2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2004 Poul-Henning Kamp
5 * Copyright (c) 1994,1997 John S. Dyson
6 * Copyright (c) 2013 The FreeBSD Foundation
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * this file contains a new buffer I/O scheme implementing a coherent
36 * VM object and buffer cache scheme. Pains have been taken to make
37 * sure that the performance degradation associated with schemes such
38 * as this is not realized.
40 * Author: John S. Dyson
41 * Significant help during the development and debugging phases
42 * had been provided by David Greenman, also of the FreeBSD core team.
44 * see man buf(9) for more info.
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/bitset.h>
55 #include <sys/boottrace.h>
58 #include <sys/counter.h>
59 #include <sys/devicestat.h>
60 #include <sys/eventhandler.h>
63 #include <sys/limits.h>
65 #include <sys/malloc.h>
66 #include <sys/mount.h>
67 #include <sys/mutex.h>
68 #include <sys/kernel.h>
69 #include <sys/kthread.h>
71 #include <sys/racct.h>
72 #include <sys/refcount.h>
73 #include <sys/resourcevar.h>
74 #include <sys/rwlock.h>
75 #include <sys/sched.h>
77 #include <sys/sysctl.h>
78 #include <sys/syscallsubr.h>
80 #include <sys/vmmeter.h>
81 #include <sys/vnode.h>
82 #include <sys/watchdog.h>
83 #include <geom/geom.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_kern.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_extern.h>
92 #include <vm/vm_map.h>
93 #include <vm/swap_pager.h>
95 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
97 struct bio_ops bioops; /* I/O operation notification */
99 struct buf_ops buf_ops_bio = {
100 .bop_name = "buf_ops_bio",
101 .bop_write = bufwrite,
102 .bop_strategy = bufstrategy,
104 .bop_bdflush = bufbdflush,
108 struct mtx_padalign bq_lock;
109 TAILQ_HEAD(, buf) bq_queue;
111 uint16_t bq_subqueue;
113 } __aligned(CACHE_LINE_SIZE);
115 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
116 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
117 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
118 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
121 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
122 struct bufqueue bd_dirtyq;
123 struct bufqueue *bd_cleanq;
124 struct mtx_padalign bd_run_lock;
129 long bd_bufspacethresh;
130 int bd_hifreebuffers;
131 int bd_lofreebuffers;
132 int bd_hidirtybuffers;
133 int bd_lodirtybuffers;
134 int bd_dirtybufthresh;
139 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
140 int __aligned(CACHE_LINE_SIZE) bd_running;
141 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
142 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
143 } __aligned(CACHE_LINE_SIZE);
145 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
146 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
147 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
148 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
149 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
150 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
151 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
152 #define BD_DOMAIN(bd) (bd - bdomain)
154 static char *buf; /* buffer header pool */
158 return ((struct buf *)(buf + (sizeof(struct buf) +
159 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
162 caddr_t __read_mostly unmapped_buf;
164 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
165 struct proc *bufdaemonproc;
167 static void vm_hold_free_pages(struct buf *bp, int newbsize);
168 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
170 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
171 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
173 static void vfs_clean_pages_dirty_buf(struct buf *bp);
174 static void vfs_setdirty_range(struct buf *bp);
175 static void vfs_vmio_invalidate(struct buf *bp);
176 static void vfs_vmio_truncate(struct buf *bp, int npages);
177 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
178 static int vfs_bio_clcheck(struct vnode *vp, int size,
179 daddr_t lblkno, daddr_t blkno);
180 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
181 void (*)(struct buf *));
182 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
183 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
184 static void buf_daemon(void);
185 static __inline void bd_wakeup(void);
186 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
187 static void bufkva_reclaim(vmem_t *, int);
188 static void bufkva_free(struct buf *);
189 static int buf_import(void *, void **, int, int, int);
190 static void buf_release(void *, void **, int);
191 static void maxbcachebuf_adjust(void);
192 static inline struct bufdomain *bufdomain(struct buf *);
193 static void bq_remove(struct bufqueue *bq, struct buf *bp);
194 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
195 static int buf_recycle(struct bufdomain *, bool kva);
196 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
197 const char *lockname);
198 static void bd_init(struct bufdomain *bd);
199 static int bd_flushall(struct bufdomain *bd);
200 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
201 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
203 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
204 int vmiodirenable = TRUE;
205 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
206 "Use the VM system for directory writes");
207 long runningbufspace;
208 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
209 "Amount of presently outstanding async buffer io");
210 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
211 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
212 static counter_u64_t bufkvaspace;
213 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
214 "Kernel virtual memory used for buffers");
215 static long maxbufspace;
216 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
217 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
218 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
219 "Maximum allowed value of bufspace (including metadata)");
220 static long bufmallocspace;
221 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
222 "Amount of malloced memory for buffers");
223 static long maxbufmallocspace;
224 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
225 0, "Maximum amount of malloced memory for buffers");
226 static long lobufspace;
227 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
228 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
229 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
230 "Minimum amount of buffers we want to have");
232 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
233 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
234 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
235 "Maximum allowed value of bufspace (excluding metadata)");
237 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
238 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
239 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
240 "Bufspace consumed before waking the daemon to free some");
241 static counter_u64_t buffreekvacnt;
242 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
243 "Number of times we have freed the KVA space from some buffer");
244 static counter_u64_t bufdefragcnt;
245 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
246 "Number of times we have had to repeat buffer allocation to defragment");
247 static long lorunningspace;
248 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
249 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
250 "Minimum preferred space used for in-progress I/O");
251 static long hirunningspace;
252 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
253 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
254 "Maximum amount of space to use for in-progress I/O");
255 int dirtybufferflushes;
256 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
257 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
259 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
260 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
261 int altbufferflushes;
262 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
263 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
264 static int recursiveflushes;
265 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
266 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
267 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
268 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
269 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
270 "Number of buffers that are dirty (has unwritten changes) at the moment");
271 static int lodirtybuffers;
272 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
273 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
274 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
275 "How many buffers we want to have free before bufdaemon can sleep");
276 static int hidirtybuffers;
277 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
278 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
279 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
280 "When the number of dirty buffers is considered severe");
282 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
283 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
284 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
285 "Number of bdwrite to bawrite conversions to clear dirty buffers");
286 static int numfreebuffers;
287 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
288 "Number of free buffers");
289 static int lofreebuffers;
290 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
291 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
292 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
293 "Target number of free buffers");
294 static int hifreebuffers;
295 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
296 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
297 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
298 "Threshold for clean buffer recycling");
299 static counter_u64_t getnewbufcalls;
300 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
301 &getnewbufcalls, "Number of calls to getnewbuf");
302 static counter_u64_t getnewbufrestarts;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
305 "Number of times getnewbuf has had to restart a buffer acquisition");
306 static counter_u64_t mappingrestarts;
307 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
309 "Number of times getblk has had to restart a buffer mapping for "
311 static counter_u64_t numbufallocfails;
312 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
313 &numbufallocfails, "Number of times buffer allocations failed");
314 static int flushbufqtarget = 100;
315 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
316 "Amount of work to do in flushbufqueues when helping bufdaemon");
317 static counter_u64_t notbufdflushes;
318 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
319 "Number of dirty buffer flushes done by the bufdaemon helpers");
320 static long barrierwrites;
321 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
322 &barrierwrites, 0, "Number of barrier writes");
323 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
324 &unmapped_buf_allowed, 0,
325 "Permit the use of the unmapped i/o");
326 int maxbcachebuf = MAXBCACHEBUF;
327 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
328 "Maximum size of a buffer cache block");
331 * This lock synchronizes access to bd_request.
333 static struct mtx_padalign __exclusive_cache_line bdlock;
336 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
337 * waitrunningbufspace().
339 static struct mtx_padalign __exclusive_cache_line rbreqlock;
342 * Lock that protects bdirtywait.
344 static struct mtx_padalign __exclusive_cache_line bdirtylock;
347 * bufdaemon shutdown request and sleep channel.
349 static bool bd_shutdown;
352 * Wakeup point for bufdaemon, as well as indicator of whether it is already
353 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
356 static int bd_request;
359 * Request for the buf daemon to write more buffers than is indicated by
360 * lodirtybuf. This may be necessary to push out excess dependencies or
361 * defragment the address space where a simple count of the number of dirty
362 * buffers is insufficient to characterize the demand for flushing them.
364 static int bd_speedupreq;
367 * Synchronization (sleep/wakeup) variable for active buffer space requests.
368 * Set when wait starts, cleared prior to wakeup().
369 * Used in runningbufwakeup() and waitrunningbufspace().
371 static int runningbufreq;
374 * Synchronization for bwillwrite() waiters.
376 static int bdirtywait;
379 * Definitions for the buffer free lists.
381 #define QUEUE_NONE 0 /* on no queue */
382 #define QUEUE_EMPTY 1 /* empty buffer headers */
383 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
384 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
385 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
387 /* Maximum number of buffer domains. */
388 #define BUF_DOMAINS 8
390 struct bufdomainset bdlodirty; /* Domains > lodirty */
391 struct bufdomainset bdhidirty; /* Domains > hidirty */
393 /* Configured number of clean queues. */
394 static int __read_mostly buf_domains;
396 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
397 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
398 struct bufqueue __exclusive_cache_line bqempty;
401 * per-cpu empty buffer cache.
406 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
411 value = *(long *)arg1;
412 error = sysctl_handle_long(oidp, &value, 0, req);
413 if (error != 0 || req->newptr == NULL)
415 mtx_lock(&rbreqlock);
416 if (arg1 == &hirunningspace) {
417 if (value < lorunningspace)
420 hirunningspace = value;
422 KASSERT(arg1 == &lorunningspace,
423 ("%s: unknown arg1", __func__));
424 if (value > hirunningspace)
427 lorunningspace = value;
429 mtx_unlock(&rbreqlock);
434 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
440 value = *(int *)arg1;
441 error = sysctl_handle_int(oidp, &value, 0, req);
442 if (error != 0 || req->newptr == NULL)
444 *(int *)arg1 = value;
445 for (i = 0; i < buf_domains; i++)
446 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
453 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
459 value = *(long *)arg1;
460 error = sysctl_handle_long(oidp, &value, 0, req);
461 if (error != 0 || req->newptr == NULL)
463 *(long *)arg1 = value;
464 for (i = 0; i < buf_domains; i++)
465 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
471 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
472 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
474 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
481 for (i = 0; i < buf_domains; i++)
482 lvalue += bdomain[i].bd_bufspace;
483 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
484 return (sysctl_handle_long(oidp, &lvalue, 0, req));
485 if (lvalue > INT_MAX)
486 /* On overflow, still write out a long to trigger ENOMEM. */
487 return (sysctl_handle_long(oidp, &lvalue, 0, req));
489 return (sysctl_handle_int(oidp, &ivalue, 0, req));
493 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
499 for (i = 0; i < buf_domains; i++)
500 lvalue += bdomain[i].bd_bufspace;
501 return (sysctl_handle_long(oidp, &lvalue, 0, req));
506 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
512 for (i = 0; i < buf_domains; i++)
513 value += bdomain[i].bd_numdirtybuffers;
514 return (sysctl_handle_int(oidp, &value, 0, req));
520 * Wakeup any bwillwrite() waiters.
525 mtx_lock(&bdirtylock);
530 mtx_unlock(&bdirtylock);
536 * Clear a domain from the appropriate bitsets when dirtybuffers
540 bd_clear(struct bufdomain *bd)
543 mtx_lock(&bdirtylock);
544 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
545 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
546 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
547 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
548 mtx_unlock(&bdirtylock);
554 * Set a domain in the appropriate bitsets when dirtybuffers
558 bd_set(struct bufdomain *bd)
561 mtx_lock(&bdirtylock);
562 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
563 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
564 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
565 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
566 mtx_unlock(&bdirtylock);
572 * Decrement the numdirtybuffers count by one and wakeup any
573 * threads blocked in bwillwrite().
576 bdirtysub(struct buf *bp)
578 struct bufdomain *bd;
582 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
583 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
585 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
592 * Increment the numdirtybuffers count by one and wakeup the buf
596 bdirtyadd(struct buf *bp)
598 struct bufdomain *bd;
602 * Only do the wakeup once as we cross the boundary. The
603 * buf daemon will keep running until the condition clears.
606 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
607 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
609 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
614 * bufspace_daemon_wakeup:
616 * Wakeup the daemons responsible for freeing clean bufs.
619 bufspace_daemon_wakeup(struct bufdomain *bd)
623 * avoid the lock if the daemon is running.
625 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
627 atomic_store_int(&bd->bd_running, 1);
628 wakeup(&bd->bd_running);
636 * Adjust the reported bufspace for a KVA managed buffer, possibly
637 * waking any waiters.
640 bufspace_adjust(struct buf *bp, int bufsize)
642 struct bufdomain *bd;
646 KASSERT((bp->b_flags & B_MALLOC) == 0,
647 ("bufspace_adjust: malloc buf %p", bp));
649 diff = bufsize - bp->b_bufsize;
651 atomic_subtract_long(&bd->bd_bufspace, -diff);
652 } else if (diff > 0) {
653 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
654 /* Wake up the daemon on the transition. */
655 if (space < bd->bd_bufspacethresh &&
656 space + diff >= bd->bd_bufspacethresh)
657 bufspace_daemon_wakeup(bd);
659 bp->b_bufsize = bufsize;
665 * Reserve bufspace before calling allocbuf(). metadata has a
666 * different space limit than data.
669 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
675 limit = bd->bd_maxbufspace;
677 limit = bd->bd_hibufspace;
678 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
681 atomic_subtract_long(&bd->bd_bufspace, size);
685 /* Wake up the daemon on the transition. */
686 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
687 bufspace_daemon_wakeup(bd);
695 * Release reserved bufspace after bufspace_adjust() has consumed it.
698 bufspace_release(struct bufdomain *bd, int size)
701 atomic_subtract_long(&bd->bd_bufspace, size);
707 * Wait for bufspace, acting as the buf daemon if a locked vnode is
708 * supplied. bd_wanted must be set prior to polling for space. The
709 * operation must be re-tried on return.
712 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
713 int slpflag, int slptimeo)
716 int error, fl, norunbuf;
718 if ((gbflags & GB_NOWAIT_BD) != 0)
723 while (bd->bd_wanted) {
724 if (vp != NULL && vp->v_type != VCHR &&
725 (td->td_pflags & TDP_BUFNEED) == 0) {
728 * getblk() is called with a vnode locked, and
729 * some majority of the dirty buffers may as
730 * well belong to the vnode. Flushing the
731 * buffers there would make a progress that
732 * cannot be achieved by the buf_daemon, that
733 * cannot lock the vnode.
735 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
736 (td->td_pflags & TDP_NORUNNINGBUF);
739 * Play bufdaemon. The getnewbuf() function
740 * may be called while the thread owns lock
741 * for another dirty buffer for the same
742 * vnode, which makes it impossible to use
743 * VOP_FSYNC() there, due to the buffer lock
746 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
747 fl = buf_flush(vp, bd, flushbufqtarget);
748 td->td_pflags &= norunbuf;
752 if (bd->bd_wanted == 0)
755 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
756 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
764 bufspace_daemon_shutdown(void *arg, int howto __unused)
766 struct bufdomain *bd = arg;
769 if (KERNEL_PANICKED())
773 bd->bd_shutdown = true;
774 wakeup(&bd->bd_running);
775 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
776 "bufspace_shutdown", 60 * hz);
779 printf("bufspacedaemon wait error: %d\n", error);
785 * buffer space management daemon. Tries to maintain some marginal
786 * amount of free buffer space so that requesting processes neither
787 * block nor work to reclaim buffers.
790 bufspace_daemon(void *arg)
792 struct bufdomain *bd = arg;
794 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
795 SHUTDOWN_PRI_LAST + 100);
798 while (!bd->bd_shutdown) {
802 * Free buffers from the clean queue until we meet our
805 * Theory of operation: The buffer cache is most efficient
806 * when some free buffer headers and space are always
807 * available to getnewbuf(). This daemon attempts to prevent
808 * the excessive blocking and synchronization associated
809 * with shortfall. It goes through three phases according
812 * 1) The daemon wakes up voluntarily once per-second
813 * during idle periods when the counters are below
814 * the wakeup thresholds (bufspacethresh, lofreebuffers).
816 * 2) The daemon wakes up as we cross the thresholds
817 * ahead of any potential blocking. This may bounce
818 * slightly according to the rate of consumption and
821 * 3) The daemon and consumers are starved for working
822 * clean buffers. This is the 'bufspace' sleep below
823 * which will inefficiently trade bufs with bqrelse
824 * until we return to condition 2.
826 while (bd->bd_bufspace > bd->bd_lobufspace ||
827 bd->bd_freebuffers < bd->bd_hifreebuffers) {
828 if (buf_recycle(bd, false) != 0) {
832 * Speedup dirty if we've run out of clean
833 * buffers. This is possible in particular
834 * because softdep may held many bufs locked
835 * pending writes to other bufs which are
836 * marked for delayed write, exhausting
837 * clean space until they are written.
842 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
843 PRIBIO|PDROP, "bufspace", hz/10);
851 * Re-check our limits and sleep. bd_running must be
852 * cleared prior to checking the limits to avoid missed
853 * wakeups. The waker will adjust one of bufspace or
854 * freebuffers prior to checking bd_running.
859 atomic_store_int(&bd->bd_running, 0);
860 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
861 bd->bd_freebuffers > bd->bd_lofreebuffers) {
862 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
865 /* Avoid spurious wakeups while running. */
866 atomic_store_int(&bd->bd_running, 1);
869 wakeup(&bd->bd_shutdown);
877 * Adjust the reported bufspace for a malloc managed buffer, possibly
878 * waking any waiters.
881 bufmallocadjust(struct buf *bp, int bufsize)
885 KASSERT((bp->b_flags & B_MALLOC) != 0,
886 ("bufmallocadjust: non-malloc buf %p", bp));
887 diff = bufsize - bp->b_bufsize;
889 atomic_subtract_long(&bufmallocspace, -diff);
891 atomic_add_long(&bufmallocspace, diff);
892 bp->b_bufsize = bufsize;
898 * Wake up processes that are waiting on asynchronous writes to fall
899 * below lorunningspace.
905 mtx_lock(&rbreqlock);
908 wakeup(&runningbufreq);
910 mtx_unlock(&rbreqlock);
916 * Decrement the outstanding write count according.
919 runningbufwakeup(struct buf *bp)
923 bspace = bp->b_runningbufspace;
926 space = atomic_fetchadd_long(&runningbufspace, -bspace);
927 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
929 bp->b_runningbufspace = 0;
931 * Only acquire the lock and wakeup on the transition from exceeding
932 * the threshold to falling below it.
934 if (space < lorunningspace)
936 if (space - bspace > lorunningspace)
942 * waitrunningbufspace()
944 * runningbufspace is a measure of the amount of I/O currently
945 * running. This routine is used in async-write situations to
946 * prevent creating huge backups of pending writes to a device.
947 * Only asynchronous writes are governed by this function.
949 * This does NOT turn an async write into a sync write. It waits
950 * for earlier writes to complete and generally returns before the
951 * caller's write has reached the device.
954 waitrunningbufspace(void)
957 mtx_lock(&rbreqlock);
958 while (runningbufspace > hirunningspace) {
960 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
962 mtx_unlock(&rbreqlock);
966 * vfs_buf_test_cache:
968 * Called when a buffer is extended. This function clears the B_CACHE
969 * bit if the newly extended portion of the buffer does not contain
973 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
974 vm_offset_t size, vm_page_t m)
978 * This function and its results are protected by higher level
979 * synchronization requiring vnode and buf locks to page in and
982 if (bp->b_flags & B_CACHE) {
983 int base = (foff + off) & PAGE_MASK;
984 if (vm_page_is_valid(m, base, size) == 0)
985 bp->b_flags &= ~B_CACHE;
989 /* Wake up the buffer daemon if necessary */
995 if (bd_request == 0) {
1003 * Adjust the maxbcachbuf tunable.
1006 maxbcachebuf_adjust(void)
1011 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1014 while (i * 2 <= maxbcachebuf)
1017 if (maxbcachebuf < MAXBSIZE)
1018 maxbcachebuf = MAXBSIZE;
1019 if (maxbcachebuf > maxphys)
1020 maxbcachebuf = maxphys;
1021 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1022 printf("maxbcachebuf=%d\n", maxbcachebuf);
1026 * bd_speedup - speedup the buffer cache flushing code
1035 if (bd_speedupreq == 0 || bd_request == 0)
1040 wakeup(&bd_request);
1041 mtx_unlock(&bdlock);
1045 #define TRANSIENT_DENOM 5
1047 #define TRANSIENT_DENOM 10
1051 * Calculating buffer cache scaling values and reserve space for buffer
1052 * headers. This is called during low level kernel initialization and
1053 * may be called more then once. We CANNOT write to the memory area
1054 * being reserved at this time.
1057 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1060 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1063 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for
1064 * this when sizing maps based on the amount of physical memory
1068 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1069 (KASAN_SHADOW_SCALE + 1);
1070 #elif defined(KMSAN)
1074 * KMSAN cannot reliably determine whether buffer data is initialized
1075 * unless it is updated through a KVA mapping.
1077 unmapped_buf_allowed = 0;
1081 * physmem_est is in pages. Convert it to kilobytes (assumes
1082 * PAGE_SIZE is >= 1K)
1084 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1086 maxbcachebuf_adjust();
1088 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1089 * For the first 64MB of ram nominally allocate sufficient buffers to
1090 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1091 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1092 * the buffer cache we limit the eventual kva reservation to
1095 * factor represents the 1/4 x ram conversion.
1098 int factor = 4 * BKVASIZE / 1024;
1101 if (physmem_est > 4096)
1102 nbuf += min((physmem_est - 4096) / factor,
1104 if (physmem_est > 65536)
1105 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1106 32 * 1024 * 1024 / (factor * 5));
1108 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1109 nbuf = maxbcache / BKVASIZE;
1114 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1115 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1116 if (nbuf > maxbuf) {
1118 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1124 * Ideal allocation size for the transient bio submap is 10%
1125 * of the maximal space buffer map. This roughly corresponds
1126 * to the amount of the buffer mapped for typical UFS load.
1128 * Clip the buffer map to reserve space for the transient
1129 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1130 * maximum buffer map extent on the platform.
1132 * The fall-back to the maxbuf in case of maxbcache unset,
1133 * allows to not trim the buffer KVA for the architectures
1134 * with ample KVA space.
1136 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1137 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1138 buf_sz = (long)nbuf * BKVASIZE;
1139 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1140 (TRANSIENT_DENOM - 1)) {
1142 * There is more KVA than memory. Do not
1143 * adjust buffer map size, and assign the rest
1144 * of maxbuf to transient map.
1146 biotmap_sz = maxbuf_sz - buf_sz;
1149 * Buffer map spans all KVA we could afford on
1150 * this platform. Give 10% (20% on i386) of
1151 * the buffer map to the transient bio map.
1153 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1154 buf_sz -= biotmap_sz;
1156 if (biotmap_sz / INT_MAX > maxphys)
1157 bio_transient_maxcnt = INT_MAX;
1159 bio_transient_maxcnt = biotmap_sz / maxphys;
1161 * Artificially limit to 1024 simultaneous in-flight I/Os
1162 * using the transient mapping.
1164 if (bio_transient_maxcnt > 1024)
1165 bio_transient_maxcnt = 1024;
1167 nbuf = buf_sz / BKVASIZE;
1171 nswbuf = min(nbuf / 4, 256);
1172 if (nswbuf < NSWBUF_MIN)
1173 nswbuf = NSWBUF_MIN;
1177 * Reserve space for the buffer cache buffers
1180 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1181 atop(maxbcachebuf)) * nbuf;
1187 * Single global constant for BUF_WMESG, to avoid getting multiple
1190 static const char buf_wmesg[] = "bufwait";
1192 /* Initialize the buffer subsystem. Called before use of any buffers. */
1199 KASSERT(maxbcachebuf >= MAXBSIZE,
1200 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1202 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1203 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1204 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1205 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1207 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1209 /* finally, initialize each buffer header and stick on empty q */
1210 for (i = 0; i < nbuf; i++) {
1212 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1213 bp->b_flags = B_INVAL;
1214 bp->b_rcred = NOCRED;
1215 bp->b_wcred = NOCRED;
1216 bp->b_qindex = QUEUE_NONE;
1218 bp->b_subqueue = mp_maxid + 1;
1220 bp->b_data = bp->b_kvabase = unmapped_buf;
1221 LIST_INIT(&bp->b_dep);
1222 BUF_LOCKINIT(bp, buf_wmesg);
1223 bq_insert(&bqempty, bp, false);
1227 * maxbufspace is the absolute maximum amount of buffer space we are
1228 * allowed to reserve in KVM and in real terms. The absolute maximum
1229 * is nominally used by metadata. hibufspace is the nominal maximum
1230 * used by most other requests. The differential is required to
1231 * ensure that metadata deadlocks don't occur.
1233 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1234 * this may result in KVM fragmentation which is not handled optimally
1235 * by the system. XXX This is less true with vmem. We could use
1238 maxbufspace = (long)nbuf * BKVASIZE;
1239 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1240 lobufspace = (hibufspace / 20) * 19; /* 95% */
1241 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1244 * Note: The 16 MiB upper limit for hirunningspace was chosen
1245 * arbitrarily and may need further tuning. It corresponds to
1246 * 128 outstanding write IO requests (if IO size is 128 KiB),
1247 * which fits with many RAID controllers' tagged queuing limits.
1248 * The lower 1 MiB limit is the historical upper limit for
1251 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1252 16 * 1024 * 1024), 1024 * 1024);
1253 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1256 * Limit the amount of malloc memory since it is wired permanently into
1257 * the kernel space. Even though this is accounted for in the buffer
1258 * allocation, we don't want the malloced region to grow uncontrolled.
1259 * The malloc scheme improves memory utilization significantly on
1260 * average (small) directories.
1262 maxbufmallocspace = hibufspace / 20;
1265 * Reduce the chance of a deadlock occurring by limiting the number
1266 * of delayed-write dirty buffers we allow to stack up.
1268 hidirtybuffers = nbuf / 4 + 20;
1269 dirtybufthresh = hidirtybuffers * 9 / 10;
1271 * To support extreme low-memory systems, make sure hidirtybuffers
1272 * cannot eat up all available buffer space. This occurs when our
1273 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1274 * buffer space assuming BKVASIZE'd buffers.
1276 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1277 hidirtybuffers >>= 1;
1279 lodirtybuffers = hidirtybuffers / 2;
1282 * lofreebuffers should be sufficient to avoid stalling waiting on
1283 * buf headers under heavy utilization. The bufs in per-cpu caches
1284 * are counted as free but will be unavailable to threads executing
1287 * hifreebuffers is the free target for the bufspace daemon. This
1288 * should be set appropriately to limit work per-iteration.
1290 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1291 hifreebuffers = (3 * lofreebuffers) / 2;
1292 numfreebuffers = nbuf;
1294 /* Setup the kva and free list allocators. */
1295 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1296 buf_zone = uma_zcache_create("buf free cache",
1297 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1298 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1301 * Size the clean queue according to the amount of buffer space.
1302 * One queue per-256mb up to the max. More queues gives better
1303 * concurrency but less accurate LRU.
1305 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1306 for (i = 0 ; i < buf_domains; i++) {
1307 struct bufdomain *bd;
1311 bd->bd_freebuffers = nbuf / buf_domains;
1312 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1313 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1314 bd->bd_bufspace = 0;
1315 bd->bd_maxbufspace = maxbufspace / buf_domains;
1316 bd->bd_hibufspace = hibufspace / buf_domains;
1317 bd->bd_lobufspace = lobufspace / buf_domains;
1318 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1319 bd->bd_numdirtybuffers = 0;
1320 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1321 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1322 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1323 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1324 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1326 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1327 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1328 mappingrestarts = counter_u64_alloc(M_WAITOK);
1329 numbufallocfails = counter_u64_alloc(M_WAITOK);
1330 notbufdflushes = counter_u64_alloc(M_WAITOK);
1331 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1332 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1333 bufkvaspace = counter_u64_alloc(M_WAITOK);
1338 vfs_buf_check_mapped(struct buf *bp)
1341 KASSERT(bp->b_kvabase != unmapped_buf,
1342 ("mapped buf: b_kvabase was not updated %p", bp));
1343 KASSERT(bp->b_data != unmapped_buf,
1344 ("mapped buf: b_data was not updated %p", bp));
1345 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1346 maxphys, ("b_data + b_offset unmapped %p", bp));
1350 vfs_buf_check_unmapped(struct buf *bp)
1353 KASSERT(bp->b_data == unmapped_buf,
1354 ("unmapped buf: corrupted b_data %p", bp));
1357 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1358 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1360 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1361 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1365 isbufbusy(struct buf *bp)
1367 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1368 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1374 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1377 bufshutdown(int show_busybufs)
1379 static int first_buf_printf = 1;
1381 int i, iter, nbusy, pbusy;
1387 * Sync filesystems for shutdown
1389 wdog_kern_pat(WD_LASTVAL);
1390 kern_sync(curthread);
1393 * With soft updates, some buffers that are
1394 * written will be remarked as dirty until other
1395 * buffers are written.
1397 for (iter = pbusy = 0; iter < 20; iter++) {
1399 for (i = nbuf - 1; i >= 0; i--) {
1405 if (first_buf_printf)
1406 printf("All buffers synced.");
1409 if (first_buf_printf) {
1410 printf("Syncing disks, buffers remaining... ");
1411 first_buf_printf = 0;
1413 printf("%d ", nbusy);
1418 wdog_kern_pat(WD_LASTVAL);
1419 kern_sync(curthread);
1423 * Spin for a while to allow interrupt threads to run.
1425 DELAY(50000 * iter);
1428 * Context switch several times to allow interrupt
1431 for (subiter = 0; subiter < 50 * iter; subiter++) {
1432 sched_relinquish(curthread);
1439 * Count only busy local buffers to prevent forcing
1440 * a fsck if we're just a client of a wedged NFS server
1443 for (i = nbuf - 1; i >= 0; i--) {
1445 if (isbufbusy(bp)) {
1447 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1448 if (bp->b_dev == NULL) {
1449 TAILQ_REMOVE(&mountlist,
1450 bp->b_vp->v_mount, mnt_list);
1455 if (show_busybufs > 0) {
1457 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1458 nbusy, bp, bp->b_vp, bp->b_flags,
1459 (intmax_t)bp->b_blkno,
1460 (intmax_t)bp->b_lblkno);
1461 BUF_LOCKPRINTINFO(bp);
1462 if (show_busybufs > 1)
1470 * Failed to sync all blocks. Indicate this and don't
1471 * unmount filesystems (thus forcing an fsck on reboot).
1473 BOOTTRACE("shutdown failed to sync buffers");
1474 printf("Giving up on %d buffers\n", nbusy);
1475 DELAY(5000000); /* 5 seconds */
1478 BOOTTRACE("shutdown sync complete");
1479 if (!first_buf_printf)
1480 printf("Final sync complete\n");
1483 * Unmount filesystems and perform swapoff, to quiesce
1484 * the system as much as possible. In particular, no
1485 * I/O should be initiated from top levels since it
1486 * might be abruptly terminated by reset, or otherwise
1487 * erronously handled because other parts of the
1488 * system are disabled.
1490 * Swapoff before unmount, because file-backed swap is
1491 * non-operational after unmount of the underlying
1494 if (!KERNEL_PANICKED()) {
1498 BOOTTRACE("shutdown unmounted all filesystems");
1500 DELAY(100000); /* wait for console output to finish */
1504 bpmap_qenter(struct buf *bp)
1507 BUF_CHECK_MAPPED(bp);
1510 * bp->b_data is relative to bp->b_offset, but
1511 * bp->b_offset may be offset into the first page.
1513 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1514 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1515 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1516 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1519 static inline struct bufdomain *
1520 bufdomain(struct buf *bp)
1523 return (&bdomain[bp->b_domain]);
1526 static struct bufqueue *
1527 bufqueue(struct buf *bp)
1530 switch (bp->b_qindex) {
1533 case QUEUE_SENTINEL:
1538 return (&bufdomain(bp)->bd_dirtyq);
1540 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1544 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1548 * Return the locked bufqueue that bp is a member of.
1550 static struct bufqueue *
1551 bufqueue_acquire(struct buf *bp)
1553 struct bufqueue *bq, *nbq;
1556 * bp can be pushed from a per-cpu queue to the
1557 * cleanq while we're waiting on the lock. Retry
1558 * if the queues don't match.
1576 * Insert the buffer into the appropriate free list. Requires a
1577 * locked buffer on entry and buffer is unlocked before return.
1580 binsfree(struct buf *bp, int qindex)
1582 struct bufdomain *bd;
1583 struct bufqueue *bq;
1585 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1586 ("binsfree: Invalid qindex %d", qindex));
1587 BUF_ASSERT_XLOCKED(bp);
1590 * Handle delayed bremfree() processing.
1592 if (bp->b_flags & B_REMFREE) {
1593 if (bp->b_qindex == qindex) {
1594 bp->b_flags |= B_REUSE;
1595 bp->b_flags &= ~B_REMFREE;
1599 bq = bufqueue_acquire(bp);
1604 if (qindex == QUEUE_CLEAN) {
1605 if (bd->bd_lim != 0)
1606 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1610 bq = &bd->bd_dirtyq;
1611 bq_insert(bq, bp, true);
1617 * Free a buffer to the buf zone once it no longer has valid contents.
1620 buf_free(struct buf *bp)
1623 if (bp->b_flags & B_REMFREE)
1625 if (bp->b_vflags & BV_BKGRDINPROG)
1626 panic("losing buffer 1");
1627 if (bp->b_rcred != NOCRED) {
1628 crfree(bp->b_rcred);
1629 bp->b_rcred = NOCRED;
1631 if (bp->b_wcred != NOCRED) {
1632 crfree(bp->b_wcred);
1633 bp->b_wcred = NOCRED;
1635 if (!LIST_EMPTY(&bp->b_dep))
1638 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1639 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1641 uma_zfree(buf_zone, bp);
1647 * Import bufs into the uma cache from the buf list. The system still
1648 * expects a static array of bufs and much of the synchronization
1649 * around bufs assumes type stable storage. As a result, UMA is used
1650 * only as a per-cpu cache of bufs still maintained on a global list.
1653 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1659 for (i = 0; i < cnt; i++) {
1660 bp = TAILQ_FIRST(&bqempty.bq_queue);
1663 bq_remove(&bqempty, bp);
1666 BQ_UNLOCK(&bqempty);
1674 * Release bufs from the uma cache back to the buffer queues.
1677 buf_release(void *arg, void **store, int cnt)
1679 struct bufqueue *bq;
1685 for (i = 0; i < cnt; i++) {
1687 /* Inline bq_insert() to batch locking. */
1688 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1689 bp->b_flags &= ~(B_AGE | B_REUSE);
1691 bp->b_qindex = bq->bq_index;
1699 * Allocate an empty buffer header.
1702 buf_alloc(struct bufdomain *bd)
1705 int freebufs, error;
1708 * We can only run out of bufs in the buf zone if the average buf
1709 * is less than BKVASIZE. In this case the actual wait/block will
1710 * come from buf_reycle() failing to flush one of these small bufs.
1713 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1715 bp = uma_zalloc(buf_zone, M_NOWAIT);
1717 atomic_add_int(&bd->bd_freebuffers, 1);
1718 bufspace_daemon_wakeup(bd);
1719 counter_u64_add(numbufallocfails, 1);
1723 * Wake-up the bufspace daemon on transition below threshold.
1725 if (freebufs == bd->bd_lofreebuffers)
1726 bufspace_daemon_wakeup(bd);
1728 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL);
1729 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1733 KASSERT(bp->b_vp == NULL,
1734 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1735 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1736 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1737 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1738 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1739 KASSERT(bp->b_npages == 0,
1740 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1741 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1742 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1743 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1745 bp->b_domain = BD_DOMAIN(bd);
1751 bp->b_blkno = bp->b_lblkno = 0;
1752 bp->b_offset = NOOFFSET;
1758 bp->b_dirtyoff = bp->b_dirtyend = 0;
1759 bp->b_bufobj = NULL;
1760 bp->b_data = bp->b_kvabase = unmapped_buf;
1761 bp->b_fsprivate1 = NULL;
1762 bp->b_fsprivate2 = NULL;
1763 bp->b_fsprivate3 = NULL;
1764 LIST_INIT(&bp->b_dep);
1772 * Free a buffer from the given bufqueue. kva controls whether the
1773 * freed buf must own some kva resources. This is used for
1777 buf_recycle(struct bufdomain *bd, bool kva)
1779 struct bufqueue *bq;
1780 struct buf *bp, *nbp;
1783 counter_u64_add(bufdefragcnt, 1);
1787 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1788 ("buf_recycle: Locks don't match"));
1789 nbp = TAILQ_FIRST(&bq->bq_queue);
1792 * Run scan, possibly freeing data and/or kva mappings on the fly
1795 while ((bp = nbp) != NULL) {
1797 * Calculate next bp (we can only use it if we do not
1798 * release the bqlock).
1800 nbp = TAILQ_NEXT(bp, b_freelist);
1803 * If we are defragging then we need a buffer with
1804 * some kva to reclaim.
1806 if (kva && bp->b_kvasize == 0)
1809 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1813 * Implement a second chance algorithm for frequently
1816 if ((bp->b_flags & B_REUSE) != 0) {
1817 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1818 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1819 bp->b_flags &= ~B_REUSE;
1825 * Skip buffers with background writes in progress.
1827 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1832 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1833 ("buf_recycle: inconsistent queue %d bp %p",
1835 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1836 ("getnewbuf: queue domain %d doesn't match request %d",
1837 bp->b_domain, (int)BD_DOMAIN(bd)));
1839 * NOTE: nbp is now entirely invalid. We can only restart
1840 * the scan from this point on.
1846 * Requeue the background write buffer with error and
1849 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1852 nbp = TAILQ_FIRST(&bq->bq_queue);
1855 bp->b_flags |= B_INVAL;
1868 * Mark the buffer for removal from the appropriate free list.
1872 bremfree(struct buf *bp)
1875 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1876 KASSERT((bp->b_flags & B_REMFREE) == 0,
1877 ("bremfree: buffer %p already marked for delayed removal.", bp));
1878 KASSERT(bp->b_qindex != QUEUE_NONE,
1879 ("bremfree: buffer %p not on a queue.", bp));
1880 BUF_ASSERT_XLOCKED(bp);
1882 bp->b_flags |= B_REMFREE;
1888 * Force an immediate removal from a free list. Used only in nfs when
1889 * it abuses the b_freelist pointer.
1892 bremfreef(struct buf *bp)
1894 struct bufqueue *bq;
1896 bq = bufqueue_acquire(bp);
1902 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1905 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1906 TAILQ_INIT(&bq->bq_queue);
1908 bq->bq_index = qindex;
1909 bq->bq_subqueue = subqueue;
1913 bd_init(struct bufdomain *bd)
1917 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1918 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1919 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1920 for (i = 0; i <= mp_maxid; i++)
1921 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1922 "bufq clean subqueue lock");
1923 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1929 * Removes a buffer from the free list, must be called with the
1930 * correct qlock held.
1933 bq_remove(struct bufqueue *bq, struct buf *bp)
1936 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1937 bp, bp->b_vp, bp->b_flags);
1938 KASSERT(bp->b_qindex != QUEUE_NONE,
1939 ("bq_remove: buffer %p not on a queue.", bp));
1940 KASSERT(bufqueue(bp) == bq,
1941 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1943 BQ_ASSERT_LOCKED(bq);
1944 if (bp->b_qindex != QUEUE_EMPTY) {
1945 BUF_ASSERT_XLOCKED(bp);
1947 KASSERT(bq->bq_len >= 1,
1948 ("queue %d underflow", bp->b_qindex));
1949 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1951 bp->b_qindex = QUEUE_NONE;
1952 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1956 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1960 BQ_ASSERT_LOCKED(bq);
1961 if (bq != bd->bd_cleanq) {
1963 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1964 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1965 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1967 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1969 bd->bd_cleanq->bq_len += bq->bq_len;
1972 if (bd->bd_wanted) {
1974 wakeup(&bd->bd_wanted);
1976 if (bq != bd->bd_cleanq)
1981 bd_flushall(struct bufdomain *bd)
1983 struct bufqueue *bq;
1987 if (bd->bd_lim == 0)
1990 for (i = 0; i <= mp_maxid; i++) {
1991 bq = &bd->bd_subq[i];
1992 if (bq->bq_len == 0)
2004 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
2006 struct bufdomain *bd;
2008 if (bp->b_qindex != QUEUE_NONE)
2009 panic("bq_insert: free buffer %p onto another queue?", bp);
2012 if (bp->b_flags & B_AGE) {
2013 /* Place this buf directly on the real queue. */
2014 if (bq->bq_index == QUEUE_CLEAN)
2017 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2020 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2022 bp->b_flags &= ~(B_AGE | B_REUSE);
2024 bp->b_qindex = bq->bq_index;
2025 bp->b_subqueue = bq->bq_subqueue;
2028 * Unlock before we notify so that we don't wakeup a waiter that
2029 * fails a trylock on the buf and sleeps again.
2034 if (bp->b_qindex == QUEUE_CLEAN) {
2036 * Flush the per-cpu queue and notify any waiters.
2038 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2039 bq->bq_len >= bd->bd_lim))
2048 * Free the kva allocation for a buffer.
2052 bufkva_free(struct buf *bp)
2056 if (bp->b_kvasize == 0) {
2057 KASSERT(bp->b_kvabase == unmapped_buf &&
2058 bp->b_data == unmapped_buf,
2059 ("Leaked KVA space on %p", bp));
2060 } else if (buf_mapped(bp))
2061 BUF_CHECK_MAPPED(bp);
2063 BUF_CHECK_UNMAPPED(bp);
2065 if (bp->b_kvasize == 0)
2068 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2069 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2070 counter_u64_add(buffreekvacnt, 1);
2071 bp->b_data = bp->b_kvabase = unmapped_buf;
2078 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2081 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2086 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2087 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2088 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2089 KASSERT(maxsize <= maxbcachebuf,
2090 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2095 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2098 * Buffer map is too fragmented. Request the caller
2099 * to defragment the map.
2103 bp->b_kvabase = (caddr_t)addr;
2104 bp->b_kvasize = maxsize;
2105 counter_u64_add(bufkvaspace, bp->b_kvasize);
2106 if ((gbflags & GB_UNMAPPED) != 0) {
2107 bp->b_data = unmapped_buf;
2108 BUF_CHECK_UNMAPPED(bp);
2110 bp->b_data = bp->b_kvabase;
2111 BUF_CHECK_MAPPED(bp);
2119 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2120 * callback that fires to avoid returning failure.
2123 bufkva_reclaim(vmem_t *vmem, int flags)
2130 for (i = 0; i < 5; i++) {
2131 for (q = 0; q < buf_domains; q++)
2132 if (buf_recycle(&bdomain[q], true) != 0)
2141 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2142 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2143 * the buffer is valid and we do not have to do anything.
2146 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2147 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2155 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2156 if (inmem(vp, *rablkno))
2158 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2159 if ((rabp->b_flags & B_CACHE) != 0) {
2166 racct_add_buf(curproc, rabp, 0);
2167 PROC_UNLOCK(curproc);
2170 td->td_ru.ru_inblock++;
2171 rabp->b_flags |= B_ASYNC;
2172 rabp->b_flags &= ~B_INVAL;
2173 if ((flags & GB_CKHASH) != 0) {
2174 rabp->b_flags |= B_CKHASH;
2175 rabp->b_ckhashcalc = ckhashfunc;
2177 rabp->b_ioflags &= ~BIO_ERROR;
2178 rabp->b_iocmd = BIO_READ;
2179 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2180 rabp->b_rcred = crhold(cred);
2181 vfs_busy_pages(rabp, 0);
2183 rabp->b_iooffset = dbtob(rabp->b_blkno);
2189 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2191 * Get a buffer with the specified data. Look in the cache first. We
2192 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2193 * is set, the buffer is valid and we do not have to do anything, see
2194 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2196 * Always return a NULL buffer pointer (in bpp) when returning an error.
2198 * The blkno parameter is the logical block being requested. Normally
2199 * the mapping of logical block number to disk block address is done
2200 * by calling VOP_BMAP(). However, if the mapping is already known, the
2201 * disk block address can be passed using the dblkno parameter. If the
2202 * disk block address is not known, then the same value should be passed
2203 * for blkno and dblkno.
2206 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2207 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2208 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2212 int error, readwait, rv;
2214 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2217 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2220 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2225 KASSERT(blkno == bp->b_lblkno,
2226 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2227 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2228 flags &= ~GB_NOSPARSE;
2232 * If not found in cache, do some I/O
2235 if ((bp->b_flags & B_CACHE) == 0) {
2238 PROC_LOCK(td->td_proc);
2239 racct_add_buf(td->td_proc, bp, 0);
2240 PROC_UNLOCK(td->td_proc);
2243 td->td_ru.ru_inblock++;
2244 bp->b_iocmd = BIO_READ;
2245 bp->b_flags &= ~B_INVAL;
2246 if ((flags & GB_CKHASH) != 0) {
2247 bp->b_flags |= B_CKHASH;
2248 bp->b_ckhashcalc = ckhashfunc;
2250 if ((flags & GB_CVTENXIO) != 0)
2251 bp->b_xflags |= BX_CVTENXIO;
2252 bp->b_ioflags &= ~BIO_ERROR;
2253 if (bp->b_rcred == NOCRED && cred != NOCRED)
2254 bp->b_rcred = crhold(cred);
2255 vfs_busy_pages(bp, 0);
2256 bp->b_iooffset = dbtob(bp->b_blkno);
2262 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2264 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2278 * Write, release buffer on completion. (Done by iodone
2279 * if async). Do not bother writing anything if the buffer
2282 * Note that we set B_CACHE here, indicating that buffer is
2283 * fully valid and thus cacheable. This is true even of NFS
2284 * now so we set it generally. This could be set either here
2285 * or in biodone() since the I/O is synchronous. We put it
2289 bufwrite(struct buf *bp)
2296 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2297 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2298 bp->b_flags |= B_INVAL | B_RELBUF;
2299 bp->b_flags &= ~B_CACHE;
2303 if (bp->b_flags & B_INVAL) {
2308 if (bp->b_flags & B_BARRIER)
2309 atomic_add_long(&barrierwrites, 1);
2311 oldflags = bp->b_flags;
2313 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2314 ("FFS background buffer should not get here %p", bp));
2318 vp_md = vp->v_vflag & VV_MD;
2323 * Mark the buffer clean. Increment the bufobj write count
2324 * before bundirty() call, to prevent other thread from seeing
2325 * empty dirty list and zero counter for writes in progress,
2326 * falsely indicating that the bufobj is clean.
2328 bufobj_wref(bp->b_bufobj);
2331 bp->b_flags &= ~B_DONE;
2332 bp->b_ioflags &= ~BIO_ERROR;
2333 bp->b_flags |= B_CACHE;
2334 bp->b_iocmd = BIO_WRITE;
2336 vfs_busy_pages(bp, 1);
2339 * Normal bwrites pipeline writes
2341 bp->b_runningbufspace = bp->b_bufsize;
2342 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2347 racct_add_buf(curproc, bp, 1);
2348 PROC_UNLOCK(curproc);
2351 curthread->td_ru.ru_oublock++;
2352 if (oldflags & B_ASYNC)
2354 bp->b_iooffset = dbtob(bp->b_blkno);
2355 buf_track(bp, __func__);
2358 if ((oldflags & B_ASYNC) == 0) {
2359 int rtval = bufwait(bp);
2362 } else if (space > hirunningspace) {
2364 * don't allow the async write to saturate the I/O
2365 * system. We will not deadlock here because
2366 * we are blocking waiting for I/O that is already in-progress
2367 * to complete. We do not block here if it is the update
2368 * or syncer daemon trying to clean up as that can lead
2371 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2372 waitrunningbufspace();
2379 bufbdflush(struct bufobj *bo, struct buf *bp)
2382 struct bufdomain *bd;
2384 bd = &bdomain[bo->bo_domain];
2385 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2386 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2388 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2391 * Try to find a buffer to flush.
2393 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2394 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2396 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2399 panic("bdwrite: found ourselves");
2401 /* Don't countdeps with the bo lock held. */
2402 if (buf_countdeps(nbp, 0)) {
2407 if (nbp->b_flags & B_CLUSTEROK) {
2408 vfs_bio_awrite(nbp);
2413 dirtybufferflushes++;
2422 * Delayed write. (Buffer is marked dirty). Do not bother writing
2423 * anything if the buffer is marked invalid.
2425 * Note that since the buffer must be completely valid, we can safely
2426 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2427 * biodone() in order to prevent getblk from writing the buffer
2428 * out synchronously.
2431 bdwrite(struct buf *bp)
2433 struct thread *td = curthread;
2437 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2438 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2439 KASSERT((bp->b_flags & B_BARRIER) == 0,
2440 ("Barrier request in delayed write %p", bp));
2442 if (bp->b_flags & B_INVAL) {
2448 * If we have too many dirty buffers, don't create any more.
2449 * If we are wildly over our limit, then force a complete
2450 * cleanup. Otherwise, just keep the situation from getting
2451 * out of control. Note that we have to avoid a recursive
2452 * disaster and not try to clean up after our own cleanup!
2456 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2457 td->td_pflags |= TDP_INBDFLUSH;
2459 td->td_pflags &= ~TDP_INBDFLUSH;
2465 * Set B_CACHE, indicating that the buffer is fully valid. This is
2466 * true even of NFS now.
2468 bp->b_flags |= B_CACHE;
2471 * This bmap keeps the system from needing to do the bmap later,
2472 * perhaps when the system is attempting to do a sync. Since it
2473 * is likely that the indirect block -- or whatever other datastructure
2474 * that the filesystem needs is still in memory now, it is a good
2475 * thing to do this. Note also, that if the pageout daemon is
2476 * requesting a sync -- there might not be enough memory to do
2477 * the bmap then... So, this is important to do.
2479 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2480 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2483 buf_track(bp, __func__);
2486 * Set the *dirty* buffer range based upon the VM system dirty
2489 * Mark the buffer pages as clean. We need to do this here to
2490 * satisfy the vnode_pager and the pageout daemon, so that it
2491 * thinks that the pages have been "cleaned". Note that since
2492 * the pages are in a delayed write buffer -- the VFS layer
2493 * "will" see that the pages get written out on the next sync,
2494 * or perhaps the cluster will be completed.
2496 vfs_clean_pages_dirty_buf(bp);
2500 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2501 * due to the softdep code.
2508 * Turn buffer into delayed write request. We must clear BIO_READ and
2509 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2510 * itself to properly update it in the dirty/clean lists. We mark it
2511 * B_DONE to ensure that any asynchronization of the buffer properly
2512 * clears B_DONE ( else a panic will occur later ).
2514 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2515 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2516 * should only be called if the buffer is known-good.
2518 * Since the buffer is not on a queue, we do not update the numfreebuffers
2521 * The buffer must be on QUEUE_NONE.
2524 bdirty(struct buf *bp)
2527 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2528 bp, bp->b_vp, bp->b_flags);
2529 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2530 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2531 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2532 bp->b_flags &= ~(B_RELBUF);
2533 bp->b_iocmd = BIO_WRITE;
2535 if ((bp->b_flags & B_DELWRI) == 0) {
2536 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2545 * Clear B_DELWRI for buffer.
2547 * Since the buffer is not on a queue, we do not update the numfreebuffers
2550 * The buffer must be on QUEUE_NONE.
2554 bundirty(struct buf *bp)
2557 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2558 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2559 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2560 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2562 if (bp->b_flags & B_DELWRI) {
2563 bp->b_flags &= ~B_DELWRI;
2568 * Since it is now being written, we can clear its deferred write flag.
2570 bp->b_flags &= ~B_DEFERRED;
2576 * Asynchronous write. Start output on a buffer, but do not wait for
2577 * it to complete. The buffer is released when the output completes.
2579 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2580 * B_INVAL buffers. Not us.
2583 bawrite(struct buf *bp)
2586 bp->b_flags |= B_ASYNC;
2593 * Asynchronous barrier write. Start output on a buffer, but do not
2594 * wait for it to complete. Place a write barrier after this write so
2595 * that this buffer and all buffers written before it are committed to
2596 * the disk before any buffers written after this write are committed
2597 * to the disk. The buffer is released when the output completes.
2600 babarrierwrite(struct buf *bp)
2603 bp->b_flags |= B_ASYNC | B_BARRIER;
2610 * Synchronous barrier write. Start output on a buffer and wait for
2611 * it to complete. Place a write barrier after this write so that
2612 * this buffer and all buffers written before it are committed to
2613 * the disk before any buffers written after this write are committed
2614 * to the disk. The buffer is released when the output completes.
2617 bbarrierwrite(struct buf *bp)
2620 bp->b_flags |= B_BARRIER;
2621 return (bwrite(bp));
2627 * Called prior to the locking of any vnodes when we are expecting to
2628 * write. We do not want to starve the buffer cache with too many
2629 * dirty buffers so we block here. By blocking prior to the locking
2630 * of any vnodes we attempt to avoid the situation where a locked vnode
2631 * prevents the various system daemons from flushing related buffers.
2637 if (buf_dirty_count_severe()) {
2638 mtx_lock(&bdirtylock);
2639 while (buf_dirty_count_severe()) {
2641 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2644 mtx_unlock(&bdirtylock);
2649 * Return true if we have too many dirty buffers.
2652 buf_dirty_count_severe(void)
2655 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2661 * Release a busy buffer and, if requested, free its resources. The
2662 * buffer will be stashed in the appropriate bufqueue[] allowing it
2663 * to be accessed later as a cache entity or reused for other purposes.
2666 brelse(struct buf *bp)
2668 struct mount *v_mnt;
2672 * Many functions erroneously call brelse with a NULL bp under rare
2673 * error conditions. Simply return when called with a NULL bp.
2677 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2678 bp, bp->b_vp, bp->b_flags);
2679 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2680 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2681 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2682 ("brelse: non-VMIO buffer marked NOREUSE"));
2684 if (BUF_LOCKRECURSED(bp)) {
2686 * Do not process, in particular, do not handle the
2687 * B_INVAL/B_RELBUF and do not release to free list.
2693 if (bp->b_flags & B_MANAGED) {
2698 if (LIST_EMPTY(&bp->b_dep)) {
2699 bp->b_flags &= ~B_IOSTARTED;
2701 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2702 ("brelse: SU io not finished bp %p", bp));
2705 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2706 BO_LOCK(bp->b_bufobj);
2707 bp->b_vflags &= ~BV_BKGRDERR;
2708 BO_UNLOCK(bp->b_bufobj);
2712 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2713 (bp->b_flags & B_INVALONERR)) {
2715 * Forced invalidation of dirty buffer contents, to be used
2716 * after a failed write in the rare case that the loss of the
2717 * contents is acceptable. The buffer is invalidated and
2720 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2721 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2724 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2725 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2726 !(bp->b_flags & B_INVAL)) {
2728 * Failed write, redirty. All errors except ENXIO (which
2729 * means the device is gone) are treated as being
2732 * XXX Treating EIO as transient is not correct; the
2733 * contract with the local storage device drivers is that
2734 * they will only return EIO once the I/O is no longer
2735 * retriable. Network I/O also respects this through the
2736 * guarantees of TCP and/or the internal retries of NFS.
2737 * ENOMEM might be transient, but we also have no way of
2738 * knowing when its ok to retry/reschedule. In general,
2739 * this entire case should be made obsolete through better
2740 * error handling/recovery and resource scheduling.
2742 * Do this also for buffers that failed with ENXIO, but have
2743 * non-empty dependencies - the soft updates code might need
2744 * to access the buffer to untangle them.
2746 * Must clear BIO_ERROR to prevent pages from being scrapped.
2748 bp->b_ioflags &= ~BIO_ERROR;
2750 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2751 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2753 * Either a failed read I/O, or we were asked to free or not
2754 * cache the buffer, or we failed to write to a device that's
2755 * no longer present.
2757 bp->b_flags |= B_INVAL;
2758 if (!LIST_EMPTY(&bp->b_dep))
2760 if (bp->b_flags & B_DELWRI)
2762 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2763 if ((bp->b_flags & B_VMIO) == 0) {
2771 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2772 * is called with B_DELWRI set, the underlying pages may wind up
2773 * getting freed causing a previous write (bdwrite()) to get 'lost'
2774 * because pages associated with a B_DELWRI bp are marked clean.
2776 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2777 * if B_DELWRI is set.
2779 if (bp->b_flags & B_DELWRI)
2780 bp->b_flags &= ~B_RELBUF;
2783 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2784 * constituted, not even NFS buffers now. Two flags effect this. If
2785 * B_INVAL, the struct buf is invalidated but the VM object is kept
2786 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2788 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2789 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2790 * buffer is also B_INVAL because it hits the re-dirtying code above.
2792 * Normally we can do this whether a buffer is B_DELWRI or not. If
2793 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2794 * the commit state and we cannot afford to lose the buffer. If the
2795 * buffer has a background write in progress, we need to keep it
2796 * around to prevent it from being reconstituted and starting a second
2800 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2802 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2803 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2804 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2805 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2806 vfs_vmio_invalidate(bp);
2810 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2811 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2813 bp->b_flags &= ~B_NOREUSE;
2814 if (bp->b_vp != NULL)
2819 * If the buffer has junk contents signal it and eventually
2820 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2823 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2824 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2825 bp->b_flags |= B_INVAL;
2826 if (bp->b_flags & B_INVAL) {
2827 if (bp->b_flags & B_DELWRI)
2833 buf_track(bp, __func__);
2835 /* buffers with no memory */
2836 if (bp->b_bufsize == 0) {
2840 /* buffers with junk contents */
2841 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2842 (bp->b_ioflags & BIO_ERROR)) {
2843 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2844 if (bp->b_vflags & BV_BKGRDINPROG)
2845 panic("losing buffer 2");
2846 qindex = QUEUE_CLEAN;
2847 bp->b_flags |= B_AGE;
2848 /* remaining buffers */
2849 } else if (bp->b_flags & B_DELWRI)
2850 qindex = QUEUE_DIRTY;
2852 qindex = QUEUE_CLEAN;
2854 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2855 panic("brelse: not dirty");
2857 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2858 bp->b_xflags &= ~(BX_CVTENXIO);
2859 /* binsfree unlocks bp. */
2860 binsfree(bp, qindex);
2864 * Release a buffer back to the appropriate queue but do not try to free
2865 * it. The buffer is expected to be used again soon.
2867 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2868 * biodone() to requeue an async I/O on completion. It is also used when
2869 * known good buffers need to be requeued but we think we may need the data
2872 * XXX we should be able to leave the B_RELBUF hint set on completion.
2875 bqrelse(struct buf *bp)
2879 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2880 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2881 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2883 qindex = QUEUE_NONE;
2884 if (BUF_LOCKRECURSED(bp)) {
2885 /* do not release to free list */
2889 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2890 bp->b_xflags &= ~(BX_CVTENXIO);
2892 if (LIST_EMPTY(&bp->b_dep)) {
2893 bp->b_flags &= ~B_IOSTARTED;
2895 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2896 ("bqrelse: SU io not finished bp %p", bp));
2899 if (bp->b_flags & B_MANAGED) {
2900 if (bp->b_flags & B_REMFREE)
2905 /* buffers with stale but valid contents */
2906 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2907 BV_BKGRDERR)) == BV_BKGRDERR) {
2908 BO_LOCK(bp->b_bufobj);
2909 bp->b_vflags &= ~BV_BKGRDERR;
2910 BO_UNLOCK(bp->b_bufobj);
2911 qindex = QUEUE_DIRTY;
2913 if ((bp->b_flags & B_DELWRI) == 0 &&
2914 (bp->b_xflags & BX_VNDIRTY))
2915 panic("bqrelse: not dirty");
2916 if ((bp->b_flags & B_NOREUSE) != 0) {
2920 qindex = QUEUE_CLEAN;
2922 buf_track(bp, __func__);
2923 /* binsfree unlocks bp. */
2924 binsfree(bp, qindex);
2928 buf_track(bp, __func__);
2934 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2935 * restore bogus pages.
2938 vfs_vmio_iodone(struct buf *bp)
2943 struct vnode *vp __unused;
2944 int i, iosize, resid;
2947 obj = bp->b_bufobj->bo_object;
2948 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2949 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2950 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2953 VNPASS(vp->v_holdcnt > 0, vp);
2954 VNPASS(vp->v_object != NULL, vp);
2956 foff = bp->b_offset;
2957 KASSERT(bp->b_offset != NOOFFSET,
2958 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2961 iosize = bp->b_bcount - bp->b_resid;
2962 for (i = 0; i < bp->b_npages; i++) {
2963 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2968 * cleanup bogus pages, restoring the originals
2971 if (m == bogus_page) {
2973 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2975 panic("biodone: page disappeared!");
2977 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2979 * In the write case, the valid and clean bits are
2980 * already changed correctly ( see bdwrite() ), so we
2981 * only need to do this here in the read case.
2983 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2984 resid)) == 0, ("vfs_vmio_iodone: page %p "
2985 "has unexpected dirty bits", m));
2986 vfs_page_set_valid(bp, foff, m);
2988 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2989 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2990 (intmax_t)foff, (uintmax_t)m->pindex));
2993 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2996 vm_object_pip_wakeupn(obj, bp->b_npages);
2997 if (bogus && buf_mapped(bp)) {
2998 BUF_CHECK_MAPPED(bp);
2999 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3000 bp->b_pages, bp->b_npages);
3005 * Perform page invalidation when a buffer is released. The fully invalid
3006 * pages will be reclaimed later in vfs_vmio_truncate().
3009 vfs_vmio_invalidate(struct buf *bp)
3013 int flags, i, resid, poffset, presid;
3015 if (buf_mapped(bp)) {
3016 BUF_CHECK_MAPPED(bp);
3017 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3019 BUF_CHECK_UNMAPPED(bp);
3021 * Get the base offset and length of the buffer. Note that
3022 * in the VMIO case if the buffer block size is not
3023 * page-aligned then b_data pointer may not be page-aligned.
3024 * But our b_pages[] array *IS* page aligned.
3026 * block sizes less then DEV_BSIZE (usually 512) are not
3027 * supported due to the page granularity bits (m->valid,
3028 * m->dirty, etc...).
3030 * See man buf(9) for more information
3032 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3033 obj = bp->b_bufobj->bo_object;
3034 resid = bp->b_bufsize;
3035 poffset = bp->b_offset & PAGE_MASK;
3036 VM_OBJECT_WLOCK(obj);
3037 for (i = 0; i < bp->b_npages; i++) {
3039 if (m == bogus_page)
3040 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3041 bp->b_pages[i] = NULL;
3043 presid = resid > (PAGE_SIZE - poffset) ?
3044 (PAGE_SIZE - poffset) : resid;
3045 KASSERT(presid >= 0, ("brelse: extra page"));
3046 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3047 if (pmap_page_wired_mappings(m) == 0)
3048 vm_page_set_invalid(m, poffset, presid);
3050 vm_page_release_locked(m, flags);
3054 VM_OBJECT_WUNLOCK(obj);
3059 * Page-granular truncation of an existing VMIO buffer.
3062 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3068 if (bp->b_npages == desiredpages)
3071 if (buf_mapped(bp)) {
3072 BUF_CHECK_MAPPED(bp);
3073 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3074 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3076 BUF_CHECK_UNMAPPED(bp);
3079 * The object lock is needed only if we will attempt to free pages.
3081 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3082 if ((bp->b_flags & B_DIRECT) != 0) {
3083 flags |= VPR_TRYFREE;
3084 obj = bp->b_bufobj->bo_object;
3085 VM_OBJECT_WLOCK(obj);
3089 for (i = desiredpages; i < bp->b_npages; i++) {
3091 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3092 bp->b_pages[i] = NULL;
3094 vm_page_release_locked(m, flags);
3096 vm_page_release(m, flags);
3099 VM_OBJECT_WUNLOCK(obj);
3100 bp->b_npages = desiredpages;
3104 * Byte granular extension of VMIO buffers.
3107 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3110 * We are growing the buffer, possibly in a
3111 * byte-granular fashion.
3119 * Step 1, bring in the VM pages from the object, allocating
3120 * them if necessary. We must clear B_CACHE if these pages
3121 * are not valid for the range covered by the buffer.
3123 obj = bp->b_bufobj->bo_object;
3124 if (bp->b_npages < desiredpages) {
3125 KASSERT(desiredpages <= atop(maxbcachebuf),
3126 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3127 bp, desiredpages, maxbcachebuf));
3130 * We must allocate system pages since blocking
3131 * here could interfere with paging I/O, no
3132 * matter which process we are.
3134 * Only exclusive busy can be tested here.
3135 * Blocking on shared busy might lead to
3136 * deadlocks once allocbuf() is called after
3137 * pages are vfs_busy_pages().
3139 (void)vm_page_grab_pages_unlocked(obj,
3140 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3141 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3142 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3143 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3144 bp->b_npages = desiredpages;
3148 * Step 2. We've loaded the pages into the buffer,
3149 * we have to figure out if we can still have B_CACHE
3150 * set. Note that B_CACHE is set according to the
3151 * byte-granular range ( bcount and size ), not the
3152 * aligned range ( newbsize ).
3154 * The VM test is against m->valid, which is DEV_BSIZE
3155 * aligned. Needless to say, the validity of the data
3156 * needs to also be DEV_BSIZE aligned. Note that this
3157 * fails with NFS if the server or some other client
3158 * extends the file's EOF. If our buffer is resized,
3159 * B_CACHE may remain set! XXX
3161 toff = bp->b_bcount;
3162 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3163 while ((bp->b_flags & B_CACHE) && toff < size) {
3166 if (tinc > (size - toff))
3168 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3169 m = bp->b_pages[pi];
3170 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3176 * Step 3, fixup the KVA pmap.
3181 BUF_CHECK_UNMAPPED(bp);
3185 * Check to see if a block at a particular lbn is available for a clustered
3189 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3196 /* If the buf isn't in core skip it */
3197 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3200 /* If the buf is busy we don't want to wait for it */
3201 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3204 /* Only cluster with valid clusterable delayed write buffers */
3205 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3206 (B_DELWRI | B_CLUSTEROK))
3209 if (bpa->b_bufsize != size)
3213 * Check to see if it is in the expected place on disk and that the
3214 * block has been mapped.
3216 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3226 * Implement clustered async writes for clearing out B_DELWRI buffers.
3227 * This is much better then the old way of writing only one buffer at
3228 * a time. Note that we may not be presented with the buffers in the
3229 * correct order, so we search for the cluster in both directions.
3232 vfs_bio_awrite(struct buf *bp)
3237 daddr_t lblkno = bp->b_lblkno;
3238 struct vnode *vp = bp->b_vp;
3246 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3248 * right now we support clustered writing only to regular files. If
3249 * we find a clusterable block we could be in the middle of a cluster
3250 * rather then at the beginning.
3252 if ((vp->v_type == VREG) &&
3253 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3254 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3255 size = vp->v_mount->mnt_stat.f_iosize;
3256 maxcl = maxphys / size;
3259 for (i = 1; i < maxcl; i++)
3260 if (vfs_bio_clcheck(vp, size, lblkno + i,
3261 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3264 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3265 if (vfs_bio_clcheck(vp, size, lblkno - j,
3266 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3272 * this is a possible cluster write
3276 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3282 bp->b_flags |= B_ASYNC;
3284 * default (old) behavior, writing out only one block
3286 * XXX returns b_bufsize instead of b_bcount for nwritten?
3288 nwritten = bp->b_bufsize;
3297 * Allocate KVA for an empty buf header according to gbflags.
3300 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3303 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3305 * In order to keep fragmentation sane we only allocate kva
3306 * in BKVASIZE chunks. XXX with vmem we can do page size.
3308 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3310 if (maxsize != bp->b_kvasize &&
3311 bufkva_alloc(bp, maxsize, gbflags))
3320 * Find and initialize a new buffer header, freeing up existing buffers
3321 * in the bufqueues as necessary. The new buffer is returned locked.
3324 * We have insufficient buffer headers
3325 * We have insufficient buffer space
3326 * buffer_arena is too fragmented ( space reservation fails )
3327 * If we have to flush dirty buffers ( but we try to avoid this )
3329 * The caller is responsible for releasing the reserved bufspace after
3330 * allocbuf() is called.
3333 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3335 struct bufdomain *bd;
3337 bool metadata, reserved;
3340 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3341 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3342 if (!unmapped_buf_allowed)
3343 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3345 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3353 bd = &bdomain[vp->v_bufobj.bo_domain];
3355 counter_u64_add(getnewbufcalls, 1);
3358 if (reserved == false &&
3359 bufspace_reserve(bd, maxsize, metadata) != 0) {
3360 counter_u64_add(getnewbufrestarts, 1);
3364 if ((bp = buf_alloc(bd)) == NULL) {
3365 counter_u64_add(getnewbufrestarts, 1);
3368 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3371 } while (buf_recycle(bd, false) == 0);
3374 bufspace_release(bd, maxsize);
3376 bp->b_flags |= B_INVAL;
3379 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3387 * buffer flushing daemon. Buffers are normally flushed by the
3388 * update daemon but if it cannot keep up this process starts to
3389 * take the load in an attempt to prevent getnewbuf() from blocking.
3391 static struct kproc_desc buf_kp = {
3396 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3399 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3403 flushed = flushbufqueues(vp, bd, target, 0);
3406 * Could not find any buffers without rollback
3407 * dependencies, so just write the first one
3408 * in the hopes of eventually making progress.
3410 if (vp != NULL && target > 2)
3412 flushbufqueues(vp, bd, target, 1);
3418 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3424 wakeup(&bd_request);
3425 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3427 mtx_unlock(&bdlock);
3429 printf("bufdaemon wait error: %d\n", error);
3435 struct bufdomain *bd;
3441 * This process needs to be suspended prior to shutdown sync.
3443 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3444 SHUTDOWN_PRI_LAST + 100);
3447 * Start the buf clean daemons as children threads.
3449 for (i = 0 ; i < buf_domains; i++) {
3452 error = kthread_add((void (*)(void *))bufspace_daemon,
3453 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3455 panic("error %d spawning bufspace daemon", error);
3459 * This process is allowed to take the buffer cache to the limit
3461 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3463 while (!bd_shutdown) {
3465 mtx_unlock(&bdlock);
3468 * Save speedupreq for this pass and reset to capture new
3471 speedupreq = bd_speedupreq;
3475 * Flush each domain sequentially according to its level and
3476 * the speedup request.
3478 for (i = 0; i < buf_domains; i++) {
3481 lodirty = bd->bd_numdirtybuffers / 2;
3483 lodirty = bd->bd_lodirtybuffers;
3484 while (bd->bd_numdirtybuffers > lodirty) {
3485 if (buf_flush(NULL, bd,
3486 bd->bd_numdirtybuffers - lodirty) == 0)
3488 kern_yield(PRI_USER);
3493 * Only clear bd_request if we have reached our low water
3494 * mark. The buf_daemon normally waits 1 second and
3495 * then incrementally flushes any dirty buffers that have
3496 * built up, within reason.
3498 * If we were unable to hit our low water mark and couldn't
3499 * find any flushable buffers, we sleep for a short period
3500 * to avoid endless loops on unlockable buffers.
3505 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3507 * We reached our low water mark, reset the
3508 * request and sleep until we are needed again.
3509 * The sleep is just so the suspend code works.
3513 * Do an extra wakeup in case dirty threshold
3514 * changed via sysctl and the explicit transition
3515 * out of shortfall was missed.
3518 if (runningbufspace <= lorunningspace)
3520 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3523 * We couldn't find any flushable dirty buffers but
3524 * still have too many dirty buffers, we
3525 * have to sleep and try again. (rare)
3527 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3530 wakeup(&bd_shutdown);
3531 mtx_unlock(&bdlock);
3538 * Try to flush a buffer in the dirty queue. We must be careful to
3539 * free up B_INVAL buffers instead of write them, which NFS is
3540 * particularly sensitive to.
3542 static int flushwithdeps = 0;
3543 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3545 "Number of buffers flushed with dependencies that require rollbacks");
3548 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3551 struct bufqueue *bq;
3552 struct buf *sentinel;
3562 bq = &bd->bd_dirtyq;
3564 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3565 sentinel->b_qindex = QUEUE_SENTINEL;
3567 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3569 while (flushed != target) {
3572 bp = TAILQ_NEXT(sentinel, b_freelist);
3574 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3575 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3582 * Skip sentinels inserted by other invocations of the
3583 * flushbufqueues(), taking care to not reorder them.
3585 * Only flush the buffers that belong to the
3586 * vnode locked by the curthread.
3588 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3593 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3599 * BKGRDINPROG can only be set with the buf and bufobj
3600 * locks both held. We tolerate a race to clear it here.
3602 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3603 (bp->b_flags & B_DELWRI) == 0) {
3607 if (bp->b_flags & B_INVAL) {
3614 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3615 if (flushdeps == 0) {
3623 * We must hold the lock on a vnode before writing
3624 * one of its buffers. Otherwise we may confuse, or
3625 * in the case of a snapshot vnode, deadlock the
3628 * The lock order here is the reverse of the normal
3629 * of vnode followed by buf lock. This is ok because
3630 * the NOWAIT will prevent deadlock.
3633 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3639 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3641 ASSERT_VOP_LOCKED(vp, "getbuf");
3643 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3644 vn_lock(vp, LK_TRYUPGRADE);
3647 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3648 bp, bp->b_vp, bp->b_flags);
3649 if (curproc == bufdaemonproc) {
3654 counter_u64_add(notbufdflushes, 1);
3656 vn_finished_write(mp);
3659 flushwithdeps += hasdeps;
3663 * Sleeping on runningbufspace while holding
3664 * vnode lock leads to deadlock.
3666 if (curproc == bufdaemonproc &&
3667 runningbufspace > hirunningspace)
3668 waitrunningbufspace();
3671 vn_finished_write(mp);
3675 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3677 free(sentinel, M_TEMP);
3682 * Check to see if a block is currently memory resident.
3685 incore(struct bufobj *bo, daddr_t blkno)
3687 return (gbincore_unlocked(bo, blkno));
3691 * Returns true if no I/O is needed to access the
3692 * associated VM object. This is like incore except
3693 * it also hunts around in the VM system for the data.
3696 inmem(struct vnode * vp, daddr_t blkno)
3699 vm_offset_t toff, tinc, size;
3704 ASSERT_VOP_LOCKED(vp, "inmem");
3706 if (incore(&vp->v_bufobj, blkno))
3708 if (vp->v_mount == NULL)
3715 if (size > vp->v_mount->mnt_stat.f_iosize)
3716 size = vp->v_mount->mnt_stat.f_iosize;
3717 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3719 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3720 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3726 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3727 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3729 * Consider page validity only if page mapping didn't change
3732 valid = vm_page_is_valid(m,
3733 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3734 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3746 * Set the dirty range for a buffer based on the status of the dirty
3747 * bits in the pages comprising the buffer. The range is limited
3748 * to the size of the buffer.
3750 * Tell the VM system that the pages associated with this buffer
3751 * are clean. This is used for delayed writes where the data is
3752 * going to go to disk eventually without additional VM intevention.
3754 * Note that while we only really need to clean through to b_bcount, we
3755 * just go ahead and clean through to b_bufsize.
3758 vfs_clean_pages_dirty_buf(struct buf *bp)
3760 vm_ooffset_t foff, noff, eoff;
3764 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3767 foff = bp->b_offset;
3768 KASSERT(bp->b_offset != NOOFFSET,
3769 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3771 vfs_busy_pages_acquire(bp);
3772 vfs_setdirty_range(bp);
3773 for (i = 0; i < bp->b_npages; i++) {
3774 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3776 if (eoff > bp->b_offset + bp->b_bufsize)
3777 eoff = bp->b_offset + bp->b_bufsize;
3779 vfs_page_set_validclean(bp, foff, m);
3780 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3783 vfs_busy_pages_release(bp);
3787 vfs_setdirty_range(struct buf *bp)
3789 vm_offset_t boffset;
3790 vm_offset_t eoffset;
3794 * test the pages to see if they have been modified directly
3795 * by users through the VM system.
3797 for (i = 0; i < bp->b_npages; i++)
3798 vm_page_test_dirty(bp->b_pages[i]);
3801 * Calculate the encompassing dirty range, boffset and eoffset,
3802 * (eoffset - boffset) bytes.
3805 for (i = 0; i < bp->b_npages; i++) {
3806 if (bp->b_pages[i]->dirty)
3809 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3811 for (i = bp->b_npages - 1; i >= 0; --i) {
3812 if (bp->b_pages[i]->dirty) {
3816 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3819 * Fit it to the buffer.
3822 if (eoffset > bp->b_bcount)
3823 eoffset = bp->b_bcount;
3826 * If we have a good dirty range, merge with the existing
3830 if (boffset < eoffset) {
3831 if (bp->b_dirtyoff > boffset)
3832 bp->b_dirtyoff = boffset;
3833 if (bp->b_dirtyend < eoffset)
3834 bp->b_dirtyend = eoffset;
3839 * Allocate the KVA mapping for an existing buffer.
3840 * If an unmapped buffer is provided but a mapped buffer is requested, take
3841 * also care to properly setup mappings between pages and KVA.
3844 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3846 int bsize, maxsize, need_mapping, need_kva;
3849 need_mapping = bp->b_data == unmapped_buf &&
3850 (gbflags & GB_UNMAPPED) == 0;
3851 need_kva = bp->b_kvabase == unmapped_buf &&
3852 bp->b_data == unmapped_buf &&
3853 (gbflags & GB_KVAALLOC) != 0;
3854 if (!need_mapping && !need_kva)
3857 BUF_CHECK_UNMAPPED(bp);
3859 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3861 * Buffer is not mapped, but the KVA was already
3862 * reserved at the time of the instantiation. Use the
3869 * Calculate the amount of the address space we would reserve
3870 * if the buffer was mapped.
3872 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3873 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3874 offset = blkno * bsize;
3875 maxsize = size + (offset & PAGE_MASK);
3876 maxsize = imax(maxsize, bsize);
3878 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3879 if ((gbflags & GB_NOWAIT_BD) != 0) {
3881 * XXXKIB: defragmentation cannot
3882 * succeed, not sure what else to do.
3884 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3886 counter_u64_add(mappingrestarts, 1);
3887 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3891 /* b_offset is handled by bpmap_qenter. */
3892 bp->b_data = bp->b_kvabase;
3893 BUF_CHECK_MAPPED(bp);
3899 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3905 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3914 * Get a block given a specified block and offset into a file/device.
3915 * The buffers B_DONE bit will be cleared on return, making it almost
3916 * ready for an I/O initiation. B_INVAL may or may not be set on
3917 * return. The caller should clear B_INVAL prior to initiating a
3920 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3921 * an existing buffer.
3923 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3924 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3925 * and then cleared based on the backing VM. If the previous buffer is
3926 * non-0-sized but invalid, B_CACHE will be cleared.
3928 * If getblk() must create a new buffer, the new buffer is returned with
3929 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3930 * case it is returned with B_INVAL clear and B_CACHE set based on the
3933 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3934 * B_CACHE bit is clear.
3936 * What this means, basically, is that the caller should use B_CACHE to
3937 * determine whether the buffer is fully valid or not and should clear
3938 * B_INVAL prior to issuing a read. If the caller intends to validate
3939 * the buffer by loading its data area with something, the caller needs
3940 * to clear B_INVAL. If the caller does this without issuing an I/O,
3941 * the caller should set B_CACHE ( as an optimization ), else the caller
3942 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3943 * a write attempt or if it was a successful read. If the caller
3944 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3945 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3947 * The blkno parameter is the logical block being requested. Normally
3948 * the mapping of logical block number to disk block address is done
3949 * by calling VOP_BMAP(). However, if the mapping is already known, the
3950 * disk block address can be passed using the dblkno parameter. If the
3951 * disk block address is not known, then the same value should be passed
3952 * for blkno and dblkno.
3955 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3956 int slptimeo, int flags, struct buf **bpp)
3961 int bsize, error, maxsize, vmio;
3964 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3965 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3966 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3967 if (vp->v_type != VCHR)
3968 ASSERT_VOP_LOCKED(vp, "getblk");
3969 if (size > maxbcachebuf)
3970 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3972 if (!unmapped_buf_allowed)
3973 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3978 /* Attempt lockless lookup first. */
3979 bp = gbincore_unlocked(bo, blkno);
3982 * With GB_NOCREAT we must be sure about not finding the buffer
3983 * as it may have been reassigned during unlocked lookup.
3985 if ((flags & GB_NOCREAT) != 0)
3987 goto newbuf_unlocked;
3990 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3995 /* Verify buf identify has not changed since lookup. */
3996 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3997 goto foundbuf_fastpath;
3999 /* It changed, fallback to locked lookup. */
4004 bp = gbincore(bo, blkno);
4009 * Buffer is in-core. If the buffer is not busy nor managed,
4010 * it must be on a queue.
4012 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
4013 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
4015 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4018 error = BUF_TIMELOCK(bp, lockflags,
4019 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4022 * If we slept and got the lock we have to restart in case
4023 * the buffer changed identities.
4025 if (error == ENOLCK)
4027 /* We timed out or were interrupted. */
4028 else if (error != 0)
4032 /* If recursed, assume caller knows the rules. */
4033 if (BUF_LOCKRECURSED(bp))
4037 * The buffer is locked. B_CACHE is cleared if the buffer is
4038 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4039 * and for a VMIO buffer B_CACHE is adjusted according to the
4042 if (bp->b_flags & B_INVAL)
4043 bp->b_flags &= ~B_CACHE;
4044 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4045 bp->b_flags |= B_CACHE;
4046 if (bp->b_flags & B_MANAGED)
4047 MPASS(bp->b_qindex == QUEUE_NONE);
4052 * check for size inconsistencies for non-VMIO case.
4054 if (bp->b_bcount != size) {
4055 if ((bp->b_flags & B_VMIO) == 0 ||
4056 (size > bp->b_kvasize)) {
4057 if (bp->b_flags & B_DELWRI) {
4058 bp->b_flags |= B_NOCACHE;
4061 if (LIST_EMPTY(&bp->b_dep)) {
4062 bp->b_flags |= B_RELBUF;
4065 bp->b_flags |= B_NOCACHE;
4074 * Handle the case of unmapped buffer which should
4075 * become mapped, or the buffer for which KVA
4076 * reservation is requested.
4078 bp_unmapped_get_kva(bp, blkno, size, flags);
4081 * If the size is inconsistent in the VMIO case, we can resize
4082 * the buffer. This might lead to B_CACHE getting set or
4083 * cleared. If the size has not changed, B_CACHE remains
4084 * unchanged from its previous state.
4088 KASSERT(bp->b_offset != NOOFFSET,
4089 ("getblk: no buffer offset"));
4092 * A buffer with B_DELWRI set and B_CACHE clear must
4093 * be committed before we can return the buffer in
4094 * order to prevent the caller from issuing a read
4095 * ( due to B_CACHE not being set ) and overwriting
4098 * Most callers, including NFS and FFS, need this to
4099 * operate properly either because they assume they
4100 * can issue a read if B_CACHE is not set, or because
4101 * ( for example ) an uncached B_DELWRI might loop due
4102 * to softupdates re-dirtying the buffer. In the latter
4103 * case, B_CACHE is set after the first write completes,
4104 * preventing further loops.
4105 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4106 * above while extending the buffer, we cannot allow the
4107 * buffer to remain with B_CACHE set after the write
4108 * completes or it will represent a corrupt state. To
4109 * deal with this we set B_NOCACHE to scrap the buffer
4112 * We might be able to do something fancy, like setting
4113 * B_CACHE in bwrite() except if B_DELWRI is already set,
4114 * so the below call doesn't set B_CACHE, but that gets real
4115 * confusing. This is much easier.
4118 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4119 bp->b_flags |= B_NOCACHE;
4123 bp->b_flags &= ~B_DONE;
4126 * Buffer is not in-core, create new buffer. The buffer
4127 * returned by getnewbuf() is locked. Note that the returned
4128 * buffer is also considered valid (not marked B_INVAL).
4133 * If the user does not want us to create the buffer, bail out
4136 if (flags & GB_NOCREAT)
4139 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4140 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4141 offset = blkno * bsize;
4142 vmio = vp->v_object != NULL;
4144 maxsize = size + (offset & PAGE_MASK);
4147 /* Do not allow non-VMIO notmapped buffers. */
4148 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4150 maxsize = imax(maxsize, bsize);
4151 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4153 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4154 KASSERT(error != EOPNOTSUPP,
4155 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4160 return (EJUSTRETURN);
4163 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4165 if (slpflag || slptimeo)
4168 * XXX This is here until the sleep path is diagnosed
4169 * enough to work under very low memory conditions.
4171 * There's an issue on low memory, 4BSD+non-preempt
4172 * systems (eg MIPS routers with 32MB RAM) where buffer
4173 * exhaustion occurs without sleeping for buffer
4174 * reclaimation. This just sticks in a loop and
4175 * constantly attempts to allocate a buffer, which
4176 * hits exhaustion and tries to wakeup bufdaemon.
4177 * This never happens because we never yield.
4179 * The real solution is to identify and fix these cases
4180 * so we aren't effectively busy-waiting in a loop
4181 * until the reclaimation path has cycles to run.
4183 kern_yield(PRI_USER);
4188 * This code is used to make sure that a buffer is not
4189 * created while the getnewbuf routine is blocked.
4190 * This can be a problem whether the vnode is locked or not.
4191 * If the buffer is created out from under us, we have to
4192 * throw away the one we just created.
4194 * Note: this must occur before we associate the buffer
4195 * with the vp especially considering limitations in
4196 * the splay tree implementation when dealing with duplicate
4200 if (gbincore(bo, blkno)) {
4202 bp->b_flags |= B_INVAL;
4203 bufspace_release(bufdomain(bp), maxsize);
4209 * Insert the buffer into the hash, so that it can
4210 * be found by incore.
4212 bp->b_lblkno = blkno;
4213 bp->b_blkno = d_blkno;
4214 bp->b_offset = offset;
4219 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4220 * buffer size starts out as 0, B_CACHE will be set by
4221 * allocbuf() for the VMIO case prior to it testing the
4222 * backing store for validity.
4226 bp->b_flags |= B_VMIO;
4227 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4228 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4229 bp, vp->v_object, bp->b_bufobj->bo_object));
4231 bp->b_flags &= ~B_VMIO;
4232 KASSERT(bp->b_bufobj->bo_object == NULL,
4233 ("ARGH! has b_bufobj->bo_object %p %p\n",
4234 bp, bp->b_bufobj->bo_object));
4235 BUF_CHECK_MAPPED(bp);
4239 bufspace_release(bufdomain(bp), maxsize);
4240 bp->b_flags &= ~B_DONE;
4242 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4244 buf_track(bp, __func__);
4245 KASSERT(bp->b_bufobj == bo,
4246 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4252 * Get an empty, disassociated buffer of given size. The buffer is initially
4256 geteblk(int size, int flags)
4261 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4262 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4263 if ((flags & GB_NOWAIT_BD) &&
4264 (curthread->td_pflags & TDP_BUFNEED) != 0)
4268 bufspace_release(bufdomain(bp), maxsize);
4269 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4274 * Truncate the backing store for a non-vmio buffer.
4277 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4280 if (bp->b_flags & B_MALLOC) {
4282 * malloced buffers are not shrunk
4284 if (newbsize == 0) {
4285 bufmallocadjust(bp, 0);
4286 free(bp->b_data, M_BIOBUF);
4287 bp->b_data = bp->b_kvabase;
4288 bp->b_flags &= ~B_MALLOC;
4292 vm_hold_free_pages(bp, newbsize);
4293 bufspace_adjust(bp, newbsize);
4297 * Extend the backing for a non-VMIO buffer.
4300 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4306 * We only use malloced memory on the first allocation.
4307 * and revert to page-allocated memory when the buffer
4310 * There is a potential smp race here that could lead
4311 * to bufmallocspace slightly passing the max. It
4312 * is probably extremely rare and not worth worrying
4315 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4316 bufmallocspace < maxbufmallocspace) {
4317 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4318 bp->b_flags |= B_MALLOC;
4319 bufmallocadjust(bp, newbsize);
4324 * If the buffer is growing on its other-than-first
4325 * allocation then we revert to the page-allocation
4330 if (bp->b_flags & B_MALLOC) {
4331 origbuf = bp->b_data;
4332 origbufsize = bp->b_bufsize;
4333 bp->b_data = bp->b_kvabase;
4334 bufmallocadjust(bp, 0);
4335 bp->b_flags &= ~B_MALLOC;
4336 newbsize = round_page(newbsize);
4338 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4339 (vm_offset_t) bp->b_data + newbsize);
4340 if (origbuf != NULL) {
4341 bcopy(origbuf, bp->b_data, origbufsize);
4342 free(origbuf, M_BIOBUF);
4344 bufspace_adjust(bp, newbsize);
4348 * This code constitutes the buffer memory from either anonymous system
4349 * memory (in the case of non-VMIO operations) or from an associated
4350 * VM object (in the case of VMIO operations). This code is able to
4351 * resize a buffer up or down.
4353 * Note that this code is tricky, and has many complications to resolve
4354 * deadlock or inconsistent data situations. Tread lightly!!!
4355 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4356 * the caller. Calling this code willy nilly can result in the loss of data.
4358 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4359 * B_CACHE for the non-VMIO case.
4362 allocbuf(struct buf *bp, int size)
4366 if (bp->b_bcount == size)
4369 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size,
4370 ("allocbuf: buffer too small %p %#x %#x",
4371 bp, bp->b_kvasize, size));
4373 newbsize = roundup2(size, DEV_BSIZE);
4374 if ((bp->b_flags & B_VMIO) == 0) {
4375 if ((bp->b_flags & B_MALLOC) == 0)
4376 newbsize = round_page(newbsize);
4378 * Just get anonymous memory from the kernel. Don't
4379 * mess with B_CACHE.
4381 if (newbsize < bp->b_bufsize)
4382 vfs_nonvmio_truncate(bp, newbsize);
4383 else if (newbsize > bp->b_bufsize)
4384 vfs_nonvmio_extend(bp, newbsize);
4388 desiredpages = size == 0 ? 0 :
4389 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4391 KASSERT((bp->b_flags & B_MALLOC) == 0,
4392 ("allocbuf: VMIO buffer can't be malloced %p", bp));
4395 * Set B_CACHE initially if buffer is 0 length or will become
4398 if (size == 0 || bp->b_bufsize == 0)
4399 bp->b_flags |= B_CACHE;
4401 if (newbsize < bp->b_bufsize)
4402 vfs_vmio_truncate(bp, desiredpages);
4403 /* XXX This looks as if it should be newbsize > b_bufsize */
4404 else if (size > bp->b_bcount)
4405 vfs_vmio_extend(bp, desiredpages, size);
4406 bufspace_adjust(bp, newbsize);
4408 bp->b_bcount = size; /* requested buffer size. */
4412 extern int inflight_transient_maps;
4414 static struct bio_queue nondump_bios;
4417 biodone(struct bio *bp)
4420 void (*done)(struct bio *);
4421 vm_offset_t start, end;
4423 biotrack(bp, __func__);
4426 * Avoid completing I/O when dumping after a panic since that may
4427 * result in a deadlock in the filesystem or pager code. Note that
4428 * this doesn't affect dumps that were started manually since we aim
4429 * to keep the system usable after it has been resumed.
4431 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4432 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4435 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4436 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4437 bp->bio_flags |= BIO_UNMAPPED;
4438 start = trunc_page((vm_offset_t)bp->bio_data);
4439 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4440 bp->bio_data = unmapped_buf;
4441 pmap_qremove(start, atop(end - start));
4442 vmem_free(transient_arena, start, end - start);
4443 atomic_add_int(&inflight_transient_maps, -1);
4445 done = bp->bio_done;
4447 * The check for done == biodone is to allow biodone to be
4448 * used as a bio_done routine.
4450 if (done == NULL || done == biodone) {
4451 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4453 bp->bio_flags |= BIO_DONE;
4461 * Wait for a BIO to finish.
4464 biowait(struct bio *bp, const char *wmesg)
4468 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4470 while ((bp->bio_flags & BIO_DONE) == 0)
4471 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4473 if (bp->bio_error != 0)
4474 return (bp->bio_error);
4475 if (!(bp->bio_flags & BIO_ERROR))
4481 biofinish(struct bio *bp, struct devstat *stat, int error)
4485 bp->bio_error = error;
4486 bp->bio_flags |= BIO_ERROR;
4489 devstat_end_transaction_bio(stat, bp);
4493 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4495 biotrack_buf(struct bio *bp, const char *location)
4498 buf_track(bp->bio_track_bp, location);
4505 * Wait for buffer I/O completion, returning error status. The buffer
4506 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4507 * error and cleared.
4510 bufwait(struct buf *bp)
4512 if (bp->b_iocmd == BIO_READ)
4513 bwait(bp, PRIBIO, "biord");
4515 bwait(bp, PRIBIO, "biowr");
4516 if (bp->b_flags & B_EINTR) {
4517 bp->b_flags &= ~B_EINTR;
4520 if (bp->b_ioflags & BIO_ERROR) {
4521 return (bp->b_error ? bp->b_error : EIO);
4530 * Finish I/O on a buffer, optionally calling a completion function.
4531 * This is usually called from an interrupt so process blocking is
4534 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4535 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4536 * assuming B_INVAL is clear.
4538 * For the VMIO case, we set B_CACHE if the op was a read and no
4539 * read error occurred, or if the op was a write. B_CACHE is never
4540 * set if the buffer is invalid or otherwise uncacheable.
4542 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4543 * initiator to leave B_INVAL set to brelse the buffer out of existence
4544 * in the biodone routine.
4547 bufdone(struct buf *bp)
4549 struct bufobj *dropobj;
4550 void (*biodone)(struct buf *);
4552 buf_track(bp, __func__);
4553 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4556 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4558 runningbufwakeup(bp);
4559 if (bp->b_iocmd == BIO_WRITE)
4560 dropobj = bp->b_bufobj;
4561 /* call optional completion function if requested */
4562 if (bp->b_iodone != NULL) {
4563 biodone = bp->b_iodone;
4564 bp->b_iodone = NULL;
4567 bufobj_wdrop(dropobj);
4570 if (bp->b_flags & B_VMIO) {
4572 * Set B_CACHE if the op was a normal read and no error
4573 * occurred. B_CACHE is set for writes in the b*write()
4576 if (bp->b_iocmd == BIO_READ &&
4577 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4578 !(bp->b_ioflags & BIO_ERROR))
4579 bp->b_flags |= B_CACHE;
4580 vfs_vmio_iodone(bp);
4582 if (!LIST_EMPTY(&bp->b_dep))
4584 if ((bp->b_flags & B_CKHASH) != 0) {
4585 KASSERT(bp->b_iocmd == BIO_READ,
4586 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4587 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4588 (*bp->b_ckhashcalc)(bp);
4591 * For asynchronous completions, release the buffer now. The brelse
4592 * will do a wakeup there if necessary - so no need to do a wakeup
4593 * here in the async case. The sync case always needs to do a wakeup.
4595 if (bp->b_flags & B_ASYNC) {
4596 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4597 (bp->b_ioflags & BIO_ERROR))
4604 bufobj_wdrop(dropobj);
4608 * This routine is called in lieu of iodone in the case of
4609 * incomplete I/O. This keeps the busy status for pages
4613 vfs_unbusy_pages(struct buf *bp)
4619 runningbufwakeup(bp);
4620 if (!(bp->b_flags & B_VMIO))
4623 obj = bp->b_bufobj->bo_object;
4624 for (i = 0; i < bp->b_npages; i++) {
4626 if (m == bogus_page) {
4627 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4629 panic("vfs_unbusy_pages: page missing\n");
4631 if (buf_mapped(bp)) {
4632 BUF_CHECK_MAPPED(bp);
4633 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4634 bp->b_pages, bp->b_npages);
4636 BUF_CHECK_UNMAPPED(bp);
4640 vm_object_pip_wakeupn(obj, bp->b_npages);
4644 * vfs_page_set_valid:
4646 * Set the valid bits in a page based on the supplied offset. The
4647 * range is restricted to the buffer's size.
4649 * This routine is typically called after a read completes.
4652 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4657 * Compute the end offset, eoff, such that [off, eoff) does not span a
4658 * page boundary and eoff is not greater than the end of the buffer.
4659 * The end of the buffer, in this case, is our file EOF, not the
4660 * allocation size of the buffer.
4662 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4663 if (eoff > bp->b_offset + bp->b_bcount)
4664 eoff = bp->b_offset + bp->b_bcount;
4667 * Set valid range. This is typically the entire buffer and thus the
4671 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4675 * vfs_page_set_validclean:
4677 * Set the valid bits and clear the dirty bits in a page based on the
4678 * supplied offset. The range is restricted to the buffer's size.
4681 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4683 vm_ooffset_t soff, eoff;
4686 * Start and end offsets in buffer. eoff - soff may not cross a
4687 * page boundary or cross the end of the buffer. The end of the
4688 * buffer, in this case, is our file EOF, not the allocation size
4692 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4693 if (eoff > bp->b_offset + bp->b_bcount)
4694 eoff = bp->b_offset + bp->b_bcount;
4697 * Set valid range. This is typically the entire buffer and thus the
4701 vm_page_set_validclean(
4703 (vm_offset_t) (soff & PAGE_MASK),
4704 (vm_offset_t) (eoff - soff)
4710 * Acquire a shared busy on all pages in the buf.
4713 vfs_busy_pages_acquire(struct buf *bp)
4717 for (i = 0; i < bp->b_npages; i++)
4718 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4722 vfs_busy_pages_release(struct buf *bp)
4726 for (i = 0; i < bp->b_npages; i++)
4727 vm_page_sunbusy(bp->b_pages[i]);
4731 * This routine is called before a device strategy routine.
4732 * It is used to tell the VM system that paging I/O is in
4733 * progress, and treat the pages associated with the buffer
4734 * almost as being exclusive busy. Also the object paging_in_progress
4735 * flag is handled to make sure that the object doesn't become
4738 * Since I/O has not been initiated yet, certain buffer flags
4739 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4740 * and should be ignored.
4743 vfs_busy_pages(struct buf *bp, int clear_modify)
4751 if (!(bp->b_flags & B_VMIO))
4754 obj = bp->b_bufobj->bo_object;
4755 foff = bp->b_offset;
4756 KASSERT(bp->b_offset != NOOFFSET,
4757 ("vfs_busy_pages: no buffer offset"));
4758 if ((bp->b_flags & B_CLUSTER) == 0) {
4759 vm_object_pip_add(obj, bp->b_npages);
4760 vfs_busy_pages_acquire(bp);
4762 if (bp->b_bufsize != 0)
4763 vfs_setdirty_range(bp);
4765 for (i = 0; i < bp->b_npages; i++) {
4767 vm_page_assert_sbusied(m);
4770 * When readying a buffer for a read ( i.e
4771 * clear_modify == 0 ), it is important to do
4772 * bogus_page replacement for valid pages in
4773 * partially instantiated buffers. Partially
4774 * instantiated buffers can, in turn, occur when
4775 * reconstituting a buffer from its VM backing store
4776 * base. We only have to do this if B_CACHE is
4777 * clear ( which causes the I/O to occur in the
4778 * first place ). The replacement prevents the read
4779 * I/O from overwriting potentially dirty VM-backed
4780 * pages. XXX bogus page replacement is, uh, bogus.
4781 * It may not work properly with small-block devices.
4782 * We need to find a better way.
4785 pmap_remove_write(m);
4786 vfs_page_set_validclean(bp, foff, m);
4787 } else if (vm_page_all_valid(m) &&
4788 (bp->b_flags & B_CACHE) == 0) {
4789 bp->b_pages[i] = bogus_page;
4792 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4794 if (bogus && buf_mapped(bp)) {
4795 BUF_CHECK_MAPPED(bp);
4796 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4797 bp->b_pages, bp->b_npages);
4802 * vfs_bio_set_valid:
4804 * Set the range within the buffer to valid. The range is
4805 * relative to the beginning of the buffer, b_offset. Note that
4806 * b_offset itself may be offset from the beginning of the first
4810 vfs_bio_set_valid(struct buf *bp, int base, int size)
4815 if (!(bp->b_flags & B_VMIO))
4819 * Fixup base to be relative to beginning of first page.
4820 * Set initial n to be the maximum number of bytes in the
4821 * first page that can be validated.
4823 base += (bp->b_offset & PAGE_MASK);
4824 n = PAGE_SIZE - (base & PAGE_MASK);
4827 * Busy may not be strictly necessary here because the pages are
4828 * unlikely to be fully valid and the vnode lock will synchronize
4829 * their access via getpages. It is grabbed for consistency with
4830 * other page validation.
4832 vfs_busy_pages_acquire(bp);
4833 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4837 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4842 vfs_busy_pages_release(bp);
4848 * If the specified buffer is a non-VMIO buffer, clear the entire
4849 * buffer. If the specified buffer is a VMIO buffer, clear and
4850 * validate only the previously invalid portions of the buffer.
4851 * This routine essentially fakes an I/O, so we need to clear
4852 * BIO_ERROR and B_INVAL.
4854 * Note that while we only theoretically need to clear through b_bcount,
4855 * we go ahead and clear through b_bufsize.
4858 vfs_bio_clrbuf(struct buf *bp)
4860 int i, j, sa, ea, slide, zbits;
4861 vm_page_bits_t mask;
4863 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4867 bp->b_flags &= ~B_INVAL;
4868 bp->b_ioflags &= ~BIO_ERROR;
4869 vfs_busy_pages_acquire(bp);
4870 sa = bp->b_offset & PAGE_MASK;
4872 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4873 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4874 ea = slide & PAGE_MASK;
4877 if (bp->b_pages[i] == bogus_page)
4880 zbits = (sizeof(vm_page_bits_t) * NBBY) -
4881 (ea - sa) / DEV_BSIZE;
4882 mask = (VM_PAGE_BITS_ALL >> zbits) << j;
4883 if ((bp->b_pages[i]->valid & mask) == mask)
4885 if ((bp->b_pages[i]->valid & mask) == 0)
4886 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4888 for (; sa < ea; sa += DEV_BSIZE, j++) {
4889 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4890 pmap_zero_page_area(bp->b_pages[i],
4895 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4896 roundup2(ea - sa, DEV_BSIZE));
4898 vfs_busy_pages_release(bp);
4903 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4908 if (buf_mapped(bp)) {
4909 BUF_CHECK_MAPPED(bp);
4910 bzero(bp->b_data + base, size);
4912 BUF_CHECK_UNMAPPED(bp);
4913 n = PAGE_SIZE - (base & PAGE_MASK);
4914 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4918 pmap_zero_page_area(m, base & PAGE_MASK, n);
4927 * Update buffer flags based on I/O request parameters, optionally releasing the
4928 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4929 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4930 * I/O). Otherwise the buffer is released to the cache.
4933 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4936 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4937 ("buf %p non-VMIO noreuse", bp));
4939 if ((ioflag & IO_DIRECT) != 0)
4940 bp->b_flags |= B_DIRECT;
4941 if ((ioflag & IO_EXT) != 0)
4942 bp->b_xflags |= BX_ALTDATA;
4943 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4944 bp->b_flags |= B_RELBUF;
4945 if ((ioflag & IO_NOREUSE) != 0)
4946 bp->b_flags |= B_NOREUSE;
4954 vfs_bio_brelse(struct buf *bp, int ioflag)
4957 b_io_dismiss(bp, ioflag, true);
4961 vfs_bio_set_flags(struct buf *bp, int ioflag)
4964 b_io_dismiss(bp, ioflag, false);
4968 * vm_hold_load_pages and vm_hold_free_pages get pages into
4969 * a buffers address space. The pages are anonymous and are
4970 * not associated with a file object.
4973 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4979 BUF_CHECK_MAPPED(bp);
4981 to = round_page(to);
4982 from = round_page(from);
4983 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4984 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4985 KASSERT(to - from <= maxbcachebuf,
4986 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4987 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4989 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4991 * note: must allocate system pages since blocking here
4992 * could interfere with paging I/O, no matter which
4995 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4996 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4997 pmap_qenter(pg, &p, 1);
4998 bp->b_pages[index] = p;
5000 bp->b_npages = index;
5003 /* Return pages associated with this buf to the vm system */
5005 vm_hold_free_pages(struct buf *bp, int newbsize)
5009 int index, newnpages;
5011 BUF_CHECK_MAPPED(bp);
5013 from = round_page((vm_offset_t)bp->b_data + newbsize);
5014 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5015 if (bp->b_npages > newnpages)
5016 pmap_qremove(from, bp->b_npages - newnpages);
5017 for (index = newnpages; index < bp->b_npages; index++) {
5018 p = bp->b_pages[index];
5019 bp->b_pages[index] = NULL;
5020 vm_page_unwire_noq(p);
5023 bp->b_npages = newnpages;
5027 * Map an IO request into kernel virtual address space.
5029 * All requests are (re)mapped into kernel VA space.
5030 * Notice that we use b_bufsize for the size of the buffer
5031 * to be mapped. b_bcount might be modified by the driver.
5033 * Note that even if the caller determines that the address space should
5034 * be valid, a race or a smaller-file mapped into a larger space may
5035 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5036 * check the return value.
5038 * This function only works with pager buffers.
5041 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5046 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5047 prot = VM_PROT_READ;
5048 if (bp->b_iocmd == BIO_READ)
5049 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5050 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5051 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5054 bp->b_bufsize = len;
5055 bp->b_npages = pidx;
5056 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5057 if (mapbuf || !unmapped_buf_allowed) {
5058 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5059 bp->b_data = bp->b_kvabase + bp->b_offset;
5061 bp->b_data = unmapped_buf;
5066 * Free the io map PTEs associated with this IO operation.
5067 * We also invalidate the TLB entries and restore the original b_addr.
5069 * This function only works with pager buffers.
5072 vunmapbuf(struct buf *bp)
5076 npages = bp->b_npages;
5078 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5079 vm_page_unhold_pages(bp->b_pages, npages);
5081 bp->b_data = unmapped_buf;
5085 bdone(struct buf *bp)
5089 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5091 bp->b_flags |= B_DONE;
5097 bwait(struct buf *bp, u_char pri, const char *wchan)
5101 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5103 while ((bp->b_flags & B_DONE) == 0)
5104 msleep(bp, mtxp, pri, wchan, 0);
5109 bufsync(struct bufobj *bo, int waitfor)
5112 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5116 bufstrategy(struct bufobj *bo, struct buf *bp)
5122 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5123 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5124 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5125 i = VOP_STRATEGY(vp, bp);
5126 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5130 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5133 bufobj_init(struct bufobj *bo, void *private)
5135 static volatile int bufobj_cleanq;
5138 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5139 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5140 bo->bo_private = private;
5141 TAILQ_INIT(&bo->bo_clean.bv_hd);
5142 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5146 bufobj_wrefl(struct bufobj *bo)
5149 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5150 ASSERT_BO_WLOCKED(bo);
5155 bufobj_wref(struct bufobj *bo)
5158 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5165 bufobj_wdrop(struct bufobj *bo)
5168 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5170 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5171 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5172 bo->bo_flag &= ~BO_WWAIT;
5173 wakeup(&bo->bo_numoutput);
5179 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5183 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5184 ASSERT_BO_WLOCKED(bo);
5186 while (bo->bo_numoutput) {
5187 bo->bo_flag |= BO_WWAIT;
5188 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5189 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5197 * Set bio_data or bio_ma for struct bio from the struct buf.
5200 bdata2bio(struct buf *bp, struct bio *bip)
5203 if (!buf_mapped(bp)) {
5204 KASSERT(unmapped_buf_allowed, ("unmapped"));
5205 bip->bio_ma = bp->b_pages;
5206 bip->bio_ma_n = bp->b_npages;
5207 bip->bio_data = unmapped_buf;
5208 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5209 bip->bio_flags |= BIO_UNMAPPED;
5210 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5211 PAGE_SIZE == bp->b_npages,
5212 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5213 (long long)bip->bio_length, bip->bio_ma_n));
5215 bip->bio_data = bp->b_data;
5220 static int buf_pager_relbuf;
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_FLAGS(bufqueues, bufqueues, DB_CMD_MEMSAFE)
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_FLAGS(lockedbufs, lockedbufs, DB_CMD_MEMSAFE)
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_FLAGS(countfreebufs, db_coundfreebufs, DB_CMD_MEMSAFE)
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);