put the process accounting for the I/O
4154 * for now.
4155 */
4156 if (lp != NULL) {
4157 if (bp->b_cmd == BUF_CMD_READ)
4158 lp->lwp_ru.ru_inblock++;
4159 else
4160 lp->lwp_ru.ru_oublock++;
4161 }
4162 }
4163
4164 /*
4165 * Tell the VM system that the pages associated with this buffer
4166 * are clean. This is used for delayed writes where the data is
4167 * going to go to disk eventually without additional VM intevention.
4168 *
4169 * NOTE: While we only really need to clean through to b_bcount, we
4170 * just go ahead and clean through to b_bufsize.
4171 */
4172 static void
4173 vfs_clean_pages(struct buf *bp)
4174 {
4175 vm_page_t m;
4176 int i;
4177
4178 if ((bp->b_flags & B_VMIO) == 0)
4179 return;
4180
4181 KASSERT(bp->b_loffset != NOOFFSET,
4182 ("vfs_clean_pages: no buffer offset"));
4183
4184 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4185 m = bp->b_xio.xio_pages[i];
4186 vfs_clean_one_page(bp, i, m);
4187 }
4188 }
4189
4190 /*
4191 * vfs_clean_one_page:
4192 *
4193 * Set the valid bits and clear the dirty bits in a page within a
4194 * buffer. The range is restricted to the buffer's size and the
4195 * buffer's logical offset might index into the first page.
4196 *
4197 * The caller has busied or soft-busied the page and it is not mapped,
4198 * test and incorporate the dirty bits into b_dirtyoff/end before
4199 * clearing them. Note that we need to clear the pmap modified bits
4200 * after determining the the page was dirty, vm_page_set_validclean()
4201 * does not do it for us.
4202 *
4203 * This routine is typically called after a read completes (dirty should
4204 * be zero in that case as we are not called on bogus-replace pages),
4205 * or before a write is initiated.
4206 */
4207 static void
4208 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4209 {
4210 int bcount;
4211 int xoff;
4212 int soff;
4213 int eoff;
4214
4215 /*
4216 * Calculate offset range within the page but relative to buffer's
4217 * loffset. loffset might be offset into the first page.
4218 */
4219 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4220 bcount = bp->b_bcount + xoff; /* offset adjusted */
4221
4222 if (pageno == 0) {
4223 soff = xoff;
4224 eoff = PAGE_SIZE;
4225 } else {
4226 soff = (pageno << PAGE_SHIFT);
4227 eoff = soff + PAGE_SIZE;
4228 }
4229 if (eoff > bcount)
4230 eoff = bcount;
4231 if (soff >= eoff)
4232 return;
4233
4234 /*
4235 * Test dirty bits and adjust b_dirtyoff/end.
4236 *
4237 * If dirty pages are incorporated into the bp any prior
4238 * B_NEEDCOMMIT state (NFS) must be cleared because the
4239 * caller has not taken into account the new dirty data.
4240 *
4241 * If the page was memory mapped the dirty bits might go beyond the
4242 * end of the buffer, but we can't really make the assumption that
4243 * a file EOF straddles the buffer (even though this is the case for
4244 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4245 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4246 * This also saves some console spam.
4247 *
4248 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4249 * NFS can handle huge commits but not huge writes.
4250 */
4251 vm_page_test_dirty(m);
4252 if (m->dirty) {
4253 if ((bp->b_flags & B_NEEDCOMMIT) &&
4254 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4255 if (debug_commit)
4256 kprintf("Warning: vfs_clean_one_page: bp %p "
4257 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4258 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4259 "doff/end %d %d\n",
4260 bp, (uintmax_t)bp->b_loffset, bp->b_bcount,
4261 bp->b_flags, bp->b_cmd,
4262 m->valid, m->dirty, xoff, soff, eoff,
4263 bp->b_dirtyoff, bp->b_dirtyend);
4264 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4265 if (debug_commit)
4266 print_backtrace(-1);
4267 }
4268 /*
4269 * Only clear the pmap modified bits if ALL the dirty bits
4270 * are set, otherwise the system might mis-clear portions
4271 * of a page.
4272 */
4273 if (m->dirty == VM_PAGE_BITS_ALL &&
4274 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4275 pmap_clear_modify(m);
4276 }
4277 if (bp->b_dirtyoff > soff - xoff)
4278 bp->b_dirtyoff = soff - xoff;
4279 if (bp->b_dirtyend < eoff - xoff)
4280 bp->b_dirtyend = eoff - xoff;
4281 }
4282
4283 /*
4284 * Set related valid bits, clear related dirty bits.
4285 * Does not mess with the pmap modified bit.
4286 *
4287 * WARNING! We cannot just clear all of m->dirty here as the
4288 * buffer cache buffers may use a DEV_BSIZE'd aligned
4289 * block size, or have an odd size (e.g. NFS at file EOF).
4290 * The putpages code can clear m->dirty to 0.
4291 *
4292 * If a VOP_WRITE generates a buffer cache buffer which
4293 * covers the same space as mapped writable pages the
4294 * buffer flush might not be able to clear all the dirty
4295 * bits and still require a putpages from the VM system
4296 * to finish it off.
4297 *
4298 * WARNING! vm_page_set_validclean() currently assumes vm_token
4299 * is held. The page might not be busied (bdwrite() case).
4300 * XXX remove this comment once we've validated that this
4301 * is no longer an issue.
4302 */
4303 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4304 }
4305
4306 #if 0
4307 /*
4308 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4309 * The page data is assumed to be valid (there is no zeroing here).
4310 */
4311 static void
4312 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4313 {
4314 int bcount;
4315 int xoff;
4316 int soff;
4317 int eoff;
4318
4319 /*
4320 * Calculate offset range within the page but relative to buffer's
4321 * loffset. loffset might be offset into the first page.
4322 */
4323 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4324 bcount = bp->b_bcount + xoff; /* offset adjusted */
4325
4326 if (pageno == 0) {
4327 soff = xoff;
4328 eoff = PAGE_SIZE;
4329 } else {
4330 soff = (pageno << PAGE_SHIFT);
4331 eoff = soff + PAGE_SIZE;
4332 }
4333 if (eoff > bcount)
4334 eoff = bcount;
4335 if (soff >= eoff)
4336 return;
4337 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4338 }
4339 #endif
4340
4341 /*
4342 * vfs_bio_clrbuf:
4343 *
4344 * Clear a buffer. This routine essentially fakes an I/O, so we need
4345 * to clear B_ERROR and B_INVAL.
4346 *
4347 * Note that while we only theoretically need to clear through b_bcount,
4348 * we go ahead and clear through b_bufsize.
4349 */
4350
4351 void
4352 vfs_bio_clrbuf(struct buf *bp)
4353 {
4354 int i, mask = 0;
4355 caddr_t sa, ea;
4356 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4357 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4358 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4359 (bp->b_loffset & PAGE_MASK) == 0) {
4360 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4361 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4362 bp->b_resid = 0;
4363 return;
4364 }
4365 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4366 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4367 bzero(bp->b_data, bp->b_bufsize);
4368 bp->b_xio.xio_pages[0]->valid |= mask;
4369 bp->b_resid = 0;
4370 return;
4371 }
4372 }
4373 sa = bp->b_data;
4374 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4375 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4376 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4377 ea = (caddr_t)(vm_offset_t)ulmin(
4378 (u_long)(vm_offset_t)ea,
4379 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4380 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4381 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4382 continue;
4383 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4384 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4385 bzero(sa, ea - sa);
4386 }
4387 } else {
4388 for (; sa < ea; sa += DEV_BSIZE, j++) {
4389 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4390 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4391 bzero(sa, DEV_BSIZE);
4392 }
4393 }
4394 bp->b_xio.xio_pages[i]->valid |= mask;
4395 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4396 }
4397 bp->b_resid = 0;
4398 } else {
4399 clrbuf(bp);
4400 }
4401 }
4402
4403 /*
4404 * vm_hold_load_pages:
4405 *
4406 * Load pages into the buffer's address space. The pages are
4407 * allocated from the kernel object in order to reduce interference
4408 * with the any VM paging I/O activity. The range of loaded
4409 * pages will be wired.
4410 *
4411 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4412 * retrieve the full range (to - from) of pages.
4413 */
4414 void
4415 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4416 {
4417 vm_offset_t pg;
4418 vm_page_t p;
4419 int index;
4420
4421 to = round_page(to);
4422 from = round_page(from);
4423 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4424
4425 pg = from;
4426 while (pg < to) {
4427 /*
4428 * Note: must allocate system pages since blocking here
4429 * could intefere with paging I/O, no matter which
4430 * process we are.
4431 */
4432 vm_object_hold(&kernel_object);
4433 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT,
4434 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4435 vm_object_drop(&kernel_object);
4436 if (p) {
4437 vm_page_wire(p);
4438 p->valid = VM_PAGE_BITS_ALL;
4439 vm_page_flag_clear(p, PG_ZERO);
4440 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4441 bp->b_xio.xio_pages[index] = p;
4442 vm_page_wakeup(p);
4443
4444 pg += PAGE_SIZE;
4445 ++index;
4446 }
4447 }
4448 bp->b_xio.xio_npages = index;
4449 }
4450
4451 /*
4452 * Allocate a page for a buffer cache buffer.
4453 *
4454 * If NULL is returned the caller is expected to retry (typically check if
4455 * the page already exists on retry before trying to allocate one).
4456 *
4457 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This
4458 * function will use the system reserve with the hope that the page
4459 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller
4460 * is done with the buffer.
4461 *
4462 * NOTE! However, TMPFS is a special case because flushing a dirty buffer
4463 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon
4464 * is capable of retiring pages (to swap). For TMPFS we don't dig
4465 * into the system reserve because doing so could stall out pretty
4466 * much every process running on the system.
4467 */
4468 static
4469 vm_page_t
4470 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit)
4471 {
4472 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK;
4473 vm_page_t p;
4474
4475 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4476
4477 /*
4478 * Try a normal allocation first.
4479 */
4480 p = vm_page_alloc(obj, pg, vmflags);
4481 if (p)
4482 return(p);
4483 if (vm_page_lookup(obj, pg))
4484 return(NULL);
4485 vm_pageout_deficit += deficit;
4486
4487 /*
4488 * Try again, digging into the system reserve.
4489 *
4490 * Trying to recover pages from the buffer cache here can deadlock
4491 * against other threads trying to busy underlying pages so we
4492 * depend on the code in brelse() and bqrelse() to free/cache the
4493 * underlying buffer cache pages when memory is low.
4494 */
4495 if (curthread->td_flags & TDF_SYSTHREAD)
4496 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
4497 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS)
4498 vmflags |= 0;
4499 else
4500 vmflags |= VM_ALLOC_SYSTEM;
4501
4502 /*recoverbufpages();*/
4503 p = vm_page_alloc(obj, pg, vmflags);
4504 if (p)
4505 return(p);
4506 if (vm_page_lookup(obj, pg))
4507 return(NULL);
4508
4509 /*
4510 * Wait for memory to free up and try again
4511 */
4512 if (vm_page_count_severe())
4513 ++lowmempgallocs;
4514 vm_wait(hz / 20 + 1);
4515
4516 p = vm_page_alloc(obj, pg, vmflags);
4517 if (p)
4518 return(p);
4519 if (vm_page_lookup(obj, pg))
4520 return(NULL);
4521
4522 /*
4523 * Ok, now we are really in trouble.
4524 */
4525 {
4526 static struct krate biokrate = { .freq = 1 };
4527 krateprintf(&biokrate,
4528 "Warning: bio_page_alloc: memory exhausted "
4529 "during bufcache page allocation from %s\n",
4530 curthread->td_comm);
4531 }
4532 if (curthread->td_flags & TDF_SYSTHREAD)
4533 vm_wait(hz / 20 + 1);
4534 else
4535 vm_wait(hz / 2 + 1);
4536 return (NULL);
4537 }
4538
4539 /*
4540 * vm_hold_free_pages:
4541 *
4542 * Return pages associated with the buffer back to the VM system.
4543 *
4544 * The range of pages underlying the buffer's address space will
4545 * be unmapped and un-wired.
4546 */
4547 void
4548 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4549 {
4550 vm_offset_t pg;
4551 vm_page_t p;
4552 int index, newnpages;
4553
4554 from = round_page(from);
4555 to = round_page(to);
4556 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4557 newnpages = index;
4558
4559 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4560 p = bp->b_xio.xio_pages[index];
4561 if (p && (index < bp->b_xio.xio_npages)) {
4562 if (p->busy) {
4563 kprintf("vm_hold_free_pages: doffset: %lld, "
4564 "loffset: %lld\n",
4565 (long long)bp->b_bio2.bio_offset,
4566 (long long)bp->b_loffset);
4567 }
4568 bp->b_xio.xio_pages[index] = NULL;
4569 pmap_kremove(pg);
4570 vm_page_busy_wait(p, FALSE, "vmhldpg");
4571 vm_page_unwire(p, 0);
4572 vm_page_free(p);
4573 }
4574 }
4575 bp->b_xio.xio_npages = newnpages;
4576 }
4577
4578 /*
4579 * vmapbuf:
4580 *
4581 * Map a user buffer into KVM via a pbuf. On return the buffer's
4582 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4583 * initialized.
4584 */
4585 int
4586 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4587 {
4588 caddr_t addr;
4589 vm_offset_t va;
4590 vm_page_t m;
4591 int vmprot;
4592 int error;
4593 int pidx;
4594 int i;
4595
4596 /*
4597 * bp had better have a command and it better be a pbuf.
4598 */
4599 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4600 KKASSERT(bp->b_flags & B_PAGING);
4601 KKASSERT(bp->b_kvabase);
4602
4603 if (bytes < 0)
4604 return (-1);
4605
4606 /*
4607 * Map the user data into KVM. Mappings have to be page-aligned.
4608 */
4609 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4610 pidx = 0;
4611
4612 vmprot = VM_PROT_READ;
4613 if (bp->b_cmd == BUF_CMD_READ)
4614 vmprot |= VM_PROT_WRITE;
4615
4616 while (addr < udata + bytes) {
4617 /*
4618 * Do the vm_fault if needed; do the copy-on-write thing
4619 * when reading stuff off device into memory.
4620 *
4621 * vm_fault_page*() returns a held VM page.
4622 */
4623 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4624 va = trunc_page(va);
4625
4626 m = vm_fault_page_quick(va, vmprot, &error);
4627 if (m == NULL) {
4628 for (i = 0; i < pidx; ++i) {
4629 vm_page_unhold(bp->b_xio.xio_pages[i]);
4630 bp->b_xio.xio_pages[i] = NULL;
4631 }
4632 return(-1);
4633 }
4634 bp->b_xio.xio_pages[pidx] = m;
4635 addr += PAGE_SIZE;
4636 ++pidx;
4637 }
4638
4639 /*
4640 * Map the page array and set the buffer fields to point to
4641 * the mapped data buffer.
4642 */
4643 if (pidx > btoc(MAXPHYS))
4644 panic("vmapbuf: mapped more than MAXPHYS");
4645 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4646
4647 bp->b_xio.xio_npages = pidx;
4648 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4649 bp->b_bcount = bytes;
4650 bp->b_bufsize = bytes;
4651 return(0);
4652 }
4653
4654 /*
4655 * vunmapbuf:
4656 *
4657 * Free the io map PTEs associated with this IO operation.
4658 * We also invalidate the TLB entries and restore the original b_addr.
4659 */
4660 void
4661 vunmapbuf(struct buf *bp)
4662 {
4663 int pidx;
4664 int npages;
4665
4666 KKASSERT(bp->b_flags & B_PAGING);
4667
4668 npages = bp->b_xio.xio_npages;
4669 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4670 for (pidx = 0; pidx < npages; ++pidx) {
4671 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4672 bp->b_xio.xio_pages[pidx] = NULL;
4673 }
4674 bp->b_xio.xio_npages = 0;
4675 bp->b_data = bp->b_kvabase;
4676 }
4677
4678 /*
4679 * Scan all buffers in the system and issue the callback.
4680 */
4681 int
4682 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4683 {
4684 int count = 0;
4685 int error;
4686 long n;
4687
4688 for (n = 0; n < nbuf; ++n) {
4689 if ((error = callback(&buf[n], info)) < 0) {
4690 count = error;
4691 break;
4692 }
4693 count += error;
4694 }
4695 return (count);
4696 }
4697
4698 /*
4699 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4700 * completion to the master buffer.
4701 */
4702 static void
4703 nestiobuf_iodone(struct bio *bio)
4704 {
4705 struct bio *mbio;
4706 struct buf *mbp, *bp;
4707 struct devstat *stats;
4708 int error;
4709 int donebytes;
4710
4711 bp = bio->bio_buf;
4712 mbio = bio->bio_caller_info1.ptr;
4713 stats = bio->bio_caller_info2.ptr;
4714 mbp = mbio->bio_buf;
4715
4716 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4717 KKASSERT(mbp != bp);
4718
4719 error = bp->b_error;
4720 if (bp->b_error == 0 &&
4721 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4722 /*
4723 * Not all got transfered, raise an error. We have no way to
4724 * propagate these conditions to mbp.
4725 */
4726 error = EIO;
4727 }
4728
4729 donebytes = bp->b_bufsize;
4730
4731 relpbuf(bp, NULL);
4732
4733 nestiobuf_done(mbio, donebytes, error, stats);
4734 }
4735
4736 void
4737 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4738 {
4739 struct buf *mbp;
4740
4741 mbp = mbio->bio_buf;
4742
4743 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4744
4745 /*
4746 * If an error occured, propagate it to the master buffer.
4747 *
4748 * Several biodone()s may wind up running concurrently so
4749 * use an atomic op to adjust b_flags.
4750 */
4751 if (error) {
4752 mbp->b_error = error;
4753 atomic_set_int(&mbp->b_flags, B_ERROR);
4754 }
4755
4756 /*
4757 * Decrement the operations in progress counter and terminate the
4758 * I/O if this was the last bit.
4759 */
4760 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4761 mbp->b_resid = 0;
4762 if (stats)
4763 devstat_end_transaction_buf(stats, mbp);
4764 biodone(mbio);
4765 }
4766 }
4767
4768 /*
4769 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4770 * the mbio from being biodone()'d while we are still adding sub-bios to
4771 * it.
4772 */
4773 void
4774 nestiobuf_init(struct bio *bio)
4775 {
4776 bio->bio_driver_info = (void *)1;
4777 }
4778
4779 /*
4780 * The BIOs added to the nestedio have already been started, remove the
4781 * count that placeheld our mbio and biodone() it if the count would
4782 * transition to 0.
4783 */
4784 void
4785 nestiobuf_start(struct bio *mbio)
4786 {
4787 struct buf *mbp = mbio->bio_buf;
4788
4789 /*
4790 * Decrement the operations in progress counter and terminate the
4791 * I/O if this was the last bit.
4792 */
4793 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4794 if (mbp->b_flags & B_ERROR)
4795 mbp->b_resid = mbp->b_bcount;
4796 else
4797 mbp->b_resid = 0;
4798 biodone(mbio);
4799 }
4800 }
4801
4802 /*
4803 * Set an intermediate error prior to calling nestiobuf_start()
4804 */
4805 void
4806 nestiobuf_error(struct bio *mbio, int error)
4807 {
4808 struct buf *mbp = mbio->bio_buf;
4809
4810 if (error) {
4811 mbp->b_error = error;
4812 atomic_set_int(&mbp->b_flags, B_ERROR);
4813 }
4814 }
4815
4816 /*
4817 * nestiobuf_add: setup a "nested" buffer.
4818 *
4819 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4820 * => 'bp' should be a buffer allocated by getiobuf.
4821 * => 'offset' is a byte offset in the master buffer.
4822 * => 'size' is a size in bytes of this nested buffer.
4823 */
4824 void
4825 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4826 {
4827 struct buf *mbp = mbio->bio_buf;
4828 struct vnode *vp = mbp->b_vp;
4829
4830 KKASSERT(mbp->b_bcount >= offset + size);
4831
4832 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4833
4834 /* kernel needs to own the lock for it to be released in biodone */
4835 BUF_KERNPROC(bp);
4836 bp->b_vp = vp;
4837 bp->b_cmd = mbp->b_cmd;
4838 bp->b_bio1.bio_done = nestiobuf_iodone;
4839 bp->b_data = (char *)mbp->b_data + offset;
4840 bp->b_resid = bp->b_bcount = size;
4841 bp->b_bufsize = bp->b_bcount;
4842
4843 bp->b_bio1.bio_track = NULL;
4844 bp->b_bio1.bio_caller_info1.ptr = mbio;
4845 bp->b_bio1.bio_caller_info2.ptr = stats;
4846 }
4847
4848 #ifdef DDB
4849
4850 DB_SHOW_COMMAND(buffer, db_show_buffer)
4851 {
4852 /* get args */
4853 struct buf *bp = (struct buf *)addr;
4854
4855 if (!have_addr) {
4856 db_printf("usage: show buffer <addr>\n");
4857 return;
4858 }
4859
4860 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4861 db_printf("b_cmd = %d\n", bp->b_cmd);
4862 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4863 "b_resid = %d\n, b_data = %p, "
4864 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4865 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4866 bp->b_data,
4867 (long long)bp->b_bio2.bio_offset,
4868 (long long)(bp->b_bio2.bio_next ?
4869 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4870 if (bp->b_xio.xio_npages) {
4871 int i;
4872 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4873 bp->b_xio.xio_npages);
4874 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4875 vm_page_t m;
4876 m = bp->b_xio.xio_pages[i];
4877 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4878 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4879 if ((i + 1) < bp->b_xio.xio_npages)
4880 db_printf(",");
4881 }
4882 db_printf("\n");
4883 }
4884 }
4885 #endif /* DDB */
Cache object: 26ea41e6f01de8de41b5099cde1e501f
|