1 /* $NetBSD: vfs_bio.c,v 1.303 2022/03/30 14:54:29 riastradh Exp $ */
2
3 /*-
4 * Copyright (c) 2007, 2008, 2009, 2019, 2020 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Andrew Doran, and by Wasabi Systems, Inc.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
30 */
31
32 /*-
33 * Copyright (c) 1982, 1986, 1989, 1993
34 * The Regents of the University of California. All rights reserved.
35 * (c) UNIX System Laboratories, Inc.
36 * All or some portions of this file are derived from material licensed
37 * to the University of California by American Telephone and Telegraph
38 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
39 * the permission of UNIX System Laboratories, Inc.
40 *
41 * Redistribution and use in source and binary forms, with or without
42 * modification, are permitted provided that the following conditions
43 * are met:
44 * 1. Redistributions of source code must retain the above copyright
45 * notice, this list of conditions and the following disclaimer.
46 * 2. Redistributions in binary form must reproduce the above copyright
47 * notice, this list of conditions and the following disclaimer in the
48 * documentation and/or other materials provided with the distribution.
49 * 3. Neither the name of the University nor the names of its contributors
50 * may be used to endorse or promote products derived from this software
51 * without specific prior written permission.
52 *
53 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63 * SUCH DAMAGE.
64 *
65 * @(#)vfs_bio.c 8.6 (Berkeley) 1/11/94
66 */
67
68 /*-
69 * Copyright (c) 1994 Christopher G. Demetriou
70 *
71 * Redistribution and use in source and binary forms, with or without
72 * modification, are permitted provided that the following conditions
73 * are met:
74 * 1. Redistributions of source code must retain the above copyright
75 * notice, this list of conditions and the following disclaimer.
76 * 2. Redistributions in binary form must reproduce the above copyright
77 * notice, this list of conditions and the following disclaimer in the
78 * documentation and/or other materials provided with the distribution.
79 * 3. All advertising materials mentioning features or use of this software
80 * must display the following acknowledgement:
81 * This product includes software developed by the University of
82 * California, Berkeley and its contributors.
83 * 4. Neither the name of the University nor the names of its contributors
84 * may be used to endorse or promote products derived from this software
85 * without specific prior written permission.
86 *
87 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
88 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
89 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
90 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
91 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
92 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
93 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
94 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
95 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
96 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
97 * SUCH DAMAGE.
98 *
99 * @(#)vfs_bio.c 8.6 (Berkeley) 1/11/94
100 */
101
102 /*
103 * The buffer cache subsystem.
104 *
105 * Some references:
106 * Bach: The Design of the UNIX Operating System (Prentice Hall, 1986)
107 * Leffler, et al.: The Design and Implementation of the 4.3BSD
108 * UNIX Operating System (Addison Welley, 1989)
109 *
110 * Locking
111 *
112 * There are three locks:
113 * - bufcache_lock: protects global buffer cache state.
114 * - BC_BUSY: a long term per-buffer lock.
115 * - buf_t::b_objlock: lock on completion (biowait vs biodone).
116 *
117 * For buffers associated with vnodes (a most common case) b_objlock points
118 * to the vnode_t::v_interlock. Otherwise, it points to generic buffer_lock.
119 *
120 * Lock order:
121 * bufcache_lock ->
122 * buf_t::b_objlock
123 */
124
125 #include <sys/cdefs.h>
126 __KERNEL_RCSID(0, "$NetBSD: vfs_bio.c,v 1.303 2022/03/30 14:54:29 riastradh Exp $");
127
128 #ifdef _KERNEL_OPT
129 #include "opt_bufcache.h"
130 #include "opt_dtrace.h"
131 #include "opt_biohist.h"
132 #endif
133
134 #include <sys/param.h>
135 #include <sys/systm.h>
136 #include <sys/kernel.h>
137 #include <sys/proc.h>
138 #include <sys/buf.h>
139 #include <sys/vnode.h>
140 #include <sys/mount.h>
141 #include <sys/resourcevar.h>
142 #include <sys/sysctl.h>
143 #include <sys/conf.h>
144 #include <sys/kauth.h>
145 #include <sys/fstrans.h>
146 #include <sys/intr.h>
147 #include <sys/cpu.h>
148 #include <sys/wapbl.h>
149 #include <sys/bitops.h>
150 #include <sys/cprng.h>
151 #include <sys/sdt.h>
152
153 #include <uvm/uvm.h> /* extern struct uvm uvm */
154
155 #include <miscfs/specfs/specdev.h>
156
157 SDT_PROVIDER_DEFINE(io);
158
159 SDT_PROBE_DEFINE4(io, kernel, , bbusy__start,
160 "struct buf *"/*bp*/,
161 "bool"/*intr*/, "int"/*timo*/, "kmutex_t *"/*interlock*/);
162 SDT_PROBE_DEFINE5(io, kernel, , bbusy__done,
163 "struct buf *"/*bp*/,
164 "bool"/*intr*/,
165 "int"/*timo*/,
166 "kmutex_t *"/*interlock*/,
167 "int"/*error*/);
168 SDT_PROBE_DEFINE0(io, kernel, , getnewbuf__start);
169 SDT_PROBE_DEFINE1(io, kernel, , getnewbuf__done, "struct buf *"/*bp*/);
170 SDT_PROBE_DEFINE3(io, kernel, , getblk__start,
171 "struct vnode *"/*vp*/, "daddr_t"/*blkno*/, "int"/*size*/);
172 SDT_PROBE_DEFINE4(io, kernel, , getblk__done,
173 "struct vnode *"/*vp*/, "daddr_t"/*blkno*/, "int"/*size*/,
174 "struct buf *"/*bp*/);
175 SDT_PROBE_DEFINE2(io, kernel, , brelse, "struct buf *"/*bp*/, "int"/*set*/);
176 SDT_PROBE_DEFINE1(io, kernel, , wait__start, "struct buf *"/*bp*/);
177 SDT_PROBE_DEFINE1(io, kernel, , wait__done, "struct buf *"/*bp*/);
178
179 #ifndef BUFPAGES
180 # define BUFPAGES 0
181 #endif
182
183 #ifdef BUFCACHE
184 # if (BUFCACHE < 5) || (BUFCACHE > 95)
185 # error BUFCACHE is not between 5 and 95
186 # endif
187 #else
188 # define BUFCACHE 15
189 #endif
190
191 u_int nbuf; /* desired number of buffer headers */
192 u_int bufpages = BUFPAGES; /* optional hardwired count */
193 u_int bufcache = BUFCACHE; /* max % of RAM to use for buffer cache */
194
195 /*
196 * Definitions for the buffer free lists.
197 */
198 #define BQUEUES 3 /* number of free buffer queues */
199
200 #define BQ_LOCKED 0 /* super-blocks &c */
201 #define BQ_LRU 1 /* lru, useful buffers */
202 #define BQ_AGE 2 /* rubbish */
203
204 struct bqueue {
205 TAILQ_HEAD(, buf) bq_queue;
206 uint64_t bq_bytes;
207 buf_t *bq_marker;
208 };
209 static struct bqueue bufqueues[BQUEUES] __cacheline_aligned;
210
211 /* Function prototypes */
212 static void buf_setwm(void);
213 static int buf_trim(void);
214 static void *bufpool_page_alloc(struct pool *, int);
215 static void bufpool_page_free(struct pool *, void *);
216 static buf_t *bio_doread(struct vnode *, daddr_t, int, int);
217 static buf_t *getnewbuf(int, int, int);
218 static int buf_lotsfree(void);
219 static int buf_canrelease(void);
220 static u_long buf_mempoolidx(u_long);
221 static u_long buf_roundsize(u_long);
222 static void *buf_alloc(size_t);
223 static void buf_mrelease(void *, size_t);
224 static void binsheadfree(buf_t *, struct bqueue *);
225 static void binstailfree(buf_t *, struct bqueue *);
226 #ifdef DEBUG
227 static int checkfreelist(buf_t *, struct bqueue *, int);
228 #endif
229 static void biointr(void *);
230 static void biodone2(buf_t *);
231 static void sysctl_kern_buf_setup(void);
232 static void sysctl_vm_buf_setup(void);
233
234 /* Initialization for biohist */
235
236 #include <sys/biohist.h>
237
238 BIOHIST_DEFINE(biohist);
239
240 void
biohist_init(void)241 biohist_init(void)
242 {
243
244 BIOHIST_INIT(biohist, BIOHIST_SIZE);
245 }
246
247 /*
248 * Definitions for the buffer hash lists.
249 */
250 #define BUFHASH(dvp, lbn) \
251 (&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash])
252 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
253 u_long bufhash;
254
255 static int bufhash_stats(struct hashstat_sysctl *, bool);
256
257 static kcondvar_t needbuffer_cv;
258
259 /*
260 * Buffer queue lock.
261 */
262 kmutex_t bufcache_lock __cacheline_aligned;
263 kmutex_t buffer_lock __cacheline_aligned;
264
265 /* Software ISR for completed transfers. */
266 static void *biodone_sih;
267
268 /* Buffer pool for I/O buffers. */
269 static pool_cache_t buf_cache;
270 static pool_cache_t bufio_cache;
271
272 #define MEMPOOL_INDEX_OFFSET (ilog2(DEV_BSIZE)) /* smallest pool is 512 bytes */
273 #define NMEMPOOLS (ilog2(MAXBSIZE) - MEMPOOL_INDEX_OFFSET + 1)
274 __CTASSERT((1 << (NMEMPOOLS + MEMPOOL_INDEX_OFFSET - 1)) == MAXBSIZE);
275
276 /* Buffer memory pools */
277 static struct pool bmempools[NMEMPOOLS];
278
279 static struct vm_map *buf_map;
280
281 /*
282 * Buffer memory pool allocator.
283 */
284 static void *
bufpool_page_alloc(struct pool * pp,int flags)285 bufpool_page_alloc(struct pool *pp, int flags)
286 {
287
288 return (void *)uvm_km_alloc(buf_map,
289 MAXBSIZE, MAXBSIZE,
290 ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT|UVM_KMF_TRYLOCK)
291 | UVM_KMF_WIRED);
292 }
293
294 static void
bufpool_page_free(struct pool * pp,void * v)295 bufpool_page_free(struct pool *pp, void *v)
296 {
297
298 uvm_km_free(buf_map, (vaddr_t)v, MAXBSIZE, UVM_KMF_WIRED);
299 }
300
301 static struct pool_allocator bufmempool_allocator = {
302 .pa_alloc = bufpool_page_alloc,
303 .pa_free = bufpool_page_free,
304 .pa_pagesz = MAXBSIZE,
305 };
306
307 /* Buffer memory management variables */
308 u_long bufmem_valimit;
309 u_long bufmem_hiwater;
310 u_long bufmem_lowater;
311 u_long bufmem;
312
313 /*
314 * MD code can call this to set a hard limit on the amount
315 * of virtual memory used by the buffer cache.
316 */
317 int
buf_setvalimit(vsize_t sz)318 buf_setvalimit(vsize_t sz)
319 {
320
321 /* We need to accommodate at least NMEMPOOLS of MAXBSIZE each */
322 if (sz < NMEMPOOLS * MAXBSIZE)
323 return EINVAL;
324
325 bufmem_valimit = sz;
326 return 0;
327 }
328
329 static void
buf_setwm(void)330 buf_setwm(void)
331 {
332
333 bufmem_hiwater = buf_memcalc();
334 /* lowater is approx. 2% of memory (with bufcache = 15) */
335 #define BUFMEM_WMSHIFT 3
336 #define BUFMEM_HIWMMIN (64 * 1024 << BUFMEM_WMSHIFT)
337 if (bufmem_hiwater < BUFMEM_HIWMMIN)
338 /* Ensure a reasonable minimum value */
339 bufmem_hiwater = BUFMEM_HIWMMIN;
340 bufmem_lowater = bufmem_hiwater >> BUFMEM_WMSHIFT;
341 }
342
343 #ifdef DEBUG
344 int debug_verify_freelist = 0;
345 static int
checkfreelist(buf_t * bp,struct bqueue * dp,int ison)346 checkfreelist(buf_t *bp, struct bqueue *dp, int ison)
347 {
348 buf_t *b;
349
350 if (!debug_verify_freelist)
351 return 1;
352
353 TAILQ_FOREACH(b, &dp->bq_queue, b_freelist) {
354 if (b == bp)
355 return ison ? 1 : 0;
356 }
357
358 return ison ? 0 : 1;
359 }
360 #endif
361
362 /*
363 * Insq/Remq for the buffer hash lists.
364 * Call with buffer queue locked.
365 */
366 static void
binsheadfree(buf_t * bp,struct bqueue * dp)367 binsheadfree(buf_t *bp, struct bqueue *dp)
368 {
369
370 KASSERT(mutex_owned(&bufcache_lock));
371 KASSERT(bp->b_freelistindex == -1);
372 TAILQ_INSERT_HEAD(&dp->bq_queue, bp, b_freelist);
373 dp->bq_bytes += bp->b_bufsize;
374 bp->b_freelistindex = dp - bufqueues;
375 }
376
377 static void
binstailfree(buf_t * bp,struct bqueue * dp)378 binstailfree(buf_t *bp, struct bqueue *dp)
379 {
380
381 KASSERT(mutex_owned(&bufcache_lock));
382 KASSERTMSG(bp->b_freelistindex == -1, "double free of buffer? "
383 "bp=%p, b_freelistindex=%d\n", bp, bp->b_freelistindex);
384 TAILQ_INSERT_TAIL(&dp->bq_queue, bp, b_freelist);
385 dp->bq_bytes += bp->b_bufsize;
386 bp->b_freelistindex = dp - bufqueues;
387 }
388
389 void
bremfree(buf_t * bp)390 bremfree(buf_t *bp)
391 {
392 struct bqueue *dp;
393 int bqidx = bp->b_freelistindex;
394
395 KASSERT(mutex_owned(&bufcache_lock));
396
397 KASSERT(bqidx != -1);
398 dp = &bufqueues[bqidx];
399 KDASSERT(checkfreelist(bp, dp, 1));
400 KASSERT(dp->bq_bytes >= bp->b_bufsize);
401 TAILQ_REMOVE(&dp->bq_queue, bp, b_freelist);
402 dp->bq_bytes -= bp->b_bufsize;
403
404 /* For the sysctl helper. */
405 if (bp == dp->bq_marker)
406 dp->bq_marker = NULL;
407
408 #if defined(DIAGNOSTIC)
409 bp->b_freelistindex = -1;
410 #endif /* defined(DIAGNOSTIC) */
411 }
412
413 /*
414 * note that for some ports this is used by pmap bootstrap code to
415 * determine kva size.
416 */
417 u_long
buf_memcalc(void)418 buf_memcalc(void)
419 {
420 u_long n;
421 vsize_t mapsz = 0;
422
423 /*
424 * Determine the upper bound of memory to use for buffers.
425 *
426 * - If bufpages is specified, use that as the number
427 * pages.
428 *
429 * - Otherwise, use bufcache as the percentage of
430 * physical memory.
431 */
432 if (bufpages != 0) {
433 n = bufpages;
434 } else {
435 if (bufcache < 5) {
436 printf("forcing bufcache %d -> 5", bufcache);
437 bufcache = 5;
438 }
439 if (bufcache > 95) {
440 printf("forcing bufcache %d -> 95", bufcache);
441 bufcache = 95;
442 }
443 if (buf_map != NULL)
444 mapsz = vm_map_max(buf_map) - vm_map_min(buf_map);
445 n = calc_cache_size(mapsz, bufcache,
446 (buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
447 / PAGE_SIZE;
448 }
449
450 n <<= PAGE_SHIFT;
451 if (bufmem_valimit != 0 && n > bufmem_valimit)
452 n = bufmem_valimit;
453
454 return (n);
455 }
456
457 /*
458 * Initialize buffers and hash links for buffers.
459 */
460 void
bufinit(void)461 bufinit(void)
462 {
463 struct bqueue *dp;
464 int use_std;
465 u_int i;
466
467 biodone_vfs = biodone;
468
469 mutex_init(&bufcache_lock, MUTEX_DEFAULT, IPL_NONE);
470 mutex_init(&buffer_lock, MUTEX_DEFAULT, IPL_NONE);
471 cv_init(&needbuffer_cv, "needbuf");
472
473 if (bufmem_valimit != 0) {
474 vaddr_t minaddr = 0, maxaddr;
475 buf_map = uvm_km_suballoc(kernel_map, &minaddr, &maxaddr,
476 bufmem_valimit, 0, false, 0);
477 if (buf_map == NULL)
478 panic("bufinit: cannot allocate submap");
479 } else
480 buf_map = kernel_map;
481
482 /*
483 * Initialize buffer cache memory parameters.
484 */
485 bufmem = 0;
486 buf_setwm();
487
488 /* On "small" machines use small pool page sizes where possible */
489 use_std = (physmem < atop(16*1024*1024));
490
491 /*
492 * Also use them on systems that can map the pool pages using
493 * a direct-mapped segment.
494 */
495 #ifdef PMAP_MAP_POOLPAGE
496 use_std = 1;
497 #endif
498
499 buf_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
500 "bufpl", NULL, IPL_SOFTBIO, NULL, NULL, NULL);
501 bufio_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
502 "biopl", NULL, IPL_BIO, NULL, NULL, NULL);
503
504 for (i = 0; i < NMEMPOOLS; i++) {
505 struct pool_allocator *pa;
506 struct pool *pp = &bmempools[i];
507 u_int size = 1 << (i + MEMPOOL_INDEX_OFFSET);
508 char *name = kmem_alloc(8, KM_SLEEP); /* XXX: never freed */
509 if (__predict_false(size >= 1048576))
510 (void)snprintf(name, 8, "buf%um", size / 1048576);
511 else if (__predict_true(size >= 1024))
512 (void)snprintf(name, 8, "buf%uk", size / 1024);
513 else
514 (void)snprintf(name, 8, "buf%ub", size);
515 pa = (size <= PAGE_SIZE && use_std)
516 ? &pool_allocator_nointr
517 : &bufmempool_allocator;
518 pool_init(pp, size, DEV_BSIZE, 0, 0, name, pa, IPL_NONE);
519 pool_setlowat(pp, 1);
520 pool_sethiwat(pp, 1);
521 }
522
523 /* Initialize the buffer queues */
524 for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
525 TAILQ_INIT(&dp->bq_queue);
526 dp->bq_bytes = 0;
527 }
528
529 /*
530 * Estimate hash table size based on the amount of memory we
531 * intend to use for the buffer cache. The average buffer
532 * size is dependent on our clients (i.e. filesystems).
533 *
534 * For now, use an empirical 3K per buffer.
535 */
536 nbuf = (bufmem_hiwater / 1024) / 3;
537 bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
538
539 sysctl_kern_buf_setup();
540 sysctl_vm_buf_setup();
541 hashstat_register("bufhash", bufhash_stats);
542 }
543
544 void
bufinit2(void)545 bufinit2(void)
546 {
547
548 biodone_sih = softint_establish(SOFTINT_BIO | SOFTINT_MPSAFE, biointr,
549 NULL);
550 if (biodone_sih == NULL)
551 panic("bufinit2: can't establish soft interrupt");
552 }
553
554 static int
buf_lotsfree(void)555 buf_lotsfree(void)
556 {
557 u_long guess;
558
559 /* Always allocate if less than the low water mark. */
560 if (bufmem < bufmem_lowater)
561 return 1;
562
563 /* Never allocate if greater than the high water mark. */
564 if (bufmem > bufmem_hiwater)
565 return 0;
566
567 /* If there's anything on the AGE list, it should be eaten. */
568 if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
569 return 0;
570
571 /*
572 * The probabily of getting a new allocation is inversely
573 * proportional to the current size of the cache above
574 * the low water mark. Divide the total first to avoid overflows
575 * in the product.
576 */
577 guess = cprng_fast32() % 16;
578
579 if ((bufmem_hiwater - bufmem_lowater) / 16 * guess >=
580 (bufmem - bufmem_lowater))
581 return 1;
582
583 /* Otherwise don't allocate. */
584 return 0;
585 }
586
587 /*
588 * Return estimate of bytes we think need to be
589 * released to help resolve low memory conditions.
590 *
591 * => called with bufcache_lock held.
592 */
593 static int
buf_canrelease(void)594 buf_canrelease(void)
595 {
596 int pagedemand, ninvalid = 0;
597
598 KASSERT(mutex_owned(&bufcache_lock));
599
600 if (bufmem < bufmem_lowater)
601 return 0;
602
603 if (bufmem > bufmem_hiwater)
604 return bufmem - bufmem_hiwater;
605
606 ninvalid += bufqueues[BQ_AGE].bq_bytes;
607
608 pagedemand = uvmexp.freetarg - uvm_availmem(false);
609 if (pagedemand < 0)
610 return ninvalid;
611 return MAX(ninvalid, MIN(2 * MAXBSIZE,
612 MIN((bufmem - bufmem_lowater) / 16, pagedemand * PAGE_SIZE)));
613 }
614
615 /*
616 * Buffer memory allocation helper functions
617 */
618 static u_long
buf_mempoolidx(u_long size)619 buf_mempoolidx(u_long size)
620 {
621 u_int n = 0;
622
623 size -= 1;
624 size >>= MEMPOOL_INDEX_OFFSET;
625 while (size) {
626 size >>= 1;
627 n += 1;
628 }
629 if (n >= NMEMPOOLS)
630 panic("buf mem pool index %d", n);
631 return n;
632 }
633
634 static u_long
buf_roundsize(u_long size)635 buf_roundsize(u_long size)
636 {
637 /* Round up to nearest power of 2 */
638 return (1 << (buf_mempoolidx(size) + MEMPOOL_INDEX_OFFSET));
639 }
640
641 static void *
buf_alloc(size_t size)642 buf_alloc(size_t size)
643 {
644 u_int n = buf_mempoolidx(size);
645 void *addr;
646
647 while (1) {
648 addr = pool_get(&bmempools[n], PR_NOWAIT);
649 if (addr != NULL)
650 break;
651
652 /* No memory, see if we can free some. If so, try again */
653 mutex_enter(&bufcache_lock);
654 if (buf_drain(1) > 0) {
655 mutex_exit(&bufcache_lock);
656 continue;
657 }
658
659 if (curlwp == uvm.pagedaemon_lwp) {
660 mutex_exit(&bufcache_lock);
661 return NULL;
662 }
663
664 /* Wait for buffers to arrive on the LRU queue */
665 cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
666 mutex_exit(&bufcache_lock);
667 }
668
669 return addr;
670 }
671
672 static void
buf_mrelease(void * addr,size_t size)673 buf_mrelease(void *addr, size_t size)
674 {
675
676 pool_put(&bmempools[buf_mempoolidx(size)], addr);
677 }
678
679 /*
680 * bread()/breadn() helper.
681 */
682 static buf_t *
bio_doread(struct vnode * vp,daddr_t blkno,int size,int async)683 bio_doread(struct vnode *vp, daddr_t blkno, int size, int async)
684 {
685 buf_t *bp;
686 struct mount *mp;
687
688 bp = getblk(vp, blkno, size, 0, 0);
689
690 /*
691 * getblk() may return NULL if we are the pagedaemon.
692 */
693 if (bp == NULL) {
694 KASSERT(curlwp == uvm.pagedaemon_lwp);
695 return NULL;
696 }
697
698 /*
699 * If buffer does not have data valid, start a read.
700 * Note that if buffer is BC_INVAL, getblk() won't return it.
701 * Therefore, it's valid if its I/O has completed or been delayed.
702 */
703 if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
704 /* Start I/O for the buffer. */
705 SET(bp->b_flags, B_READ | async);
706 if (async)
707 BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
708 else
709 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
710 VOP_STRATEGY(vp, bp);
711
712 /* Pay for the read. */
713 curlwp->l_ru.ru_inblock++;
714 } else if (async)
715 brelse(bp, 0);
716
717 if (vp->v_type == VBLK)
718 mp = spec_node_getmountedfs(vp);
719 else
720 mp = vp->v_mount;
721
722 /*
723 * Collect statistics on synchronous and asynchronous reads.
724 * Reads from block devices are charged to their associated
725 * filesystem (if any).
726 */
727 if (mp != NULL) {
728 if (async == 0)
729 mp->mnt_stat.f_syncreads++;
730 else
731 mp->mnt_stat.f_asyncreads++;
732 }
733
734 return (bp);
735 }
736
737 /*
738 * Read a disk block.
739 * This algorithm described in Bach (p.54).
740 */
741 int
bread(struct vnode * vp,daddr_t blkno,int size,int flags,buf_t ** bpp)742 bread(struct vnode *vp, daddr_t blkno, int size, int flags, buf_t **bpp)
743 {
744 buf_t *bp;
745 int error;
746
747 BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
748
749 /* Get buffer for block. */
750 bp = *bpp = bio_doread(vp, blkno, size, 0);
751 if (bp == NULL)
752 return ENOMEM;
753
754 /* Wait for the read to complete, and return result. */
755 error = biowait(bp);
756 if (error == 0 && (flags & B_MODIFY) != 0)
757 error = fscow_run(bp, true);
758 if (error) {
759 brelse(bp, 0);
760 *bpp = NULL;
761 }
762
763 return error;
764 }
765
766 /*
767 * Read-ahead multiple disk blocks. The first is sync, the rest async.
768 * Trivial modification to the breada algorithm presented in Bach (p.55).
769 */
770 int
breadn(struct vnode * vp,daddr_t blkno,int size,daddr_t * rablks,int * rasizes,int nrablks,int flags,buf_t ** bpp)771 breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
772 int *rasizes, int nrablks, int flags, buf_t **bpp)
773 {
774 buf_t *bp;
775 int error, i;
776
777 BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
778
779 bp = *bpp = bio_doread(vp, blkno, size, 0);
780 if (bp == NULL)
781 return ENOMEM;
782
783 /*
784 * For each of the read-ahead blocks, start a read, if necessary.
785 */
786 mutex_enter(&bufcache_lock);
787 for (i = 0; i < nrablks; i++) {
788 /* If it's in the cache, just go on to next one. */
789 if (incore(vp, rablks[i]))
790 continue;
791
792 /* Get a buffer for the read-ahead block */
793 mutex_exit(&bufcache_lock);
794 (void) bio_doread(vp, rablks[i], rasizes[i], B_ASYNC);
795 mutex_enter(&bufcache_lock);
796 }
797 mutex_exit(&bufcache_lock);
798
799 /* Otherwise, we had to start a read for it; wait until it's valid. */
800 error = biowait(bp);
801 if (error == 0 && (flags & B_MODIFY) != 0)
802 error = fscow_run(bp, true);
803 if (error) {
804 brelse(bp, 0);
805 *bpp = NULL;
806 }
807
808 return error;
809 }
810
811 /*
812 * Block write. Described in Bach (p.56)
813 */
814 int
bwrite(buf_t * bp)815 bwrite(buf_t *bp)
816 {
817 int rv, sync, wasdelayed;
818 struct vnode *vp;
819 struct mount *mp;
820
821 BIOHIST_FUNC(__func__); BIOHIST_CALLARGS(biohist, "bp=%#jx",
822 (uintptr_t)bp, 0, 0, 0);
823
824 KASSERT(ISSET(bp->b_cflags, BC_BUSY));
825 KASSERT(!cv_has_waiters(&bp->b_done));
826
827 vp = bp->b_vp;
828
829 /*
830 * dholland 20160728 AFAICT vp==NULL must be impossible as it
831 * will crash upon reaching VOP_STRATEGY below... see further
832 * analysis on tech-kern.
833 */
834 KASSERTMSG(vp != NULL, "bwrite given buffer with null vnode");
835
836 if (vp != NULL) {
837 KASSERT(bp->b_objlock == vp->v_interlock);
838 if (vp->v_type == VBLK)
839 mp = spec_node_getmountedfs(vp);
840 else
841 mp = vp->v_mount;
842 } else {
843 mp = NULL;
844 }
845
846 if (mp && mp->mnt_wapbl) {
847 if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
848 bdwrite(bp);
849 return 0;
850 }
851 }
852
853 /*
854 * Remember buffer type, to switch on it later. If the write was
855 * synchronous, but the file system was mounted with MNT_ASYNC,
856 * convert it to a delayed write.
857 * XXX note that this relies on delayed tape writes being converted
858 * to async, not sync writes (which is safe, but ugly).
859 */
860 sync = !ISSET(bp->b_flags, B_ASYNC);
861 if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
862 bdwrite(bp);
863 return (0);
864 }
865
866 /*
867 * Collect statistics on synchronous and asynchronous writes.
868 * Writes to block devices are charged to their associated
869 * filesystem (if any).
870 */
871 if (mp != NULL) {
872 if (sync)
873 mp->mnt_stat.f_syncwrites++;
874 else
875 mp->mnt_stat.f_asyncwrites++;
876 }
877
878 /*
879 * Pay for the I/O operation and make sure the buf is on the correct
880 * vnode queue.
881 */
882 bp->b_error = 0;
883 wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
884 CLR(bp->b_flags, B_READ);
885 if (wasdelayed) {
886 mutex_enter(&bufcache_lock);
887 mutex_enter(bp->b_objlock);
888 CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
889 reassignbuf(bp, bp->b_vp);
890 /* Wake anyone trying to busy the buffer via vnode's lists. */
891 cv_broadcast(&bp->b_busy);
892 mutex_exit(&bufcache_lock);
893 } else {
894 curlwp->l_ru.ru_oublock++;
895 mutex_enter(bp->b_objlock);
896 CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
897 }
898 if (vp != NULL)
899 vp->v_numoutput++;
900 mutex_exit(bp->b_objlock);
901
902 /* Initiate disk write. */
903 if (sync)
904 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
905 else
906 BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
907
908 VOP_STRATEGY(vp, bp);
909
910 if (sync) {
911 /* If I/O was synchronous, wait for it to complete. */
912 rv = biowait(bp);
913
914 /* Release the buffer. */
915 brelse(bp, 0);
916
917 return (rv);
918 } else {
919 return (0);
920 }
921 }
922
923 int
vn_bwrite(void * v)924 vn_bwrite(void *v)
925 {
926 struct vop_bwrite_args *ap = v;
927
928 return (bwrite(ap->a_bp));
929 }
930
931 /*
932 * Delayed write.
933 *
934 * The buffer is marked dirty, but is not queued for I/O.
935 * This routine should be used when the buffer is expected
936 * to be modified again soon, typically a small write that
937 * partially fills a buffer.
938 *
939 * NB: magnetic tapes cannot be delayed; they must be
940 * written in the order that the writes are requested.
941 *
942 * Described in Leffler, et al. (pp. 208-213).
943 */
944 void
bdwrite(buf_t * bp)945 bdwrite(buf_t *bp)
946 {
947
948 BIOHIST_FUNC(__func__); BIOHIST_CALLARGS(biohist, "bp=%#jx",
949 (uintptr_t)bp, 0, 0, 0);
950
951 KASSERT(bp->b_vp == NULL || bp->b_vp->v_tag != VT_UFS ||
952 bp->b_vp->v_type == VBLK || ISSET(bp->b_flags, B_COWDONE));
953 KASSERT(ISSET(bp->b_cflags, BC_BUSY));
954 KASSERT(!cv_has_waiters(&bp->b_done));
955
956 /* If this is a tape block, write the block now. */
957 if (bdev_type(bp->b_dev) == D_TAPE) {
958 bawrite(bp);
959 return;
960 }
961
962 if (wapbl_vphaswapbl(bp->b_vp)) {
963 struct mount *mp = wapbl_vptomp(bp->b_vp);
964
965 if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
966 WAPBL_ADD_BUF(mp, bp);
967 }
968 }
969
970 /*
971 * If the block hasn't been seen before:
972 * (1) Mark it as having been seen,
973 * (2) Charge for the write,
974 * (3) Make sure it's on its vnode's correct block list.
975 */
976 KASSERT(bp->b_vp == NULL || bp->b_objlock == bp->b_vp->v_interlock);
977
978 if (!ISSET(bp->b_oflags, BO_DELWRI)) {
979 mutex_enter(&bufcache_lock);
980 mutex_enter(bp->b_objlock);
981 SET(bp->b_oflags, BO_DELWRI);
982 curlwp->l_ru.ru_oublock++;
983 reassignbuf(bp, bp->b_vp);
984 /* Wake anyone trying to busy the buffer via vnode's lists. */
985 cv_broadcast(&bp->b_busy);
986 mutex_exit(&bufcache_lock);
987 } else {
988 mutex_enter(bp->b_objlock);
989 }
990 /* Otherwise, the "write" is done, so mark and release the buffer. */
991 CLR(bp->b_oflags, BO_DONE);
992 mutex_exit(bp->b_objlock);
993
994 brelse(bp, 0);
995 }
996
997 /*
998 * Asynchronous block write; just an asynchronous bwrite().
999 */
1000 void
bawrite(buf_t * bp)1001 bawrite(buf_t *bp)
1002 {
1003
1004 KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1005 KASSERT(bp->b_vp != NULL);
1006
1007 SET(bp->b_flags, B_ASYNC);
1008 VOP_BWRITE(bp->b_vp, bp);
1009 }
1010
1011 /*
1012 * Release a buffer on to the free lists.
1013 * Described in Bach (p. 46).
1014 */
1015 void
brelsel(buf_t * bp,int set)1016 brelsel(buf_t *bp, int set)
1017 {
1018 struct bqueue *bufq;
1019 struct vnode *vp;
1020
1021 SDT_PROBE2(io, kernel, , brelse, bp, set);
1022
1023 KASSERT(bp != NULL);
1024 KASSERT(mutex_owned(&bufcache_lock));
1025 KASSERT(!cv_has_waiters(&bp->b_done));
1026
1027 SET(bp->b_cflags, set);
1028
1029 KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1030 KASSERT(bp->b_iodone == NULL);
1031
1032 /* Wake up any processes waiting for any buffer to become free. */
1033 cv_signal(&needbuffer_cv);
1034
1035 /* Wake up any proceeses waiting for _this_ buffer to become free */
1036 if (ISSET(bp->b_cflags, BC_WANTED))
1037 CLR(bp->b_cflags, BC_WANTED|BC_AGE);
1038
1039 /* If it's clean clear the copy-on-write flag. */
1040 if (ISSET(bp->b_flags, B_COWDONE)) {
1041 mutex_enter(bp->b_objlock);
1042 if (!ISSET(bp->b_oflags, BO_DELWRI))
1043 CLR(bp->b_flags, B_COWDONE);
1044 mutex_exit(bp->b_objlock);
1045 }
1046
1047 /*
1048 * Determine which queue the buffer should be on, then put it there.
1049 */
1050
1051 /* If it's locked, don't report an error; try again later. */
1052 if (ISSET(bp->b_flags, B_LOCKED))
1053 bp->b_error = 0;
1054
1055 /* If it's not cacheable, or an error, mark it invalid. */
1056 if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
1057 SET(bp->b_cflags, BC_INVAL);
1058
1059 if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1060 /*
1061 * This is a delayed write buffer that was just flushed to
1062 * disk. It is still on the LRU queue. If it's become
1063 * invalid, then we need to move it to a different queue;
1064 * otherwise leave it in its current position.
1065 */
1066 CLR(bp->b_cflags, BC_VFLUSH);
1067 if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
1068 !ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
1069 KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
1070 goto already_queued;
1071 } else {
1072 bremfree(bp);
1073 }
1074 }
1075
1076 KDASSERT(checkfreelist(bp, &bufqueues[BQ_AGE], 0));
1077 KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 0));
1078 KDASSERT(checkfreelist(bp, &bufqueues[BQ_LOCKED], 0));
1079
1080 if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
1081 /*
1082 * If it's invalid or empty, dissociate it from its vnode
1083 * and put on the head of the appropriate queue.
1084 */
1085 if (ISSET(bp->b_flags, B_LOCKED)) {
1086 if (wapbl_vphaswapbl(vp = bp->b_vp)) {
1087 struct mount *mp = wapbl_vptomp(vp);
1088
1089 KASSERT(bp->b_iodone
1090 != mp->mnt_wapbl_op->wo_wapbl_biodone);
1091 WAPBL_REMOVE_BUF(mp, bp);
1092 }
1093 }
1094
1095 mutex_enter(bp->b_objlock);
1096 CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
1097 if ((vp = bp->b_vp) != NULL) {
1098 KASSERT(bp->b_objlock == vp->v_interlock);
1099 reassignbuf(bp, bp->b_vp);
1100 brelvp(bp);
1101 mutex_exit(vp->v_interlock);
1102 } else {
1103 KASSERT(bp->b_objlock == &buffer_lock);
1104 mutex_exit(bp->b_objlock);
1105 }
1106 /* We want to dispose of the buffer, so wake everybody. */
1107 cv_broadcast(&bp->b_busy);
1108 if (bp->b_bufsize <= 0)
1109 /* no data */
1110 goto already_queued;
1111 else
1112 /* invalid data */
1113 bufq = &bufqueues[BQ_AGE];
1114 binsheadfree(bp, bufq);
1115 } else {
1116 /*
1117 * It has valid data. Put it on the end of the appropriate
1118 * queue, so that it'll stick around for as long as possible.
1119 * If buf is AGE, but has dependencies, must put it on last
1120 * bufqueue to be scanned, ie LRU. This protects against the
1121 * livelock where BQ_AGE only has buffers with dependencies,
1122 * and we thus never get to the dependent buffers in BQ_LRU.
1123 */
1124 if (ISSET(bp->b_flags, B_LOCKED)) {
1125 /* locked in core */
1126 bufq = &bufqueues[BQ_LOCKED];
1127 } else if (!ISSET(bp->b_cflags, BC_AGE)) {
1128 /* valid data */
1129 bufq = &bufqueues[BQ_LRU];
1130 } else {
1131 /* stale but valid data */
1132 bufq = &bufqueues[BQ_AGE];
1133 }
1134 binstailfree(bp, bufq);
1135 }
1136 already_queued:
1137 /* Unlock the buffer. */
1138 CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
1139 CLR(bp->b_flags, B_ASYNC);
1140
1141 /*
1142 * Wake only the highest priority waiter on the lock, in order to
1143 * prevent a thundering herd: many LWPs simultaneously awakening and
1144 * competing for the buffer's lock. Testing in 2019 revealed this
1145 * to reduce contention on bufcache_lock tenfold during a kernel
1146 * compile. Here and elsewhere, when the buffer is changing
1147 * identity, being disposed of, or moving from one list to another,
1148 * we wake all lock requestors.
1149 */
1150 if (bp->b_bufsize <= 0) {
1151 cv_broadcast(&bp->b_busy);
1152 buf_destroy(bp);
1153 #ifdef DEBUG
1154 memset((char *)bp, 0, sizeof(*bp));
1155 #endif
1156 pool_cache_put(buf_cache, bp);
1157 } else
1158 cv_signal(&bp->b_busy);
1159 }
1160
1161 void
brelse(buf_t * bp,int set)1162 brelse(buf_t *bp, int set)
1163 {
1164
1165 mutex_enter(&bufcache_lock);
1166 brelsel(bp, set);
1167 mutex_exit(&bufcache_lock);
1168 }
1169
1170 /*
1171 * Determine if a block is in the cache.
1172 * Just look on what would be its hash chain. If it's there, return
1173 * a pointer to it, unless it's marked invalid. If it's marked invalid,
1174 * we normally don't return the buffer, unless the caller explicitly
1175 * wants us to.
1176 */
1177 buf_t *
incore(struct vnode * vp,daddr_t blkno)1178 incore(struct vnode *vp, daddr_t blkno)
1179 {
1180 buf_t *bp;
1181
1182 KASSERT(mutex_owned(&bufcache_lock));
1183
1184 /* Search hash chain */
1185 LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
1186 if (bp->b_lblkno == blkno && bp->b_vp == vp &&
1187 !ISSET(bp->b_cflags, BC_INVAL)) {
1188 KASSERT(bp->b_objlock == vp->v_interlock);
1189 return (bp);
1190 }
1191 }
1192
1193 return (NULL);
1194 }
1195
1196 /*
1197 * Get a block of requested size that is associated with
1198 * a given vnode and block offset. If it is found in the
1199 * block cache, mark it as having been found, make it busy
1200 * and return it. Otherwise, return an empty block of the
1201 * correct size. It is up to the caller to insure that the
1202 * cached blocks be of the correct size.
1203 */
1204 buf_t *
getblk(struct vnode * vp,daddr_t blkno,int size,int slpflag,int slptimeo)1205 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
1206 {
1207 int err, preserve;
1208 buf_t *bp;
1209
1210 mutex_enter(&bufcache_lock);
1211 SDT_PROBE3(io, kernel, , getblk__start, vp, blkno, size);
1212 loop:
1213 bp = incore(vp, blkno);
1214 if (bp != NULL) {
1215 err = bbusy(bp, ((slpflag & PCATCH) != 0), slptimeo, NULL);
1216 if (err != 0) {
1217 if (err == EPASSTHROUGH)
1218 goto loop;
1219 mutex_exit(&bufcache_lock);
1220 SDT_PROBE4(io, kernel, , getblk__done,
1221 vp, blkno, size, NULL);
1222 return (NULL);
1223 }
1224 KASSERT(!cv_has_waiters(&bp->b_done));
1225 #ifdef DIAGNOSTIC
1226 if (ISSET(bp->b_oflags, BO_DONE|BO_DELWRI) &&
1227 bp->b_bcount < size && vp->v_type != VBLK)
1228 panic("getblk: block size invariant failed");
1229 #endif
1230 bremfree(bp);
1231 preserve = 1;
1232 } else {
1233 if ((bp = getnewbuf(slpflag, slptimeo, 0)) == NULL)
1234 goto loop;
1235
1236 if (incore(vp, blkno) != NULL) {
1237 /* The block has come into memory in the meantime. */
1238 brelsel(bp, 0);
1239 goto loop;
1240 }
1241
1242 LIST_INSERT_HEAD(BUFHASH(vp, blkno), bp, b_hash);
1243 bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
1244 mutex_enter(vp->v_interlock);
1245 bgetvp(vp, bp);
1246 mutex_exit(vp->v_interlock);
1247 preserve = 0;
1248 }
1249 mutex_exit(&bufcache_lock);
1250
1251 /*
1252 * LFS can't track total size of B_LOCKED buffer (locked_queue_bytes)
1253 * if we re-size buffers here.
1254 */
1255 if (ISSET(bp->b_flags, B_LOCKED)) {
1256 KASSERT(bp->b_bufsize >= size);
1257 } else {
1258 if (allocbuf(bp, size, preserve)) {
1259 mutex_enter(&bufcache_lock);
1260 LIST_REMOVE(bp, b_hash);
1261 brelsel(bp, BC_INVAL);
1262 mutex_exit(&bufcache_lock);
1263 SDT_PROBE4(io, kernel, , getblk__done,
1264 vp, blkno, size, NULL);
1265 return NULL;
1266 }
1267 }
1268 BIO_SETPRIO(bp, BPRIO_DEFAULT);
1269 SDT_PROBE4(io, kernel, , getblk__done, vp, blkno, size, bp);
1270 return (bp);
1271 }
1272
1273 /*
1274 * Get an empty, disassociated buffer of given size.
1275 */
1276 buf_t *
geteblk(int size)1277 geteblk(int size)
1278 {
1279 buf_t *bp;
1280 int error __diagused;
1281
1282 mutex_enter(&bufcache_lock);
1283 while ((bp = getnewbuf(0, 0, 0)) == NULL)
1284 ;
1285
1286 SET(bp->b_cflags, BC_INVAL);
1287 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1288 mutex_exit(&bufcache_lock);
1289 BIO_SETPRIO(bp, BPRIO_DEFAULT);
1290 error = allocbuf(bp, size, 0);
1291 KASSERT(error == 0);
1292 return (bp);
1293 }
1294
1295 /*
1296 * Expand or contract the actual memory allocated to a buffer.
1297 *
1298 * If the buffer shrinks, data is lost, so it's up to the
1299 * caller to have written it out *first*; this routine will not
1300 * start a write. If the buffer grows, it's the callers
1301 * responsibility to fill out the buffer's additional contents.
1302 */
1303 int
allocbuf(buf_t * bp,int size,int preserve)1304 allocbuf(buf_t *bp, int size, int preserve)
1305 {
1306 void *addr;
1307 vsize_t oldsize, desired_size;
1308 int oldcount;
1309 int delta;
1310
1311 desired_size = buf_roundsize(size);
1312 if (desired_size > MAXBSIZE)
1313 printf("allocbuf: buffer larger than MAXBSIZE requested");
1314
1315 oldcount = bp->b_bcount;
1316
1317 bp->b_bcount = size;
1318
1319 oldsize = bp->b_bufsize;
1320 if (oldsize == desired_size) {
1321 /*
1322 * Do not short cut the WAPBL resize, as the buffer length
1323 * could still have changed and this would corrupt the
1324 * tracking of the transaction length.
1325 */
1326 goto out;
1327 }
1328
1329 /*
1330 * If we want a buffer of a different size, re-allocate the
1331 * buffer's memory; copy old content only if needed.
1332 */
1333 addr = buf_alloc(desired_size);
1334 if (addr == NULL)
1335 return ENOMEM;
1336 if (preserve)
1337 memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
1338 if (bp->b_data != NULL)
1339 buf_mrelease(bp->b_data, oldsize);
1340 bp->b_data = addr;
1341 bp->b_bufsize = desired_size;
1342
1343 /*
1344 * Update overall buffer memory counter (protected by bufcache_lock)
1345 */
1346 delta = (long)desired_size - (long)oldsize;
1347
1348 mutex_enter(&bufcache_lock);
1349 if ((bufmem += delta) > bufmem_hiwater) {
1350 /*
1351 * Need to trim overall memory usage.
1352 */
1353 while (buf_canrelease()) {
1354 if (preempt_needed()) {
1355 mutex_exit(&bufcache_lock);
1356 preempt();
1357 mutex_enter(&bufcache_lock);
1358 }
1359 if (buf_trim() == 0)
1360 break;
1361 }
1362 }
1363 mutex_exit(&bufcache_lock);
1364
1365 out:
1366 if (wapbl_vphaswapbl(bp->b_vp))
1367 WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp, oldsize, oldcount);
1368
1369 return 0;
1370 }
1371
1372 /*
1373 * Find a buffer which is available for use.
1374 * Select something from a free list.
1375 * Preference is to AGE list, then LRU list.
1376 *
1377 * Called with the buffer queues locked.
1378 * Return buffer locked.
1379 */
1380 static buf_t *
getnewbuf(int slpflag,int slptimeo,int from_bufq)1381 getnewbuf(int slpflag, int slptimeo, int from_bufq)
1382 {
1383 buf_t *bp;
1384 struct vnode *vp;
1385 struct mount *transmp = NULL;
1386
1387 SDT_PROBE0(io, kernel, , getnewbuf__start);
1388
1389 start:
1390 KASSERT(mutex_owned(&bufcache_lock));
1391
1392 /*
1393 * Get a new buffer from the pool.
1394 */
1395 if (!from_bufq && buf_lotsfree()) {
1396 mutex_exit(&bufcache_lock);
1397 bp = pool_cache_get(buf_cache, PR_NOWAIT);
1398 if (bp != NULL) {
1399 memset((char *)bp, 0, sizeof(*bp));
1400 buf_init(bp);
1401 SET(bp->b_cflags, BC_BUSY); /* mark buffer busy */
1402 mutex_enter(&bufcache_lock);
1403 #if defined(DIAGNOSTIC)
1404 bp->b_freelistindex = -1;
1405 #endif /* defined(DIAGNOSTIC) */
1406 SDT_PROBE1(io, kernel, , getnewbuf__done, bp);
1407 return (bp);
1408 }
1409 mutex_enter(&bufcache_lock);
1410 }
1411
1412 KASSERT(mutex_owned(&bufcache_lock));
1413 if ((bp = TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue)) != NULL) {
1414 KASSERT(!ISSET(bp->b_oflags, BO_DELWRI));
1415 } else {
1416 TAILQ_FOREACH(bp, &bufqueues[BQ_LRU].bq_queue, b_freelist) {
1417 if (ISSET(bp->b_cflags, BC_VFLUSH) ||
1418 !ISSET(bp->b_oflags, BO_DELWRI))
1419 break;
1420 if (fstrans_start_nowait(bp->b_vp->v_mount) == 0) {
1421 KASSERT(transmp == NULL);
1422 transmp = bp->b_vp->v_mount;
1423 break;
1424 }
1425 }
1426 }
1427 if (bp != NULL) {
1428 KASSERT(!ISSET(bp->b_cflags, BC_BUSY) || ISSET(bp->b_cflags, BC_VFLUSH));
1429 bremfree(bp);
1430
1431 /* Buffer is no longer on free lists. */
1432 SET(bp->b_cflags, BC_BUSY);
1433
1434 /* Wake anyone trying to lock the old identity. */
1435 cv_broadcast(&bp->b_busy);
1436 } else {
1437 /*
1438 * XXX: !from_bufq should be removed.
1439 */
1440 if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
1441 /* wait for a free buffer of any kind */
1442 if ((slpflag & PCATCH) != 0)
1443 (void)cv_timedwait_sig(&needbuffer_cv,
1444 &bufcache_lock, slptimeo);
1445 else
1446 (void)cv_timedwait(&needbuffer_cv,
1447 &bufcache_lock, slptimeo);
1448 }
1449 SDT_PROBE1(io, kernel, , getnewbuf__done, NULL);
1450 return (NULL);
1451 }
1452
1453 #ifdef DIAGNOSTIC
1454 if (bp->b_bufsize <= 0)
1455 panic("buffer %p: on queue but empty", bp);
1456 #endif
1457
1458 if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1459 /*
1460 * This is a delayed write buffer being flushed to disk. Make
1461 * sure it gets aged out of the queue when it's finished, and
1462 * leave it off the LRU queue.
1463 */
1464 CLR(bp->b_cflags, BC_VFLUSH);
1465 SET(bp->b_cflags, BC_AGE);
1466 goto start;
1467 }
1468
1469 KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1470 KASSERT(!cv_has_waiters(&bp->b_done));
1471
1472 /*
1473 * If buffer was a delayed write, start it and return NULL
1474 * (since we might sleep while starting the write).
1475 */
1476 if (ISSET(bp->b_oflags, BO_DELWRI)) {
1477 /*
1478 * This buffer has gone through the LRU, so make sure it gets
1479 * reused ASAP.
1480 */
1481 SET(bp->b_cflags, BC_AGE);
1482 mutex_exit(&bufcache_lock);
1483 bawrite(bp);
1484 KASSERT(transmp != NULL);
1485 fstrans_done(transmp);
1486 mutex_enter(&bufcache_lock);
1487 SDT_PROBE1(io, kernel, , getnewbuf__done, NULL);
1488 return (NULL);
1489 }
1490
1491 KASSERT(transmp == NULL);
1492
1493 vp = bp->b_vp;
1494
1495 /* clear out various other fields */
1496 bp->b_cflags = BC_BUSY;
1497 bp->b_oflags = 0;
1498 bp->b_flags = 0;
1499 bp->b_dev = NODEV;
1500 bp->b_blkno = 0;
1501 bp->b_lblkno = 0;
1502 bp->b_rawblkno = 0;
1503 bp->b_iodone = 0;
1504 bp->b_error = 0;
1505 bp->b_resid = 0;
1506 bp->b_bcount = 0;
1507
1508 LIST_REMOVE(bp, b_hash);
1509
1510 /* Disassociate us from our vnode, if we had one... */
1511 if (vp != NULL) {
1512 mutex_enter(vp->v_interlock);
1513 brelvp(bp);
1514 mutex_exit(vp->v_interlock);
1515 }
1516
1517 SDT_PROBE1(io, kernel, , getnewbuf__done, bp);
1518 return (bp);
1519 }
1520
1521 /*
1522 * Invalidate the specified buffer if it exists.
1523 */
1524 void
binvalbuf(struct vnode * vp,daddr_t blkno)1525 binvalbuf(struct vnode *vp, daddr_t blkno)
1526 {
1527 buf_t *bp;
1528 int err;
1529
1530 mutex_enter(&bufcache_lock);
1531
1532 loop:
1533 bp = incore(vp, blkno);
1534 if (bp != NULL) {
1535 err = bbusy(bp, 0, 0, NULL);
1536 if (err == EPASSTHROUGH)
1537 goto loop;
1538 bremfree(bp);
1539 if (ISSET(bp->b_oflags, BO_DELWRI)) {
1540 SET(bp->b_cflags, BC_NOCACHE);
1541 mutex_exit(&bufcache_lock);
1542 bwrite(bp);
1543 } else {
1544 brelsel(bp, BC_INVAL);
1545 mutex_exit(&bufcache_lock);
1546 }
1547 } else
1548 mutex_exit(&bufcache_lock);
1549 }
1550
1551 /*
1552 * Attempt to free an aged buffer off the queues.
1553 * Called with queue lock held.
1554 * Returns the amount of buffer memory freed.
1555 */
1556 static int
buf_trim(void)1557 buf_trim(void)
1558 {
1559 buf_t *bp;
1560 long size;
1561
1562 KASSERT(mutex_owned(&bufcache_lock));
1563
1564 /* Instruct getnewbuf() to get buffers off the queues */
1565 if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
1566 return 0;
1567
1568 KASSERT((bp->b_cflags & BC_WANTED) == 0);
1569 size = bp->b_bufsize;
1570 bufmem -= size;
1571 if (size > 0) {
1572 buf_mrelease(bp->b_data, size);
1573 bp->b_bcount = bp->b_bufsize = 0;
1574 }
1575 /* brelse() will return the buffer to the global buffer pool */
1576 brelsel(bp, 0);
1577 return size;
1578 }
1579
1580 int
buf_drain(int n)1581 buf_drain(int n)
1582 {
1583 int size = 0, sz;
1584
1585 KASSERT(mutex_owned(&bufcache_lock));
1586
1587 while (size < n && bufmem > bufmem_lowater) {
1588 sz = buf_trim();
1589 if (sz <= 0)
1590 break;
1591 size += sz;
1592 }
1593
1594 return size;
1595 }
1596
1597 /*
1598 * Wait for operations on the buffer to complete.
1599 * When they do, extract and return the I/O's error value.
1600 */
1601 int
biowait(buf_t * bp)1602 biowait(buf_t *bp)
1603 {
1604
1605 BIOHIST_FUNC(__func__);
1606
1607 KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1608
1609 SDT_PROBE1(io, kernel, , wait__start, bp);
1610
1611 mutex_enter(bp->b_objlock);
1612
1613 BIOHIST_CALLARGS(biohist, "bp=%#jx, oflags=0x%jx, ret_addr=%#jx",
1614 (uintptr_t)bp, bp->b_oflags,
1615 (uintptr_t)__builtin_return_address(0), 0);
1616
1617 while (!ISSET(bp->b_oflags, BO_DONE | BO_DELWRI)) {
1618 BIOHIST_LOG(biohist, "waiting bp=%#jx", (uintptr_t)bp, 0, 0, 0);
1619 cv_wait(&bp->b_done, bp->b_objlock);
1620 }
1621 mutex_exit(bp->b_objlock);
1622
1623 SDT_PROBE1(io, kernel, , wait__done, bp);
1624
1625 BIOHIST_LOG(biohist, "return %jd", bp->b_error, 0, 0, 0);
1626
1627 return bp->b_error;
1628 }
1629
1630 /*
1631 * Mark I/O complete on a buffer.
1632 *
1633 * If a callback has been requested, e.g. the pageout
1634 * daemon, do so. Otherwise, awaken waiting processes.
1635 *
1636 * [ Leffler, et al., says on p.247:
1637 * "This routine wakes up the blocked process, frees the buffer
1638 * for an asynchronous write, or, for a request by the pagedaemon
1639 * process, invokes a procedure specified in the buffer structure" ]
1640 *
1641 * In real life, the pagedaemon (or other system processes) wants
1642 * to do async stuff too, and doesn't want the buffer brelse()'d.
1643 * (for swap pager, that puts swap buffers on the free lists (!!!),
1644 * for the vn device, that puts allocated buffers on the free lists!)
1645 */
1646 void
biodone(buf_t * bp)1647 biodone(buf_t *bp)
1648 {
1649 int s;
1650
1651 BIOHIST_FUNC(__func__);
1652
1653 KASSERT(!ISSET(bp->b_oflags, BO_DONE));
1654
1655 if (cpu_intr_p()) {
1656 /* From interrupt mode: defer to a soft interrupt. */
1657 s = splvm();
1658 TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
1659
1660 BIOHIST_CALLARGS(biohist, "bp=%#jx, softint scheduled",
1661 (uintptr_t)bp, 0, 0, 0);
1662 softint_schedule(biodone_sih);
1663 splx(s);
1664 } else {
1665 /* Process now - the buffer may be freed soon. */
1666 biodone2(bp);
1667 }
1668 }
1669
1670 SDT_PROBE_DEFINE1(io, kernel, , done, "struct buf *"/*bp*/);
1671
1672 static void
biodone2(buf_t * bp)1673 biodone2(buf_t *bp)
1674 {
1675 void (*callout)(buf_t *);
1676
1677 SDT_PROBE1(io, kernel, ,done, bp);
1678
1679 BIOHIST_FUNC(__func__);
1680 BIOHIST_CALLARGS(biohist, "bp=%#jx", (uintptr_t)bp, 0, 0, 0);
1681
1682 mutex_enter(bp->b_objlock);
1683 /* Note that the transfer is done. */
1684 if (ISSET(bp->b_oflags, BO_DONE))
1685 panic("biodone2 already");
1686 CLR(bp->b_flags, B_COWDONE);
1687 SET(bp->b_oflags, BO_DONE);
1688 BIO_SETPRIO(bp, BPRIO_DEFAULT);
1689
1690 /* Wake up waiting writers. */
1691 if (!ISSET(bp->b_flags, B_READ))
1692 vwakeup(bp);
1693
1694 if ((callout = bp->b_iodone) != NULL) {
1695 BIOHIST_LOG(biohist, "callout %#jx", (uintptr_t)callout,
1696 0, 0, 0);
1697
1698 /* Note callout done, then call out. */
1699 KASSERT(!cv_has_waiters(&bp->b_done));
1700 bp->b_iodone = NULL;
1701 mutex_exit(bp->b_objlock);
1702 (*callout)(bp);
1703 } else if (ISSET(bp->b_flags, B_ASYNC)) {
1704 /* If async, release. */
1705 BIOHIST_LOG(biohist, "async", 0, 0, 0, 0);
1706 KASSERT(!cv_has_waiters(&bp->b_done));
1707 mutex_exit(bp->b_objlock);
1708 brelse(bp, 0);
1709 } else {
1710 /* Otherwise just wake up waiters in biowait(). */
1711 BIOHIST_LOG(biohist, "wake-up", 0, 0, 0, 0);
1712 cv_broadcast(&bp->b_done);
1713 mutex_exit(bp->b_objlock);
1714 }
1715 }
1716
1717 static void
biointr(void * cookie)1718 biointr(void *cookie)
1719 {
1720 struct cpu_info *ci;
1721 buf_t *bp;
1722 int s;
1723
1724 BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
1725
1726 ci = curcpu();
1727
1728 s = splvm();
1729 while (!TAILQ_EMPTY(&ci->ci_data.cpu_biodone)) {
1730 KASSERT(curcpu() == ci);
1731
1732 bp = TAILQ_FIRST(&ci->ci_data.cpu_biodone);
1733 TAILQ_REMOVE(&ci->ci_data.cpu_biodone, bp, b_actq);
1734 splx(s);
1735
1736 BIOHIST_LOG(biohist, "bp=%#jx", (uintptr_t)bp, 0, 0, 0);
1737 biodone2(bp);
1738
1739 s = splvm();
1740 }
1741 splx(s);
1742 }
1743
1744 static void
sysctl_fillbuf(const buf_t * i,struct buf_sysctl * o)1745 sysctl_fillbuf(const buf_t *i, struct buf_sysctl *o)
1746 {
1747 const bool allowaddr = get_expose_address(curproc);
1748
1749 memset(o, 0, sizeof(*o));
1750
1751 o->b_flags = i->b_flags | i->b_cflags | i->b_oflags;
1752 o->b_error = i->b_error;
1753 o->b_prio = i->b_prio;
1754 o->b_dev = i->b_dev;
1755 o->b_bufsize = i->b_bufsize;
1756 o->b_bcount = i->b_bcount;
1757 o->b_resid = i->b_resid;
1758 COND_SET_VALUE(o->b_addr, PTRTOUINT64(i->b_data), allowaddr);
1759 o->b_blkno = i->b_blkno;
1760 o->b_rawblkno = i->b_rawblkno;
1761 COND_SET_VALUE(o->b_iodone, PTRTOUINT64(i->b_iodone), allowaddr);
1762 COND_SET_VALUE(o->b_proc, PTRTOUINT64(i->b_proc), allowaddr);
1763 COND_SET_VALUE(o->b_vp, PTRTOUINT64(i->b_vp), allowaddr);
1764 COND_SET_VALUE(o->b_saveaddr, PTRTOUINT64(i->b_saveaddr), allowaddr);
1765 o->b_lblkno = i->b_lblkno;
1766 }
1767
1768 static int
sysctl_dobuf(SYSCTLFN_ARGS)1769 sysctl_dobuf(SYSCTLFN_ARGS)
1770 {
1771 buf_t *bp;
1772 struct buf_sysctl bs;
1773 struct bqueue *bq;
1774 char *dp;
1775 u_int i, op, arg;
1776 size_t len, needed, elem_size, out_size;
1777 int error, elem_count, retries;
1778
1779 if (namelen == 1 && name[0] == CTL_QUERY)
1780 return (sysctl_query(SYSCTLFN_CALL(rnode)));
1781
1782 if (namelen != 4)
1783 return (EINVAL);
1784
1785 retries = 100;
1786 retry:
1787 dp = oldp;
1788 len = (oldp != NULL) ? *oldlenp : 0;
1789 op = name[0];
1790 arg = name[1];
1791 elem_size = name[2];
1792 elem_count = name[3];
1793 out_size = MIN(sizeof(bs), elem_size);
1794
1795 /*
1796 * at the moment, these are just "placeholders" to make the
1797 * API for retrieving kern.buf data more extensible in the
1798 * future.
1799 *
1800 * XXX kern.buf currently has "netbsd32" issues. hopefully
1801 * these will be resolved at a later point.
1802 */
1803 if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
1804 elem_size < 1 || elem_count < 0)
1805 return (EINVAL);
1806
1807 if (oldp == NULL) {
1808 /* count only, don't run through the buffer queues */
1809 needed = pool_cache_nget(buf_cache) - pool_cache_nput(buf_cache);
1810 *oldlenp = (needed + KERN_BUFSLOP) * elem_size;
1811
1812 return 0;
1813 }
1814
1815 error = 0;
1816 needed = 0;
1817 sysctl_unlock();
1818 mutex_enter(&bufcache_lock);
1819 for (i = 0; i < BQUEUES; i++) {
1820 bq = &bufqueues[i];
1821 TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
1822 bq->bq_marker = bp;
1823 if (len >= elem_size && elem_count > 0) {
1824 sysctl_fillbuf(bp, &bs);
1825 mutex_exit(&bufcache_lock);
1826 error = copyout(&bs, dp, out_size);
1827 mutex_enter(&bufcache_lock);
1828 if (error)
1829 break;
1830 if (bq->bq_marker != bp) {
1831 /*
1832 * This sysctl node is only for
1833 * statistics. Retry; if the
1834 * queue keeps changing, then
1835 * bail out.
1836 */
1837 if (retries-- == 0) {
1838 error = EAGAIN;
1839 break;
1840 }
1841 mutex_exit(&bufcache_lock);
1842 sysctl_relock();
1843 goto retry;
1844 }
1845 dp += elem_size;
1846 len -= elem_size;
1847 }
1848 needed += elem_size;
1849 if (elem_count > 0 && elem_count != INT_MAX)
1850 elem_count--;
1851 }
1852 if (error != 0)
1853 break;
1854 }
1855 mutex_exit(&bufcache_lock);
1856 sysctl_relock();
1857
1858 *oldlenp = needed;
1859
1860 return (error);
1861 }
1862
1863 static int
sysctl_bufvm_update(SYSCTLFN_ARGS)1864 sysctl_bufvm_update(SYSCTLFN_ARGS)
1865 {
1866 int error, rv;
1867 struct sysctlnode node;
1868 unsigned int temp_bufcache;
1869 unsigned long temp_water;
1870
1871 /* Take a copy of the supplied node and its data */
1872 node = *rnode;
1873 if (node.sysctl_data == &bufcache) {
1874 node.sysctl_data = &temp_bufcache;
1875 temp_bufcache = *(unsigned int *)rnode->sysctl_data;
1876 } else {
1877 node.sysctl_data = &temp_water;
1878 temp_water = *(unsigned long *)rnode->sysctl_data;
1879 }
1880
1881 /* Update the copy */
1882 error = sysctl_lookup(SYSCTLFN_CALL(&node));
1883 if (error || newp == NULL)
1884 return (error);
1885
1886 if (rnode->sysctl_data == &bufcache) {
1887 if (temp_bufcache > 100)
1888 return (EINVAL);
1889 bufcache = temp_bufcache;
1890 buf_setwm();
1891 } else if (rnode->sysctl_data == &bufmem_lowater) {
1892 if (bufmem_hiwater - temp_water < 16)
1893 return (EINVAL);
1894 bufmem_lowater = temp_water;
1895 } else if (rnode->sysctl_data == &bufmem_hiwater) {
1896 if (temp_water - bufmem_lowater < 16)
1897 return (EINVAL);
1898 bufmem_hiwater = temp_water;
1899 } else
1900 return (EINVAL);
1901
1902 /* Drain until below new high water mark */
1903 sysctl_unlock();
1904 mutex_enter(&bufcache_lock);
1905 while (bufmem > bufmem_hiwater) {
1906 rv = buf_drain((bufmem - bufmem_hiwater) / (2 * 1024));
1907 if (rv <= 0)
1908 break;
1909 }
1910 mutex_exit(&bufcache_lock);
1911 sysctl_relock();
1912
1913 return 0;
1914 }
1915
1916 static struct sysctllog *vfsbio_sysctllog;
1917
1918 static void
sysctl_kern_buf_setup(void)1919 sysctl_kern_buf_setup(void)
1920 {
1921
1922 sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1923 CTLFLAG_PERMANENT,
1924 CTLTYPE_NODE, "buf",
1925 SYSCTL_DESCR("Kernel buffer cache information"),
1926 sysctl_dobuf, 0, NULL, 0,
1927 CTL_KERN, KERN_BUF, CTL_EOL);
1928 }
1929
1930 static void
sysctl_vm_buf_setup(void)1931 sysctl_vm_buf_setup(void)
1932 {
1933
1934 sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1935 CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1936 CTLTYPE_INT, "bufcache",
1937 SYSCTL_DESCR("Percentage of physical memory to use for "
1938 "buffer cache"),
1939 sysctl_bufvm_update, 0, &bufcache, 0,
1940 CTL_VM, CTL_CREATE, CTL_EOL);
1941 sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1942 CTLFLAG_PERMANENT|CTLFLAG_READONLY,
1943 CTLTYPE_LONG, "bufmem",
1944 SYSCTL_DESCR("Amount of kernel memory used by buffer "
1945 "cache"),
1946 NULL, 0, &bufmem, 0,
1947 CTL_VM, CTL_CREATE, CTL_EOL);
1948 sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1949 CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1950 CTLTYPE_LONG, "bufmem_lowater",
1951 SYSCTL_DESCR("Minimum amount of kernel memory to "
1952 "reserve for buffer cache"),
1953 sysctl_bufvm_update, 0, &bufmem_lowater, 0,
1954 CTL_VM, CTL_CREATE, CTL_EOL);
1955 sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1956 CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1957 CTLTYPE_LONG, "bufmem_hiwater",
1958 SYSCTL_DESCR("Maximum amount of kernel memory to use "
1959 "for buffer cache"),
1960 sysctl_bufvm_update, 0, &bufmem_hiwater, 0,
1961 CTL_VM, CTL_CREATE, CTL_EOL);
1962 }
1963
1964 static int
bufhash_stats(struct hashstat_sysctl * hs,bool fill)1965 bufhash_stats(struct hashstat_sysctl *hs, bool fill)
1966 {
1967 buf_t *bp;
1968 uint64_t chain;
1969
1970 strlcpy(hs->hash_name, "bufhash", sizeof(hs->hash_name));
1971 strlcpy(hs->hash_desc, "buffer hash", sizeof(hs->hash_desc));
1972 if (!fill)
1973 return 0;
1974
1975 hs->hash_size = bufhash + 1;
1976
1977 for (size_t i = 0; i < hs->hash_size; i++) {
1978 chain = 0;
1979
1980 mutex_enter(&bufcache_lock);
1981 LIST_FOREACH(bp, &bufhashtbl[i], b_hash) {
1982 chain++;
1983 }
1984 mutex_exit(&bufcache_lock);
1985
1986 if (chain > 0) {
1987 hs->hash_used++;
1988 hs->hash_items += chain;
1989 if (chain > hs->hash_maxchain)
1990 hs->hash_maxchain = chain;
1991 }
1992 preempt_point();
1993 }
1994
1995 return 0;
1996 }
1997
1998 #ifdef DEBUG
1999 /*
2000 * Print out statistics on the current allocation of the buffer pool.
2001 * Can be enabled to print out on every ``sync'' by setting "syncprt"
2002 * in vfs_syscalls.c using sysctl.
2003 */
2004 void
vfs_bufstats(void)2005 vfs_bufstats(void)
2006 {
2007 int i, j, count;
2008 buf_t *bp;
2009 struct bqueue *dp;
2010 int counts[MAXBSIZE / MIN_PAGE_SIZE + 1];
2011 static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
2012
2013 for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
2014 count = 0;
2015 memset(counts, 0, sizeof(counts));
2016 TAILQ_FOREACH(bp, &dp->bq_queue, b_freelist) {
2017 counts[bp->b_bufsize / PAGE_SIZE]++;
2018 count++;
2019 }
2020 printf("%s: total-%d", bname[i], count);
2021 for (j = 0; j <= MAXBSIZE / PAGE_SIZE; j++)
2022 if (counts[j] != 0)
2023 printf(", %d-%d", j * PAGE_SIZE, counts[j]);
2024 printf("\n");
2025 }
2026 }
2027 #endif /* DEBUG */
2028
2029 /* ------------------------------ */
2030
2031 buf_t *
getiobuf(struct vnode * vp,bool waitok)2032 getiobuf(struct vnode *vp, bool waitok)
2033 {
2034 buf_t *bp;
2035
2036 bp = pool_cache_get(bufio_cache, (waitok ? PR_WAITOK : PR_NOWAIT));
2037 if (bp == NULL)
2038 return bp;
2039
2040 buf_init(bp);
2041
2042 if ((bp->b_vp = vp) != NULL) {
2043 bp->b_objlock = vp->v_interlock;
2044 } else {
2045 KASSERT(bp->b_objlock == &buffer_lock);
2046 }
2047
2048 return bp;
2049 }
2050
2051 void
putiobuf(buf_t * bp)2052 putiobuf(buf_t *bp)
2053 {
2054
2055 buf_destroy(bp);
2056 pool_cache_put(bufio_cache, bp);
2057 }
2058
2059 /*
2060 * nestiobuf_iodone: b_iodone callback for nested buffers.
2061 */
2062
2063 void
nestiobuf_iodone(buf_t * bp)2064 nestiobuf_iodone(buf_t *bp)
2065 {
2066 buf_t *mbp = bp->b_private;
2067 int error;
2068 int donebytes;
2069
2070 KASSERT(bp->b_bcount <= bp->b_bufsize);
2071 KASSERT(mbp != bp);
2072
2073 error = bp->b_error;
2074 if (bp->b_error == 0 &&
2075 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
2076 /*
2077 * Not all got transferred, raise an error. We have no way to
2078 * propagate these conditions to mbp.
2079 */
2080 error = EIO;
2081 }
2082
2083 donebytes = bp->b_bufsize;
2084
2085 putiobuf(bp);
2086 nestiobuf_done(mbp, donebytes, error);
2087 }
2088
2089 /*
2090 * nestiobuf_setup: setup a "nested" buffer.
2091 *
2092 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
2093 * => 'bp' should be a buffer allocated by getiobuf.
2094 * => 'offset' is a byte offset in the master buffer.
2095 * => 'size' is a size in bytes of this nested buffer.
2096 */
2097
2098 void
nestiobuf_setup(buf_t * mbp,buf_t * bp,int offset,size_t size)2099 nestiobuf_setup(buf_t *mbp, buf_t *bp, int offset, size_t size)
2100 {
2101 const int b_pass = mbp->b_flags & (B_READ|B_PHYS|B_RAW|B_MEDIA_FLAGS);
2102 struct vnode *vp = mbp->b_vp;
2103
2104 KASSERT(mbp->b_bcount >= offset + size);
2105 bp->b_vp = vp;
2106 bp->b_dev = mbp->b_dev;
2107 bp->b_objlock = mbp->b_objlock;
2108 bp->b_cflags = BC_BUSY;
2109 bp->b_flags = B_ASYNC | b_pass;
2110 bp->b_iodone = nestiobuf_iodone;
2111 bp->b_data = (char *)mbp->b_data + offset;
2112 bp->b_resid = bp->b_bcount = size;
2113 bp->b_bufsize = bp->b_bcount;
2114 bp->b_private = mbp;
2115 BIO_COPYPRIO(bp, mbp);
2116 if (BUF_ISWRITE(bp) && vp != NULL) {
2117 mutex_enter(vp->v_interlock);
2118 vp->v_numoutput++;
2119 mutex_exit(vp->v_interlock);
2120 }
2121 }
2122
2123 /*
2124 * nestiobuf_done: propagate completion to the master buffer.
2125 *
2126 * => 'donebytes' specifies how many bytes in the 'mbp' is completed.
2127 * => 'error' is an errno(2) that 'donebytes' has been completed with.
2128 */
2129
2130 void
nestiobuf_done(buf_t * mbp,int donebytes,int error)2131 nestiobuf_done(buf_t *mbp, int donebytes, int error)
2132 {
2133
2134 if (donebytes == 0) {
2135 return;
2136 }
2137 mutex_enter(mbp->b_objlock);
2138 KASSERT(mbp->b_resid >= donebytes);
2139 mbp->b_resid -= donebytes;
2140 if (error)
2141 mbp->b_error = error;
2142 if (mbp->b_resid == 0) {
2143 if (mbp->b_error)
2144 mbp->b_resid = mbp->b_bcount;
2145 mutex_exit(mbp->b_objlock);
2146 biodone(mbp);
2147 } else
2148 mutex_exit(mbp->b_objlock);
2149 }
2150
2151 void
buf_init(buf_t * bp)2152 buf_init(buf_t *bp)
2153 {
2154
2155 cv_init(&bp->b_busy, "biolock");
2156 cv_init(&bp->b_done, "biowait");
2157 bp->b_dev = NODEV;
2158 bp->b_error = 0;
2159 bp->b_flags = 0;
2160 bp->b_cflags = 0;
2161 bp->b_oflags = 0;
2162 bp->b_objlock = &buffer_lock;
2163 bp->b_iodone = NULL;
2164 bp->b_dev = NODEV;
2165 bp->b_vnbufs.le_next = NOLIST;
2166 BIO_SETPRIO(bp, BPRIO_DEFAULT);
2167 }
2168
2169 void
buf_destroy(buf_t * bp)2170 buf_destroy(buf_t *bp)
2171 {
2172
2173 cv_destroy(&bp->b_done);
2174 cv_destroy(&bp->b_busy);
2175 }
2176
2177 int
bbusy(buf_t * bp,bool intr,int timo,kmutex_t * interlock)2178 bbusy(buf_t *bp, bool intr, int timo, kmutex_t *interlock)
2179 {
2180 int error;
2181
2182 KASSERT(mutex_owned(&bufcache_lock));
2183
2184 SDT_PROBE4(io, kernel, , bbusy__start, bp, intr, timo, interlock);
2185
2186 if ((bp->b_cflags & BC_BUSY) != 0) {
2187 if (curlwp == uvm.pagedaemon_lwp) {
2188 error = EDEADLK;
2189 goto out;
2190 }
2191 bp->b_cflags |= BC_WANTED;
2192 if (interlock != NULL)
2193 mutex_exit(interlock);
2194 if (intr) {
2195 error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
2196 timo);
2197 } else {
2198 error = cv_timedwait(&bp->b_busy, &bufcache_lock,
2199 timo);
2200 }
2201 /*
2202 * At this point the buffer may be gone: don't touch it
2203 * again. The caller needs to find it again and retry.
2204 */
2205 if (interlock != NULL)
2206 mutex_enter(interlock);
2207 if (error == 0)
2208 error = EPASSTHROUGH;
2209 } else {
2210 bp->b_cflags |= BC_BUSY;
2211 error = 0;
2212 }
2213
2214 out: SDT_PROBE5(io, kernel, , bbusy__done,
2215 bp, intr, timo, interlock, error);
2216 return error;
2217 }
2218
2219 /*
2220 * Nothing outside this file should really need to know about nbuf,
2221 * but a few things still want to read it, so give them a way to do that.
2222 */
2223 u_int
buf_nbuf(void)2224 buf_nbuf(void)
2225 {
2226
2227 return nbuf;
2228 }
2229