2 * SPDX-License-Identifier: BSD-2-Clause
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 #include <sys/param.h>
49 #include <sys/systm.h>
52 #include <sys/bitset.h>
54 #include <sys/counter.h>
56 #include <sys/devicestat.h>
57 #include <sys/eventhandler.h>
60 #include <sys/limits.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
68 #include <sys/racct.h>
69 #include <sys/refcount.h>
70 #include <sys/resourcevar.h>
71 #include <sys/rwlock.h>
73 #include <sys/sysctl.h>
74 #include <sys/syscallsubr.h>
76 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/watchdog.h>
79 #include <geom/geom.h>
81 #include <vm/vm_param.h>
82 #include <vm/vm_kern.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_extern.h>
88 #include <vm/vm_map.h>
89 #include <vm/swap_pager.h>
91 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
93 struct bio_ops bioops; /* I/O operation notification */
95 struct buf_ops buf_ops_bio = {
96 .bop_name = "buf_ops_bio",
97 .bop_write = bufwrite,
98 .bop_strategy = bufstrategy,
100 .bop_bdflush = bufbdflush,
104 struct mtx_padalign bq_lock;
105 TAILQ_HEAD(, buf) bq_queue;
107 uint16_t bq_subqueue;
109 } __aligned(CACHE_LINE_SIZE);
111 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
112 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
113 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
114 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
117 struct bufqueue *bd_subq;
118 struct bufqueue bd_dirtyq;
119 struct bufqueue *bd_cleanq;
120 struct mtx_padalign bd_run_lock;
125 long bd_bufspacethresh;
126 int bd_hifreebuffers;
127 int bd_lofreebuffers;
128 int bd_hidirtybuffers;
129 int bd_lodirtybuffers;
130 int bd_dirtybufthresh;
135 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
136 int __aligned(CACHE_LINE_SIZE) bd_running;
137 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
138 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
139 } __aligned(CACHE_LINE_SIZE);
141 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
142 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
143 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
144 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
145 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
146 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
147 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
148 #define BD_DOMAIN(bd) (bd - bdomain)
150 static char *buf; /* buffer header pool */
154 return ((struct buf *)(buf + (sizeof(struct buf) +
155 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
158 caddr_t __read_mostly unmapped_buf;
160 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
161 struct proc *bufdaemonproc;
163 static void vm_hold_free_pages(struct buf *bp, int newbsize);
164 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
166 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
167 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
169 static void vfs_clean_pages_dirty_buf(struct buf *bp);
170 static void vfs_setdirty_range(struct buf *bp);
171 static void vfs_vmio_invalidate(struct buf *bp);
172 static void vfs_vmio_truncate(struct buf *bp, int npages);
173 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
174 static int vfs_bio_clcheck(struct vnode *vp, int size,
175 daddr_t lblkno, daddr_t blkno);
176 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
177 void (*)(struct buf *));
178 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
179 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
180 static void buf_daemon(void);
181 static __inline void bd_wakeup(void);
182 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
183 static void bufkva_reclaim(vmem_t *, int);
184 static void bufkva_free(struct buf *);
185 static int buf_import(void *, void **, int, int, int);
186 static void buf_release(void *, void **, int);
187 static void maxbcachebuf_adjust(void);
188 static inline struct bufdomain *bufdomain(struct buf *);
189 static void bq_remove(struct bufqueue *bq, struct buf *bp);
190 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
191 static int buf_recycle(struct bufdomain *, bool kva);
192 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
193 const char *lockname);
194 static void bd_init(struct bufdomain *bd);
195 static int bd_flushall(struct bufdomain *bd);
196 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
197 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
199 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
200 int vmiodirenable = TRUE;
201 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
202 "Use the VM system for directory writes");
203 long runningbufspace;
204 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
205 "Amount of presently outstanding async buffer io");
206 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
207 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
208 static counter_u64_t bufkvaspace;
209 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
210 "Kernel virtual memory used for buffers");
211 static long maxbufspace;
212 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
213 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
214 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
215 "Maximum allowed value of bufspace (including metadata)");
216 static long bufmallocspace;
217 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
218 "Amount of malloced memory for buffers");
219 static long maxbufmallocspace;
220 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
221 0, "Maximum amount of malloced memory for buffers");
222 static long lobufspace;
223 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
224 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
225 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
226 "Minimum amount of buffers we want to have");
228 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
229 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
230 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
231 "Maximum allowed value of bufspace (excluding metadata)");
233 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
234 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
235 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
236 "Bufspace consumed before waking the daemon to free some");
237 static counter_u64_t buffreekvacnt;
238 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
239 "Number of times we have freed the KVA space from some buffer");
240 static counter_u64_t bufdefragcnt;
241 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
242 "Number of times we have had to repeat buffer allocation to defragment");
243 static long lorunningspace;
244 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
245 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
246 "Minimum preferred space used for in-progress I/O");
247 static long hirunningspace;
248 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
249 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
250 "Maximum amount of space to use for in-progress I/O");
251 int dirtybufferflushes;
252 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
253 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
255 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
256 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
257 int altbufferflushes;
258 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
259 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
260 static int recursiveflushes;
261 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
262 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
263 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
264 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
265 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
266 "Number of buffers that are dirty (has unwritten changes) at the moment");
267 static int lodirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
269 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
270 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
271 "How many buffers we want to have free before bufdaemon can sleep");
272 static int hidirtybuffers;
273 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
274 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
275 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
276 "When the number of dirty buffers is considered severe");
278 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
279 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
280 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
281 "Number of bdwrite to bawrite conversions to clear dirty buffers");
282 static int numfreebuffers;
283 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
284 "Number of free buffers");
285 static int lofreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
287 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
288 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
289 "Target number of free buffers");
290 static int hifreebuffers;
291 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
292 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
293 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
294 "Threshold for clean buffer recycling");
295 static counter_u64_t getnewbufcalls;
296 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
297 &getnewbufcalls, "Number of calls to getnewbuf");
298 static counter_u64_t getnewbufrestarts;
299 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
301 "Number of times getnewbuf has had to restart a buffer acquisition");
302 static counter_u64_t mappingrestarts;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
305 "Number of times getblk has had to restart a buffer mapping for "
307 static counter_u64_t numbufallocfails;
308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
309 &numbufallocfails, "Number of times buffer allocations failed");
310 static int flushbufqtarget = 100;
311 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
312 "Amount of work to do in flushbufqueues when helping bufdaemon");
313 static counter_u64_t notbufdflushes;
314 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
315 "Number of dirty buffer flushes done by the bufdaemon helpers");
316 static long barrierwrites;
317 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
318 &barrierwrites, 0, "Number of barrier writes");
319 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed,
320 CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
321 &unmapped_buf_allowed, 0,
322 "Permit the use of the unmapped i/o");
323 int maxbcachebuf = MAXBCACHEBUF;
324 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
325 "Maximum size of a buffer cache block");
328 * This lock synchronizes access to bd_request.
330 static struct mtx_padalign __exclusive_cache_line bdlock;
333 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
334 * waitrunningbufspace().
336 static struct mtx_padalign __exclusive_cache_line rbreqlock;
339 * Lock that protects bdirtywait.
341 static struct mtx_padalign __exclusive_cache_line bdirtylock;
344 * bufdaemon shutdown request and sleep channel.
346 static bool bd_shutdown;
349 * Wakeup point for bufdaemon, as well as indicator of whether it is already
350 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
353 static int bd_request;
356 * Request for the buf daemon to write more buffers than is indicated by
357 * lodirtybuf. This may be necessary to push out excess dependencies or
358 * defragment the address space where a simple count of the number of dirty
359 * buffers is insufficient to characterize the demand for flushing them.
361 static int bd_speedupreq;
364 * Synchronization (sleep/wakeup) variable for active buffer space requests.
365 * Set when wait starts, cleared prior to wakeup().
366 * Used in runningbufwakeup() and waitrunningbufspace().
368 static int runningbufreq;
371 * Synchronization for bwillwrite() waiters.
373 static int bdirtywait;
376 * Definitions for the buffer free lists.
378 #define QUEUE_NONE 0 /* on no queue */
379 #define QUEUE_EMPTY 1 /* empty buffer headers */
380 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
381 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
382 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
384 /* Maximum number of buffer domains. */
385 #define BUF_DOMAINS 8
387 struct bufdomainset bdlodirty; /* Domains > lodirty */
388 struct bufdomainset bdhidirty; /* Domains > hidirty */
390 /* Configured number of clean queues. */
391 static int __read_mostly buf_domains;
393 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
394 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
395 struct bufqueue __exclusive_cache_line bqempty;
398 * per-cpu empty buffer cache.
403 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
408 value = *(long *)arg1;
409 error = sysctl_handle_long(oidp, &value, 0, req);
410 if (error != 0 || req->newptr == NULL)
412 mtx_lock(&rbreqlock);
413 if (arg1 == &hirunningspace) {
414 if (value < lorunningspace)
417 hirunningspace = value;
419 KASSERT(arg1 == &lorunningspace,
420 ("%s: unknown arg1", __func__));
421 if (value > hirunningspace)
424 lorunningspace = value;
426 mtx_unlock(&rbreqlock);
431 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
437 value = *(int *)arg1;
438 error = sysctl_handle_int(oidp, &value, 0, req);
439 if (error != 0 || req->newptr == NULL)
441 *(int *)arg1 = value;
442 for (i = 0; i < buf_domains; i++)
443 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
450 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
456 value = *(long *)arg1;
457 error = sysctl_handle_long(oidp, &value, 0, req);
458 if (error != 0 || req->newptr == NULL)
460 *(long *)arg1 = value;
461 for (i = 0; i < buf_domains; i++)
462 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
468 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
469 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
471 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
478 for (i = 0; i < buf_domains; i++)
479 lvalue += bdomain[i].bd_bufspace;
480 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
481 return (sysctl_handle_long(oidp, &lvalue, 0, req));
482 if (lvalue > INT_MAX)
483 /* On overflow, still write out a long to trigger ENOMEM. */
484 return (sysctl_handle_long(oidp, &lvalue, 0, req));
486 return (sysctl_handle_int(oidp, &ivalue, 0, req));
490 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
496 for (i = 0; i < buf_domains; i++)
497 lvalue += bdomain[i].bd_bufspace;
498 return (sysctl_handle_long(oidp, &lvalue, 0, req));
503 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
509 for (i = 0; i < buf_domains; i++)
510 value += bdomain[i].bd_numdirtybuffers;
511 return (sysctl_handle_int(oidp, &value, 0, req));
517 * Wakeup any bwillwrite() waiters.
522 mtx_lock(&bdirtylock);
527 mtx_unlock(&bdirtylock);
533 * Clear a domain from the appropriate bitsets when dirtybuffers
537 bd_clear(struct bufdomain *bd)
540 mtx_lock(&bdirtylock);
541 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
542 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
543 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
544 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
545 mtx_unlock(&bdirtylock);
551 * Set a domain in the appropriate bitsets when dirtybuffers
555 bd_set(struct bufdomain *bd)
558 mtx_lock(&bdirtylock);
559 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
560 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
561 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
562 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
563 mtx_unlock(&bdirtylock);
569 * Decrement the numdirtybuffers count by one and wakeup any
570 * threads blocked in bwillwrite().
573 bdirtysub(struct buf *bp)
575 struct bufdomain *bd;
579 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
580 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
582 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
589 * Increment the numdirtybuffers count by one and wakeup the buf
593 bdirtyadd(struct buf *bp)
595 struct bufdomain *bd;
599 * Only do the wakeup once as we cross the boundary. The
600 * buf daemon will keep running until the condition clears.
603 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
604 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
606 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
611 * bufspace_daemon_wakeup:
613 * Wakeup the daemons responsible for freeing clean bufs.
616 bufspace_daemon_wakeup(struct bufdomain *bd)
620 * avoid the lock if the daemon is running.
622 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
624 atomic_store_int(&bd->bd_running, 1);
625 wakeup(&bd->bd_running);
633 * Adjust the reported bufspace for a KVA managed buffer, possibly
634 * waking any waiters.
637 bufspace_adjust(struct buf *bp, int bufsize)
639 struct bufdomain *bd;
643 KASSERT((bp->b_flags & B_MALLOC) == 0,
644 ("bufspace_adjust: malloc buf %p", bp));
646 diff = bufsize - bp->b_bufsize;
648 atomic_subtract_long(&bd->bd_bufspace, -diff);
649 } else if (diff > 0) {
650 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
651 /* Wake up the daemon on the transition. */
652 if (space < bd->bd_bufspacethresh &&
653 space + diff >= bd->bd_bufspacethresh)
654 bufspace_daemon_wakeup(bd);
656 bp->b_bufsize = bufsize;
662 * Reserve bufspace before calling allocbuf(). metadata has a
663 * different space limit than data.
666 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
672 limit = bd->bd_maxbufspace;
674 limit = bd->bd_hibufspace;
675 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
678 atomic_subtract_long(&bd->bd_bufspace, size);
682 /* Wake up the daemon on the transition. */
683 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
684 bufspace_daemon_wakeup(bd);
692 * Release reserved bufspace after bufspace_adjust() has consumed it.
695 bufspace_release(struct bufdomain *bd, int size)
698 atomic_subtract_long(&bd->bd_bufspace, size);
704 * Wait for bufspace, acting as the buf daemon if a locked vnode is
705 * supplied. bd_wanted must be set prior to polling for space. The
706 * operation must be re-tried on return.
709 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
710 int slpflag, int slptimeo)
713 int error, fl, norunbuf;
715 if ((gbflags & GB_NOWAIT_BD) != 0)
720 while (bd->bd_wanted) {
721 if (vp != NULL && vp->v_type != VCHR &&
722 (td->td_pflags & TDP_BUFNEED) == 0) {
725 * getblk() is called with a vnode locked, and
726 * some majority of the dirty buffers may as
727 * well belong to the vnode. Flushing the
728 * buffers there would make a progress that
729 * cannot be achieved by the buf_daemon, that
730 * cannot lock the vnode.
732 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
733 (td->td_pflags & TDP_NORUNNINGBUF);
736 * Play bufdaemon. The getnewbuf() function
737 * may be called while the thread owns lock
738 * for another dirty buffer for the same
739 * vnode, which makes it impossible to use
740 * VOP_FSYNC() there, due to the buffer lock
743 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
744 fl = buf_flush(vp, bd, flushbufqtarget);
745 td->td_pflags &= norunbuf;
749 if (bd->bd_wanted == 0)
752 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
753 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
761 bufspace_daemon_shutdown(void *arg, int howto __unused)
763 struct bufdomain *bd = arg;
766 if (KERNEL_PANICKED())
770 bd->bd_shutdown = true;
771 wakeup(&bd->bd_running);
772 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
773 "bufspace_shutdown", 60 * hz);
776 printf("bufspacedaemon wait error: %d\n", error);
782 * buffer space management daemon. Tries to maintain some marginal
783 * amount of free buffer space so that requesting processes neither
784 * block nor work to reclaim buffers.
787 bufspace_daemon(void *arg)
789 struct bufdomain *bd = arg;
791 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
792 SHUTDOWN_PRI_LAST + 100);
795 while (!bd->bd_shutdown) {
799 * Free buffers from the clean queue until we meet our
802 * Theory of operation: The buffer cache is most efficient
803 * when some free buffer headers and space are always
804 * available to getnewbuf(). This daemon attempts to prevent
805 * the excessive blocking and synchronization associated
806 * with shortfall. It goes through three phases according
809 * 1) The daemon wakes up voluntarily once per-second
810 * during idle periods when the counters are below
811 * the wakeup thresholds (bufspacethresh, lofreebuffers).
813 * 2) The daemon wakes up as we cross the thresholds
814 * ahead of any potential blocking. This may bounce
815 * slightly according to the rate of consumption and
818 * 3) The daemon and consumers are starved for working
819 * clean buffers. This is the 'bufspace' sleep below
820 * which will inefficiently trade bufs with bqrelse
821 * until we return to condition 2.
823 while (bd->bd_bufspace > bd->bd_lobufspace ||
824 bd->bd_freebuffers < bd->bd_hifreebuffers) {
825 if (buf_recycle(bd, false) != 0) {
829 * Speedup dirty if we've run out of clean
830 * buffers. This is possible in particular
831 * because softdep may held many bufs locked
832 * pending writes to other bufs which are
833 * marked for delayed write, exhausting
834 * clean space until they are written.
839 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
840 PRIBIO|PDROP, "bufspace", hz/10);
848 * Re-check our limits and sleep. bd_running must be
849 * cleared prior to checking the limits to avoid missed
850 * wakeups. The waker will adjust one of bufspace or
851 * freebuffers prior to checking bd_running.
856 atomic_store_int(&bd->bd_running, 0);
857 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
858 bd->bd_freebuffers > bd->bd_lofreebuffers) {
859 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
862 /* Avoid spurious wakeups while running. */
863 atomic_store_int(&bd->bd_running, 1);
866 wakeup(&bd->bd_shutdown);
874 * Adjust the reported bufspace for a malloc managed buffer, possibly
875 * waking any waiters.
878 bufmallocadjust(struct buf *bp, int bufsize)
882 KASSERT((bp->b_flags & B_MALLOC) != 0,
883 ("bufmallocadjust: non-malloc buf %p", bp));
884 diff = bufsize - bp->b_bufsize;
886 atomic_subtract_long(&bufmallocspace, -diff);
888 atomic_add_long(&bufmallocspace, diff);
889 bp->b_bufsize = bufsize;
895 * Wake up processes that are waiting on asynchronous writes to fall
896 * below lorunningspace.
902 mtx_lock(&rbreqlock);
905 wakeup(&runningbufreq);
907 mtx_unlock(&rbreqlock);
913 * Decrement the outstanding write count according.
916 runningbufwakeup(struct buf *bp)
920 bspace = bp->b_runningbufspace;
923 space = atomic_fetchadd_long(&runningbufspace, -bspace);
924 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
926 bp->b_runningbufspace = 0;
928 * Only acquire the lock and wakeup on the transition from exceeding
929 * the threshold to falling below it.
931 if (space < lorunningspace)
933 if (space - bspace > lorunningspace)
939 * waitrunningbufspace()
941 * runningbufspace is a measure of the amount of I/O currently
942 * running. This routine is used in async-write situations to
943 * prevent creating huge backups of pending writes to a device.
944 * Only asynchronous writes are governed by this function.
946 * This does NOT turn an async write into a sync write. It waits
947 * for earlier writes to complete and generally returns before the
948 * caller's write has reached the device.
951 waitrunningbufspace(void)
954 mtx_lock(&rbreqlock);
955 while (runningbufspace > hirunningspace) {
957 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
959 mtx_unlock(&rbreqlock);
963 * vfs_buf_test_cache:
965 * Called when a buffer is extended. This function clears the B_CACHE
966 * bit if the newly extended portion of the buffer does not contain
970 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
971 vm_offset_t size, vm_page_t m)
975 * This function and its results are protected by higher level
976 * synchronization requiring vnode and buf locks to page in and
979 if (bp->b_flags & B_CACHE) {
980 int base = (foff + off) & PAGE_MASK;
981 if (vm_page_is_valid(m, base, size) == 0)
982 bp->b_flags &= ~B_CACHE;
986 /* Wake up the buffer daemon if necessary */
992 if (bd_request == 0) {
1000 * Adjust the maxbcachbuf tunable.
1003 maxbcachebuf_adjust(void)
1008 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1011 while (i * 2 <= maxbcachebuf)
1014 if (maxbcachebuf < MAXBSIZE)
1015 maxbcachebuf = MAXBSIZE;
1016 if (maxbcachebuf > maxphys)
1017 maxbcachebuf = maxphys;
1018 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1019 printf("maxbcachebuf=%d\n", maxbcachebuf);
1023 * bd_speedup - speedup the buffer cache flushing code
1032 if (bd_speedupreq == 0 || bd_request == 0)
1037 wakeup(&bd_request);
1038 mtx_unlock(&bdlock);
1042 #define TRANSIENT_DENOM 5
1044 #define TRANSIENT_DENOM 10
1048 * Calculating buffer cache scaling values and reserve space for buffer
1049 * headers. This is called during low level kernel initialization and
1050 * may be called more then once. We CANNOT write to the memory area
1051 * being reserved at this time.
1054 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1057 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1061 * With KASAN enabled, the kernel map is shadowed. Account for this
1062 * when sizing maps based on the amount of physical memory available.
1064 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1065 (KASAN_SHADOW_SCALE + 1);
1069 * physmem_est is in pages. Convert it to kilobytes (assumes
1070 * PAGE_SIZE is >= 1K)
1072 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1074 maxbcachebuf_adjust();
1076 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1077 * For the first 64MB of ram nominally allocate sufficient buffers to
1078 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1079 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1080 * the buffer cache we limit the eventual kva reservation to
1083 * factor represents the 1/4 x ram conversion.
1086 int factor = 4 * BKVASIZE / 1024;
1089 if (physmem_est > 4096)
1090 nbuf += min((physmem_est - 4096) / factor,
1092 if (physmem_est > 65536)
1093 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1094 32 * 1024 * 1024 / (factor * 5));
1096 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1097 nbuf = maxbcache / BKVASIZE;
1102 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1103 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1104 if (nbuf > maxbuf) {
1106 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1112 * Ideal allocation size for the transient bio submap is 10%
1113 * of the maximal space buffer map. This roughly corresponds
1114 * to the amount of the buffer mapped for typical UFS load.
1116 * Clip the buffer map to reserve space for the transient
1117 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1118 * maximum buffer map extent on the platform.
1120 * The fall-back to the maxbuf in case of maxbcache unset,
1121 * allows to not trim the buffer KVA for the architectures
1122 * with ample KVA space.
1124 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1125 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1126 buf_sz = (long)nbuf * BKVASIZE;
1127 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1128 (TRANSIENT_DENOM - 1)) {
1130 * There is more KVA than memory. Do not
1131 * adjust buffer map size, and assign the rest
1132 * of maxbuf to transient map.
1134 biotmap_sz = maxbuf_sz - buf_sz;
1137 * Buffer map spans all KVA we could afford on
1138 * this platform. Give 10% (20% on i386) of
1139 * the buffer map to the transient bio map.
1141 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1142 buf_sz -= biotmap_sz;
1144 if (biotmap_sz / INT_MAX > maxphys)
1145 bio_transient_maxcnt = INT_MAX;
1147 bio_transient_maxcnt = biotmap_sz / maxphys;
1149 * Artificially limit to 1024 simultaneous in-flight I/Os
1150 * using the transient mapping.
1152 if (bio_transient_maxcnt > 1024)
1153 bio_transient_maxcnt = 1024;
1155 nbuf = buf_sz / BKVASIZE;
1160 * Pager buffers are allocated for short periods, so scale the
1161 * number of reserved buffers based on the number of CPUs rather
1162 * than amount of memory.
1164 nswbuf = min(nbuf / 4, 32 * mp_ncpus);
1165 if (nswbuf < NSWBUF_MIN)
1166 nswbuf = NSWBUF_MIN;
1170 * Reserve space for the buffer cache buffers
1173 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1174 atop(maxbcachebuf)) * nbuf;
1180 * Single global constant for BUF_WMESG, to avoid getting multiple
1183 static const char buf_wmesg[] = "bufwait";
1185 /* Initialize the buffer subsystem. Called before use of any buffers. */
1192 KASSERT(maxbcachebuf >= MAXBSIZE,
1193 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1195 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1196 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1197 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1198 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1200 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1202 /* finally, initialize each buffer header and stick on empty q */
1203 for (i = 0; i < nbuf; i++) {
1205 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1206 bp->b_flags = B_INVAL;
1207 bp->b_rcred = NOCRED;
1208 bp->b_wcred = NOCRED;
1209 bp->b_qindex = QUEUE_NONE;
1211 bp->b_subqueue = mp_maxid + 1;
1213 bp->b_data = bp->b_kvabase = unmapped_buf;
1214 LIST_INIT(&bp->b_dep);
1215 BUF_LOCKINIT(bp, buf_wmesg);
1216 bq_insert(&bqempty, bp, false);
1220 * maxbufspace is the absolute maximum amount of buffer space we are
1221 * allowed to reserve in KVM and in real terms. The absolute maximum
1222 * is nominally used by metadata. hibufspace is the nominal maximum
1223 * used by most other requests. The differential is required to
1224 * ensure that metadata deadlocks don't occur.
1226 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1227 * this may result in KVM fragmentation which is not handled optimally
1228 * by the system. XXX This is less true with vmem. We could use
1231 maxbufspace = (long)nbuf * BKVASIZE;
1232 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1233 lobufspace = (hibufspace / 20) * 19; /* 95% */
1234 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1237 * Note: The 16 MiB upper limit for hirunningspace was chosen
1238 * arbitrarily and may need further tuning. It corresponds to
1239 * 128 outstanding write IO requests (if IO size is 128 KiB),
1240 * which fits with many RAID controllers' tagged queuing limits.
1241 * The lower 1 MiB limit is the historical upper limit for
1244 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1245 16 * 1024 * 1024), 1024 * 1024);
1246 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1249 * Limit the amount of malloc memory since it is wired permanently into
1250 * the kernel space. Even though this is accounted for in the buffer
1251 * allocation, we don't want the malloced region to grow uncontrolled.
1252 * The malloc scheme improves memory utilization significantly on
1253 * average (small) directories.
1255 maxbufmallocspace = hibufspace / 20;
1258 * Reduce the chance of a deadlock occurring by limiting the number
1259 * of delayed-write dirty buffers we allow to stack up.
1261 hidirtybuffers = nbuf / 4 + 20;
1262 dirtybufthresh = hidirtybuffers * 9 / 10;
1264 * To support extreme low-memory systems, make sure hidirtybuffers
1265 * cannot eat up all available buffer space. This occurs when our
1266 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1267 * buffer space assuming BKVASIZE'd buffers.
1269 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1270 hidirtybuffers >>= 1;
1272 lodirtybuffers = hidirtybuffers / 2;
1275 * lofreebuffers should be sufficient to avoid stalling waiting on
1276 * buf headers under heavy utilization. The bufs in per-cpu caches
1277 * are counted as free but will be unavailable to threads executing
1280 * hifreebuffers is the free target for the bufspace daemon. This
1281 * should be set appropriately to limit work per-iteration.
1283 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1284 hifreebuffers = (3 * lofreebuffers) / 2;
1285 numfreebuffers = nbuf;
1287 /* Setup the kva and free list allocators. */
1288 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1289 buf_zone = uma_zcache_create("buf free cache",
1290 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1291 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1294 * Size the clean queue according to the amount of buffer space.
1295 * One queue per-256mb up to the max. More queues gives better
1296 * concurrency but less accurate LRU.
1298 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1299 for (i = 0 ; i < buf_domains; i++) {
1300 struct bufdomain *bd;
1304 bd->bd_freebuffers = nbuf / buf_domains;
1305 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1306 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1307 bd->bd_bufspace = 0;
1308 bd->bd_maxbufspace = maxbufspace / buf_domains;
1309 bd->bd_hibufspace = hibufspace / buf_domains;
1310 bd->bd_lobufspace = lobufspace / buf_domains;
1311 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1312 bd->bd_numdirtybuffers = 0;
1313 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1314 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1315 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1316 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1317 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1319 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1320 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1321 mappingrestarts = counter_u64_alloc(M_WAITOK);
1322 numbufallocfails = counter_u64_alloc(M_WAITOK);
1323 notbufdflushes = counter_u64_alloc(M_WAITOK);
1324 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1325 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1326 bufkvaspace = counter_u64_alloc(M_WAITOK);
1331 vfs_buf_check_mapped(struct buf *bp)
1334 KASSERT(bp->b_kvabase != unmapped_buf,
1335 ("mapped buf: b_kvabase was not updated %p", bp));
1336 KASSERT(bp->b_data != unmapped_buf,
1337 ("mapped buf: b_data was not updated %p", bp));
1338 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1339 maxphys, ("b_data + b_offset unmapped %p", bp));
1343 vfs_buf_check_unmapped(struct buf *bp)
1346 KASSERT(bp->b_data == unmapped_buf,
1347 ("unmapped buf: corrupted b_data %p", bp));
1350 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1351 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1353 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1354 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1358 isbufbusy(struct buf *bp)
1360 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1361 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1367 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1370 bufshutdown(int show_busybufs)
1372 static int first_buf_printf = 1;
1374 int i, iter, nbusy, pbusy;
1380 * Sync filesystems for shutdown
1382 wdog_kern_pat(WD_LASTVAL);
1383 kern_sync(curthread);
1386 * With soft updates, some buffers that are
1387 * written will be remarked as dirty until other
1388 * buffers are written.
1390 for (iter = pbusy = 0; iter < 20; iter++) {
1392 for (i = nbuf - 1; i >= 0; i--) {
1398 if (first_buf_printf)
1399 printf("All buffers synced.");
1402 if (first_buf_printf) {
1403 printf("Syncing disks, buffers remaining... ");
1404 first_buf_printf = 0;
1406 printf("%d ", nbusy);
1411 wdog_kern_pat(WD_LASTVAL);
1412 kern_sync(curthread);
1416 * Spin for a while to allow interrupt threads to run.
1418 DELAY(50000 * iter);
1421 * Context switch several times to allow interrupt
1424 for (subiter = 0; subiter < 50 * iter; subiter++) {
1425 thread_lock(curthread);
1433 * Count only busy local buffers to prevent forcing
1434 * a fsck if we're just a client of a wedged NFS server
1437 for (i = nbuf - 1; i >= 0; i--) {
1439 if (isbufbusy(bp)) {
1441 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1442 if (bp->b_dev == NULL) {
1443 TAILQ_REMOVE(&mountlist,
1444 bp->b_vp->v_mount, mnt_list);
1449 if (show_busybufs > 0) {
1451 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1452 nbusy, bp, bp->b_vp, bp->b_flags,
1453 (intmax_t)bp->b_blkno,
1454 (intmax_t)bp->b_lblkno);
1455 BUF_LOCKPRINTINFO(bp);
1456 if (show_busybufs > 1)
1464 * Failed to sync all blocks. Indicate this and don't
1465 * unmount filesystems (thus forcing an fsck on reboot).
1467 printf("Giving up on %d buffers\n", nbusy);
1468 DELAY(5000000); /* 5 seconds */
1471 if (!first_buf_printf)
1472 printf("Final sync complete\n");
1475 * Unmount filesystems and perform swapoff, to quiesce
1476 * the system as much as possible. In particular, no
1477 * I/O should be initiated from top levels since it
1478 * might be abruptly terminated by reset, or otherwise
1479 * erronously handled because other parts of the
1480 * system are disabled.
1482 * Swapoff before unmount, because file-backed swap is
1483 * non-operational after unmount of the underlying
1486 if (!KERNEL_PANICKED()) {
1491 DELAY(100000); /* wait for console output to finish */
1495 bpmap_qenter(struct buf *bp)
1498 BUF_CHECK_MAPPED(bp);
1501 * bp->b_data is relative to bp->b_offset, but
1502 * bp->b_offset may be offset into the first page.
1504 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1505 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1506 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1507 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1510 static inline struct bufdomain *
1511 bufdomain(struct buf *bp)
1514 return (&bdomain[bp->b_domain]);
1517 static struct bufqueue *
1518 bufqueue(struct buf *bp)
1521 switch (bp->b_qindex) {
1524 case QUEUE_SENTINEL:
1529 return (&bufdomain(bp)->bd_dirtyq);
1531 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1535 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1539 * Return the locked bufqueue that bp is a member of.
1541 static struct bufqueue *
1542 bufqueue_acquire(struct buf *bp)
1544 struct bufqueue *bq, *nbq;
1547 * bp can be pushed from a per-cpu queue to the
1548 * cleanq while we're waiting on the lock. Retry
1549 * if the queues don't match.
1567 * Insert the buffer into the appropriate free list. Requires a
1568 * locked buffer on entry and buffer is unlocked before return.
1571 binsfree(struct buf *bp, int qindex)
1573 struct bufdomain *bd;
1574 struct bufqueue *bq;
1576 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1577 ("binsfree: Invalid qindex %d", qindex));
1578 BUF_ASSERT_XLOCKED(bp);
1581 * Handle delayed bremfree() processing.
1583 if (bp->b_flags & B_REMFREE) {
1584 if (bp->b_qindex == qindex) {
1585 bp->b_flags |= B_REUSE;
1586 bp->b_flags &= ~B_REMFREE;
1590 bq = bufqueue_acquire(bp);
1595 if (qindex == QUEUE_CLEAN) {
1596 if (bd->bd_lim != 0)
1597 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1601 bq = &bd->bd_dirtyq;
1602 bq_insert(bq, bp, true);
1608 * Free a buffer to the buf zone once it no longer has valid contents.
1611 buf_free(struct buf *bp)
1614 if (bp->b_flags & B_REMFREE)
1616 if (bp->b_vflags & BV_BKGRDINPROG)
1617 panic("losing buffer 1");
1618 if (bp->b_rcred != NOCRED) {
1619 crfree(bp->b_rcred);
1620 bp->b_rcred = NOCRED;
1622 if (bp->b_wcred != NOCRED) {
1623 crfree(bp->b_wcred);
1624 bp->b_wcred = NOCRED;
1626 if (!LIST_EMPTY(&bp->b_dep))
1629 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1630 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1632 uma_zfree(buf_zone, bp);
1638 * Import bufs into the uma cache from the buf list. The system still
1639 * expects a static array of bufs and much of the synchronization
1640 * around bufs assumes type stable storage. As a result, UMA is used
1641 * only as a per-cpu cache of bufs still maintained on a global list.
1644 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1650 for (i = 0; i < cnt; i++) {
1651 bp = TAILQ_FIRST(&bqempty.bq_queue);
1654 bq_remove(&bqempty, bp);
1657 BQ_UNLOCK(&bqempty);
1665 * Release bufs from the uma cache back to the buffer queues.
1668 buf_release(void *arg, void **store, int cnt)
1670 struct bufqueue *bq;
1676 for (i = 0; i < cnt; i++) {
1678 /* Inline bq_insert() to batch locking. */
1679 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1680 bp->b_flags &= ~(B_AGE | B_REUSE);
1682 bp->b_qindex = bq->bq_index;
1690 * Allocate an empty buffer header.
1693 buf_alloc(struct bufdomain *bd)
1696 int freebufs, error;
1699 * We can only run out of bufs in the buf zone if the average buf
1700 * is less than BKVASIZE. In this case the actual wait/block will
1701 * come from buf_reycle() failing to flush one of these small bufs.
1704 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1706 bp = uma_zalloc(buf_zone, M_NOWAIT);
1708 atomic_add_int(&bd->bd_freebuffers, 1);
1709 bufspace_daemon_wakeup(bd);
1710 counter_u64_add(numbufallocfails, 1);
1714 * Wake-up the bufspace daemon on transition below threshold.
1716 if (freebufs == bd->bd_lofreebuffers)
1717 bufspace_daemon_wakeup(bd);
1719 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL);
1720 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1724 KASSERT(bp->b_vp == NULL,
1725 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1726 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1727 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1728 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1729 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1730 KASSERT(bp->b_npages == 0,
1731 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1732 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1733 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1734 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1736 bp->b_domain = BD_DOMAIN(bd);
1742 bp->b_blkno = bp->b_lblkno = 0;
1743 bp->b_offset = NOOFFSET;
1749 bp->b_dirtyoff = bp->b_dirtyend = 0;
1750 bp->b_bufobj = NULL;
1751 bp->b_data = bp->b_kvabase = unmapped_buf;
1752 bp->b_fsprivate1 = NULL;
1753 bp->b_fsprivate2 = NULL;
1754 bp->b_fsprivate3 = NULL;
1755 LIST_INIT(&bp->b_dep);
1763 * Free a buffer from the given bufqueue. kva controls whether the
1764 * freed buf must own some kva resources. This is used for
1768 buf_recycle(struct bufdomain *bd, bool kva)
1770 struct bufqueue *bq;
1771 struct buf *bp, *nbp;
1774 counter_u64_add(bufdefragcnt, 1);
1778 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1779 ("buf_recycle: Locks don't match"));
1780 nbp = TAILQ_FIRST(&bq->bq_queue);
1783 * Run scan, possibly freeing data and/or kva mappings on the fly
1786 while ((bp = nbp) != NULL) {
1788 * Calculate next bp (we can only use it if we do not
1789 * release the bqlock).
1791 nbp = TAILQ_NEXT(bp, b_freelist);
1794 * If we are defragging then we need a buffer with
1795 * some kva to reclaim.
1797 if (kva && bp->b_kvasize == 0)
1800 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1804 * Implement a second chance algorithm for frequently
1807 if ((bp->b_flags & B_REUSE) != 0) {
1808 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1809 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1810 bp->b_flags &= ~B_REUSE;
1816 * Skip buffers with background writes in progress.
1818 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1823 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1824 ("buf_recycle: inconsistent queue %d bp %p",
1826 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1827 ("getnewbuf: queue domain %d doesn't match request %d",
1828 bp->b_domain, (int)BD_DOMAIN(bd)));
1830 * NOTE: nbp is now entirely invalid. We can only restart
1831 * the scan from this point on.
1837 * Requeue the background write buffer with error and
1840 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1843 nbp = TAILQ_FIRST(&bq->bq_queue);
1846 bp->b_flags |= B_INVAL;
1859 * Mark the buffer for removal from the appropriate free list.
1863 bremfree(struct buf *bp)
1866 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1867 KASSERT((bp->b_flags & B_REMFREE) == 0,
1868 ("bremfree: buffer %p already marked for delayed removal.", bp));
1869 KASSERT(bp->b_qindex != QUEUE_NONE,
1870 ("bremfree: buffer %p not on a queue.", bp));
1871 BUF_ASSERT_XLOCKED(bp);
1873 bp->b_flags |= B_REMFREE;
1879 * Force an immediate removal from a free list. Used only in nfs when
1880 * it abuses the b_freelist pointer.
1883 bremfreef(struct buf *bp)
1885 struct bufqueue *bq;
1887 bq = bufqueue_acquire(bp);
1893 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1896 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1897 TAILQ_INIT(&bq->bq_queue);
1899 bq->bq_index = qindex;
1900 bq->bq_subqueue = subqueue;
1904 bd_init(struct bufdomain *bd)
1908 /* Per-CPU clean buf queues, plus one global queue. */
1909 bd->bd_subq = mallocarray(mp_maxid + 2, sizeof(struct bufqueue),
1910 M_BIOBUF, M_WAITOK | M_ZERO);
1911 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1912 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1913 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1914 for (i = 0; i <= mp_maxid; i++)
1915 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1916 "bufq clean subqueue lock");
1917 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1923 * Removes a buffer from the free list, must be called with the
1924 * correct qlock held.
1927 bq_remove(struct bufqueue *bq, struct buf *bp)
1930 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1931 bp, bp->b_vp, bp->b_flags);
1932 KASSERT(bp->b_qindex != QUEUE_NONE,
1933 ("bq_remove: buffer %p not on a queue.", bp));
1934 KASSERT(bufqueue(bp) == bq,
1935 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1937 BQ_ASSERT_LOCKED(bq);
1938 if (bp->b_qindex != QUEUE_EMPTY) {
1939 BUF_ASSERT_XLOCKED(bp);
1941 KASSERT(bq->bq_len >= 1,
1942 ("queue %d underflow", bp->b_qindex));
1943 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1945 bp->b_qindex = QUEUE_NONE;
1946 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1950 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1954 BQ_ASSERT_LOCKED(bq);
1955 if (bq != bd->bd_cleanq) {
1957 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1958 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1959 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1961 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1963 bd->bd_cleanq->bq_len += bq->bq_len;
1966 if (bd->bd_wanted) {
1968 wakeup(&bd->bd_wanted);
1970 if (bq != bd->bd_cleanq)
1975 bd_flushall(struct bufdomain *bd)
1977 struct bufqueue *bq;
1981 if (bd->bd_lim == 0)
1984 for (i = 0; i <= mp_maxid; i++) {
1985 bq = &bd->bd_subq[i];
1986 if (bq->bq_len == 0)
1998 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
2000 struct bufdomain *bd;
2002 if (bp->b_qindex != QUEUE_NONE)
2003 panic("bq_insert: free buffer %p onto another queue?", bp);
2006 if (bp->b_flags & B_AGE) {
2007 /* Place this buf directly on the real queue. */
2008 if (bq->bq_index == QUEUE_CLEAN)
2011 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2014 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2016 bp->b_flags &= ~(B_AGE | B_REUSE);
2018 bp->b_qindex = bq->bq_index;
2019 bp->b_subqueue = bq->bq_subqueue;
2022 * Unlock before we notify so that we don't wakeup a waiter that
2023 * fails a trylock on the buf and sleeps again.
2028 if (bp->b_qindex == QUEUE_CLEAN) {
2030 * Flush the per-cpu queue and notify any waiters.
2032 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2033 bq->bq_len >= bd->bd_lim))
2042 * Free the kva allocation for a buffer.
2046 bufkva_free(struct buf *bp)
2050 if (bp->b_kvasize == 0) {
2051 KASSERT(bp->b_kvabase == unmapped_buf &&
2052 bp->b_data == unmapped_buf,
2053 ("Leaked KVA space on %p", bp));
2054 } else if (buf_mapped(bp))
2055 BUF_CHECK_MAPPED(bp);
2057 BUF_CHECK_UNMAPPED(bp);
2059 if (bp->b_kvasize == 0)
2062 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2063 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2064 counter_u64_add(buffreekvacnt, 1);
2065 bp->b_data = bp->b_kvabase = unmapped_buf;
2072 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2075 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2080 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2081 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2082 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2083 KASSERT(maxsize <= maxbcachebuf,
2084 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2089 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2092 * Buffer map is too fragmented. Request the caller
2093 * to defragment the map.
2097 bp->b_kvabase = (caddr_t)addr;
2098 bp->b_kvasize = maxsize;
2099 counter_u64_add(bufkvaspace, bp->b_kvasize);
2100 if ((gbflags & GB_UNMAPPED) != 0) {
2101 bp->b_data = unmapped_buf;
2102 BUF_CHECK_UNMAPPED(bp);
2104 bp->b_data = bp->b_kvabase;
2105 BUF_CHECK_MAPPED(bp);
2113 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2114 * callback that fires to avoid returning failure.
2117 bufkva_reclaim(vmem_t *vmem, int flags)
2124 for (i = 0; i < 5; i++) {
2125 for (q = 0; q < buf_domains; q++)
2126 if (buf_recycle(&bdomain[q], true) != 0)
2135 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2136 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2137 * the buffer is valid and we do not have to do anything.
2140 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2141 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2149 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2150 if (inmem(vp, *rablkno))
2152 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2153 if ((rabp->b_flags & B_CACHE) != 0) {
2160 racct_add_buf(curproc, rabp, 0);
2161 PROC_UNLOCK(curproc);
2164 td->td_ru.ru_inblock++;
2165 rabp->b_flags |= B_ASYNC;
2166 rabp->b_flags &= ~B_INVAL;
2167 if ((flags & GB_CKHASH) != 0) {
2168 rabp->b_flags |= B_CKHASH;
2169 rabp->b_ckhashcalc = ckhashfunc;
2171 rabp->b_ioflags &= ~BIO_ERROR;
2172 rabp->b_iocmd = BIO_READ;
2173 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2174 rabp->b_rcred = crhold(cred);
2175 vfs_busy_pages(rabp, 0);
2177 rabp->b_iooffset = dbtob(rabp->b_blkno);
2183 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2185 * Get a buffer with the specified data. Look in the cache first. We
2186 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2187 * is set, the buffer is valid and we do not have to do anything, see
2188 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2190 * Always return a NULL buffer pointer (in bpp) when returning an error.
2192 * The blkno parameter is the logical block being requested. Normally
2193 * the mapping of logical block number to disk block address is done
2194 * by calling VOP_BMAP(). However, if the mapping is already known, the
2195 * disk block address can be passed using the dblkno parameter. If the
2196 * disk block address is not known, then the same value should be passed
2197 * for blkno and dblkno.
2200 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2201 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2202 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2206 int error, readwait, rv;
2208 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2211 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2214 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2219 KASSERT(blkno == bp->b_lblkno,
2220 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2221 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2222 flags &= ~GB_NOSPARSE;
2226 * If not found in cache, do some I/O
2229 if ((bp->b_flags & B_CACHE) == 0) {
2232 PROC_LOCK(td->td_proc);
2233 racct_add_buf(td->td_proc, bp, 0);
2234 PROC_UNLOCK(td->td_proc);
2237 td->td_ru.ru_inblock++;
2238 bp->b_iocmd = BIO_READ;
2239 bp->b_flags &= ~B_INVAL;
2240 if ((flags & GB_CKHASH) != 0) {
2241 bp->b_flags |= B_CKHASH;
2242 bp->b_ckhashcalc = ckhashfunc;
2244 if ((flags & GB_CVTENXIO) != 0)
2245 bp->b_xflags |= BX_CVTENXIO;
2246 bp->b_ioflags &= ~BIO_ERROR;
2247 if (bp->b_rcred == NOCRED && cred != NOCRED)
2248 bp->b_rcred = crhold(cred);
2249 vfs_busy_pages(bp, 0);
2250 bp->b_iooffset = dbtob(bp->b_blkno);
2256 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2258 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2272 * Write, release buffer on completion. (Done by iodone
2273 * if async). Do not bother writing anything if the buffer
2276 * Note that we set B_CACHE here, indicating that buffer is
2277 * fully valid and thus cacheable. This is true even of NFS
2278 * now so we set it generally. This could be set either here
2279 * or in biodone() since the I/O is synchronous. We put it
2283 bufwrite(struct buf *bp)
2290 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2291 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2292 bp->b_flags |= B_INVAL | B_RELBUF;
2293 bp->b_flags &= ~B_CACHE;
2297 if (bp->b_flags & B_INVAL) {
2302 if (bp->b_flags & B_BARRIER)
2303 atomic_add_long(&barrierwrites, 1);
2305 oldflags = bp->b_flags;
2307 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2308 ("FFS background buffer should not get here %p", bp));
2312 vp_md = vp->v_vflag & VV_MD;
2317 * Mark the buffer clean. Increment the bufobj write count
2318 * before bundirty() call, to prevent other thread from seeing
2319 * empty dirty list and zero counter for writes in progress,
2320 * falsely indicating that the bufobj is clean.
2322 bufobj_wref(bp->b_bufobj);
2325 bp->b_flags &= ~B_DONE;
2326 bp->b_ioflags &= ~BIO_ERROR;
2327 bp->b_flags |= B_CACHE;
2328 bp->b_iocmd = BIO_WRITE;
2330 vfs_busy_pages(bp, 1);
2333 * Normal bwrites pipeline writes
2335 bp->b_runningbufspace = bp->b_bufsize;
2336 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2341 racct_add_buf(curproc, bp, 1);
2342 PROC_UNLOCK(curproc);
2345 curthread->td_ru.ru_oublock++;
2346 if (oldflags & B_ASYNC)
2348 bp->b_iooffset = dbtob(bp->b_blkno);
2349 buf_track(bp, __func__);
2352 if ((oldflags & B_ASYNC) == 0) {
2353 int rtval = bufwait(bp);
2356 } else if (space > hirunningspace) {
2358 * don't allow the async write to saturate the I/O
2359 * system. We will not deadlock here because
2360 * we are blocking waiting for I/O that is already in-progress
2361 * to complete. We do not block here if it is the update
2362 * or syncer daemon trying to clean up as that can lead
2365 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2366 waitrunningbufspace();
2373 bufbdflush(struct bufobj *bo, struct buf *bp)
2376 struct bufdomain *bd;
2378 bd = &bdomain[bo->bo_domain];
2379 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2380 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2382 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2385 * Try to find a buffer to flush.
2387 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2388 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2390 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2393 panic("bdwrite: found ourselves");
2395 /* Don't countdeps with the bo lock held. */
2396 if (buf_countdeps(nbp, 0)) {
2401 if (nbp->b_flags & B_CLUSTEROK) {
2402 vfs_bio_awrite(nbp);
2407 dirtybufferflushes++;
2416 * Delayed write. (Buffer is marked dirty). Do not bother writing
2417 * anything if the buffer is marked invalid.
2419 * Note that since the buffer must be completely valid, we can safely
2420 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2421 * biodone() in order to prevent getblk from writing the buffer
2422 * out synchronously.
2425 bdwrite(struct buf *bp)
2427 struct thread *td = curthread;
2431 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2432 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2433 KASSERT((bp->b_flags & B_BARRIER) == 0,
2434 ("Barrier request in delayed write %p", bp));
2436 if (bp->b_flags & B_INVAL) {
2442 * If we have too many dirty buffers, don't create any more.
2443 * If we are wildly over our limit, then force a complete
2444 * cleanup. Otherwise, just keep the situation from getting
2445 * out of control. Note that we have to avoid a recursive
2446 * disaster and not try to clean up after our own cleanup!
2450 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2451 td->td_pflags |= TDP_INBDFLUSH;
2453 td->td_pflags &= ~TDP_INBDFLUSH;
2459 * Set B_CACHE, indicating that the buffer is fully valid. This is
2460 * true even of NFS now.
2462 bp->b_flags |= B_CACHE;
2465 * This bmap keeps the system from needing to do the bmap later,
2466 * perhaps when the system is attempting to do a sync. Since it
2467 * is likely that the indirect block -- or whatever other datastructure
2468 * that the filesystem needs is still in memory now, it is a good
2469 * thing to do this. Note also, that if the pageout daemon is
2470 * requesting a sync -- there might not be enough memory to do
2471 * the bmap then... So, this is important to do.
2473 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2474 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2477 buf_track(bp, __func__);
2480 * Set the *dirty* buffer range based upon the VM system dirty
2483 * Mark the buffer pages as clean. We need to do this here to
2484 * satisfy the vnode_pager and the pageout daemon, so that it
2485 * thinks that the pages have been "cleaned". Note that since
2486 * the pages are in a delayed write buffer -- the VFS layer
2487 * "will" see that the pages get written out on the next sync,
2488 * or perhaps the cluster will be completed.
2490 vfs_clean_pages_dirty_buf(bp);
2494 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2495 * due to the softdep code.
2502 * Turn buffer into delayed write request. We must clear BIO_READ and
2503 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2504 * itself to properly update it in the dirty/clean lists. We mark it
2505 * B_DONE to ensure that any asynchronization of the buffer properly
2506 * clears B_DONE ( else a panic will occur later ).
2508 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2509 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2510 * should only be called if the buffer is known-good.
2512 * Since the buffer is not on a queue, we do not update the numfreebuffers
2515 * The buffer must be on QUEUE_NONE.
2518 bdirty(struct buf *bp)
2521 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2522 bp, bp->b_vp, bp->b_flags);
2523 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2524 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2525 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2526 bp->b_flags &= ~(B_RELBUF);
2527 bp->b_iocmd = BIO_WRITE;
2529 if ((bp->b_flags & B_DELWRI) == 0) {
2530 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2539 * Clear B_DELWRI for buffer.
2541 * Since the buffer is not on a queue, we do not update the numfreebuffers
2544 * The buffer must be on QUEUE_NONE.
2548 bundirty(struct buf *bp)
2551 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2552 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2553 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2554 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2556 if (bp->b_flags & B_DELWRI) {
2557 bp->b_flags &= ~B_DELWRI;
2562 * Since it is now being written, we can clear its deferred write flag.
2564 bp->b_flags &= ~B_DEFERRED;
2570 * Asynchronous write. Start output on a buffer, but do not wait for
2571 * it to complete. The buffer is released when the output completes.
2573 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2574 * B_INVAL buffers. Not us.
2577 bawrite(struct buf *bp)
2580 bp->b_flags |= B_ASYNC;
2587 * Asynchronous barrier write. Start output on a buffer, but do not
2588 * wait for it to complete. Place a write barrier after this write so
2589 * that this buffer and all buffers written before it are committed to
2590 * the disk before any buffers written after this write are committed
2591 * to the disk. The buffer is released when the output completes.
2594 babarrierwrite(struct buf *bp)
2597 bp->b_flags |= B_ASYNC | B_BARRIER;
2604 * Synchronous barrier write. Start output on a buffer and wait for
2605 * it to complete. Place a write barrier after this write so that
2606 * this buffer and all buffers written before it are committed to
2607 * the disk before any buffers written after this write are committed
2608 * to the disk. The buffer is released when the output completes.
2611 bbarrierwrite(struct buf *bp)
2614 bp->b_flags |= B_BARRIER;
2615 return (bwrite(bp));
2621 * Called prior to the locking of any vnodes when we are expecting to
2622 * write. We do not want to starve the buffer cache with too many
2623 * dirty buffers so we block here. By blocking prior to the locking
2624 * of any vnodes we attempt to avoid the situation where a locked vnode
2625 * prevents the various system daemons from flushing related buffers.
2631 if (buf_dirty_count_severe()) {
2632 mtx_lock(&bdirtylock);
2633 while (buf_dirty_count_severe()) {
2635 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2638 mtx_unlock(&bdirtylock);
2643 * Return true if we have too many dirty buffers.
2646 buf_dirty_count_severe(void)
2649 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2655 * Release a busy buffer and, if requested, free its resources. The
2656 * buffer will be stashed in the appropriate bufqueue[] allowing it
2657 * to be accessed later as a cache entity or reused for other purposes.
2660 brelse(struct buf *bp)
2662 struct mount *v_mnt;
2666 * Many functions erroneously call brelse with a NULL bp under rare
2667 * error conditions. Simply return when called with a NULL bp.
2671 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2672 bp, bp->b_vp, bp->b_flags);
2673 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2674 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2675 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2676 ("brelse: non-VMIO buffer marked NOREUSE"));
2678 if (BUF_LOCKRECURSED(bp)) {
2680 * Do not process, in particular, do not handle the
2681 * B_INVAL/B_RELBUF and do not release to free list.
2687 if (bp->b_flags & B_MANAGED) {
2692 if (LIST_EMPTY(&bp->b_dep)) {
2693 bp->b_flags &= ~B_IOSTARTED;
2695 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2696 ("brelse: SU io not finished bp %p", bp));
2699 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2700 BO_LOCK(bp->b_bufobj);
2701 bp->b_vflags &= ~BV_BKGRDERR;
2702 BO_UNLOCK(bp->b_bufobj);
2706 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2707 (bp->b_flags & B_INVALONERR)) {
2709 * Forced invalidation of dirty buffer contents, to be used
2710 * after a failed write in the rare case that the loss of the
2711 * contents is acceptable. The buffer is invalidated and
2714 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2715 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2718 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2719 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2720 !(bp->b_flags & B_INVAL)) {
2722 * Failed write, redirty. All errors except ENXIO (which
2723 * means the device is gone) are treated as being
2726 * XXX Treating EIO as transient is not correct; the
2727 * contract with the local storage device drivers is that
2728 * they will only return EIO once the I/O is no longer
2729 * retriable. Network I/O also respects this through the
2730 * guarantees of TCP and/or the internal retries of NFS.
2731 * ENOMEM might be transient, but we also have no way of
2732 * knowing when its ok to retry/reschedule. In general,
2733 * this entire case should be made obsolete through better
2734 * error handling/recovery and resource scheduling.
2736 * Do this also for buffers that failed with ENXIO, but have
2737 * non-empty dependencies - the soft updates code might need
2738 * to access the buffer to untangle them.
2740 * Must clear BIO_ERROR to prevent pages from being scrapped.
2742 bp->b_ioflags &= ~BIO_ERROR;
2744 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2745 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2747 * Either a failed read I/O, or we were asked to free or not
2748 * cache the buffer, or we failed to write to a device that's
2749 * no longer present.
2751 bp->b_flags |= B_INVAL;
2752 if (!LIST_EMPTY(&bp->b_dep))
2754 if (bp->b_flags & B_DELWRI)
2756 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2757 if ((bp->b_flags & B_VMIO) == 0) {
2765 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2766 * is called with B_DELWRI set, the underlying pages may wind up
2767 * getting freed causing a previous write (bdwrite()) to get 'lost'
2768 * because pages associated with a B_DELWRI bp are marked clean.
2770 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2771 * if B_DELWRI is set.
2773 if (bp->b_flags & B_DELWRI)
2774 bp->b_flags &= ~B_RELBUF;
2777 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2778 * constituted, not even NFS buffers now. Two flags effect this. If
2779 * B_INVAL, the struct buf is invalidated but the VM object is kept
2780 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2782 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2783 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2784 * buffer is also B_INVAL because it hits the re-dirtying code above.
2786 * Normally we can do this whether a buffer is B_DELWRI or not. If
2787 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2788 * the commit state and we cannot afford to lose the buffer. If the
2789 * buffer has a background write in progress, we need to keep it
2790 * around to prevent it from being reconstituted and starting a second
2794 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2796 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2797 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2798 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2799 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2800 vfs_vmio_invalidate(bp);
2804 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2805 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2807 bp->b_flags &= ~B_NOREUSE;
2808 if (bp->b_vp != NULL)
2813 * If the buffer has junk contents signal it and eventually
2814 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2817 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2818 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2819 bp->b_flags |= B_INVAL;
2820 if (bp->b_flags & B_INVAL) {
2821 if (bp->b_flags & B_DELWRI)
2827 buf_track(bp, __func__);
2829 /* buffers with no memory */
2830 if (bp->b_bufsize == 0) {
2834 /* buffers with junk contents */
2835 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2836 (bp->b_ioflags & BIO_ERROR)) {
2837 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2838 if (bp->b_vflags & BV_BKGRDINPROG)
2839 panic("losing buffer 2");
2840 qindex = QUEUE_CLEAN;
2841 bp->b_flags |= B_AGE;
2842 /* remaining buffers */
2843 } else if (bp->b_flags & B_DELWRI)
2844 qindex = QUEUE_DIRTY;
2846 qindex = QUEUE_CLEAN;
2848 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2849 panic("brelse: not dirty");
2851 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2852 bp->b_xflags &= ~(BX_CVTENXIO);
2853 /* binsfree unlocks bp. */
2854 binsfree(bp, qindex);
2858 * Release a buffer back to the appropriate queue but do not try to free
2859 * it. The buffer is expected to be used again soon.
2861 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2862 * biodone() to requeue an async I/O on completion. It is also used when
2863 * known good buffers need to be requeued but we think we may need the data
2866 * XXX we should be able to leave the B_RELBUF hint set on completion.
2869 bqrelse(struct buf *bp)
2873 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2874 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2875 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2877 qindex = QUEUE_NONE;
2878 if (BUF_LOCKRECURSED(bp)) {
2879 /* do not release to free list */
2883 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2884 bp->b_xflags &= ~(BX_CVTENXIO);
2886 if (LIST_EMPTY(&bp->b_dep)) {
2887 bp->b_flags &= ~B_IOSTARTED;
2889 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2890 ("bqrelse: SU io not finished bp %p", bp));
2893 if (bp->b_flags & B_MANAGED) {
2894 if (bp->b_flags & B_REMFREE)
2899 /* buffers with stale but valid contents */
2900 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2901 BV_BKGRDERR)) == BV_BKGRDERR) {
2902 BO_LOCK(bp->b_bufobj);
2903 bp->b_vflags &= ~BV_BKGRDERR;
2904 BO_UNLOCK(bp->b_bufobj);
2905 qindex = QUEUE_DIRTY;
2907 if ((bp->b_flags & B_DELWRI) == 0 &&
2908 (bp->b_xflags & BX_VNDIRTY))
2909 panic("bqrelse: not dirty");
2910 if ((bp->b_flags & B_NOREUSE) != 0) {
2914 qindex = QUEUE_CLEAN;
2916 buf_track(bp, __func__);
2917 /* binsfree unlocks bp. */
2918 binsfree(bp, qindex);
2922 buf_track(bp, __func__);
2928 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2929 * restore bogus pages.
2932 vfs_vmio_iodone(struct buf *bp)
2937 struct vnode *vp __unused;
2938 int i, iosize, resid;
2941 obj = bp->b_bufobj->bo_object;
2942 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2943 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2944 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2947 VNPASS(vp->v_holdcnt > 0, vp);
2948 VNPASS(vp->v_object != NULL, vp);
2950 foff = bp->b_offset;
2951 KASSERT(bp->b_offset != NOOFFSET,
2952 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2955 iosize = bp->b_bcount - bp->b_resid;
2956 for (i = 0; i < bp->b_npages; i++) {
2957 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2962 * cleanup bogus pages, restoring the originals
2965 if (m == bogus_page) {
2967 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2969 panic("biodone: page disappeared!");
2971 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2973 * In the write case, the valid and clean bits are
2974 * already changed correctly ( see bdwrite() ), so we
2975 * only need to do this here in the read case.
2977 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2978 resid)) == 0, ("vfs_vmio_iodone: page %p "
2979 "has unexpected dirty bits", m));
2980 vfs_page_set_valid(bp, foff, m);
2982 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2983 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2984 (intmax_t)foff, (uintmax_t)m->pindex));
2987 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2990 vm_object_pip_wakeupn(obj, bp->b_npages);
2991 if (bogus && buf_mapped(bp)) {
2992 BUF_CHECK_MAPPED(bp);
2993 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2994 bp->b_pages, bp->b_npages);
2999 * Perform page invalidation when a buffer is released. The fully invalid
3000 * pages will be reclaimed later in vfs_vmio_truncate().
3003 vfs_vmio_invalidate(struct buf *bp)
3007 int flags, i, resid, poffset, presid;
3009 if (buf_mapped(bp)) {
3010 BUF_CHECK_MAPPED(bp);
3011 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3013 BUF_CHECK_UNMAPPED(bp);
3015 * Get the base offset and length of the buffer. Note that
3016 * in the VMIO case if the buffer block size is not
3017 * page-aligned then b_data pointer may not be page-aligned.
3018 * But our b_pages[] array *IS* page aligned.
3020 * block sizes less then DEV_BSIZE (usually 512) are not
3021 * supported due to the page granularity bits (m->valid,
3022 * m->dirty, etc...).
3024 * See man buf(9) for more information
3026 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3027 obj = bp->b_bufobj->bo_object;
3028 resid = bp->b_bufsize;
3029 poffset = bp->b_offset & PAGE_MASK;
3030 VM_OBJECT_WLOCK(obj);
3031 for (i = 0; i < bp->b_npages; i++) {
3033 if (m == bogus_page)
3034 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3035 bp->b_pages[i] = NULL;
3037 presid = resid > (PAGE_SIZE - poffset) ?
3038 (PAGE_SIZE - poffset) : resid;
3039 KASSERT(presid >= 0, ("brelse: extra page"));
3040 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3041 if (pmap_page_wired_mappings(m) == 0)
3042 vm_page_set_invalid(m, poffset, presid);
3044 vm_page_release_locked(m, flags);
3048 VM_OBJECT_WUNLOCK(obj);
3053 * Page-granular truncation of an existing VMIO buffer.
3056 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3062 if (bp->b_npages == desiredpages)
3065 if (buf_mapped(bp)) {
3066 BUF_CHECK_MAPPED(bp);
3067 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3068 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3070 BUF_CHECK_UNMAPPED(bp);
3073 * The object lock is needed only if we will attempt to free pages.
3075 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3076 if ((bp->b_flags & B_DIRECT) != 0) {
3077 flags |= VPR_TRYFREE;
3078 obj = bp->b_bufobj->bo_object;
3079 VM_OBJECT_WLOCK(obj);
3083 for (i = desiredpages; i < bp->b_npages; i++) {
3085 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3086 bp->b_pages[i] = NULL;
3088 vm_page_release_locked(m, flags);
3090 vm_page_release(m, flags);
3093 VM_OBJECT_WUNLOCK(obj);
3094 bp->b_npages = desiredpages;
3098 * Byte granular extension of VMIO buffers.
3101 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3104 * We are growing the buffer, possibly in a
3105 * byte-granular fashion.
3113 * Step 1, bring in the VM pages from the object, allocating
3114 * them if necessary. We must clear B_CACHE if these pages
3115 * are not valid for the range covered by the buffer.
3117 obj = bp->b_bufobj->bo_object;
3118 if (bp->b_npages < desiredpages) {
3119 KASSERT(desiredpages <= atop(maxbcachebuf),
3120 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3121 bp, desiredpages, maxbcachebuf));
3124 * We must allocate system pages since blocking
3125 * here could interfere with paging I/O, no
3126 * matter which process we are.
3128 * Only exclusive busy can be tested here.
3129 * Blocking on shared busy might lead to
3130 * deadlocks once allocbuf() is called after
3131 * pages are vfs_busy_pages().
3133 (void)vm_page_grab_pages_unlocked(obj,
3134 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3135 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3136 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3137 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3138 bp->b_npages = desiredpages;
3142 * Step 2. We've loaded the pages into the buffer,
3143 * we have to figure out if we can still have B_CACHE
3144 * set. Note that B_CACHE is set according to the
3145 * byte-granular range ( bcount and size ), not the
3146 * aligned range ( newbsize ).
3148 * The VM test is against m->valid, which is DEV_BSIZE
3149 * aligned. Needless to say, the validity of the data
3150 * needs to also be DEV_BSIZE aligned. Note that this
3151 * fails with NFS if the server or some other client
3152 * extends the file's EOF. If our buffer is resized,
3153 * B_CACHE may remain set! XXX
3155 toff = bp->b_bcount;
3156 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3157 while ((bp->b_flags & B_CACHE) && toff < size) {
3160 if (tinc > (size - toff))
3162 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3163 m = bp->b_pages[pi];
3164 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3170 * Step 3, fixup the KVA pmap.
3175 BUF_CHECK_UNMAPPED(bp);
3179 * Check to see if a block at a particular lbn is available for a clustered
3183 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3190 /* If the buf isn't in core skip it */
3191 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3194 /* If the buf is busy we don't want to wait for it */
3195 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3198 /* Only cluster with valid clusterable delayed write buffers */
3199 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3200 (B_DELWRI | B_CLUSTEROK))
3203 if (bpa->b_bufsize != size)
3207 * Check to see if it is in the expected place on disk and that the
3208 * block has been mapped.
3210 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3220 * Implement clustered async writes for clearing out B_DELWRI buffers.
3221 * This is much better then the old way of writing only one buffer at
3222 * a time. Note that we may not be presented with the buffers in the
3223 * correct order, so we search for the cluster in both directions.
3226 vfs_bio_awrite(struct buf *bp)
3231 daddr_t lblkno = bp->b_lblkno;
3232 struct vnode *vp = bp->b_vp;
3240 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3242 * right now we support clustered writing only to regular files. If
3243 * we find a clusterable block we could be in the middle of a cluster
3244 * rather then at the beginning.
3246 if ((vp->v_type == VREG) &&
3247 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3248 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3249 size = vp->v_mount->mnt_stat.f_iosize;
3250 maxcl = maxphys / size;
3253 for (i = 1; i < maxcl; i++)
3254 if (vfs_bio_clcheck(vp, size, lblkno + i,
3255 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3258 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3259 if (vfs_bio_clcheck(vp, size, lblkno - j,
3260 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3266 * this is a possible cluster write
3270 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3276 bp->b_flags |= B_ASYNC;
3278 * default (old) behavior, writing out only one block
3280 * XXX returns b_bufsize instead of b_bcount for nwritten?
3282 nwritten = bp->b_bufsize;
3291 * Allocate KVA for an empty buf header according to gbflags.
3294 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3297 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3299 * In order to keep fragmentation sane we only allocate kva
3300 * in BKVASIZE chunks. XXX with vmem we can do page size.
3302 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3304 if (maxsize != bp->b_kvasize &&
3305 bufkva_alloc(bp, maxsize, gbflags))
3314 * Find and initialize a new buffer header, freeing up existing buffers
3315 * in the bufqueues as necessary. The new buffer is returned locked.
3318 * We have insufficient buffer headers
3319 * We have insufficient buffer space
3320 * buffer_arena is too fragmented ( space reservation fails )
3321 * If we have to flush dirty buffers ( but we try to avoid this )
3323 * The caller is responsible for releasing the reserved bufspace after
3324 * allocbuf() is called.
3327 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3329 struct bufdomain *bd;
3331 bool metadata, reserved;
3334 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3335 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3336 if (!unmapped_buf_allowed)
3337 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3339 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3347 bd = &bdomain[vp->v_bufobj.bo_domain];
3349 counter_u64_add(getnewbufcalls, 1);
3352 if (reserved == false &&
3353 bufspace_reserve(bd, maxsize, metadata) != 0) {
3354 counter_u64_add(getnewbufrestarts, 1);
3358 if ((bp = buf_alloc(bd)) == NULL) {
3359 counter_u64_add(getnewbufrestarts, 1);
3362 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3365 } while (buf_recycle(bd, false) == 0);
3368 bufspace_release(bd, maxsize);
3370 bp->b_flags |= B_INVAL;
3373 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3381 * buffer flushing daemon. Buffers are normally flushed by the
3382 * update daemon but if it cannot keep up this process starts to
3383 * take the load in an attempt to prevent getnewbuf() from blocking.
3385 static struct kproc_desc buf_kp = {
3390 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3393 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3397 flushed = flushbufqueues(vp, bd, target, 0);
3400 * Could not find any buffers without rollback
3401 * dependencies, so just write the first one
3402 * in the hopes of eventually making progress.
3404 if (vp != NULL && target > 2)
3406 flushbufqueues(vp, bd, target, 1);
3412 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3416 if (KERNEL_PANICKED())
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);
3978 goto newbuf_unlocked;
3980 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3985 /* Verify buf identify has not changed since lookup. */
3986 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3987 goto foundbuf_fastpath;
3989 /* It changed, fallback to locked lookup. */
3994 bp = gbincore(bo, blkno);
3999 * Buffer is in-core. If the buffer is not busy nor managed,
4000 * it must be on a queue.
4002 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
4003 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
4005 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4008 error = BUF_TIMELOCK(bp, lockflags,
4009 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4012 * If we slept and got the lock we have to restart in case
4013 * the buffer changed identities.
4015 if (error == ENOLCK)
4017 /* We timed out or were interrupted. */
4018 else if (error != 0)
4022 /* If recursed, assume caller knows the rules. */
4023 if (BUF_LOCKRECURSED(bp))
4027 * The buffer is locked. B_CACHE is cleared if the buffer is
4028 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4029 * and for a VMIO buffer B_CACHE is adjusted according to the
4032 if (bp->b_flags & B_INVAL)
4033 bp->b_flags &= ~B_CACHE;
4034 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4035 bp->b_flags |= B_CACHE;
4036 if (bp->b_flags & B_MANAGED)
4037 MPASS(bp->b_qindex == QUEUE_NONE);
4042 * check for size inconsistencies for non-VMIO case.
4044 if (bp->b_bcount != size) {
4045 if ((bp->b_flags & B_VMIO) == 0 ||
4046 (size > bp->b_kvasize)) {
4047 if (bp->b_flags & B_DELWRI) {
4048 bp->b_flags |= B_NOCACHE;
4051 if (LIST_EMPTY(&bp->b_dep)) {
4052 bp->b_flags |= B_RELBUF;
4055 bp->b_flags |= B_NOCACHE;
4064 * Handle the case of unmapped buffer which should
4065 * become mapped, or the buffer for which KVA
4066 * reservation is requested.
4068 bp_unmapped_get_kva(bp, blkno, size, flags);
4071 * If the size is inconsistent in the VMIO case, we can resize
4072 * the buffer. This might lead to B_CACHE getting set or
4073 * cleared. If the size has not changed, B_CACHE remains
4074 * unchanged from its previous state.
4078 KASSERT(bp->b_offset != NOOFFSET,
4079 ("getblk: no buffer offset"));
4082 * A buffer with B_DELWRI set and B_CACHE clear must
4083 * be committed before we can return the buffer in
4084 * order to prevent the caller from issuing a read
4085 * ( due to B_CACHE not being set ) and overwriting
4088 * Most callers, including NFS and FFS, need this to
4089 * operate properly either because they assume they
4090 * can issue a read if B_CACHE is not set, or because
4091 * ( for example ) an uncached B_DELWRI might loop due
4092 * to softupdates re-dirtying the buffer. In the latter
4093 * case, B_CACHE is set after the first write completes,
4094 * preventing further loops.
4095 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4096 * above while extending the buffer, we cannot allow the
4097 * buffer to remain with B_CACHE set after the write
4098 * completes or it will represent a corrupt state. To
4099 * deal with this we set B_NOCACHE to scrap the buffer
4102 * We might be able to do something fancy, like setting
4103 * B_CACHE in bwrite() except if B_DELWRI is already set,
4104 * so the below call doesn't set B_CACHE, but that gets real
4105 * confusing. This is much easier.
4108 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4109 bp->b_flags |= B_NOCACHE;
4113 bp->b_flags &= ~B_DONE;
4116 * Buffer is not in-core, create new buffer. The buffer
4117 * returned by getnewbuf() is locked. Note that the returned
4118 * buffer is also considered valid (not marked B_INVAL).
4123 * If the user does not want us to create the buffer, bail out
4126 if (flags & GB_NOCREAT)
4129 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4130 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4131 offset = blkno * bsize;
4132 vmio = vp->v_object != NULL;
4134 maxsize = size + (offset & PAGE_MASK);
4137 /* Do not allow non-VMIO notmapped buffers. */
4138 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4140 maxsize = imax(maxsize, bsize);
4141 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4143 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4144 KASSERT(error != EOPNOTSUPP,
4145 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4150 return (EJUSTRETURN);
4153 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4155 if (slpflag || slptimeo)
4158 * XXX This is here until the sleep path is diagnosed
4159 * enough to work under very low memory conditions.
4161 * There's an issue on low memory, 4BSD+non-preempt
4162 * systems (eg MIPS routers with 32MB RAM) where buffer
4163 * exhaustion occurs without sleeping for buffer
4164 * reclaimation. This just sticks in a loop and
4165 * constantly attempts to allocate a buffer, which
4166 * hits exhaustion and tries to wakeup bufdaemon.
4167 * This never happens because we never yield.
4169 * The real solution is to identify and fix these cases
4170 * so we aren't effectively busy-waiting in a loop
4171 * until the reclaimation path has cycles to run.
4173 kern_yield(PRI_USER);
4178 * This code is used to make sure that a buffer is not
4179 * created while the getnewbuf routine is blocked.
4180 * This can be a problem whether the vnode is locked or not.
4181 * If the buffer is created out from under us, we have to
4182 * throw away the one we just created.
4184 * Note: this must occur before we associate the buffer
4185 * with the vp especially considering limitations in
4186 * the splay tree implementation when dealing with duplicate
4190 if (gbincore(bo, blkno)) {
4192 bp->b_flags |= B_INVAL;
4193 bufspace_release(bufdomain(bp), maxsize);
4199 * Insert the buffer into the hash, so that it can
4200 * be found by incore.
4202 bp->b_lblkno = blkno;
4203 bp->b_blkno = d_blkno;
4204 bp->b_offset = offset;
4209 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4210 * buffer size starts out as 0, B_CACHE will be set by
4211 * allocbuf() for the VMIO case prior to it testing the
4212 * backing store for validity.
4216 bp->b_flags |= B_VMIO;
4217 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4218 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4219 bp, vp->v_object, bp->b_bufobj->bo_object));
4221 bp->b_flags &= ~B_VMIO;
4222 KASSERT(bp->b_bufobj->bo_object == NULL,
4223 ("ARGH! has b_bufobj->bo_object %p %p\n",
4224 bp, bp->b_bufobj->bo_object));
4225 BUF_CHECK_MAPPED(bp);
4229 bufspace_release(bufdomain(bp), maxsize);
4230 bp->b_flags &= ~B_DONE;
4232 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4234 buf_track(bp, __func__);
4235 KASSERT(bp->b_bufobj == bo,
4236 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4242 * Get an empty, disassociated buffer of given size. The buffer is initially
4246 geteblk(int size, int flags)
4251 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4252 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4253 if ((flags & GB_NOWAIT_BD) &&
4254 (curthread->td_pflags & TDP_BUFNEED) != 0)
4258 bufspace_release(bufdomain(bp), maxsize);
4259 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4264 * Truncate the backing store for a non-vmio buffer.
4267 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4270 if (bp->b_flags & B_MALLOC) {
4272 * malloced buffers are not shrunk
4274 if (newbsize == 0) {
4275 bufmallocadjust(bp, 0);
4276 free(bp->b_data, M_BIOBUF);
4277 bp->b_data = bp->b_kvabase;
4278 bp->b_flags &= ~B_MALLOC;
4282 vm_hold_free_pages(bp, newbsize);
4283 bufspace_adjust(bp, newbsize);
4287 * Extend the backing for a non-VMIO buffer.
4290 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4296 * We only use malloced memory on the first allocation.
4297 * and revert to page-allocated memory when the buffer
4300 * There is a potential smp race here that could lead
4301 * to bufmallocspace slightly passing the max. It
4302 * is probably extremely rare and not worth worrying
4305 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4306 bufmallocspace < maxbufmallocspace) {
4307 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4308 bp->b_flags |= B_MALLOC;
4309 bufmallocadjust(bp, newbsize);
4314 * If the buffer is growing on its other-than-first
4315 * allocation then we revert to the page-allocation
4320 if (bp->b_flags & B_MALLOC) {
4321 origbuf = bp->b_data;
4322 origbufsize = bp->b_bufsize;
4323 bp->b_data = bp->b_kvabase;
4324 bufmallocadjust(bp, 0);
4325 bp->b_flags &= ~B_MALLOC;
4326 newbsize = round_page(newbsize);
4328 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4329 (vm_offset_t) bp->b_data + newbsize);
4330 if (origbuf != NULL) {
4331 bcopy(origbuf, bp->b_data, origbufsize);
4332 free(origbuf, M_BIOBUF);
4334 bufspace_adjust(bp, newbsize);
4338 * This code constitutes the buffer memory from either anonymous system
4339 * memory (in the case of non-VMIO operations) or from an associated
4340 * VM object (in the case of VMIO operations). This code is able to
4341 * resize a buffer up or down.
4343 * Note that this code is tricky, and has many complications to resolve
4344 * deadlock or inconsistent data situations. Tread lightly!!!
4345 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4346 * the caller. Calling this code willy nilly can result in the loss of data.
4348 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4349 * B_CACHE for the non-VMIO case.
4352 allocbuf(struct buf *bp, int size)
4356 if (bp->b_bcount == size)
4359 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size,
4360 ("allocbuf: buffer too small %p %#x %#x",
4361 bp, bp->b_kvasize, size));
4363 newbsize = roundup2(size, DEV_BSIZE);
4364 if ((bp->b_flags & B_VMIO) == 0) {
4365 if ((bp->b_flags & B_MALLOC) == 0)
4366 newbsize = round_page(newbsize);
4368 * Just get anonymous memory from the kernel. Don't
4369 * mess with B_CACHE.
4371 if (newbsize < bp->b_bufsize)
4372 vfs_nonvmio_truncate(bp, newbsize);
4373 else if (newbsize > bp->b_bufsize)
4374 vfs_nonvmio_extend(bp, newbsize);
4378 desiredpages = size == 0 ? 0 :
4379 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4381 KASSERT((bp->b_flags & B_MALLOC) == 0,
4382 ("allocbuf: VMIO buffer can't be malloced %p", bp));
4385 * Set B_CACHE initially if buffer is 0 length or will become
4388 if (size == 0 || bp->b_bufsize == 0)
4389 bp->b_flags |= B_CACHE;
4391 if (newbsize < bp->b_bufsize)
4392 vfs_vmio_truncate(bp, desiredpages);
4393 /* XXX This looks as if it should be newbsize > b_bufsize */
4394 else if (size > bp->b_bcount)
4395 vfs_vmio_extend(bp, desiredpages, size);
4396 bufspace_adjust(bp, newbsize);
4398 bp->b_bcount = size; /* requested buffer size. */
4402 extern int inflight_transient_maps;
4404 static struct bio_queue nondump_bios;
4407 biodone(struct bio *bp)
4410 void (*done)(struct bio *);
4411 vm_offset_t start, end;
4413 biotrack(bp, __func__);
4416 * Avoid completing I/O when dumping after a panic since that may
4417 * result in a deadlock in the filesystem or pager code. Note that
4418 * this doesn't affect dumps that were started manually since we aim
4419 * to keep the system usable after it has been resumed.
4421 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4422 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4425 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4426 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4427 bp->bio_flags |= BIO_UNMAPPED;
4428 start = trunc_page((vm_offset_t)bp->bio_data);
4429 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4430 bp->bio_data = unmapped_buf;
4431 pmap_qremove(start, atop(end - start));
4432 vmem_free(transient_arena, start, end - start);
4433 atomic_add_int(&inflight_transient_maps, -1);
4435 done = bp->bio_done;
4437 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4439 bp->bio_flags |= BIO_DONE;
4447 * Wait for a BIO to finish.
4450 biowait(struct bio *bp, const char *wmesg)
4454 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4456 while ((bp->bio_flags & BIO_DONE) == 0)
4457 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4459 if (bp->bio_error != 0)
4460 return (bp->bio_error);
4461 if (!(bp->bio_flags & BIO_ERROR))
4467 biofinish(struct bio *bp, struct devstat *stat, int error)
4471 bp->bio_error = error;
4472 bp->bio_flags |= BIO_ERROR;
4475 devstat_end_transaction_bio(stat, bp);
4479 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4481 biotrack_buf(struct bio *bp, const char *location)
4484 buf_track(bp->bio_track_bp, location);
4491 * Wait for buffer I/O completion, returning error status. The buffer
4492 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4493 * error and cleared.
4496 bufwait(struct buf *bp)
4498 if (bp->b_iocmd == BIO_READ)
4499 bwait(bp, PRIBIO, "biord");
4501 bwait(bp, PRIBIO, "biowr");
4502 if (bp->b_flags & B_EINTR) {
4503 bp->b_flags &= ~B_EINTR;
4506 if (bp->b_ioflags & BIO_ERROR) {
4507 return (bp->b_error ? bp->b_error : EIO);
4516 * Finish I/O on a buffer, optionally calling a completion function.
4517 * This is usually called from an interrupt so process blocking is
4520 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4521 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4522 * assuming B_INVAL is clear.
4524 * For the VMIO case, we set B_CACHE if the op was a read and no
4525 * read error occurred, or if the op was a write. B_CACHE is never
4526 * set if the buffer is invalid or otherwise uncacheable.
4528 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4529 * initiator to leave B_INVAL set to brelse the buffer out of existence
4530 * in the biodone routine.
4533 bufdone(struct buf *bp)
4535 struct bufobj *dropobj;
4536 void (*biodone)(struct buf *);
4538 buf_track(bp, __func__);
4539 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4542 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4544 runningbufwakeup(bp);
4545 if (bp->b_iocmd == BIO_WRITE)
4546 dropobj = bp->b_bufobj;
4547 /* call optional completion function if requested */
4548 if (bp->b_iodone != NULL) {
4549 biodone = bp->b_iodone;
4550 bp->b_iodone = NULL;
4553 bufobj_wdrop(dropobj);
4556 if (bp->b_flags & B_VMIO) {
4558 * Set B_CACHE if the op was a normal read and no error
4559 * occurred. B_CACHE is set for writes in the b*write()
4562 if (bp->b_iocmd == BIO_READ &&
4563 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4564 !(bp->b_ioflags & BIO_ERROR))
4565 bp->b_flags |= B_CACHE;
4566 vfs_vmio_iodone(bp);
4568 if (!LIST_EMPTY(&bp->b_dep))
4570 if ((bp->b_flags & B_CKHASH) != 0) {
4571 KASSERT(bp->b_iocmd == BIO_READ,
4572 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4573 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4574 (*bp->b_ckhashcalc)(bp);
4577 * For asynchronous completions, release the buffer now. The brelse
4578 * will do a wakeup there if necessary - so no need to do a wakeup
4579 * here in the async case. The sync case always needs to do a wakeup.
4581 if (bp->b_flags & B_ASYNC) {
4582 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4583 (bp->b_ioflags & BIO_ERROR))
4590 bufobj_wdrop(dropobj);
4594 * This routine is called in lieu of iodone in the case of
4595 * incomplete I/O. This keeps the busy status for pages
4599 vfs_unbusy_pages(struct buf *bp)
4605 runningbufwakeup(bp);
4606 if (!(bp->b_flags & B_VMIO))
4609 obj = bp->b_bufobj->bo_object;
4610 for (i = 0; i < bp->b_npages; i++) {
4612 if (m == bogus_page) {
4613 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4615 panic("vfs_unbusy_pages: page missing\n");
4617 if (buf_mapped(bp)) {
4618 BUF_CHECK_MAPPED(bp);
4619 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4620 bp->b_pages, bp->b_npages);
4622 BUF_CHECK_UNMAPPED(bp);
4626 vm_object_pip_wakeupn(obj, bp->b_npages);
4630 * vfs_page_set_valid:
4632 * Set the valid bits in a page based on the supplied offset. The
4633 * range is restricted to the buffer's size.
4635 * This routine is typically called after a read completes.
4638 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4643 * Compute the end offset, eoff, such that [off, eoff) does not span a
4644 * page boundary and eoff is not greater than the end of the buffer.
4645 * The end of the buffer, in this case, is our file EOF, not the
4646 * allocation size of the buffer.
4648 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4649 if (eoff > bp->b_offset + bp->b_bcount)
4650 eoff = bp->b_offset + bp->b_bcount;
4653 * Set valid range. This is typically the entire buffer and thus the
4657 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4661 * vfs_page_set_validclean:
4663 * Set the valid bits and clear the dirty bits in a page based on the
4664 * supplied offset. The range is restricted to the buffer's size.
4667 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4669 vm_ooffset_t soff, eoff;
4672 * Start and end offsets in buffer. eoff - soff may not cross a
4673 * page boundary or cross the end of the buffer. The end of the
4674 * buffer, in this case, is our file EOF, not the allocation size
4678 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4679 if (eoff > bp->b_offset + bp->b_bcount)
4680 eoff = bp->b_offset + bp->b_bcount;
4683 * Set valid range. This is typically the entire buffer and thus the
4687 vm_page_set_validclean(
4689 (vm_offset_t) (soff & PAGE_MASK),
4690 (vm_offset_t) (eoff - soff)
4696 * Acquire a shared busy on all pages in the buf.
4699 vfs_busy_pages_acquire(struct buf *bp)
4703 for (i = 0; i < bp->b_npages; i++)
4704 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4708 vfs_busy_pages_release(struct buf *bp)
4712 for (i = 0; i < bp->b_npages; i++)
4713 vm_page_sunbusy(bp->b_pages[i]);
4717 * This routine is called before a device strategy routine.
4718 * It is used to tell the VM system that paging I/O is in
4719 * progress, and treat the pages associated with the buffer
4720 * almost as being exclusive busy. Also the object paging_in_progress
4721 * flag is handled to make sure that the object doesn't become
4724 * Since I/O has not been initiated yet, certain buffer flags
4725 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4726 * and should be ignored.
4729 vfs_busy_pages(struct buf *bp, int clear_modify)
4737 if (!(bp->b_flags & B_VMIO))
4740 obj = bp->b_bufobj->bo_object;
4741 foff = bp->b_offset;
4742 KASSERT(bp->b_offset != NOOFFSET,
4743 ("vfs_busy_pages: no buffer offset"));
4744 if ((bp->b_flags & B_CLUSTER) == 0) {
4745 vm_object_pip_add(obj, bp->b_npages);
4746 vfs_busy_pages_acquire(bp);
4748 if (bp->b_bufsize != 0)
4749 vfs_setdirty_range(bp);
4751 for (i = 0; i < bp->b_npages; i++) {
4753 vm_page_assert_sbusied(m);
4756 * When readying a buffer for a read ( i.e
4757 * clear_modify == 0 ), it is important to do
4758 * bogus_page replacement for valid pages in
4759 * partially instantiated buffers. Partially
4760 * instantiated buffers can, in turn, occur when
4761 * reconstituting a buffer from its VM backing store
4762 * base. We only have to do this if B_CACHE is
4763 * clear ( which causes the I/O to occur in the
4764 * first place ). The replacement prevents the read
4765 * I/O from overwriting potentially dirty VM-backed
4766 * pages. XXX bogus page replacement is, uh, bogus.
4767 * It may not work properly with small-block devices.
4768 * We need to find a better way.
4771 pmap_remove_write(m);
4772 vfs_page_set_validclean(bp, foff, m);
4773 } else if (vm_page_all_valid(m) &&
4774 (bp->b_flags & B_CACHE) == 0) {
4775 bp->b_pages[i] = bogus_page;
4778 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4780 if (bogus && buf_mapped(bp)) {
4781 BUF_CHECK_MAPPED(bp);
4782 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4783 bp->b_pages, bp->b_npages);
4788 * vfs_bio_set_valid:
4790 * Set the range within the buffer to valid. The range is
4791 * relative to the beginning of the buffer, b_offset. Note that
4792 * b_offset itself may be offset from the beginning of the first
4796 vfs_bio_set_valid(struct buf *bp, int base, int size)
4801 if (!(bp->b_flags & B_VMIO))
4805 * Fixup base to be relative to beginning of first page.
4806 * Set initial n to be the maximum number of bytes in the
4807 * first page that can be validated.
4809 base += (bp->b_offset & PAGE_MASK);
4810 n = PAGE_SIZE - (base & PAGE_MASK);
4813 * Busy may not be strictly necessary here because the pages are
4814 * unlikely to be fully valid and the vnode lock will synchronize
4815 * their access via getpages. It is grabbed for consistency with
4816 * other page validation.
4818 vfs_busy_pages_acquire(bp);
4819 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4823 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4828 vfs_busy_pages_release(bp);
4834 * If the specified buffer is a non-VMIO buffer, clear the entire
4835 * buffer. If the specified buffer is a VMIO buffer, clear and
4836 * validate only the previously invalid portions of the buffer.
4837 * This routine essentially fakes an I/O, so we need to clear
4838 * BIO_ERROR and B_INVAL.
4840 * Note that while we only theoretically need to clear through b_bcount,
4841 * we go ahead and clear through b_bufsize.
4844 vfs_bio_clrbuf(struct buf *bp)
4846 int i, j, mask, sa, ea, slide;
4848 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4852 bp->b_flags &= ~B_INVAL;
4853 bp->b_ioflags &= ~BIO_ERROR;
4854 vfs_busy_pages_acquire(bp);
4855 sa = bp->b_offset & PAGE_MASK;
4857 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4858 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4859 ea = slide & PAGE_MASK;
4862 if (bp->b_pages[i] == bogus_page)
4865 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4866 if ((bp->b_pages[i]->valid & mask) == mask)
4868 if ((bp->b_pages[i]->valid & mask) == 0)
4869 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4871 for (; sa < ea; sa += DEV_BSIZE, j++) {
4872 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4873 pmap_zero_page_area(bp->b_pages[i],
4878 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4879 roundup2(ea - sa, DEV_BSIZE));
4881 vfs_busy_pages_release(bp);
4886 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4891 if (buf_mapped(bp)) {
4892 BUF_CHECK_MAPPED(bp);
4893 bzero(bp->b_data + base, size);
4895 BUF_CHECK_UNMAPPED(bp);
4896 n = PAGE_SIZE - (base & PAGE_MASK);
4897 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4901 pmap_zero_page_area(m, base & PAGE_MASK, n);
4910 * Update buffer flags based on I/O request parameters, optionally releasing the
4911 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4912 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4913 * I/O). Otherwise the buffer is released to the cache.
4916 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4919 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4920 ("buf %p non-VMIO noreuse", bp));
4922 if ((ioflag & IO_DIRECT) != 0)
4923 bp->b_flags |= B_DIRECT;
4924 if ((ioflag & IO_EXT) != 0)
4925 bp->b_xflags |= BX_ALTDATA;
4926 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4927 bp->b_flags |= B_RELBUF;
4928 if ((ioflag & IO_NOREUSE) != 0)
4929 bp->b_flags |= B_NOREUSE;
4937 vfs_bio_brelse(struct buf *bp, int ioflag)
4940 b_io_dismiss(bp, ioflag, true);
4944 vfs_bio_set_flags(struct buf *bp, int ioflag)
4947 b_io_dismiss(bp, ioflag, false);
4951 * vm_hold_load_pages and vm_hold_free_pages get pages into
4952 * a buffers address space. The pages are anonymous and are
4953 * not associated with a file object.
4956 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4962 BUF_CHECK_MAPPED(bp);
4964 to = round_page(to);
4965 from = round_page(from);
4966 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4967 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4968 KASSERT(to - from <= maxbcachebuf,
4969 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4970 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4972 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4974 * note: must allocate system pages since blocking here
4975 * could interfere with paging I/O, no matter which
4978 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4979 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4980 pmap_qenter(pg, &p, 1);
4981 bp->b_pages[index] = p;
4983 bp->b_npages = index;
4986 /* Return pages associated with this buf to the vm system */
4988 vm_hold_free_pages(struct buf *bp, int newbsize)
4992 int index, newnpages;
4994 BUF_CHECK_MAPPED(bp);
4996 from = round_page((vm_offset_t)bp->b_data + newbsize);
4997 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4998 if (bp->b_npages > newnpages)
4999 pmap_qremove(from, bp->b_npages - newnpages);
5000 for (index = newnpages; index < bp->b_npages; index++) {
5001 p = bp->b_pages[index];
5002 bp->b_pages[index] = NULL;
5003 vm_page_unwire_noq(p);
5006 bp->b_npages = newnpages;
5010 * Map an IO request into kernel virtual address space.
5012 * All requests are (re)mapped into kernel VA space.
5013 * Notice that we use b_bufsize for the size of the buffer
5014 * to be mapped. b_bcount might be modified by the driver.
5016 * Note that even if the caller determines that the address space should
5017 * be valid, a race or a smaller-file mapped into a larger space may
5018 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5019 * check the return value.
5021 * This function only works with pager buffers.
5024 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5029 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5030 prot = VM_PROT_READ;
5031 if (bp->b_iocmd == BIO_READ)
5032 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5033 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5034 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5037 bp->b_bufsize = len;
5038 bp->b_npages = pidx;
5039 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5040 if (mapbuf || !unmapped_buf_allowed) {
5041 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5042 bp->b_data = bp->b_kvabase + bp->b_offset;
5044 bp->b_data = unmapped_buf;
5049 * Free the io map PTEs associated with this IO operation.
5050 * We also invalidate the TLB entries and restore the original b_addr.
5052 * This function only works with pager buffers.
5055 vunmapbuf(struct buf *bp)
5059 npages = bp->b_npages;
5061 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5062 vm_page_unhold_pages(bp->b_pages, npages);
5064 bp->b_data = unmapped_buf;
5068 bdone(struct buf *bp)
5072 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5074 bp->b_flags |= B_DONE;
5080 bwait(struct buf *bp, u_char pri, const char *wchan)
5084 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5086 while ((bp->b_flags & B_DONE) == 0)
5087 msleep(bp, mtxp, pri, wchan, 0);
5092 bufsync(struct bufobj *bo, int waitfor)
5095 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5099 bufstrategy(struct bufobj *bo, struct buf *bp)
5105 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5106 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5107 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5108 i = VOP_STRATEGY(vp, bp);
5109 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5113 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5116 bufobj_init(struct bufobj *bo, void *private)
5118 static volatile int bufobj_cleanq;
5121 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5122 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5123 bo->bo_private = private;
5124 TAILQ_INIT(&bo->bo_clean.bv_hd);
5125 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5129 bufobj_wrefl(struct bufobj *bo)
5132 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5133 ASSERT_BO_WLOCKED(bo);
5138 bufobj_wref(struct bufobj *bo)
5141 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5148 bufobj_wdrop(struct bufobj *bo)
5151 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5153 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5154 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5155 bo->bo_flag &= ~BO_WWAIT;
5156 wakeup(&bo->bo_numoutput);
5162 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5166 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5167 ASSERT_BO_WLOCKED(bo);
5169 while (bo->bo_numoutput) {
5170 bo->bo_flag |= BO_WWAIT;
5171 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5172 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5180 * Set bio_data or bio_ma for struct bio from the struct buf.
5183 bdata2bio(struct buf *bp, struct bio *bip)
5186 if (!buf_mapped(bp)) {
5187 KASSERT(unmapped_buf_allowed, ("unmapped"));
5188 bip->bio_ma = bp->b_pages;
5189 bip->bio_ma_n = bp->b_npages;
5190 bip->bio_data = unmapped_buf;
5191 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5192 bip->bio_flags |= BIO_UNMAPPED;
5193 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5194 PAGE_SIZE == bp->b_npages,
5195 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5196 (long long)bip->bio_length, bip->bio_ma_n));
5198 bip->bio_data = bp->b_data;
5204 * The MIPS pmap code currently doesn't handle aliased pages.
5205 * The VIPT caches may not handle page aliasing themselves, leading
5206 * to data corruption.
5208 * As such, this code makes a system extremely unhappy if said
5209 * system doesn't support unaliasing the above situation in hardware.
5210 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5211 * this feature at build time, so it has to be handled in software.
5213 * Once the MIPS pmap/cache code grows to support this function on
5214 * earlier chips, it should be flipped back off.
5217 static int buf_pager_relbuf = 1;
5219 static int buf_pager_relbuf = 0;
5221 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5222 &buf_pager_relbuf, 0,
5223 "Make buffer pager release buffers after reading");
5226 * The buffer pager. It uses buffer reads to validate pages.
5228 * In contrast to the generic local pager from vm/vnode_pager.c, this
5229 * pager correctly and easily handles volumes where the underlying
5230 * device block size is greater than the machine page size. The
5231 * buffer cache transparently extends the requested page run to be
5232 * aligned at the block boundary, and does the necessary bogus page
5233 * replacements in the addends to avoid obliterating already valid
5236 * The only non-trivial issue is that the exclusive busy state for
5237 * pages, which is assumed by the vm_pager_getpages() interface, is
5238 * incompatible with the VMIO buffer cache's desire to share-busy the
5239 * pages. This function performs a trivial downgrade of the pages'
5240 * state before reading buffers, and a less trivial upgrade from the
5241 * shared-busy to excl-busy state after the read.
5244 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5245 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5246 vbg_get_blksize_t get_blksize)
5253 vm_ooffset_t la, lb, poff, poffe;
5255 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5258 object = vp->v_object;
5261 la = IDX_TO_OFF(ma[count - 1]->pindex);
5262 if (la >= object->un_pager.vnp.vnp_size)
5263 return (VM_PAGER_BAD);
5266 * Change the meaning of la from where the last requested page starts
5267 * to where it ends, because that's the end of the requested region
5268 * and the start of the potential read-ahead region.
5271 lpart = la > object->un_pager.vnp.vnp_size;
5272 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5275 return (VM_PAGER_ERROR);
5278 * Calculate read-ahead, behind and total pages.
5281 lb = IDX_TO_OFF(ma[0]->pindex);
5282 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5284 if (rbehind != NULL)
5286 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5287 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5288 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5293 VM_CNT_INC(v_vnodein);
5294 VM_CNT_ADD(v_vnodepgsin, pgsin);
5296 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5297 != 0) ? GB_UNMAPPED : 0;
5299 for (i = 0; i < count; i++) {
5300 if (ma[i] != bogus_page)
5301 vm_page_busy_downgrade(ma[i]);
5305 for (i = 0; i < count; i++) {
5307 if (m == bogus_page)
5311 * Pages are shared busy and the object lock is not
5312 * owned, which together allow for the pages'
5313 * invalidation. The racy test for validity avoids
5314 * useless creation of the buffer for the most typical
5315 * case when invalidation is not used in redo or for
5316 * parallel read. The shared->excl upgrade loop at
5317 * the end of the function catches the race in a
5318 * reliable way (protected by the object lock).
5320 if (vm_page_all_valid(m))
5323 poff = IDX_TO_OFF(m->pindex);
5324 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5325 for (; poff < poffe; poff += bsize) {
5326 lbn = get_lblkno(vp, poff);
5331 error = get_blksize(vp, lbn, &bsize);
5333 error = bread_gb(vp, lbn, bsize,
5334 curthread->td_ucred, br_flags, &bp);
5337 if (bp->b_rcred == curthread->td_ucred) {
5338 crfree(bp->b_rcred);
5339 bp->b_rcred = NOCRED;
5341 if (LIST_EMPTY(&bp->b_dep)) {
5343 * Invalidation clears m->valid, but
5344 * may leave B_CACHE flag if the
5345 * buffer existed at the invalidation
5346 * time. In this case, recycle the
5347 * buffer to do real read on next
5348 * bread() after redo.
5350 * Otherwise B_RELBUF is not strictly
5351 * necessary, enable to reduce buf
5354 if (buf_pager_relbuf ||
5355 !vm_page_all_valid(m))
5356 bp->b_flags |= B_RELBUF;
5358 bp->b_flags &= ~B_NOCACHE;
5364 KASSERT(1 /* racy, enable for debugging */ ||
5365 vm_page_all_valid(m) || i == count - 1,
5366 ("buf %d %p invalid", i, m));
5367 if (i == count - 1 && lpart) {
5368 if (!vm_page_none_valid(m) &&
5369 !vm_page_all_valid(m))
5370 vm_page_zero_invalid(m, TRUE);
5377 for (i = 0; i < count; i++) {
5378 if (ma[i] == bogus_page)
5380 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5381 vm_page_sunbusy(ma[i]);
5382 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5387 * Since the pages were only sbusy while neither the
5388 * buffer nor the object lock was held by us, or
5389 * reallocated while vm_page_grab() slept for busy
5390 * relinguish, they could have been invalidated.
5391 * Recheck the valid bits and re-read as needed.
5393 * Note that the last page is made fully valid in the
5394 * read loop, and partial validity for the page at
5395 * index count - 1 could mean that the page was
5396 * invalidated or removed, so we must restart for
5399 if (!vm_page_all_valid(ma[i]))
5402 if (redo && error == 0)
5404 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5407 #include "opt_ddb.h"
5409 #include <ddb/ddb.h>
5411 /* DDB command to show buffer data */
5412 DB_SHOW_COMMAND(buffer, db_show_buffer)
5415 struct buf *bp = (struct buf *)addr;
5416 #ifdef FULL_BUF_TRACKING
5421 db_printf("usage: show buffer <addr>\n");
5425 db_printf("buf at %p\n", bp);
5426 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5427 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5428 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5429 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5430 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5431 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5433 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5434 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5435 "b_vp = %p, b_dep = %p\n",
5436 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5437 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5438 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5439 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5440 bp->b_kvabase, bp->b_kvasize);
5443 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5444 for (i = 0; i < bp->b_npages; i++) {
5448 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5450 (u_long)VM_PAGE_TO_PHYS(m));
5452 db_printf("( ??? )");
5453 if ((i + 1) < bp->b_npages)
5458 BUF_LOCKPRINTINFO(bp);
5459 #if defined(FULL_BUF_TRACKING)
5460 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5462 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5463 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5464 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5466 db_printf(" %2u: %s\n", j,
5467 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5469 #elif defined(BUF_TRACKING)
5470 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5475 DB_SHOW_COMMAND(bufqueues, bufqueues)
5477 struct bufdomain *bd;
5482 db_printf("bqempty: %d\n", bqempty.bq_len);
5484 for (i = 0; i < buf_domains; i++) {
5486 db_printf("Buf domain %d\n", i);
5487 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5488 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5489 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5491 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5492 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5493 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5494 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5495 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5497 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5498 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5499 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5500 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5503 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5504 total += bp->b_bufsize;
5505 db_printf("\tcleanq count\t%d (%ld)\n",
5506 bd->bd_cleanq->bq_len, total);
5508 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5509 total += bp->b_bufsize;
5510 db_printf("\tdirtyq count\t%d (%ld)\n",
5511 bd->bd_dirtyq.bq_len, total);
5512 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5513 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5514 db_printf("\tCPU ");
5515 for (j = 0; j <= mp_maxid; j++)
5516 db_printf("%d, ", bd->bd_subq[j].bq_len);
5520 for (j = 0; j < nbuf; j++) {
5522 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5524 total += bp->b_bufsize;
5527 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5530 for (j = 0; j < nbuf; j++) {
5532 if (bp->b_domain == i) {
5534 total += bp->b_bufsize;
5537 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5541 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5546 for (i = 0; i < nbuf; i++) {
5548 if (BUF_ISLOCKED(bp)) {
5549 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5557 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5563 db_printf("usage: show vnodebufs <addr>\n");
5566 vp = (struct vnode *)addr;
5567 db_printf("Clean buffers:\n");
5568 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5569 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5572 db_printf("Dirty buffers:\n");
5573 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5574 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5579 DB_COMMAND(countfreebufs, db_coundfreebufs)
5582 int i, used = 0, nfree = 0;
5585 db_printf("usage: countfreebufs\n");
5589 for (i = 0; i < nbuf; i++) {
5591 if (bp->b_qindex == QUEUE_EMPTY)
5597 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5599 db_printf("numfreebuffers is %d\n", numfreebuffers);