2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2004 Poul-Henning Kamp
5 * Copyright (c) 1994,1997 John S. Dyson
6 * Copyright (c) 2013 The FreeBSD Foundation
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * this file contains a new buffer I/O scheme implementing a coherent
36 * VM object and buffer cache scheme. Pains have been taken to make
37 * sure that the performance degradation associated with schemes such
38 * as this is not realized.
40 * Author: John S. Dyson
41 * Significant help during the development and debugging phases
42 * had been provided by David Greenman, also of the FreeBSD core team.
44 * see man buf(9) for more info.
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/bitset.h>
55 #include <sys/boottrace.h>
58 #include <sys/counter.h>
59 #include <sys/devicestat.h>
60 #include <sys/eventhandler.h>
63 #include <sys/limits.h>
65 #include <sys/malloc.h>
66 #include <sys/mount.h>
67 #include <sys/mutex.h>
68 #include <sys/kernel.h>
69 #include <sys/kthread.h>
71 #include <sys/racct.h>
72 #include <sys/refcount.h>
73 #include <sys/resourcevar.h>
74 #include <sys/rwlock.h>
76 #include <sys/sysctl.h>
77 #include <sys/syscallsubr.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/watchdog.h>
82 #include <geom/geom.h>
84 #include <vm/vm_param.h>
85 #include <vm/vm_kern.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_pager.h>
90 #include <vm/vm_extern.h>
91 #include <vm/vm_map.h>
92 #include <vm/swap_pager.h>
94 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
96 struct bio_ops bioops; /* I/O operation notification */
98 struct buf_ops buf_ops_bio = {
99 .bop_name = "buf_ops_bio",
100 .bop_write = bufwrite,
101 .bop_strategy = bufstrategy,
103 .bop_bdflush = bufbdflush,
107 struct mtx_padalign bq_lock;
108 TAILQ_HEAD(, buf) bq_queue;
110 uint16_t bq_subqueue;
112 } __aligned(CACHE_LINE_SIZE);
114 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
115 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
116 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
117 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
120 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
121 struct bufqueue bd_dirtyq;
122 struct bufqueue *bd_cleanq;
123 struct mtx_padalign bd_run_lock;
128 long bd_bufspacethresh;
129 int bd_hifreebuffers;
130 int bd_lofreebuffers;
131 int bd_hidirtybuffers;
132 int bd_lodirtybuffers;
133 int bd_dirtybufthresh;
138 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
139 int __aligned(CACHE_LINE_SIZE) bd_running;
140 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
141 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
142 } __aligned(CACHE_LINE_SIZE);
144 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
145 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
146 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
147 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
148 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
149 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
150 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
151 #define BD_DOMAIN(bd) (bd - bdomain)
153 static char *buf; /* buffer header pool */
157 return ((struct buf *)(buf + (sizeof(struct buf) +
158 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
161 caddr_t __read_mostly unmapped_buf;
163 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
164 struct proc *bufdaemonproc;
166 static void vm_hold_free_pages(struct buf *bp, int newbsize);
167 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
169 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
170 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
172 static void vfs_clean_pages_dirty_buf(struct buf *bp);
173 static void vfs_setdirty_range(struct buf *bp);
174 static void vfs_vmio_invalidate(struct buf *bp);
175 static void vfs_vmio_truncate(struct buf *bp, int npages);
176 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
177 static int vfs_bio_clcheck(struct vnode *vp, int size,
178 daddr_t lblkno, daddr_t blkno);
179 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
180 void (*)(struct buf *));
181 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
182 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
183 static void buf_daemon(void);
184 static __inline void bd_wakeup(void);
185 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
186 static void bufkva_reclaim(vmem_t *, int);
187 static void bufkva_free(struct buf *);
188 static int buf_import(void *, void **, int, int, int);
189 static void buf_release(void *, void **, int);
190 static void maxbcachebuf_adjust(void);
191 static inline struct bufdomain *bufdomain(struct buf *);
192 static void bq_remove(struct bufqueue *bq, struct buf *bp);
193 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
194 static int buf_recycle(struct bufdomain *, bool kva);
195 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
196 const char *lockname);
197 static void bd_init(struct bufdomain *bd);
198 static int bd_flushall(struct bufdomain *bd);
199 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
200 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
202 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
203 int vmiodirenable = TRUE;
204 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
205 "Use the VM system for directory writes");
206 long runningbufspace;
207 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
208 "Amount of presently outstanding async buffer io");
209 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
210 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
211 static counter_u64_t bufkvaspace;
212 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
213 "Kernel virtual memory used for buffers");
214 static long maxbufspace;
215 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
216 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
217 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
218 "Maximum allowed value of bufspace (including metadata)");
219 static long bufmallocspace;
220 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
221 "Amount of malloced memory for buffers");
222 static long maxbufmallocspace;
223 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
224 0, "Maximum amount of malloced memory for buffers");
225 static long lobufspace;
226 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
227 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
228 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
229 "Minimum amount of buffers we want to have");
231 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
232 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
233 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
234 "Maximum allowed value of bufspace (excluding metadata)");
236 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
237 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
238 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
239 "Bufspace consumed before waking the daemon to free some");
240 static counter_u64_t buffreekvacnt;
241 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
242 "Number of times we have freed the KVA space from some buffer");
243 static counter_u64_t bufdefragcnt;
244 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
245 "Number of times we have had to repeat buffer allocation to defragment");
246 static long lorunningspace;
247 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
248 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
249 "Minimum preferred space used for in-progress I/O");
250 static long hirunningspace;
251 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
252 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
253 "Maximum amount of space to use for in-progress I/O");
254 int dirtybufferflushes;
255 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
256 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
258 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
259 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
260 int altbufferflushes;
261 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
262 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
263 static int recursiveflushes;
264 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
265 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
266 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
267 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
268 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
269 "Number of buffers that are dirty (has unwritten changes) at the moment");
270 static int lodirtybuffers;
271 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
272 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
273 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
274 "How many buffers we want to have free before bufdaemon can sleep");
275 static int hidirtybuffers;
276 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
277 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
278 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
279 "When the number of dirty buffers is considered severe");
281 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
282 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
283 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
284 "Number of bdwrite to bawrite conversions to clear dirty buffers");
285 static int numfreebuffers;
286 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
287 "Number of free buffers");
288 static int lofreebuffers;
289 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
290 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
291 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
292 "Target number of free buffers");
293 static int hifreebuffers;
294 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
295 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
296 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
297 "Threshold for clean buffer recycling");
298 static counter_u64_t getnewbufcalls;
299 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
300 &getnewbufcalls, "Number of calls to getnewbuf");
301 static counter_u64_t getnewbufrestarts;
302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
304 "Number of times getnewbuf has had to restart a buffer acquisition");
305 static counter_u64_t mappingrestarts;
306 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
308 "Number of times getblk has had to restart a buffer mapping for "
310 static counter_u64_t numbufallocfails;
311 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
312 &numbufallocfails, "Number of times buffer allocations failed");
313 static int flushbufqtarget = 100;
314 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
315 "Amount of work to do in flushbufqueues when helping bufdaemon");
316 static counter_u64_t notbufdflushes;
317 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
318 "Number of dirty buffer flushes done by the bufdaemon helpers");
319 static long barrierwrites;
320 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
321 &barrierwrites, 0, "Number of barrier writes");
322 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
323 &unmapped_buf_allowed, 0,
324 "Permit the use of the unmapped i/o");
325 int maxbcachebuf = MAXBCACHEBUF;
326 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
327 "Maximum size of a buffer cache block");
330 * This lock synchronizes access to bd_request.
332 static struct mtx_padalign __exclusive_cache_line bdlock;
335 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
336 * waitrunningbufspace().
338 static struct mtx_padalign __exclusive_cache_line rbreqlock;
341 * Lock that protects bdirtywait.
343 static struct mtx_padalign __exclusive_cache_line bdirtylock;
346 * bufdaemon shutdown request and sleep channel.
348 static bool bd_shutdown;
351 * Wakeup point for bufdaemon, as well as indicator of whether it is already
352 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
355 static int bd_request;
358 * Request for the buf daemon to write more buffers than is indicated by
359 * lodirtybuf. This may be necessary to push out excess dependencies or
360 * defragment the address space where a simple count of the number of dirty
361 * buffers is insufficient to characterize the demand for flushing them.
363 static int bd_speedupreq;
366 * Synchronization (sleep/wakeup) variable for active buffer space requests.
367 * Set when wait starts, cleared prior to wakeup().
368 * Used in runningbufwakeup() and waitrunningbufspace().
370 static int runningbufreq;
373 * Synchronization for bwillwrite() waiters.
375 static int bdirtywait;
378 * Definitions for the buffer free lists.
380 #define QUEUE_NONE 0 /* on no queue */
381 #define QUEUE_EMPTY 1 /* empty buffer headers */
382 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
383 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
384 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
386 /* Maximum number of buffer domains. */
387 #define BUF_DOMAINS 8
389 struct bufdomainset bdlodirty; /* Domains > lodirty */
390 struct bufdomainset bdhidirty; /* Domains > hidirty */
392 /* Configured number of clean queues. */
393 static int __read_mostly buf_domains;
395 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
396 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
397 struct bufqueue __exclusive_cache_line bqempty;
400 * per-cpu empty buffer cache.
405 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
410 value = *(long *)arg1;
411 error = sysctl_handle_long(oidp, &value, 0, req);
412 if (error != 0 || req->newptr == NULL)
414 mtx_lock(&rbreqlock);
415 if (arg1 == &hirunningspace) {
416 if (value < lorunningspace)
419 hirunningspace = value;
421 KASSERT(arg1 == &lorunningspace,
422 ("%s: unknown arg1", __func__));
423 if (value > hirunningspace)
426 lorunningspace = value;
428 mtx_unlock(&rbreqlock);
433 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
439 value = *(int *)arg1;
440 error = sysctl_handle_int(oidp, &value, 0, req);
441 if (error != 0 || req->newptr == NULL)
443 *(int *)arg1 = value;
444 for (i = 0; i < buf_domains; i++)
445 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
452 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
458 value = *(long *)arg1;
459 error = sysctl_handle_long(oidp, &value, 0, req);
460 if (error != 0 || req->newptr == NULL)
462 *(long *)arg1 = value;
463 for (i = 0; i < buf_domains; i++)
464 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
470 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
471 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
473 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
480 for (i = 0; i < buf_domains; i++)
481 lvalue += bdomain[i].bd_bufspace;
482 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
483 return (sysctl_handle_long(oidp, &lvalue, 0, req));
484 if (lvalue > INT_MAX)
485 /* On overflow, still write out a long to trigger ENOMEM. */
486 return (sysctl_handle_long(oidp, &lvalue, 0, req));
488 return (sysctl_handle_int(oidp, &ivalue, 0, req));
492 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
498 for (i = 0; i < buf_domains; i++)
499 lvalue += bdomain[i].bd_bufspace;
500 return (sysctl_handle_long(oidp, &lvalue, 0, req));
505 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
511 for (i = 0; i < buf_domains; i++)
512 value += bdomain[i].bd_numdirtybuffers;
513 return (sysctl_handle_int(oidp, &value, 0, req));
519 * Wakeup any bwillwrite() waiters.
524 mtx_lock(&bdirtylock);
529 mtx_unlock(&bdirtylock);
535 * Clear a domain from the appropriate bitsets when dirtybuffers
539 bd_clear(struct bufdomain *bd)
542 mtx_lock(&bdirtylock);
543 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
544 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
545 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
546 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
547 mtx_unlock(&bdirtylock);
553 * Set a domain in the appropriate bitsets when dirtybuffers
557 bd_set(struct bufdomain *bd)
560 mtx_lock(&bdirtylock);
561 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
562 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
563 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
564 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
565 mtx_unlock(&bdirtylock);
571 * Decrement the numdirtybuffers count by one and wakeup any
572 * threads blocked in bwillwrite().
575 bdirtysub(struct buf *bp)
577 struct bufdomain *bd;
581 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
582 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
584 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
591 * Increment the numdirtybuffers count by one and wakeup the buf
595 bdirtyadd(struct buf *bp)
597 struct bufdomain *bd;
601 * Only do the wakeup once as we cross the boundary. The
602 * buf daemon will keep running until the condition clears.
605 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
606 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
608 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
613 * bufspace_daemon_wakeup:
615 * Wakeup the daemons responsible for freeing clean bufs.
618 bufspace_daemon_wakeup(struct bufdomain *bd)
622 * avoid the lock if the daemon is running.
624 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
626 atomic_store_int(&bd->bd_running, 1);
627 wakeup(&bd->bd_running);
635 * Adjust the reported bufspace for a KVA managed buffer, possibly
636 * waking any waiters.
639 bufspace_adjust(struct buf *bp, int bufsize)
641 struct bufdomain *bd;
645 KASSERT((bp->b_flags & B_MALLOC) == 0,
646 ("bufspace_adjust: malloc buf %p", bp));
648 diff = bufsize - bp->b_bufsize;
650 atomic_subtract_long(&bd->bd_bufspace, -diff);
651 } else if (diff > 0) {
652 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
653 /* Wake up the daemon on the transition. */
654 if (space < bd->bd_bufspacethresh &&
655 space + diff >= bd->bd_bufspacethresh)
656 bufspace_daemon_wakeup(bd);
658 bp->b_bufsize = bufsize;
664 * Reserve bufspace before calling allocbuf(). metadata has a
665 * different space limit than data.
668 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
674 limit = bd->bd_maxbufspace;
676 limit = bd->bd_hibufspace;
677 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
680 atomic_subtract_long(&bd->bd_bufspace, size);
684 /* Wake up the daemon on the transition. */
685 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
686 bufspace_daemon_wakeup(bd);
694 * Release reserved bufspace after bufspace_adjust() has consumed it.
697 bufspace_release(struct bufdomain *bd, int size)
700 atomic_subtract_long(&bd->bd_bufspace, size);
706 * Wait for bufspace, acting as the buf daemon if a locked vnode is
707 * supplied. bd_wanted must be set prior to polling for space. The
708 * operation must be re-tried on return.
711 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
712 int slpflag, int slptimeo)
715 int error, fl, norunbuf;
717 if ((gbflags & GB_NOWAIT_BD) != 0)
722 while (bd->bd_wanted) {
723 if (vp != NULL && vp->v_type != VCHR &&
724 (td->td_pflags & TDP_BUFNEED) == 0) {
727 * getblk() is called with a vnode locked, and
728 * some majority of the dirty buffers may as
729 * well belong to the vnode. Flushing the
730 * buffers there would make a progress that
731 * cannot be achieved by the buf_daemon, that
732 * cannot lock the vnode.
734 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
735 (td->td_pflags & TDP_NORUNNINGBUF);
738 * Play bufdaemon. The getnewbuf() function
739 * may be called while the thread owns lock
740 * for another dirty buffer for the same
741 * vnode, which makes it impossible to use
742 * VOP_FSYNC() there, due to the buffer lock
745 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
746 fl = buf_flush(vp, bd, flushbufqtarget);
747 td->td_pflags &= norunbuf;
751 if (bd->bd_wanted == 0)
754 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
755 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
763 bufspace_daemon_shutdown(void *arg, int howto __unused)
765 struct bufdomain *bd = arg;
769 bd->bd_shutdown = true;
770 wakeup(&bd->bd_running);
771 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
772 "bufspace_shutdown", 60 * hz);
775 printf("bufspacedaemon wait error: %d\n", error);
781 * buffer space management daemon. Tries to maintain some marginal
782 * amount of free buffer space so that requesting processes neither
783 * block nor work to reclaim buffers.
786 bufspace_daemon(void *arg)
788 struct bufdomain *bd = arg;
790 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
791 SHUTDOWN_PRI_LAST + 100);
794 while (!bd->bd_shutdown) {
798 * Free buffers from the clean queue until we meet our
801 * Theory of operation: The buffer cache is most efficient
802 * when some free buffer headers and space are always
803 * available to getnewbuf(). This daemon attempts to prevent
804 * the excessive blocking and synchronization associated
805 * with shortfall. It goes through three phases according
808 * 1) The daemon wakes up voluntarily once per-second
809 * during idle periods when the counters are below
810 * the wakeup thresholds (bufspacethresh, lofreebuffers).
812 * 2) The daemon wakes up as we cross the thresholds
813 * ahead of any potential blocking. This may bounce
814 * slightly according to the rate of consumption and
817 * 3) The daemon and consumers are starved for working
818 * clean buffers. This is the 'bufspace' sleep below
819 * which will inefficiently trade bufs with bqrelse
820 * until we return to condition 2.
822 while (bd->bd_bufspace > bd->bd_lobufspace ||
823 bd->bd_freebuffers < bd->bd_hifreebuffers) {
824 if (buf_recycle(bd, false) != 0) {
828 * Speedup dirty if we've run out of clean
829 * buffers. This is possible in particular
830 * because softdep may held many bufs locked
831 * pending writes to other bufs which are
832 * marked for delayed write, exhausting
833 * clean space until they are written.
838 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
839 PRIBIO|PDROP, "bufspace", hz/10);
847 * Re-check our limits and sleep. bd_running must be
848 * cleared prior to checking the limits to avoid missed
849 * wakeups. The waker will adjust one of bufspace or
850 * freebuffers prior to checking bd_running.
855 atomic_store_int(&bd->bd_running, 0);
856 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
857 bd->bd_freebuffers > bd->bd_lofreebuffers) {
858 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
861 /* Avoid spurious wakeups while running. */
862 atomic_store_int(&bd->bd_running, 1);
865 wakeup(&bd->bd_shutdown);
873 * Adjust the reported bufspace for a malloc managed buffer, possibly
874 * waking any waiters.
877 bufmallocadjust(struct buf *bp, int bufsize)
881 KASSERT((bp->b_flags & B_MALLOC) != 0,
882 ("bufmallocadjust: non-malloc buf %p", bp));
883 diff = bufsize - bp->b_bufsize;
885 atomic_subtract_long(&bufmallocspace, -diff);
887 atomic_add_long(&bufmallocspace, diff);
888 bp->b_bufsize = bufsize;
894 * Wake up processes that are waiting on asynchronous writes to fall
895 * below lorunningspace.
901 mtx_lock(&rbreqlock);
904 wakeup(&runningbufreq);
906 mtx_unlock(&rbreqlock);
912 * Decrement the outstanding write count according.
915 runningbufwakeup(struct buf *bp)
919 bspace = bp->b_runningbufspace;
922 space = atomic_fetchadd_long(&runningbufspace, -bspace);
923 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
925 bp->b_runningbufspace = 0;
927 * Only acquire the lock and wakeup on the transition from exceeding
928 * the threshold to falling below it.
930 if (space < lorunningspace)
932 if (space - bspace > lorunningspace)
938 * waitrunningbufspace()
940 * runningbufspace is a measure of the amount of I/O currently
941 * running. This routine is used in async-write situations to
942 * prevent creating huge backups of pending writes to a device.
943 * Only asynchronous writes are governed by this function.
945 * This does NOT turn an async write into a sync write. It waits
946 * for earlier writes to complete and generally returns before the
947 * caller's write has reached the device.
950 waitrunningbufspace(void)
953 mtx_lock(&rbreqlock);
954 while (runningbufspace > hirunningspace) {
956 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
958 mtx_unlock(&rbreqlock);
962 * vfs_buf_test_cache:
964 * Called when a buffer is extended. This function clears the B_CACHE
965 * bit if the newly extended portion of the buffer does not contain
969 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
970 vm_offset_t size, vm_page_t m)
974 * This function and its results are protected by higher level
975 * synchronization requiring vnode and buf locks to page in and
978 if (bp->b_flags & B_CACHE) {
979 int base = (foff + off) & PAGE_MASK;
980 if (vm_page_is_valid(m, base, size) == 0)
981 bp->b_flags &= ~B_CACHE;
985 /* Wake up the buffer daemon if necessary */
991 if (bd_request == 0) {
999 * Adjust the maxbcachbuf tunable.
1002 maxbcachebuf_adjust(void)
1007 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1010 while (i * 2 <= maxbcachebuf)
1013 if (maxbcachebuf < MAXBSIZE)
1014 maxbcachebuf = MAXBSIZE;
1015 if (maxbcachebuf > maxphys)
1016 maxbcachebuf = maxphys;
1017 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1018 printf("maxbcachebuf=%d\n", maxbcachebuf);
1022 * bd_speedup - speedup the buffer cache flushing code
1031 if (bd_speedupreq == 0 || bd_request == 0)
1036 wakeup(&bd_request);
1037 mtx_unlock(&bdlock);
1041 #define TRANSIENT_DENOM 5
1043 #define TRANSIENT_DENOM 10
1047 * Calculating buffer cache scaling values and reserve space for buffer
1048 * headers. This is called during low level kernel initialization and
1049 * may be called more then once. We CANNOT write to the memory area
1050 * being reserved at this time.
1053 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1056 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1059 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for
1060 * this when sizing maps based on the amount of physical memory
1064 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1065 (KASAN_SHADOW_SCALE + 1);
1066 #elif defined(KMSAN)
1070 * KMSAN cannot reliably determine whether buffer data is initialized
1071 * unless it is updated through a KVA mapping.
1073 unmapped_buf_allowed = 0;
1077 * physmem_est is in pages. Convert it to kilobytes (assumes
1078 * PAGE_SIZE is >= 1K)
1080 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1082 maxbcachebuf_adjust();
1084 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1085 * For the first 64MB of ram nominally allocate sufficient buffers to
1086 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1087 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1088 * the buffer cache we limit the eventual kva reservation to
1091 * factor represents the 1/4 x ram conversion.
1094 int factor = 4 * BKVASIZE / 1024;
1097 if (physmem_est > 4096)
1098 nbuf += min((physmem_est - 4096) / factor,
1100 if (physmem_est > 65536)
1101 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1102 32 * 1024 * 1024 / (factor * 5));
1104 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1105 nbuf = maxbcache / BKVASIZE;
1110 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1111 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1112 if (nbuf > maxbuf) {
1114 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1120 * Ideal allocation size for the transient bio submap is 10%
1121 * of the maximal space buffer map. This roughly corresponds
1122 * to the amount of the buffer mapped for typical UFS load.
1124 * Clip the buffer map to reserve space for the transient
1125 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1126 * maximum buffer map extent on the platform.
1128 * The fall-back to the maxbuf in case of maxbcache unset,
1129 * allows to not trim the buffer KVA for the architectures
1130 * with ample KVA space.
1132 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1133 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1134 buf_sz = (long)nbuf * BKVASIZE;
1135 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1136 (TRANSIENT_DENOM - 1)) {
1138 * There is more KVA than memory. Do not
1139 * adjust buffer map size, and assign the rest
1140 * of maxbuf to transient map.
1142 biotmap_sz = maxbuf_sz - buf_sz;
1145 * Buffer map spans all KVA we could afford on
1146 * this platform. Give 10% (20% on i386) of
1147 * the buffer map to the transient bio map.
1149 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1150 buf_sz -= biotmap_sz;
1152 if (biotmap_sz / INT_MAX > maxphys)
1153 bio_transient_maxcnt = INT_MAX;
1155 bio_transient_maxcnt = biotmap_sz / maxphys;
1157 * Artificially limit to 1024 simultaneous in-flight I/Os
1158 * using the transient mapping.
1160 if (bio_transient_maxcnt > 1024)
1161 bio_transient_maxcnt = 1024;
1163 nbuf = buf_sz / BKVASIZE;
1167 nswbuf = min(nbuf / 4, 256);
1168 if (nswbuf < NSWBUF_MIN)
1169 nswbuf = NSWBUF_MIN;
1173 * Reserve space for the buffer cache buffers
1176 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1177 atop(maxbcachebuf)) * nbuf;
1183 * Single global constant for BUF_WMESG, to avoid getting multiple
1186 static const char buf_wmesg[] = "bufwait";
1188 /* Initialize the buffer subsystem. Called before use of any buffers. */
1195 KASSERT(maxbcachebuf >= MAXBSIZE,
1196 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1198 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1199 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1200 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1201 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1203 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1205 /* finally, initialize each buffer header and stick on empty q */
1206 for (i = 0; i < nbuf; i++) {
1208 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1209 bp->b_flags = B_INVAL;
1210 bp->b_rcred = NOCRED;
1211 bp->b_wcred = NOCRED;
1212 bp->b_qindex = QUEUE_NONE;
1214 bp->b_subqueue = mp_maxid + 1;
1216 bp->b_data = bp->b_kvabase = unmapped_buf;
1217 LIST_INIT(&bp->b_dep);
1218 BUF_LOCKINIT(bp, buf_wmesg);
1219 bq_insert(&bqempty, bp, false);
1223 * maxbufspace is the absolute maximum amount of buffer space we are
1224 * allowed to reserve in KVM and in real terms. The absolute maximum
1225 * is nominally used by metadata. hibufspace is the nominal maximum
1226 * used by most other requests. The differential is required to
1227 * ensure that metadata deadlocks don't occur.
1229 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1230 * this may result in KVM fragmentation which is not handled optimally
1231 * by the system. XXX This is less true with vmem. We could use
1234 maxbufspace = (long)nbuf * BKVASIZE;
1235 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1236 lobufspace = (hibufspace / 20) * 19; /* 95% */
1237 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1240 * Note: The 16 MiB upper limit for hirunningspace was chosen
1241 * arbitrarily and may need further tuning. It corresponds to
1242 * 128 outstanding write IO requests (if IO size is 128 KiB),
1243 * which fits with many RAID controllers' tagged queuing limits.
1244 * The lower 1 MiB limit is the historical upper limit for
1247 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1248 16 * 1024 * 1024), 1024 * 1024);
1249 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1252 * Limit the amount of malloc memory since it is wired permanently into
1253 * the kernel space. Even though this is accounted for in the buffer
1254 * allocation, we don't want the malloced region to grow uncontrolled.
1255 * The malloc scheme improves memory utilization significantly on
1256 * average (small) directories.
1258 maxbufmallocspace = hibufspace / 20;
1261 * Reduce the chance of a deadlock occurring by limiting the number
1262 * of delayed-write dirty buffers we allow to stack up.
1264 hidirtybuffers = nbuf / 4 + 20;
1265 dirtybufthresh = hidirtybuffers * 9 / 10;
1267 * To support extreme low-memory systems, make sure hidirtybuffers
1268 * cannot eat up all available buffer space. This occurs when our
1269 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1270 * buffer space assuming BKVASIZE'd buffers.
1272 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1273 hidirtybuffers >>= 1;
1275 lodirtybuffers = hidirtybuffers / 2;
1278 * lofreebuffers should be sufficient to avoid stalling waiting on
1279 * buf headers under heavy utilization. The bufs in per-cpu caches
1280 * are counted as free but will be unavailable to threads executing
1283 * hifreebuffers is the free target for the bufspace daemon. This
1284 * should be set appropriately to limit work per-iteration.
1286 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1287 hifreebuffers = (3 * lofreebuffers) / 2;
1288 numfreebuffers = nbuf;
1290 /* Setup the kva and free list allocators. */
1291 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1292 buf_zone = uma_zcache_create("buf free cache",
1293 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1294 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1297 * Size the clean queue according to the amount of buffer space.
1298 * One queue per-256mb up to the max. More queues gives better
1299 * concurrency but less accurate LRU.
1301 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1302 for (i = 0 ; i < buf_domains; i++) {
1303 struct bufdomain *bd;
1307 bd->bd_freebuffers = nbuf / buf_domains;
1308 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1309 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1310 bd->bd_bufspace = 0;
1311 bd->bd_maxbufspace = maxbufspace / buf_domains;
1312 bd->bd_hibufspace = hibufspace / buf_domains;
1313 bd->bd_lobufspace = lobufspace / buf_domains;
1314 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1315 bd->bd_numdirtybuffers = 0;
1316 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1317 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1318 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1319 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1320 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1322 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1323 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1324 mappingrestarts = counter_u64_alloc(M_WAITOK);
1325 numbufallocfails = counter_u64_alloc(M_WAITOK);
1326 notbufdflushes = counter_u64_alloc(M_WAITOK);
1327 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1328 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1329 bufkvaspace = counter_u64_alloc(M_WAITOK);
1334 vfs_buf_check_mapped(struct buf *bp)
1337 KASSERT(bp->b_kvabase != unmapped_buf,
1338 ("mapped buf: b_kvabase was not updated %p", bp));
1339 KASSERT(bp->b_data != unmapped_buf,
1340 ("mapped buf: b_data was not updated %p", bp));
1341 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1342 maxphys, ("b_data + b_offset unmapped %p", bp));
1346 vfs_buf_check_unmapped(struct buf *bp)
1349 KASSERT(bp->b_data == unmapped_buf,
1350 ("unmapped buf: corrupted b_data %p", bp));
1353 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1354 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1356 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1357 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1361 isbufbusy(struct buf *bp)
1363 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1364 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1370 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1373 bufshutdown(int show_busybufs)
1375 static int first_buf_printf = 1;
1377 int i, iter, nbusy, pbusy;
1383 * Sync filesystems for shutdown
1385 wdog_kern_pat(WD_LASTVAL);
1386 kern_sync(curthread);
1389 * With soft updates, some buffers that are
1390 * written will be remarked as dirty until other
1391 * buffers are written.
1393 for (iter = pbusy = 0; iter < 20; iter++) {
1395 for (i = nbuf - 1; i >= 0; i--) {
1401 if (first_buf_printf)
1402 printf("All buffers synced.");
1405 if (first_buf_printf) {
1406 printf("Syncing disks, buffers remaining... ");
1407 first_buf_printf = 0;
1409 printf("%d ", nbusy);
1414 wdog_kern_pat(WD_LASTVAL);
1415 kern_sync(curthread);
1419 * Spin for a while to allow interrupt threads to run.
1421 DELAY(50000 * iter);
1424 * Context switch several times to allow interrupt
1427 for (subiter = 0; subiter < 50 * iter; subiter++) {
1428 thread_lock(curthread);
1436 * Count only busy local buffers to prevent forcing
1437 * a fsck if we're just a client of a wedged NFS server
1440 for (i = nbuf - 1; i >= 0; i--) {
1442 if (isbufbusy(bp)) {
1444 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1445 if (bp->b_dev == NULL) {
1446 TAILQ_REMOVE(&mountlist,
1447 bp->b_vp->v_mount, mnt_list);
1452 if (show_busybufs > 0) {
1454 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1455 nbusy, bp, bp->b_vp, bp->b_flags,
1456 (intmax_t)bp->b_blkno,
1457 (intmax_t)bp->b_lblkno);
1458 BUF_LOCKPRINTINFO(bp);
1459 if (show_busybufs > 1)
1467 * Failed to sync all blocks. Indicate this and don't
1468 * unmount filesystems (thus forcing an fsck on reboot).
1470 BOOTTRACE("shutdown failed to sync buffers");
1471 printf("Giving up on %d buffers\n", nbusy);
1472 DELAY(5000000); /* 5 seconds */
1475 BOOTTRACE("shutdown sync complete");
1476 if (!first_buf_printf)
1477 printf("Final sync complete\n");
1480 * Unmount filesystems and perform swapoff, to quiesce
1481 * the system as much as possible. In particular, no
1482 * I/O should be initiated from top levels since it
1483 * might be abruptly terminated by reset, or otherwise
1484 * erronously handled because other parts of the
1485 * system are disabled.
1487 * Swapoff before unmount, because file-backed swap is
1488 * non-operational after unmount of the underlying
1491 if (!KERNEL_PANICKED()) {
1495 BOOTTRACE("shutdown unmounted all filesystems");
1497 DELAY(100000); /* wait for console output to finish */
1501 bpmap_qenter(struct buf *bp)
1504 BUF_CHECK_MAPPED(bp);
1507 * bp->b_data is relative to bp->b_offset, but
1508 * bp->b_offset may be offset into the first page.
1510 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1511 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1512 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1513 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1516 static inline struct bufdomain *
1517 bufdomain(struct buf *bp)
1520 return (&bdomain[bp->b_domain]);
1523 static struct bufqueue *
1524 bufqueue(struct buf *bp)
1527 switch (bp->b_qindex) {
1530 case QUEUE_SENTINEL:
1535 return (&bufdomain(bp)->bd_dirtyq);
1537 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1541 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1545 * Return the locked bufqueue that bp is a member of.
1547 static struct bufqueue *
1548 bufqueue_acquire(struct buf *bp)
1550 struct bufqueue *bq, *nbq;
1553 * bp can be pushed from a per-cpu queue to the
1554 * cleanq while we're waiting on the lock. Retry
1555 * if the queues don't match.
1573 * Insert the buffer into the appropriate free list. Requires a
1574 * locked buffer on entry and buffer is unlocked before return.
1577 binsfree(struct buf *bp, int qindex)
1579 struct bufdomain *bd;
1580 struct bufqueue *bq;
1582 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1583 ("binsfree: Invalid qindex %d", qindex));
1584 BUF_ASSERT_XLOCKED(bp);
1587 * Handle delayed bremfree() processing.
1589 if (bp->b_flags & B_REMFREE) {
1590 if (bp->b_qindex == qindex) {
1591 bp->b_flags |= B_REUSE;
1592 bp->b_flags &= ~B_REMFREE;
1596 bq = bufqueue_acquire(bp);
1601 if (qindex == QUEUE_CLEAN) {
1602 if (bd->bd_lim != 0)
1603 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1607 bq = &bd->bd_dirtyq;
1608 bq_insert(bq, bp, true);
1614 * Free a buffer to the buf zone once it no longer has valid contents.
1617 buf_free(struct buf *bp)
1620 if (bp->b_flags & B_REMFREE)
1622 if (bp->b_vflags & BV_BKGRDINPROG)
1623 panic("losing buffer 1");
1624 if (bp->b_rcred != NOCRED) {
1625 crfree(bp->b_rcred);
1626 bp->b_rcred = NOCRED;
1628 if (bp->b_wcred != NOCRED) {
1629 crfree(bp->b_wcred);
1630 bp->b_wcred = NOCRED;
1632 if (!LIST_EMPTY(&bp->b_dep))
1635 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1636 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1638 uma_zfree(buf_zone, bp);
1644 * Import bufs into the uma cache from the buf list. The system still
1645 * expects a static array of bufs and much of the synchronization
1646 * around bufs assumes type stable storage. As a result, UMA is used
1647 * only as a per-cpu cache of bufs still maintained on a global list.
1650 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1656 for (i = 0; i < cnt; i++) {
1657 bp = TAILQ_FIRST(&bqempty.bq_queue);
1660 bq_remove(&bqempty, bp);
1663 BQ_UNLOCK(&bqempty);
1671 * Release bufs from the uma cache back to the buffer queues.
1674 buf_release(void *arg, void **store, int cnt)
1676 struct bufqueue *bq;
1682 for (i = 0; i < cnt; i++) {
1684 /* Inline bq_insert() to batch locking. */
1685 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1686 bp->b_flags &= ~(B_AGE | B_REUSE);
1688 bp->b_qindex = bq->bq_index;
1696 * Allocate an empty buffer header.
1699 buf_alloc(struct bufdomain *bd)
1702 int freebufs, error;
1705 * We can only run out of bufs in the buf zone if the average buf
1706 * is less than BKVASIZE. In this case the actual wait/block will
1707 * come from buf_reycle() failing to flush one of these small bufs.
1710 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1712 bp = uma_zalloc(buf_zone, M_NOWAIT);
1714 atomic_add_int(&bd->bd_freebuffers, 1);
1715 bufspace_daemon_wakeup(bd);
1716 counter_u64_add(numbufallocfails, 1);
1720 * Wake-up the bufspace daemon on transition below threshold.
1722 if (freebufs == bd->bd_lofreebuffers)
1723 bufspace_daemon_wakeup(bd);
1725 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
1726 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1730 KASSERT(bp->b_vp == NULL,
1731 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1732 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1733 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1734 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1735 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1736 KASSERT(bp->b_npages == 0,
1737 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1738 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1739 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1740 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1742 bp->b_domain = BD_DOMAIN(bd);
1748 bp->b_blkno = bp->b_lblkno = 0;
1749 bp->b_offset = NOOFFSET;
1755 bp->b_dirtyoff = bp->b_dirtyend = 0;
1756 bp->b_bufobj = NULL;
1757 bp->b_data = bp->b_kvabase = unmapped_buf;
1758 bp->b_fsprivate1 = NULL;
1759 bp->b_fsprivate2 = NULL;
1760 bp->b_fsprivate3 = NULL;
1761 LIST_INIT(&bp->b_dep);
1769 * Free a buffer from the given bufqueue. kva controls whether the
1770 * freed buf must own some kva resources. This is used for
1774 buf_recycle(struct bufdomain *bd, bool kva)
1776 struct bufqueue *bq;
1777 struct buf *bp, *nbp;
1780 counter_u64_add(bufdefragcnt, 1);
1784 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1785 ("buf_recycle: Locks don't match"));
1786 nbp = TAILQ_FIRST(&bq->bq_queue);
1789 * Run scan, possibly freeing data and/or kva mappings on the fly
1792 while ((bp = nbp) != NULL) {
1794 * Calculate next bp (we can only use it if we do not
1795 * release the bqlock).
1797 nbp = TAILQ_NEXT(bp, b_freelist);
1800 * If we are defragging then we need a buffer with
1801 * some kva to reclaim.
1803 if (kva && bp->b_kvasize == 0)
1806 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1810 * Implement a second chance algorithm for frequently
1813 if ((bp->b_flags & B_REUSE) != 0) {
1814 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1815 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1816 bp->b_flags &= ~B_REUSE;
1822 * Skip buffers with background writes in progress.
1824 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1829 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1830 ("buf_recycle: inconsistent queue %d bp %p",
1832 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1833 ("getnewbuf: queue domain %d doesn't match request %d",
1834 bp->b_domain, (int)BD_DOMAIN(bd)));
1836 * NOTE: nbp is now entirely invalid. We can only restart
1837 * the scan from this point on.
1843 * Requeue the background write buffer with error and
1846 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1849 nbp = TAILQ_FIRST(&bq->bq_queue);
1852 bp->b_flags |= B_INVAL;
1865 * Mark the buffer for removal from the appropriate free list.
1869 bremfree(struct buf *bp)
1872 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1873 KASSERT((bp->b_flags & B_REMFREE) == 0,
1874 ("bremfree: buffer %p already marked for delayed removal.", bp));
1875 KASSERT(bp->b_qindex != QUEUE_NONE,
1876 ("bremfree: buffer %p not on a queue.", bp));
1877 BUF_ASSERT_XLOCKED(bp);
1879 bp->b_flags |= B_REMFREE;
1885 * Force an immediate removal from a free list. Used only in nfs when
1886 * it abuses the b_freelist pointer.
1889 bremfreef(struct buf *bp)
1891 struct bufqueue *bq;
1893 bq = bufqueue_acquire(bp);
1899 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1902 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1903 TAILQ_INIT(&bq->bq_queue);
1905 bq->bq_index = qindex;
1906 bq->bq_subqueue = subqueue;
1910 bd_init(struct bufdomain *bd)
1914 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1915 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1916 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1917 for (i = 0; i <= mp_maxid; i++)
1918 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1919 "bufq clean subqueue lock");
1920 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1926 * Removes a buffer from the free list, must be called with the
1927 * correct qlock held.
1930 bq_remove(struct bufqueue *bq, struct buf *bp)
1933 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1934 bp, bp->b_vp, bp->b_flags);
1935 KASSERT(bp->b_qindex != QUEUE_NONE,
1936 ("bq_remove: buffer %p not on a queue.", bp));
1937 KASSERT(bufqueue(bp) == bq,
1938 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1940 BQ_ASSERT_LOCKED(bq);
1941 if (bp->b_qindex != QUEUE_EMPTY) {
1942 BUF_ASSERT_XLOCKED(bp);
1944 KASSERT(bq->bq_len >= 1,
1945 ("queue %d underflow", bp->b_qindex));
1946 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1948 bp->b_qindex = QUEUE_NONE;
1949 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1953 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1957 BQ_ASSERT_LOCKED(bq);
1958 if (bq != bd->bd_cleanq) {
1960 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1961 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1962 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1964 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1966 bd->bd_cleanq->bq_len += bq->bq_len;
1969 if (bd->bd_wanted) {
1971 wakeup(&bd->bd_wanted);
1973 if (bq != bd->bd_cleanq)
1978 bd_flushall(struct bufdomain *bd)
1980 struct bufqueue *bq;
1984 if (bd->bd_lim == 0)
1987 for (i = 0; i <= mp_maxid; i++) {
1988 bq = &bd->bd_subq[i];
1989 if (bq->bq_len == 0)
2001 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
2003 struct bufdomain *bd;
2005 if (bp->b_qindex != QUEUE_NONE)
2006 panic("bq_insert: free buffer %p onto another queue?", bp);
2009 if (bp->b_flags & B_AGE) {
2010 /* Place this buf directly on the real queue. */
2011 if (bq->bq_index == QUEUE_CLEAN)
2014 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2017 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2019 bp->b_flags &= ~(B_AGE | B_REUSE);
2021 bp->b_qindex = bq->bq_index;
2022 bp->b_subqueue = bq->bq_subqueue;
2025 * Unlock before we notify so that we don't wakeup a waiter that
2026 * fails a trylock on the buf and sleeps again.
2031 if (bp->b_qindex == QUEUE_CLEAN) {
2033 * Flush the per-cpu queue and notify any waiters.
2035 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2036 bq->bq_len >= bd->bd_lim))
2045 * Free the kva allocation for a buffer.
2049 bufkva_free(struct buf *bp)
2053 if (bp->b_kvasize == 0) {
2054 KASSERT(bp->b_kvabase == unmapped_buf &&
2055 bp->b_data == unmapped_buf,
2056 ("Leaked KVA space on %p", bp));
2057 } else if (buf_mapped(bp))
2058 BUF_CHECK_MAPPED(bp);
2060 BUF_CHECK_UNMAPPED(bp);
2062 if (bp->b_kvasize == 0)
2065 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2066 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2067 counter_u64_add(buffreekvacnt, 1);
2068 bp->b_data = bp->b_kvabase = unmapped_buf;
2075 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2078 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2083 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2084 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2085 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2086 KASSERT(maxsize <= maxbcachebuf,
2087 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2092 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2095 * Buffer map is too fragmented. Request the caller
2096 * to defragment the map.
2100 bp->b_kvabase = (caddr_t)addr;
2101 bp->b_kvasize = maxsize;
2102 counter_u64_add(bufkvaspace, bp->b_kvasize);
2103 if ((gbflags & GB_UNMAPPED) != 0) {
2104 bp->b_data = unmapped_buf;
2105 BUF_CHECK_UNMAPPED(bp);
2107 bp->b_data = bp->b_kvabase;
2108 BUF_CHECK_MAPPED(bp);
2116 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2117 * callback that fires to avoid returning failure.
2120 bufkva_reclaim(vmem_t *vmem, int flags)
2127 for (i = 0; i < 5; i++) {
2128 for (q = 0; q < buf_domains; q++)
2129 if (buf_recycle(&bdomain[q], true) != 0)
2138 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2139 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2140 * the buffer is valid and we do not have to do anything.
2143 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2144 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2152 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2153 if (inmem(vp, *rablkno))
2155 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2156 if ((rabp->b_flags & B_CACHE) != 0) {
2163 racct_add_buf(curproc, rabp, 0);
2164 PROC_UNLOCK(curproc);
2167 td->td_ru.ru_inblock++;
2168 rabp->b_flags |= B_ASYNC;
2169 rabp->b_flags &= ~B_INVAL;
2170 if ((flags & GB_CKHASH) != 0) {
2171 rabp->b_flags |= B_CKHASH;
2172 rabp->b_ckhashcalc = ckhashfunc;
2174 rabp->b_ioflags &= ~BIO_ERROR;
2175 rabp->b_iocmd = BIO_READ;
2176 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2177 rabp->b_rcred = crhold(cred);
2178 vfs_busy_pages(rabp, 0);
2180 rabp->b_iooffset = dbtob(rabp->b_blkno);
2186 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2188 * Get a buffer with the specified data. Look in the cache first. We
2189 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2190 * is set, the buffer is valid and we do not have to do anything, see
2191 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2193 * Always return a NULL buffer pointer (in bpp) when returning an error.
2195 * The blkno parameter is the logical block being requested. Normally
2196 * the mapping of logical block number to disk block address is done
2197 * by calling VOP_BMAP(). However, if the mapping is already known, the
2198 * disk block address can be passed using the dblkno parameter. If the
2199 * disk block address is not known, then the same value should be passed
2200 * for blkno and dblkno.
2203 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2204 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2205 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2209 int error, readwait, rv;
2211 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2214 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2217 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2222 KASSERT(blkno == bp->b_lblkno,
2223 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2224 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2225 flags &= ~GB_NOSPARSE;
2229 * If not found in cache, do some I/O
2232 if ((bp->b_flags & B_CACHE) == 0) {
2235 PROC_LOCK(td->td_proc);
2236 racct_add_buf(td->td_proc, bp, 0);
2237 PROC_UNLOCK(td->td_proc);
2240 td->td_ru.ru_inblock++;
2241 bp->b_iocmd = BIO_READ;
2242 bp->b_flags &= ~B_INVAL;
2243 if ((flags & GB_CKHASH) != 0) {
2244 bp->b_flags |= B_CKHASH;
2245 bp->b_ckhashcalc = ckhashfunc;
2247 if ((flags & GB_CVTENXIO) != 0)
2248 bp->b_xflags |= BX_CVTENXIO;
2249 bp->b_ioflags &= ~BIO_ERROR;
2250 if (bp->b_rcred == NOCRED && cred != NOCRED)
2251 bp->b_rcred = crhold(cred);
2252 vfs_busy_pages(bp, 0);
2253 bp->b_iooffset = dbtob(bp->b_blkno);
2259 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2261 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2275 * Write, release buffer on completion. (Done by iodone
2276 * if async). Do not bother writing anything if the buffer
2279 * Note that we set B_CACHE here, indicating that buffer is
2280 * fully valid and thus cacheable. This is true even of NFS
2281 * now so we set it generally. This could be set either here
2282 * or in biodone() since the I/O is synchronous. We put it
2286 bufwrite(struct buf *bp)
2293 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2294 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2295 bp->b_flags |= B_INVAL | B_RELBUF;
2296 bp->b_flags &= ~B_CACHE;
2300 if (bp->b_flags & B_INVAL) {
2305 if (bp->b_flags & B_BARRIER)
2306 atomic_add_long(&barrierwrites, 1);
2308 oldflags = bp->b_flags;
2310 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2311 ("FFS background buffer should not get here %p", bp));
2315 vp_md = vp->v_vflag & VV_MD;
2320 * Mark the buffer clean. Increment the bufobj write count
2321 * before bundirty() call, to prevent other thread from seeing
2322 * empty dirty list and zero counter for writes in progress,
2323 * falsely indicating that the bufobj is clean.
2325 bufobj_wref(bp->b_bufobj);
2328 bp->b_flags &= ~B_DONE;
2329 bp->b_ioflags &= ~BIO_ERROR;
2330 bp->b_flags |= B_CACHE;
2331 bp->b_iocmd = BIO_WRITE;
2333 vfs_busy_pages(bp, 1);
2336 * Normal bwrites pipeline writes
2338 bp->b_runningbufspace = bp->b_bufsize;
2339 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2344 racct_add_buf(curproc, bp, 1);
2345 PROC_UNLOCK(curproc);
2348 curthread->td_ru.ru_oublock++;
2349 if (oldflags & B_ASYNC)
2351 bp->b_iooffset = dbtob(bp->b_blkno);
2352 buf_track(bp, __func__);
2355 if ((oldflags & B_ASYNC) == 0) {
2356 int rtval = bufwait(bp);
2359 } else if (space > hirunningspace) {
2361 * don't allow the async write to saturate the I/O
2362 * system. We will not deadlock here because
2363 * we are blocking waiting for I/O that is already in-progress
2364 * to complete. We do not block here if it is the update
2365 * or syncer daemon trying to clean up as that can lead
2368 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2369 waitrunningbufspace();
2376 bufbdflush(struct bufobj *bo, struct buf *bp)
2379 struct bufdomain *bd;
2381 bd = &bdomain[bo->bo_domain];
2382 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2383 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2385 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2388 * Try to find a buffer to flush.
2390 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2391 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2393 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2396 panic("bdwrite: found ourselves");
2398 /* Don't countdeps with the bo lock held. */
2399 if (buf_countdeps(nbp, 0)) {
2404 if (nbp->b_flags & B_CLUSTEROK) {
2405 vfs_bio_awrite(nbp);
2410 dirtybufferflushes++;
2419 * Delayed write. (Buffer is marked dirty). Do not bother writing
2420 * anything if the buffer is marked invalid.
2422 * Note that since the buffer must be completely valid, we can safely
2423 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2424 * biodone() in order to prevent getblk from writing the buffer
2425 * out synchronously.
2428 bdwrite(struct buf *bp)
2430 struct thread *td = curthread;
2434 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2435 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2436 KASSERT((bp->b_flags & B_BARRIER) == 0,
2437 ("Barrier request in delayed write %p", bp));
2439 if (bp->b_flags & B_INVAL) {
2445 * If we have too many dirty buffers, don't create any more.
2446 * If we are wildly over our limit, then force a complete
2447 * cleanup. Otherwise, just keep the situation from getting
2448 * out of control. Note that we have to avoid a recursive
2449 * disaster and not try to clean up after our own cleanup!
2453 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2454 td->td_pflags |= TDP_INBDFLUSH;
2456 td->td_pflags &= ~TDP_INBDFLUSH;
2462 * Set B_CACHE, indicating that the buffer is fully valid. This is
2463 * true even of NFS now.
2465 bp->b_flags |= B_CACHE;
2468 * This bmap keeps the system from needing to do the bmap later,
2469 * perhaps when the system is attempting to do a sync. Since it
2470 * is likely that the indirect block -- or whatever other datastructure
2471 * that the filesystem needs is still in memory now, it is a good
2472 * thing to do this. Note also, that if the pageout daemon is
2473 * requesting a sync -- there might not be enough memory to do
2474 * the bmap then... So, this is important to do.
2476 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2477 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2480 buf_track(bp, __func__);
2483 * Set the *dirty* buffer range based upon the VM system dirty
2486 * Mark the buffer pages as clean. We need to do this here to
2487 * satisfy the vnode_pager and the pageout daemon, so that it
2488 * thinks that the pages have been "cleaned". Note that since
2489 * the pages are in a delayed write buffer -- the VFS layer
2490 * "will" see that the pages get written out on the next sync,
2491 * or perhaps the cluster will be completed.
2493 vfs_clean_pages_dirty_buf(bp);
2497 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2498 * due to the softdep code.
2505 * Turn buffer into delayed write request. We must clear BIO_READ and
2506 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2507 * itself to properly update it in the dirty/clean lists. We mark it
2508 * B_DONE to ensure that any asynchronization of the buffer properly
2509 * clears B_DONE ( else a panic will occur later ).
2511 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2512 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2513 * should only be called if the buffer is known-good.
2515 * Since the buffer is not on a queue, we do not update the numfreebuffers
2518 * The buffer must be on QUEUE_NONE.
2521 bdirty(struct buf *bp)
2524 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2525 bp, bp->b_vp, bp->b_flags);
2526 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2527 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2528 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2529 bp->b_flags &= ~(B_RELBUF);
2530 bp->b_iocmd = BIO_WRITE;
2532 if ((bp->b_flags & B_DELWRI) == 0) {
2533 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2542 * Clear B_DELWRI for buffer.
2544 * Since the buffer is not on a queue, we do not update the numfreebuffers
2547 * The buffer must be on QUEUE_NONE.
2551 bundirty(struct buf *bp)
2554 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2555 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2556 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2557 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2559 if (bp->b_flags & B_DELWRI) {
2560 bp->b_flags &= ~B_DELWRI;
2565 * Since it is now being written, we can clear its deferred write flag.
2567 bp->b_flags &= ~B_DEFERRED;
2573 * Asynchronous write. Start output on a buffer, but do not wait for
2574 * it to complete. The buffer is released when the output completes.
2576 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2577 * B_INVAL buffers. Not us.
2580 bawrite(struct buf *bp)
2583 bp->b_flags |= B_ASYNC;
2590 * Asynchronous barrier write. Start output on a buffer, but do not
2591 * wait for it to complete. Place a write barrier after this write so
2592 * that this buffer and all buffers written before it are committed to
2593 * the disk before any buffers written after this write are committed
2594 * to the disk. The buffer is released when the output completes.
2597 babarrierwrite(struct buf *bp)
2600 bp->b_flags |= B_ASYNC | B_BARRIER;
2607 * Synchronous barrier write. Start output on a buffer and wait for
2608 * it to complete. Place a write barrier after this write so that
2609 * this buffer and all buffers written before it are committed to
2610 * the disk before any buffers written after this write are committed
2611 * to the disk. The buffer is released when the output completes.
2614 bbarrierwrite(struct buf *bp)
2617 bp->b_flags |= B_BARRIER;
2618 return (bwrite(bp));
2624 * Called prior to the locking of any vnodes when we are expecting to
2625 * write. We do not want to starve the buffer cache with too many
2626 * dirty buffers so we block here. By blocking prior to the locking
2627 * of any vnodes we attempt to avoid the situation where a locked vnode
2628 * prevents the various system daemons from flushing related buffers.
2634 if (buf_dirty_count_severe()) {
2635 mtx_lock(&bdirtylock);
2636 while (buf_dirty_count_severe()) {
2638 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2641 mtx_unlock(&bdirtylock);
2646 * Return true if we have too many dirty buffers.
2649 buf_dirty_count_severe(void)
2652 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2658 * Release a busy buffer and, if requested, free its resources. The
2659 * buffer will be stashed in the appropriate bufqueue[] allowing it
2660 * to be accessed later as a cache entity or reused for other purposes.
2663 brelse(struct buf *bp)
2665 struct mount *v_mnt;
2669 * Many functions erroneously call brelse with a NULL bp under rare
2670 * error conditions. Simply return when called with a NULL bp.
2674 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2675 bp, bp->b_vp, bp->b_flags);
2676 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2677 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2678 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2679 ("brelse: non-VMIO buffer marked NOREUSE"));
2681 if (BUF_LOCKRECURSED(bp)) {
2683 * Do not process, in particular, do not handle the
2684 * B_INVAL/B_RELBUF and do not release to free list.
2690 if (bp->b_flags & B_MANAGED) {
2695 if (LIST_EMPTY(&bp->b_dep)) {
2696 bp->b_flags &= ~B_IOSTARTED;
2698 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2699 ("brelse: SU io not finished bp %p", bp));
2702 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2703 BO_LOCK(bp->b_bufobj);
2704 bp->b_vflags &= ~BV_BKGRDERR;
2705 BO_UNLOCK(bp->b_bufobj);
2709 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2710 (bp->b_flags & B_INVALONERR)) {
2712 * Forced invalidation of dirty buffer contents, to be used
2713 * after a failed write in the rare case that the loss of the
2714 * contents is acceptable. The buffer is invalidated and
2717 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2718 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2721 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2722 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2723 !(bp->b_flags & B_INVAL)) {
2725 * Failed write, redirty. All errors except ENXIO (which
2726 * means the device is gone) are treated as being
2729 * XXX Treating EIO as transient is not correct; the
2730 * contract with the local storage device drivers is that
2731 * they will only return EIO once the I/O is no longer
2732 * retriable. Network I/O also respects this through the
2733 * guarantees of TCP and/or the internal retries of NFS.
2734 * ENOMEM might be transient, but we also have no way of
2735 * knowing when its ok to retry/reschedule. In general,
2736 * this entire case should be made obsolete through better
2737 * error handling/recovery and resource scheduling.
2739 * Do this also for buffers that failed with ENXIO, but have
2740 * non-empty dependencies - the soft updates code might need
2741 * to access the buffer to untangle them.
2743 * Must clear BIO_ERROR to prevent pages from being scrapped.
2745 bp->b_ioflags &= ~BIO_ERROR;
2747 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2748 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2750 * Either a failed read I/O, or we were asked to free or not
2751 * cache the buffer, or we failed to write to a device that's
2752 * no longer present.
2754 bp->b_flags |= B_INVAL;
2755 if (!LIST_EMPTY(&bp->b_dep))
2757 if (bp->b_flags & B_DELWRI)
2759 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2760 if ((bp->b_flags & B_VMIO) == 0) {
2768 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2769 * is called with B_DELWRI set, the underlying pages may wind up
2770 * getting freed causing a previous write (bdwrite()) to get 'lost'
2771 * because pages associated with a B_DELWRI bp are marked clean.
2773 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2774 * if B_DELWRI is set.
2776 if (bp->b_flags & B_DELWRI)
2777 bp->b_flags &= ~B_RELBUF;
2780 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2781 * constituted, not even NFS buffers now. Two flags effect this. If
2782 * B_INVAL, the struct buf is invalidated but the VM object is kept
2783 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2785 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2786 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2787 * buffer is also B_INVAL because it hits the re-dirtying code above.
2789 * Normally we can do this whether a buffer is B_DELWRI or not. If
2790 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2791 * the commit state and we cannot afford to lose the buffer. If the
2792 * buffer has a background write in progress, we need to keep it
2793 * around to prevent it from being reconstituted and starting a second
2797 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2799 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2800 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2801 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2802 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2803 vfs_vmio_invalidate(bp);
2807 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2808 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2810 bp->b_flags &= ~B_NOREUSE;
2811 if (bp->b_vp != NULL)
2816 * If the buffer has junk contents signal it and eventually
2817 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2820 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2821 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2822 bp->b_flags |= B_INVAL;
2823 if (bp->b_flags & B_INVAL) {
2824 if (bp->b_flags & B_DELWRI)
2830 buf_track(bp, __func__);
2832 /* buffers with no memory */
2833 if (bp->b_bufsize == 0) {
2837 /* buffers with junk contents */
2838 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2839 (bp->b_ioflags & BIO_ERROR)) {
2840 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2841 if (bp->b_vflags & BV_BKGRDINPROG)
2842 panic("losing buffer 2");
2843 qindex = QUEUE_CLEAN;
2844 bp->b_flags |= B_AGE;
2845 /* remaining buffers */
2846 } else if (bp->b_flags & B_DELWRI)
2847 qindex = QUEUE_DIRTY;
2849 qindex = QUEUE_CLEAN;
2851 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2852 panic("brelse: not dirty");
2854 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2855 bp->b_xflags &= ~(BX_CVTENXIO);
2856 /* binsfree unlocks bp. */
2857 binsfree(bp, qindex);
2861 * Release a buffer back to the appropriate queue but do not try to free
2862 * it. The buffer is expected to be used again soon.
2864 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2865 * biodone() to requeue an async I/O on completion. It is also used when
2866 * known good buffers need to be requeued but we think we may need the data
2869 * XXX we should be able to leave the B_RELBUF hint set on completion.
2872 bqrelse(struct buf *bp)
2876 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2877 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2878 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2880 qindex = QUEUE_NONE;
2881 if (BUF_LOCKRECURSED(bp)) {
2882 /* do not release to free list */
2886 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2887 bp->b_xflags &= ~(BX_CVTENXIO);
2889 if (LIST_EMPTY(&bp->b_dep)) {
2890 bp->b_flags &= ~B_IOSTARTED;
2892 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2893 ("bqrelse: SU io not finished bp %p", bp));
2896 if (bp->b_flags & B_MANAGED) {
2897 if (bp->b_flags & B_REMFREE)
2902 /* buffers with stale but valid contents */
2903 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2904 BV_BKGRDERR)) == BV_BKGRDERR) {
2905 BO_LOCK(bp->b_bufobj);
2906 bp->b_vflags &= ~BV_BKGRDERR;
2907 BO_UNLOCK(bp->b_bufobj);
2908 qindex = QUEUE_DIRTY;
2910 if ((bp->b_flags & B_DELWRI) == 0 &&
2911 (bp->b_xflags & BX_VNDIRTY))
2912 panic("bqrelse: not dirty");
2913 if ((bp->b_flags & B_NOREUSE) != 0) {
2917 qindex = QUEUE_CLEAN;
2919 buf_track(bp, __func__);
2920 /* binsfree unlocks bp. */
2921 binsfree(bp, qindex);
2925 buf_track(bp, __func__);
2931 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2932 * restore bogus pages.
2935 vfs_vmio_iodone(struct buf *bp)
2940 struct vnode *vp __unused;
2941 int i, iosize, resid;
2944 obj = bp->b_bufobj->bo_object;
2945 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2946 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2947 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2950 VNPASS(vp->v_holdcnt > 0, vp);
2951 VNPASS(vp->v_object != NULL, vp);
2953 foff = bp->b_offset;
2954 KASSERT(bp->b_offset != NOOFFSET,
2955 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2958 iosize = bp->b_bcount - bp->b_resid;
2959 for (i = 0; i < bp->b_npages; i++) {
2960 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2965 * cleanup bogus pages, restoring the originals
2968 if (m == bogus_page) {
2970 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2972 panic("biodone: page disappeared!");
2974 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2976 * In the write case, the valid and clean bits are
2977 * already changed correctly ( see bdwrite() ), so we
2978 * only need to do this here in the read case.
2980 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2981 resid)) == 0, ("vfs_vmio_iodone: page %p "
2982 "has unexpected dirty bits", m));
2983 vfs_page_set_valid(bp, foff, m);
2985 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2986 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2987 (intmax_t)foff, (uintmax_t)m->pindex));
2990 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2993 vm_object_pip_wakeupn(obj, bp->b_npages);
2994 if (bogus && buf_mapped(bp)) {
2995 BUF_CHECK_MAPPED(bp);
2996 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2997 bp->b_pages, bp->b_npages);
3002 * Perform page invalidation when a buffer is released. The fully invalid
3003 * pages will be reclaimed later in vfs_vmio_truncate().
3006 vfs_vmio_invalidate(struct buf *bp)
3010 int flags, i, resid, poffset, presid;
3012 if (buf_mapped(bp)) {
3013 BUF_CHECK_MAPPED(bp);
3014 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3016 BUF_CHECK_UNMAPPED(bp);
3018 * Get the base offset and length of the buffer. Note that
3019 * in the VMIO case if the buffer block size is not
3020 * page-aligned then b_data pointer may not be page-aligned.
3021 * But our b_pages[] array *IS* page aligned.
3023 * block sizes less then DEV_BSIZE (usually 512) are not
3024 * supported due to the page granularity bits (m->valid,
3025 * m->dirty, etc...).
3027 * See man buf(9) for more information
3029 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3030 obj = bp->b_bufobj->bo_object;
3031 resid = bp->b_bufsize;
3032 poffset = bp->b_offset & PAGE_MASK;
3033 VM_OBJECT_WLOCK(obj);
3034 for (i = 0; i < bp->b_npages; i++) {
3036 if (m == bogus_page)
3037 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3038 bp->b_pages[i] = NULL;
3040 presid = resid > (PAGE_SIZE - poffset) ?
3041 (PAGE_SIZE - poffset) : resid;
3042 KASSERT(presid >= 0, ("brelse: extra page"));
3043 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3044 if (pmap_page_wired_mappings(m) == 0)
3045 vm_page_set_invalid(m, poffset, presid);
3047 vm_page_release_locked(m, flags);
3051 VM_OBJECT_WUNLOCK(obj);
3056 * Page-granular truncation of an existing VMIO buffer.
3059 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3065 if (bp->b_npages == desiredpages)
3068 if (buf_mapped(bp)) {
3069 BUF_CHECK_MAPPED(bp);
3070 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3071 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3073 BUF_CHECK_UNMAPPED(bp);
3076 * The object lock is needed only if we will attempt to free pages.
3078 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3079 if ((bp->b_flags & B_DIRECT) != 0) {
3080 flags |= VPR_TRYFREE;
3081 obj = bp->b_bufobj->bo_object;
3082 VM_OBJECT_WLOCK(obj);
3086 for (i = desiredpages; i < bp->b_npages; i++) {
3088 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3089 bp->b_pages[i] = NULL;
3091 vm_page_release_locked(m, flags);
3093 vm_page_release(m, flags);
3096 VM_OBJECT_WUNLOCK(obj);
3097 bp->b_npages = desiredpages;
3101 * Byte granular extension of VMIO buffers.
3104 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3107 * We are growing the buffer, possibly in a
3108 * byte-granular fashion.
3116 * Step 1, bring in the VM pages from the object, allocating
3117 * them if necessary. We must clear B_CACHE if these pages
3118 * are not valid for the range covered by the buffer.
3120 obj = bp->b_bufobj->bo_object;
3121 if (bp->b_npages < desiredpages) {
3122 KASSERT(desiredpages <= atop(maxbcachebuf),
3123 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3124 bp, desiredpages, maxbcachebuf));
3127 * We must allocate system pages since blocking
3128 * here could interfere with paging I/O, no
3129 * matter which process we are.
3131 * Only exclusive busy can be tested here.
3132 * Blocking on shared busy might lead to
3133 * deadlocks once allocbuf() is called after
3134 * pages are vfs_busy_pages().
3136 (void)vm_page_grab_pages_unlocked(obj,
3137 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3138 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3139 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3140 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3141 bp->b_npages = desiredpages;
3145 * Step 2. We've loaded the pages into the buffer,
3146 * we have to figure out if we can still have B_CACHE
3147 * set. Note that B_CACHE is set according to the
3148 * byte-granular range ( bcount and size ), not the
3149 * aligned range ( newbsize ).
3151 * The VM test is against m->valid, which is DEV_BSIZE
3152 * aligned. Needless to say, the validity of the data
3153 * needs to also be DEV_BSIZE aligned. Note that this
3154 * fails with NFS if the server or some other client
3155 * extends the file's EOF. If our buffer is resized,
3156 * B_CACHE may remain set! XXX
3158 toff = bp->b_bcount;
3159 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3160 while ((bp->b_flags & B_CACHE) && toff < size) {
3163 if (tinc > (size - toff))
3165 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3166 m = bp->b_pages[pi];
3167 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3173 * Step 3, fixup the KVA pmap.
3178 BUF_CHECK_UNMAPPED(bp);
3182 * Check to see if a block at a particular lbn is available for a clustered
3186 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3193 /* If the buf isn't in core skip it */
3194 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3197 /* If the buf is busy we don't want to wait for it */
3198 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3201 /* Only cluster with valid clusterable delayed write buffers */
3202 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3203 (B_DELWRI | B_CLUSTEROK))
3206 if (bpa->b_bufsize != size)
3210 * Check to see if it is in the expected place on disk and that the
3211 * block has been mapped.
3213 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3223 * Implement clustered async writes for clearing out B_DELWRI buffers.
3224 * This is much better then the old way of writing only one buffer at
3225 * a time. Note that we may not be presented with the buffers in the
3226 * correct order, so we search for the cluster in both directions.
3229 vfs_bio_awrite(struct buf *bp)
3234 daddr_t lblkno = bp->b_lblkno;
3235 struct vnode *vp = bp->b_vp;
3243 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3245 * right now we support clustered writing only to regular files. If
3246 * we find a clusterable block we could be in the middle of a cluster
3247 * rather then at the beginning.
3249 if ((vp->v_type == VREG) &&
3250 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3251 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3252 size = vp->v_mount->mnt_stat.f_iosize;
3253 maxcl = maxphys / size;
3256 for (i = 1; i < maxcl; i++)
3257 if (vfs_bio_clcheck(vp, size, lblkno + i,
3258 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3261 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3262 if (vfs_bio_clcheck(vp, size, lblkno - j,
3263 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3269 * this is a possible cluster write
3273 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3279 bp->b_flags |= B_ASYNC;
3281 * default (old) behavior, writing out only one block
3283 * XXX returns b_bufsize instead of b_bcount for nwritten?
3285 nwritten = bp->b_bufsize;
3294 * Allocate KVA for an empty buf header according to gbflags.
3297 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3300 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3302 * In order to keep fragmentation sane we only allocate kva
3303 * in BKVASIZE chunks. XXX with vmem we can do page size.
3305 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3307 if (maxsize != bp->b_kvasize &&
3308 bufkva_alloc(bp, maxsize, gbflags))
3317 * Find and initialize a new buffer header, freeing up existing buffers
3318 * in the bufqueues as necessary. The new buffer is returned locked.
3321 * We have insufficient buffer headers
3322 * We have insufficient buffer space
3323 * buffer_arena is too fragmented ( space reservation fails )
3324 * If we have to flush dirty buffers ( but we try to avoid this )
3326 * The caller is responsible for releasing the reserved bufspace after
3327 * allocbuf() is called.
3330 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3332 struct bufdomain *bd;
3334 bool metadata, reserved;
3337 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3338 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3339 if (!unmapped_buf_allowed)
3340 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3342 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3350 bd = &bdomain[vp->v_bufobj.bo_domain];
3352 counter_u64_add(getnewbufcalls, 1);
3355 if (reserved == false &&
3356 bufspace_reserve(bd, maxsize, metadata) != 0) {
3357 counter_u64_add(getnewbufrestarts, 1);
3361 if ((bp = buf_alloc(bd)) == NULL) {
3362 counter_u64_add(getnewbufrestarts, 1);
3365 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3368 } while (buf_recycle(bd, false) == 0);
3371 bufspace_release(bd, maxsize);
3373 bp->b_flags |= B_INVAL;
3376 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3384 * buffer flushing daemon. Buffers are normally flushed by the
3385 * update daemon but if it cannot keep up this process starts to
3386 * take the load in an attempt to prevent getnewbuf() from blocking.
3388 static struct kproc_desc buf_kp = {
3393 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3396 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3400 flushed = flushbufqueues(vp, bd, target, 0);
3403 * Could not find any buffers without rollback
3404 * dependencies, so just write the first one
3405 * in the hopes of eventually making progress.
3407 if (vp != NULL && target > 2)
3409 flushbufqueues(vp, bd, target, 1);
3415 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3421 wakeup(&bd_request);
3422 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3424 mtx_unlock(&bdlock);
3426 printf("bufdaemon wait error: %d\n", error);
3432 struct bufdomain *bd;
3438 * This process needs to be suspended prior to shutdown sync.
3440 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3441 SHUTDOWN_PRI_LAST + 100);
3444 * Start the buf clean daemons as children threads.
3446 for (i = 0 ; i < buf_domains; i++) {
3449 error = kthread_add((void (*)(void *))bufspace_daemon,
3450 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3452 panic("error %d spawning bufspace daemon", error);
3456 * This process is allowed to take the buffer cache to the limit
3458 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3460 while (!bd_shutdown) {
3462 mtx_unlock(&bdlock);
3465 * Save speedupreq for this pass and reset to capture new
3468 speedupreq = bd_speedupreq;
3472 * Flush each domain sequentially according to its level and
3473 * the speedup request.
3475 for (i = 0; i < buf_domains; i++) {
3478 lodirty = bd->bd_numdirtybuffers / 2;
3480 lodirty = bd->bd_lodirtybuffers;
3481 while (bd->bd_numdirtybuffers > lodirty) {
3482 if (buf_flush(NULL, bd,
3483 bd->bd_numdirtybuffers - lodirty) == 0)
3485 kern_yield(PRI_USER);
3490 * Only clear bd_request if we have reached our low water
3491 * mark. The buf_daemon normally waits 1 second and
3492 * then incrementally flushes any dirty buffers that have
3493 * built up, within reason.
3495 * If we were unable to hit our low water mark and couldn't
3496 * find any flushable buffers, we sleep for a short period
3497 * to avoid endless loops on unlockable buffers.
3502 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3504 * We reached our low water mark, reset the
3505 * request and sleep until we are needed again.
3506 * The sleep is just so the suspend code works.
3510 * Do an extra wakeup in case dirty threshold
3511 * changed via sysctl and the explicit transition
3512 * out of shortfall was missed.
3515 if (runningbufspace <= lorunningspace)
3517 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3520 * We couldn't find any flushable dirty buffers but
3521 * still have too many dirty buffers, we
3522 * have to sleep and try again. (rare)
3524 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3527 wakeup(&bd_shutdown);
3528 mtx_unlock(&bdlock);
3535 * Try to flush a buffer in the dirty queue. We must be careful to
3536 * free up B_INVAL buffers instead of write them, which NFS is
3537 * particularly sensitive to.
3539 static int flushwithdeps = 0;
3540 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3542 "Number of buffers flushed with dependencies that require rollbacks");
3545 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3548 struct bufqueue *bq;
3549 struct buf *sentinel;
3559 bq = &bd->bd_dirtyq;
3561 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3562 sentinel->b_qindex = QUEUE_SENTINEL;
3564 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3566 while (flushed != target) {
3569 bp = TAILQ_NEXT(sentinel, b_freelist);
3571 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3572 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3579 * Skip sentinels inserted by other invocations of the
3580 * flushbufqueues(), taking care to not reorder them.
3582 * Only flush the buffers that belong to the
3583 * vnode locked by the curthread.
3585 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3590 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3596 * BKGRDINPROG can only be set with the buf and bufobj
3597 * locks both held. We tolerate a race to clear it here.
3599 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3600 (bp->b_flags & B_DELWRI) == 0) {
3604 if (bp->b_flags & B_INVAL) {
3611 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3612 if (flushdeps == 0) {
3620 * We must hold the lock on a vnode before writing
3621 * one of its buffers. Otherwise we may confuse, or
3622 * in the case of a snapshot vnode, deadlock the
3625 * The lock order here is the reverse of the normal
3626 * of vnode followed by buf lock. This is ok because
3627 * the NOWAIT will prevent deadlock.
3630 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3636 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3638 ASSERT_VOP_LOCKED(vp, "getbuf");
3640 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3641 vn_lock(vp, LK_TRYUPGRADE);
3644 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3645 bp, bp->b_vp, bp->b_flags);
3646 if (curproc == bufdaemonproc) {
3651 counter_u64_add(notbufdflushes, 1);
3653 vn_finished_write(mp);
3656 flushwithdeps += hasdeps;
3660 * Sleeping on runningbufspace while holding
3661 * vnode lock leads to deadlock.
3663 if (curproc == bufdaemonproc &&
3664 runningbufspace > hirunningspace)
3665 waitrunningbufspace();
3668 vn_finished_write(mp);
3672 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3674 free(sentinel, M_TEMP);
3679 * Check to see if a block is currently memory resident.
3682 incore(struct bufobj *bo, daddr_t blkno)
3684 return (gbincore_unlocked(bo, blkno));
3688 * Returns true if no I/O is needed to access the
3689 * associated VM object. This is like incore except
3690 * it also hunts around in the VM system for the data.
3693 inmem(struct vnode * vp, daddr_t blkno)
3696 vm_offset_t toff, tinc, size;
3701 ASSERT_VOP_LOCKED(vp, "inmem");
3703 if (incore(&vp->v_bufobj, blkno))
3705 if (vp->v_mount == NULL)
3712 if (size > vp->v_mount->mnt_stat.f_iosize)
3713 size = vp->v_mount->mnt_stat.f_iosize;
3714 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3716 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3717 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3723 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3724 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3726 * Consider page validity only if page mapping didn't change
3729 valid = vm_page_is_valid(m,
3730 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3731 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3743 * Set the dirty range for a buffer based on the status of the dirty
3744 * bits in the pages comprising the buffer. The range is limited
3745 * to the size of the buffer.
3747 * Tell the VM system that the pages associated with this buffer
3748 * are clean. This is used for delayed writes where the data is
3749 * going to go to disk eventually without additional VM intevention.
3751 * Note that while we only really need to clean through to b_bcount, we
3752 * just go ahead and clean through to b_bufsize.
3755 vfs_clean_pages_dirty_buf(struct buf *bp)
3757 vm_ooffset_t foff, noff, eoff;
3761 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3764 foff = bp->b_offset;
3765 KASSERT(bp->b_offset != NOOFFSET,
3766 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3768 vfs_busy_pages_acquire(bp);
3769 vfs_setdirty_range(bp);
3770 for (i = 0; i < bp->b_npages; i++) {
3771 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3773 if (eoff > bp->b_offset + bp->b_bufsize)
3774 eoff = bp->b_offset + bp->b_bufsize;
3776 vfs_page_set_validclean(bp, foff, m);
3777 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3780 vfs_busy_pages_release(bp);
3784 vfs_setdirty_range(struct buf *bp)
3786 vm_offset_t boffset;
3787 vm_offset_t eoffset;
3791 * test the pages to see if they have been modified directly
3792 * by users through the VM system.
3794 for (i = 0; i < bp->b_npages; i++)
3795 vm_page_test_dirty(bp->b_pages[i]);
3798 * Calculate the encompassing dirty range, boffset and eoffset,
3799 * (eoffset - boffset) bytes.
3802 for (i = 0; i < bp->b_npages; i++) {
3803 if (bp->b_pages[i]->dirty)
3806 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3808 for (i = bp->b_npages - 1; i >= 0; --i) {
3809 if (bp->b_pages[i]->dirty) {
3813 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3816 * Fit it to the buffer.
3819 if (eoffset > bp->b_bcount)
3820 eoffset = bp->b_bcount;
3823 * If we have a good dirty range, merge with the existing
3827 if (boffset < eoffset) {
3828 if (bp->b_dirtyoff > boffset)
3829 bp->b_dirtyoff = boffset;
3830 if (bp->b_dirtyend < eoffset)
3831 bp->b_dirtyend = eoffset;
3836 * Allocate the KVA mapping for an existing buffer.
3837 * If an unmapped buffer is provided but a mapped buffer is requested, take
3838 * also care to properly setup mappings between pages and KVA.
3841 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3843 int bsize, maxsize, need_mapping, need_kva;
3846 need_mapping = bp->b_data == unmapped_buf &&
3847 (gbflags & GB_UNMAPPED) == 0;
3848 need_kva = bp->b_kvabase == unmapped_buf &&
3849 bp->b_data == unmapped_buf &&
3850 (gbflags & GB_KVAALLOC) != 0;
3851 if (!need_mapping && !need_kva)
3854 BUF_CHECK_UNMAPPED(bp);
3856 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3858 * Buffer is not mapped, but the KVA was already
3859 * reserved at the time of the instantiation. Use the
3866 * Calculate the amount of the address space we would reserve
3867 * if the buffer was mapped.
3869 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3870 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3871 offset = blkno * bsize;
3872 maxsize = size + (offset & PAGE_MASK);
3873 maxsize = imax(maxsize, bsize);
3875 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3876 if ((gbflags & GB_NOWAIT_BD) != 0) {
3878 * XXXKIB: defragmentation cannot
3879 * succeed, not sure what else to do.
3881 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3883 counter_u64_add(mappingrestarts, 1);
3884 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3888 /* b_offset is handled by bpmap_qenter. */
3889 bp->b_data = bp->b_kvabase;
3890 BUF_CHECK_MAPPED(bp);
3896 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3902 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3911 * Get a block given a specified block and offset into a file/device.
3912 * The buffers B_DONE bit will be cleared on return, making it almost
3913 * ready for an I/O initiation. B_INVAL may or may not be set on
3914 * return. The caller should clear B_INVAL prior to initiating a
3917 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3918 * an existing buffer.
3920 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3921 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3922 * and then cleared based on the backing VM. If the previous buffer is
3923 * non-0-sized but invalid, B_CACHE will be cleared.
3925 * If getblk() must create a new buffer, the new buffer is returned with
3926 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3927 * case it is returned with B_INVAL clear and B_CACHE set based on the
3930 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3931 * B_CACHE bit is clear.
3933 * What this means, basically, is that the caller should use B_CACHE to
3934 * determine whether the buffer is fully valid or not and should clear
3935 * B_INVAL prior to issuing a read. If the caller intends to validate
3936 * the buffer by loading its data area with something, the caller needs
3937 * to clear B_INVAL. If the caller does this without issuing an I/O,
3938 * the caller should set B_CACHE ( as an optimization ), else the caller
3939 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3940 * a write attempt or if it was a successful read. If the caller
3941 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3942 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3944 * The blkno parameter is the logical block being requested. Normally
3945 * the mapping of logical block number to disk block address is done
3946 * by calling VOP_BMAP(). However, if the mapping is already known, the
3947 * disk block address can be passed using the dblkno parameter. If the
3948 * disk block address is not known, then the same value should be passed
3949 * for blkno and dblkno.
3952 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3953 int slptimeo, int flags, struct buf **bpp)
3958 int bsize, error, maxsize, vmio;
3961 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3962 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3963 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3964 if (vp->v_type != VCHR)
3965 ASSERT_VOP_LOCKED(vp, "getblk");
3966 if (size > maxbcachebuf)
3967 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3969 if (!unmapped_buf_allowed)
3970 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3975 /* Attempt lockless lookup first. */
3976 bp = gbincore_unlocked(bo, blkno);
3979 * With GB_NOCREAT we must be sure about not finding the buffer
3980 * as it may have been reassigned during unlocked lookup.
3982 if ((flags & GB_NOCREAT) != 0)
3984 goto newbuf_unlocked;
3987 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3992 /* Verify buf identify has not changed since lookup. */
3993 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3994 goto foundbuf_fastpath;
3996 /* It changed, fallback to locked lookup. */
4001 bp = gbincore(bo, blkno);
4006 * Buffer is in-core. If the buffer is not busy nor managed,
4007 * it must be on a queue.
4009 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
4010 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
4012 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4015 error = BUF_TIMELOCK(bp, lockflags,
4016 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4019 * If we slept and got the lock we have to restart in case
4020 * the buffer changed identities.
4022 if (error == ENOLCK)
4024 /* We timed out or were interrupted. */
4025 else if (error != 0)
4029 /* If recursed, assume caller knows the rules. */
4030 if (BUF_LOCKRECURSED(bp))
4034 * The buffer is locked. B_CACHE is cleared if the buffer is
4035 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4036 * and for a VMIO buffer B_CACHE is adjusted according to the
4039 if (bp->b_flags & B_INVAL)
4040 bp->b_flags &= ~B_CACHE;
4041 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4042 bp->b_flags |= B_CACHE;
4043 if (bp->b_flags & B_MANAGED)
4044 MPASS(bp->b_qindex == QUEUE_NONE);
4049 * check for size inconsistencies for non-VMIO case.
4051 if (bp->b_bcount != size) {
4052 if ((bp->b_flags & B_VMIO) == 0 ||
4053 (size > bp->b_kvasize)) {
4054 if (bp->b_flags & B_DELWRI) {
4055 bp->b_flags |= B_NOCACHE;
4058 if (LIST_EMPTY(&bp->b_dep)) {
4059 bp->b_flags |= B_RELBUF;
4062 bp->b_flags |= B_NOCACHE;
4071 * Handle the case of unmapped buffer which should
4072 * become mapped, or the buffer for which KVA
4073 * reservation is requested.
4075 bp_unmapped_get_kva(bp, blkno, size, flags);
4078 * If the size is inconsistent in the VMIO case, we can resize
4079 * the buffer. This might lead to B_CACHE getting set or
4080 * cleared. If the size has not changed, B_CACHE remains
4081 * unchanged from its previous state.
4085 KASSERT(bp->b_offset != NOOFFSET,
4086 ("getblk: no buffer offset"));
4089 * A buffer with B_DELWRI set and B_CACHE clear must
4090 * be committed before we can return the buffer in
4091 * order to prevent the caller from issuing a read
4092 * ( due to B_CACHE not being set ) and overwriting
4095 * Most callers, including NFS and FFS, need this to
4096 * operate properly either because they assume they
4097 * can issue a read if B_CACHE is not set, or because
4098 * ( for example ) an uncached B_DELWRI might loop due
4099 * to softupdates re-dirtying the buffer. In the latter
4100 * case, B_CACHE is set after the first write completes,
4101 * preventing further loops.
4102 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4103 * above while extending the buffer, we cannot allow the
4104 * buffer to remain with B_CACHE set after the write
4105 * completes or it will represent a corrupt state. To
4106 * deal with this we set B_NOCACHE to scrap the buffer
4109 * We might be able to do something fancy, like setting
4110 * B_CACHE in bwrite() except if B_DELWRI is already set,
4111 * so the below call doesn't set B_CACHE, but that gets real
4112 * confusing. This is much easier.
4115 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4116 bp->b_flags |= B_NOCACHE;
4120 bp->b_flags &= ~B_DONE;
4123 * Buffer is not in-core, create new buffer. The buffer
4124 * returned by getnewbuf() is locked. Note that the returned
4125 * buffer is also considered valid (not marked B_INVAL).
4130 * If the user does not want us to create the buffer, bail out
4133 if (flags & GB_NOCREAT)
4136 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4137 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4138 offset = blkno * bsize;
4139 vmio = vp->v_object != NULL;
4141 maxsize = size + (offset & PAGE_MASK);
4144 /* Do not allow non-VMIO notmapped buffers. */
4145 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4147 maxsize = imax(maxsize, bsize);
4148 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4150 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4151 KASSERT(error != EOPNOTSUPP,
4152 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4157 return (EJUSTRETURN);
4160 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4162 if (slpflag || slptimeo)
4165 * XXX This is here until the sleep path is diagnosed
4166 * enough to work under very low memory conditions.
4168 * There's an issue on low memory, 4BSD+non-preempt
4169 * systems (eg MIPS routers with 32MB RAM) where buffer
4170 * exhaustion occurs without sleeping for buffer
4171 * reclaimation. This just sticks in a loop and
4172 * constantly attempts to allocate a buffer, which
4173 * hits exhaustion and tries to wakeup bufdaemon.
4174 * This never happens because we never yield.
4176 * The real solution is to identify and fix these cases
4177 * so we aren't effectively busy-waiting in a loop
4178 * until the reclaimation path has cycles to run.
4180 kern_yield(PRI_USER);
4185 * This code is used to make sure that a buffer is not
4186 * created while the getnewbuf routine is blocked.
4187 * This can be a problem whether the vnode is locked or not.
4188 * If the buffer is created out from under us, we have to
4189 * throw away the one we just created.
4191 * Note: this must occur before we associate the buffer
4192 * with the vp especially considering limitations in
4193 * the splay tree implementation when dealing with duplicate
4197 if (gbincore(bo, blkno)) {
4199 bp->b_flags |= B_INVAL;
4200 bufspace_release(bufdomain(bp), maxsize);
4206 * Insert the buffer into the hash, so that it can
4207 * be found by incore.
4209 bp->b_lblkno = blkno;
4210 bp->b_blkno = d_blkno;
4211 bp->b_offset = offset;
4216 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4217 * buffer size starts out as 0, B_CACHE will be set by
4218 * allocbuf() for the VMIO case prior to it testing the
4219 * backing store for validity.
4223 bp->b_flags |= B_VMIO;
4224 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4225 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4226 bp, vp->v_object, bp->b_bufobj->bo_object));
4228 bp->b_flags &= ~B_VMIO;
4229 KASSERT(bp->b_bufobj->bo_object == NULL,
4230 ("ARGH! has b_bufobj->bo_object %p %p\n",
4231 bp, bp->b_bufobj->bo_object));
4232 BUF_CHECK_MAPPED(bp);
4236 bufspace_release(bufdomain(bp), maxsize);
4237 bp->b_flags &= ~B_DONE;
4239 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4241 buf_track(bp, __func__);
4242 KASSERT(bp->b_bufobj == bo,
4243 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4249 * Get an empty, disassociated buffer of given size. The buffer is initially
4253 geteblk(int size, int flags)
4258 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4259 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4260 if ((flags & GB_NOWAIT_BD) &&
4261 (curthread->td_pflags & TDP_BUFNEED) != 0)
4265 bufspace_release(bufdomain(bp), maxsize);
4266 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4271 * Truncate the backing store for a non-vmio buffer.
4274 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4277 if (bp->b_flags & B_MALLOC) {
4279 * malloced buffers are not shrunk
4281 if (newbsize == 0) {
4282 bufmallocadjust(bp, 0);
4283 free(bp->b_data, M_BIOBUF);
4284 bp->b_data = bp->b_kvabase;
4285 bp->b_flags &= ~B_MALLOC;
4289 vm_hold_free_pages(bp, newbsize);
4290 bufspace_adjust(bp, newbsize);
4294 * Extend the backing for a non-VMIO buffer.
4297 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4303 * We only use malloced memory on the first allocation.
4304 * and revert to page-allocated memory when the buffer
4307 * There is a potential smp race here that could lead
4308 * to bufmallocspace slightly passing the max. It
4309 * is probably extremely rare and not worth worrying
4312 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4313 bufmallocspace < maxbufmallocspace) {
4314 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4315 bp->b_flags |= B_MALLOC;
4316 bufmallocadjust(bp, newbsize);
4321 * If the buffer is growing on its other-than-first
4322 * allocation then we revert to the page-allocation
4327 if (bp->b_flags & B_MALLOC) {
4328 origbuf = bp->b_data;
4329 origbufsize = bp->b_bufsize;
4330 bp->b_data = bp->b_kvabase;
4331 bufmallocadjust(bp, 0);
4332 bp->b_flags &= ~B_MALLOC;
4333 newbsize = round_page(newbsize);
4335 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4336 (vm_offset_t) bp->b_data + newbsize);
4337 if (origbuf != NULL) {
4338 bcopy(origbuf, bp->b_data, origbufsize);
4339 free(origbuf, M_BIOBUF);
4341 bufspace_adjust(bp, newbsize);
4345 * This code constitutes the buffer memory from either anonymous system
4346 * memory (in the case of non-VMIO operations) or from an associated
4347 * VM object (in the case of VMIO operations). This code is able to
4348 * resize a buffer up or down.
4350 * Note that this code is tricky, and has many complications to resolve
4351 * deadlock or inconsistent data situations. Tread lightly!!!
4352 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4353 * the caller. Calling this code willy nilly can result in the loss of data.
4355 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4356 * B_CACHE for the non-VMIO case.
4359 allocbuf(struct buf *bp, int size)
4363 if (bp->b_bcount == size)
4366 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4367 panic("allocbuf: buffer too small");
4369 newbsize = roundup2(size, DEV_BSIZE);
4370 if ((bp->b_flags & B_VMIO) == 0) {
4371 if ((bp->b_flags & B_MALLOC) == 0)
4372 newbsize = round_page(newbsize);
4374 * Just get anonymous memory from the kernel. Don't
4375 * mess with B_CACHE.
4377 if (newbsize < bp->b_bufsize)
4378 vfs_nonvmio_truncate(bp, newbsize);
4379 else if (newbsize > bp->b_bufsize)
4380 vfs_nonvmio_extend(bp, newbsize);
4384 desiredpages = (size == 0) ? 0 :
4385 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4387 if (bp->b_flags & B_MALLOC)
4388 panic("allocbuf: VMIO buffer can't be malloced");
4390 * Set B_CACHE initially if buffer is 0 length or will become
4393 if (size == 0 || bp->b_bufsize == 0)
4394 bp->b_flags |= B_CACHE;
4396 if (newbsize < bp->b_bufsize)
4397 vfs_vmio_truncate(bp, desiredpages);
4398 /* XXX This looks as if it should be newbsize > b_bufsize */
4399 else if (size > bp->b_bcount)
4400 vfs_vmio_extend(bp, desiredpages, size);
4401 bufspace_adjust(bp, newbsize);
4403 bp->b_bcount = size; /* requested buffer size. */
4407 extern int inflight_transient_maps;
4409 static struct bio_queue nondump_bios;
4412 biodone(struct bio *bp)
4415 void (*done)(struct bio *);
4416 vm_offset_t start, end;
4418 biotrack(bp, __func__);
4421 * Avoid completing I/O when dumping after a panic since that may
4422 * result in a deadlock in the filesystem or pager code. Note that
4423 * this doesn't affect dumps that were started manually since we aim
4424 * to keep the system usable after it has been resumed.
4426 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4427 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4430 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4431 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4432 bp->bio_flags |= BIO_UNMAPPED;
4433 start = trunc_page((vm_offset_t)bp->bio_data);
4434 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4435 bp->bio_data = unmapped_buf;
4436 pmap_qremove(start, atop(end - start));
4437 vmem_free(transient_arena, start, end - start);
4438 atomic_add_int(&inflight_transient_maps, -1);
4440 done = bp->bio_done;
4442 * The check for done == biodone is to allow biodone to be
4443 * used as a bio_done routine.
4445 if (done == NULL || done == biodone) {
4446 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4448 bp->bio_flags |= BIO_DONE;
4456 * Wait for a BIO to finish.
4459 biowait(struct bio *bp, const char *wmesg)
4463 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4465 while ((bp->bio_flags & BIO_DONE) == 0)
4466 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4468 if (bp->bio_error != 0)
4469 return (bp->bio_error);
4470 if (!(bp->bio_flags & BIO_ERROR))
4476 biofinish(struct bio *bp, struct devstat *stat, int error)
4480 bp->bio_error = error;
4481 bp->bio_flags |= BIO_ERROR;
4484 devstat_end_transaction_bio(stat, bp);
4488 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4490 biotrack_buf(struct bio *bp, const char *location)
4493 buf_track(bp->bio_track_bp, location);
4500 * Wait for buffer I/O completion, returning error status. The buffer
4501 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4502 * error and cleared.
4505 bufwait(struct buf *bp)
4507 if (bp->b_iocmd == BIO_READ)
4508 bwait(bp, PRIBIO, "biord");
4510 bwait(bp, PRIBIO, "biowr");
4511 if (bp->b_flags & B_EINTR) {
4512 bp->b_flags &= ~B_EINTR;
4515 if (bp->b_ioflags & BIO_ERROR) {
4516 return (bp->b_error ? bp->b_error : EIO);
4525 * Finish I/O on a buffer, optionally calling a completion function.
4526 * This is usually called from an interrupt so process blocking is
4529 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4530 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4531 * assuming B_INVAL is clear.
4533 * For the VMIO case, we set B_CACHE if the op was a read and no
4534 * read error occurred, or if the op was a write. B_CACHE is never
4535 * set if the buffer is invalid or otherwise uncacheable.
4537 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4538 * initiator to leave B_INVAL set to brelse the buffer out of existence
4539 * in the biodone routine.
4542 bufdone(struct buf *bp)
4544 struct bufobj *dropobj;
4545 void (*biodone)(struct buf *);
4547 buf_track(bp, __func__);
4548 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4551 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4553 runningbufwakeup(bp);
4554 if (bp->b_iocmd == BIO_WRITE)
4555 dropobj = bp->b_bufobj;
4556 /* call optional completion function if requested */
4557 if (bp->b_iodone != NULL) {
4558 biodone = bp->b_iodone;
4559 bp->b_iodone = NULL;
4562 bufobj_wdrop(dropobj);
4565 if (bp->b_flags & B_VMIO) {
4567 * Set B_CACHE if the op was a normal read and no error
4568 * occurred. B_CACHE is set for writes in the b*write()
4571 if (bp->b_iocmd == BIO_READ &&
4572 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4573 !(bp->b_ioflags & BIO_ERROR))
4574 bp->b_flags |= B_CACHE;
4575 vfs_vmio_iodone(bp);
4577 if (!LIST_EMPTY(&bp->b_dep))
4579 if ((bp->b_flags & B_CKHASH) != 0) {
4580 KASSERT(bp->b_iocmd == BIO_READ,
4581 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4582 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4583 (*bp->b_ckhashcalc)(bp);
4586 * For asynchronous completions, release the buffer now. The brelse
4587 * will do a wakeup there if necessary - so no need to do a wakeup
4588 * here in the async case. The sync case always needs to do a wakeup.
4590 if (bp->b_flags & B_ASYNC) {
4591 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4592 (bp->b_ioflags & BIO_ERROR))
4599 bufobj_wdrop(dropobj);
4603 * This routine is called in lieu of iodone in the case of
4604 * incomplete I/O. This keeps the busy status for pages
4608 vfs_unbusy_pages(struct buf *bp)
4614 runningbufwakeup(bp);
4615 if (!(bp->b_flags & B_VMIO))
4618 obj = bp->b_bufobj->bo_object;
4619 for (i = 0; i < bp->b_npages; i++) {
4621 if (m == bogus_page) {
4622 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4624 panic("vfs_unbusy_pages: page missing\n");
4626 if (buf_mapped(bp)) {
4627 BUF_CHECK_MAPPED(bp);
4628 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4629 bp->b_pages, bp->b_npages);
4631 BUF_CHECK_UNMAPPED(bp);
4635 vm_object_pip_wakeupn(obj, bp->b_npages);
4639 * vfs_page_set_valid:
4641 * Set the valid bits in a page based on the supplied offset. The
4642 * range is restricted to the buffer's size.
4644 * This routine is typically called after a read completes.
4647 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4652 * Compute the end offset, eoff, such that [off, eoff) does not span a
4653 * page boundary and eoff is not greater than the end of the buffer.
4654 * The end of the buffer, in this case, is our file EOF, not the
4655 * allocation size of the buffer.
4657 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4658 if (eoff > bp->b_offset + bp->b_bcount)
4659 eoff = bp->b_offset + bp->b_bcount;
4662 * Set valid range. This is typically the entire buffer and thus the
4666 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4670 * vfs_page_set_validclean:
4672 * Set the valid bits and clear the dirty bits in a page based on the
4673 * supplied offset. The range is restricted to the buffer's size.
4676 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4678 vm_ooffset_t soff, eoff;
4681 * Start and end offsets in buffer. eoff - soff may not cross a
4682 * page boundary or cross the end of the buffer. The end of the
4683 * buffer, in this case, is our file EOF, not the allocation size
4687 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4688 if (eoff > bp->b_offset + bp->b_bcount)
4689 eoff = bp->b_offset + bp->b_bcount;
4692 * Set valid range. This is typically the entire buffer and thus the
4696 vm_page_set_validclean(
4698 (vm_offset_t) (soff & PAGE_MASK),
4699 (vm_offset_t) (eoff - soff)
4705 * Acquire a shared busy on all pages in the buf.
4708 vfs_busy_pages_acquire(struct buf *bp)
4712 for (i = 0; i < bp->b_npages; i++)
4713 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4717 vfs_busy_pages_release(struct buf *bp)
4721 for (i = 0; i < bp->b_npages; i++)
4722 vm_page_sunbusy(bp->b_pages[i]);
4726 * This routine is called before a device strategy routine.
4727 * It is used to tell the VM system that paging I/O is in
4728 * progress, and treat the pages associated with the buffer
4729 * almost as being exclusive busy. Also the object paging_in_progress
4730 * flag is handled to make sure that the object doesn't become
4733 * Since I/O has not been initiated yet, certain buffer flags
4734 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4735 * and should be ignored.
4738 vfs_busy_pages(struct buf *bp, int clear_modify)
4746 if (!(bp->b_flags & B_VMIO))
4749 obj = bp->b_bufobj->bo_object;
4750 foff = bp->b_offset;
4751 KASSERT(bp->b_offset != NOOFFSET,
4752 ("vfs_busy_pages: no buffer offset"));
4753 if ((bp->b_flags & B_CLUSTER) == 0) {
4754 vm_object_pip_add(obj, bp->b_npages);
4755 vfs_busy_pages_acquire(bp);
4757 if (bp->b_bufsize != 0)
4758 vfs_setdirty_range(bp);
4760 for (i = 0; i < bp->b_npages; i++) {
4762 vm_page_assert_sbusied(m);
4765 * When readying a buffer for a read ( i.e
4766 * clear_modify == 0 ), it is important to do
4767 * bogus_page replacement for valid pages in
4768 * partially instantiated buffers. Partially
4769 * instantiated buffers can, in turn, occur when
4770 * reconstituting a buffer from its VM backing store
4771 * base. We only have to do this if B_CACHE is
4772 * clear ( which causes the I/O to occur in the
4773 * first place ). The replacement prevents the read
4774 * I/O from overwriting potentially dirty VM-backed
4775 * pages. XXX bogus page replacement is, uh, bogus.
4776 * It may not work properly with small-block devices.
4777 * We need to find a better way.
4780 pmap_remove_write(m);
4781 vfs_page_set_validclean(bp, foff, m);
4782 } else if (vm_page_all_valid(m) &&
4783 (bp->b_flags & B_CACHE) == 0) {
4784 bp->b_pages[i] = bogus_page;
4787 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4789 if (bogus && buf_mapped(bp)) {
4790 BUF_CHECK_MAPPED(bp);
4791 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4792 bp->b_pages, bp->b_npages);
4797 * vfs_bio_set_valid:
4799 * Set the range within the buffer to valid. The range is
4800 * relative to the beginning of the buffer, b_offset. Note that
4801 * b_offset itself may be offset from the beginning of the first
4805 vfs_bio_set_valid(struct buf *bp, int base, int size)
4810 if (!(bp->b_flags & B_VMIO))
4814 * Fixup base to be relative to beginning of first page.
4815 * Set initial n to be the maximum number of bytes in the
4816 * first page that can be validated.
4818 base += (bp->b_offset & PAGE_MASK);
4819 n = PAGE_SIZE - (base & PAGE_MASK);
4822 * Busy may not be strictly necessary here because the pages are
4823 * unlikely to be fully valid and the vnode lock will synchronize
4824 * their access via getpages. It is grabbed for consistency with
4825 * other page validation.
4827 vfs_busy_pages_acquire(bp);
4828 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4832 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4837 vfs_busy_pages_release(bp);
4843 * If the specified buffer is a non-VMIO buffer, clear the entire
4844 * buffer. If the specified buffer is a VMIO buffer, clear and
4845 * validate only the previously invalid portions of the buffer.
4846 * This routine essentially fakes an I/O, so we need to clear
4847 * BIO_ERROR and B_INVAL.
4849 * Note that while we only theoretically need to clear through b_bcount,
4850 * we go ahead and clear through b_bufsize.
4853 vfs_bio_clrbuf(struct buf *bp)
4855 int i, j, mask, sa, ea, slide;
4857 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4861 bp->b_flags &= ~B_INVAL;
4862 bp->b_ioflags &= ~BIO_ERROR;
4863 vfs_busy_pages_acquire(bp);
4864 sa = bp->b_offset & PAGE_MASK;
4866 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4867 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4868 ea = slide & PAGE_MASK;
4871 if (bp->b_pages[i] == bogus_page)
4874 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4875 if ((bp->b_pages[i]->valid & mask) == mask)
4877 if ((bp->b_pages[i]->valid & mask) == 0)
4878 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4880 for (; sa < ea; sa += DEV_BSIZE, j++) {
4881 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4882 pmap_zero_page_area(bp->b_pages[i],
4887 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4888 roundup2(ea - sa, DEV_BSIZE));
4890 vfs_busy_pages_release(bp);
4895 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4900 if (buf_mapped(bp)) {
4901 BUF_CHECK_MAPPED(bp);
4902 bzero(bp->b_data + base, size);
4904 BUF_CHECK_UNMAPPED(bp);
4905 n = PAGE_SIZE - (base & PAGE_MASK);
4906 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4910 pmap_zero_page_area(m, base & PAGE_MASK, n);
4919 * Update buffer flags based on I/O request parameters, optionally releasing the
4920 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4921 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4922 * I/O). Otherwise the buffer is released to the cache.
4925 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4928 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4929 ("buf %p non-VMIO noreuse", bp));
4931 if ((ioflag & IO_DIRECT) != 0)
4932 bp->b_flags |= B_DIRECT;
4933 if ((ioflag & IO_EXT) != 0)
4934 bp->b_xflags |= BX_ALTDATA;
4935 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4936 bp->b_flags |= B_RELBUF;
4937 if ((ioflag & IO_NOREUSE) != 0)
4938 bp->b_flags |= B_NOREUSE;
4946 vfs_bio_brelse(struct buf *bp, int ioflag)
4949 b_io_dismiss(bp, ioflag, true);
4953 vfs_bio_set_flags(struct buf *bp, int ioflag)
4956 b_io_dismiss(bp, ioflag, false);
4960 * vm_hold_load_pages and vm_hold_free_pages get pages into
4961 * a buffers address space. The pages are anonymous and are
4962 * not associated with a file object.
4965 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4971 BUF_CHECK_MAPPED(bp);
4973 to = round_page(to);
4974 from = round_page(from);
4975 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4976 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4977 KASSERT(to - from <= maxbcachebuf,
4978 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4979 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4981 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4983 * note: must allocate system pages since blocking here
4984 * could interfere with paging I/O, no matter which
4987 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4988 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4989 pmap_qenter(pg, &p, 1);
4990 bp->b_pages[index] = p;
4992 bp->b_npages = index;
4995 /* Return pages associated with this buf to the vm system */
4997 vm_hold_free_pages(struct buf *bp, int newbsize)
5001 int index, newnpages;
5003 BUF_CHECK_MAPPED(bp);
5005 from = round_page((vm_offset_t)bp->b_data + newbsize);
5006 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5007 if (bp->b_npages > newnpages)
5008 pmap_qremove(from, bp->b_npages - newnpages);
5009 for (index = newnpages; index < bp->b_npages; index++) {
5010 p = bp->b_pages[index];
5011 bp->b_pages[index] = NULL;
5012 vm_page_unwire_noq(p);
5015 bp->b_npages = newnpages;
5019 * Map an IO request into kernel virtual address space.
5021 * All requests are (re)mapped into kernel VA space.
5022 * Notice that we use b_bufsize for the size of the buffer
5023 * to be mapped. b_bcount might be modified by the driver.
5025 * Note that even if the caller determines that the address space should
5026 * be valid, a race or a smaller-file mapped into a larger space may
5027 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5028 * check the return value.
5030 * This function only works with pager buffers.
5033 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5038 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5039 prot = VM_PROT_READ;
5040 if (bp->b_iocmd == BIO_READ)
5041 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5042 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5043 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5046 bp->b_bufsize = len;
5047 bp->b_npages = pidx;
5048 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5049 if (mapbuf || !unmapped_buf_allowed) {
5050 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5051 bp->b_data = bp->b_kvabase + bp->b_offset;
5053 bp->b_data = unmapped_buf;
5058 * Free the io map PTEs associated with this IO operation.
5059 * We also invalidate the TLB entries and restore the original b_addr.
5061 * This function only works with pager buffers.
5064 vunmapbuf(struct buf *bp)
5068 npages = bp->b_npages;
5070 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5071 vm_page_unhold_pages(bp->b_pages, npages);
5073 bp->b_data = unmapped_buf;
5077 bdone(struct buf *bp)
5081 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5083 bp->b_flags |= B_DONE;
5089 bwait(struct buf *bp, u_char pri, const char *wchan)
5093 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5095 while ((bp->b_flags & B_DONE) == 0)
5096 msleep(bp, mtxp, pri, wchan, 0);
5101 bufsync(struct bufobj *bo, int waitfor)
5104 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5108 bufstrategy(struct bufobj *bo, struct buf *bp)
5114 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5115 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5116 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5117 i = VOP_STRATEGY(vp, bp);
5118 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5122 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5125 bufobj_init(struct bufobj *bo, void *private)
5127 static volatile int bufobj_cleanq;
5130 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5131 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5132 bo->bo_private = private;
5133 TAILQ_INIT(&bo->bo_clean.bv_hd);
5134 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5138 bufobj_wrefl(struct bufobj *bo)
5141 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5142 ASSERT_BO_WLOCKED(bo);
5147 bufobj_wref(struct bufobj *bo)
5150 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5157 bufobj_wdrop(struct bufobj *bo)
5160 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5162 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5163 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5164 bo->bo_flag &= ~BO_WWAIT;
5165 wakeup(&bo->bo_numoutput);
5171 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5175 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5176 ASSERT_BO_WLOCKED(bo);
5178 while (bo->bo_numoutput) {
5179 bo->bo_flag |= BO_WWAIT;
5180 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5181 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5189 * Set bio_data or bio_ma for struct bio from the struct buf.
5192 bdata2bio(struct buf *bp, struct bio *bip)
5195 if (!buf_mapped(bp)) {
5196 KASSERT(unmapped_buf_allowed, ("unmapped"));
5197 bip->bio_ma = bp->b_pages;
5198 bip->bio_ma_n = bp->b_npages;
5199 bip->bio_data = unmapped_buf;
5200 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5201 bip->bio_flags |= BIO_UNMAPPED;
5202 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5203 PAGE_SIZE == bp->b_npages,
5204 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5205 (long long)bip->bio_length, bip->bio_ma_n));
5207 bip->bio_data = bp->b_data;
5213 * The MIPS pmap code currently doesn't handle aliased pages.
5214 * The VIPT caches may not handle page aliasing themselves, leading
5215 * to data corruption.
5217 * As such, this code makes a system extremely unhappy if said
5218 * system doesn't support unaliasing the above situation in hardware.
5219 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5220 * this feature at build time, so it has to be handled in software.
5222 * Once the MIPS pmap/cache code grows to support this function on
5223 * earlier chips, it should be flipped back off.
5226 static int buf_pager_relbuf = 1;
5228 static int buf_pager_relbuf = 0;
5230 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5231 &buf_pager_relbuf, 0,
5232 "Make buffer pager release buffers after reading");
5235 * The buffer pager. It uses buffer reads to validate pages.
5237 * In contrast to the generic local pager from vm/vnode_pager.c, this
5238 * pager correctly and easily handles volumes where the underlying
5239 * device block size is greater than the machine page size. The
5240 * buffer cache transparently extends the requested page run to be
5241 * aligned at the block boundary, and does the necessary bogus page
5242 * replacements in the addends to avoid obliterating already valid
5245 * The only non-trivial issue is that the exclusive busy state for
5246 * pages, which is assumed by the vm_pager_getpages() interface, is
5247 * incompatible with the VMIO buffer cache's desire to share-busy the
5248 * pages. This function performs a trivial downgrade of the pages'
5249 * state before reading buffers, and a less trivial upgrade from the
5250 * shared-busy to excl-busy state after the read.
5253 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5254 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5255 vbg_get_blksize_t get_blksize)
5262 vm_ooffset_t la, lb, poff, poffe;
5264 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5267 object = vp->v_object;
5270 la = IDX_TO_OFF(ma[count - 1]->pindex);
5271 if (la >= object->un_pager.vnp.vnp_size)
5272 return (VM_PAGER_BAD);
5275 * Change the meaning of la from where the last requested page starts
5276 * to where it ends, because that's the end of the requested region
5277 * and the start of the potential read-ahead region.
5280 lpart = la > object->un_pager.vnp.vnp_size;
5281 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5284 return (VM_PAGER_ERROR);
5287 * Calculate read-ahead, behind and total pages.
5290 lb = IDX_TO_OFF(ma[0]->pindex);
5291 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5293 if (rbehind != NULL)
5295 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5296 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5297 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5302 VM_CNT_INC(v_vnodein);
5303 VM_CNT_ADD(v_vnodepgsin, pgsin);
5305 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5306 != 0) ? GB_UNMAPPED : 0;
5308 for (i = 0; i < count; i++) {
5309 if (ma[i] != bogus_page)
5310 vm_page_busy_downgrade(ma[i]);
5314 for (i = 0; i < count; i++) {
5316 if (m == bogus_page)
5320 * Pages are shared busy and the object lock is not
5321 * owned, which together allow for the pages'
5322 * invalidation. The racy test for validity avoids
5323 * useless creation of the buffer for the most typical
5324 * case when invalidation is not used in redo or for
5325 * parallel read. The shared->excl upgrade loop at
5326 * the end of the function catches the race in a
5327 * reliable way (protected by the object lock).
5329 if (vm_page_all_valid(m))
5332 poff = IDX_TO_OFF(m->pindex);
5333 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5334 for (; poff < poffe; poff += bsize) {
5335 lbn = get_lblkno(vp, poff);
5340 error = get_blksize(vp, lbn, &bsize);
5342 error = bread_gb(vp, lbn, bsize,
5343 curthread->td_ucred, br_flags, &bp);
5346 if (bp->b_rcred == curthread->td_ucred) {
5347 crfree(bp->b_rcred);
5348 bp->b_rcred = NOCRED;
5350 if (LIST_EMPTY(&bp->b_dep)) {
5352 * Invalidation clears m->valid, but
5353 * may leave B_CACHE flag if the
5354 * buffer existed at the invalidation
5355 * time. In this case, recycle the
5356 * buffer to do real read on next
5357 * bread() after redo.
5359 * Otherwise B_RELBUF is not strictly
5360 * necessary, enable to reduce buf
5363 if (buf_pager_relbuf ||
5364 !vm_page_all_valid(m))
5365 bp->b_flags |= B_RELBUF;
5367 bp->b_flags &= ~B_NOCACHE;
5373 KASSERT(1 /* racy, enable for debugging */ ||
5374 vm_page_all_valid(m) || i == count - 1,
5375 ("buf %d %p invalid", i, m));
5376 if (i == count - 1 && lpart) {
5377 if (!vm_page_none_valid(m) &&
5378 !vm_page_all_valid(m))
5379 vm_page_zero_invalid(m, TRUE);
5386 for (i = 0; i < count; i++) {
5387 if (ma[i] == bogus_page)
5389 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5390 vm_page_sunbusy(ma[i]);
5391 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5396 * Since the pages were only sbusy while neither the
5397 * buffer nor the object lock was held by us, or
5398 * reallocated while vm_page_grab() slept for busy
5399 * relinguish, they could have been invalidated.
5400 * Recheck the valid bits and re-read as needed.
5402 * Note that the last page is made fully valid in the
5403 * read loop, and partial validity for the page at
5404 * index count - 1 could mean that the page was
5405 * invalidated or removed, so we must restart for
5408 if (!vm_page_all_valid(ma[i]))
5411 if (redo && error == 0)
5413 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5416 #include "opt_ddb.h"
5418 #include <ddb/ddb.h>
5420 /* DDB command to show buffer data */
5421 DB_SHOW_COMMAND(buffer, db_show_buffer)
5424 struct buf *bp = (struct buf *)addr;
5425 #ifdef FULL_BUF_TRACKING
5430 db_printf("usage: show buffer <addr>\n");
5434 db_printf("buf at %p\n", bp);
5435 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5436 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5437 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5438 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5439 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5440 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5442 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5443 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5444 "b_vp = %p, b_dep = %p\n",
5445 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5446 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5447 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5448 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5449 bp->b_kvabase, bp->b_kvasize);
5452 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5453 for (i = 0; i < bp->b_npages; i++) {
5457 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5459 (u_long)VM_PAGE_TO_PHYS(m));
5461 db_printf("( ??? )");
5462 if ((i + 1) < bp->b_npages)
5467 BUF_LOCKPRINTINFO(bp);
5468 #if defined(FULL_BUF_TRACKING)
5469 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5471 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5472 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5473 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5475 db_printf(" %2u: %s\n", j,
5476 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5478 #elif defined(BUF_TRACKING)
5479 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5484 DB_SHOW_COMMAND(bufqueues, bufqueues)
5486 struct bufdomain *bd;
5491 db_printf("bqempty: %d\n", bqempty.bq_len);
5493 for (i = 0; i < buf_domains; i++) {
5495 db_printf("Buf domain %d\n", i);
5496 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5497 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5498 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5500 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5501 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5502 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5503 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5504 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5506 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5507 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5508 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5509 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5512 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5513 total += bp->b_bufsize;
5514 db_printf("\tcleanq count\t%d (%ld)\n",
5515 bd->bd_cleanq->bq_len, total);
5517 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5518 total += bp->b_bufsize;
5519 db_printf("\tdirtyq count\t%d (%ld)\n",
5520 bd->bd_dirtyq.bq_len, total);
5521 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5522 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5523 db_printf("\tCPU ");
5524 for (j = 0; j <= mp_maxid; j++)
5525 db_printf("%d, ", bd->bd_subq[j].bq_len);
5529 for (j = 0; j < nbuf; j++) {
5531 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5533 total += bp->b_bufsize;
5536 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5539 for (j = 0; j < nbuf; j++) {
5541 if (bp->b_domain == i) {
5543 total += bp->b_bufsize;
5546 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5550 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5555 for (i = 0; i < nbuf; i++) {
5557 if (BUF_ISLOCKED(bp)) {
5558 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5566 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5572 db_printf("usage: show vnodebufs <addr>\n");
5575 vp = (struct vnode *)addr;
5576 db_printf("Clean buffers:\n");
5577 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5578 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5581 db_printf("Dirty buffers:\n");
5582 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5583 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5588 DB_COMMAND(countfreebufs, db_coundfreebufs)
5591 int i, used = 0, nfree = 0;
5594 db_printf("usage: countfreebufs\n");
5598 for (i = 0; i < nbuf; i++) {
5600 if (bp->b_qindex == QUEUE_EMPTY)
5606 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5608 db_printf("numfreebuffers is %d\n", numfreebuffers);