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>
53 #include <sys/bitset.h>
55 #include <sys/counter.h>
57 #include <sys/devicestat.h>
58 #include <sys/eventhandler.h>
61 #include <sys/limits.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
69 #include <sys/racct.h>
70 #include <sys/refcount.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
74 #include <sys/sysctl.h>
75 #include <sys/syscallsubr.h>
77 #include <sys/vmmeter.h>
78 #include <sys/vnode.h>
79 #include <sys/watchdog.h>
80 #include <geom/geom.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/vm_map.h>
90 #include <vm/swap_pager.h>
92 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
94 struct bio_ops bioops; /* I/O operation notification */
96 struct buf_ops buf_ops_bio = {
97 .bop_name = "buf_ops_bio",
98 .bop_write = bufwrite,
99 .bop_strategy = bufstrategy,
101 .bop_bdflush = bufbdflush,
105 struct mtx_padalign bq_lock;
106 TAILQ_HEAD(, buf) bq_queue;
108 uint16_t bq_subqueue;
110 } __aligned(CACHE_LINE_SIZE);
112 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
113 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
114 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
115 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
118 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
119 struct bufqueue bd_dirtyq;
120 struct bufqueue *bd_cleanq;
121 struct mtx_padalign bd_run_lock;
126 long bd_bufspacethresh;
127 int bd_hifreebuffers;
128 int bd_lofreebuffers;
129 int bd_hidirtybuffers;
130 int bd_lodirtybuffers;
131 int bd_dirtybufthresh;
135 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
136 int __aligned(CACHE_LINE_SIZE) bd_running;
137 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
138 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
139 } __aligned(CACHE_LINE_SIZE);
141 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
142 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
143 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
144 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
145 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
146 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
147 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
148 #define BD_DOMAIN(bd) (bd - bdomain)
150 static char *buf; /* buffer header pool */
154 return ((struct buf *)(buf + (sizeof(struct buf) +
155 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
158 caddr_t __read_mostly unmapped_buf;
160 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
161 struct proc *bufdaemonproc;
163 static void vm_hold_free_pages(struct buf *bp, int newbsize);
164 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
166 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
167 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
169 static void vfs_clean_pages_dirty_buf(struct buf *bp);
170 static void vfs_setdirty_range(struct buf *bp);
171 static void vfs_vmio_invalidate(struct buf *bp);
172 static void vfs_vmio_truncate(struct buf *bp, int npages);
173 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
174 static int vfs_bio_clcheck(struct vnode *vp, int size,
175 daddr_t lblkno, daddr_t blkno);
176 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
177 void (*)(struct buf *));
178 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
179 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
180 static void buf_daemon(void);
181 static __inline void bd_wakeup(void);
182 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
183 static void bufkva_reclaim(vmem_t *, int);
184 static void bufkva_free(struct buf *);
185 static int buf_import(void *, void **, int, int, int);
186 static void buf_release(void *, void **, int);
187 static void maxbcachebuf_adjust(void);
188 static inline struct bufdomain *bufdomain(struct buf *);
189 static void bq_remove(struct bufqueue *bq, struct buf *bp);
190 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
191 static int buf_recycle(struct bufdomain *, bool kva);
192 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
193 const char *lockname);
194 static void bd_init(struct bufdomain *bd);
195 static int bd_flushall(struct bufdomain *bd);
196 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
197 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
199 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
200 int vmiodirenable = TRUE;
201 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
202 "Use the VM system for directory writes");
203 long runningbufspace;
204 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
205 "Amount of presently outstanding async buffer io");
206 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
207 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
208 static counter_u64_t bufkvaspace;
209 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
210 "Kernel virtual memory used for buffers");
211 static long maxbufspace;
212 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
213 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
214 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
215 "Maximum allowed value of bufspace (including metadata)");
216 static long bufmallocspace;
217 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
218 "Amount of malloced memory for buffers");
219 static long maxbufmallocspace;
220 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
221 0, "Maximum amount of malloced memory for buffers");
222 static long lobufspace;
223 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
224 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
225 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
226 "Minimum amount of buffers we want to have");
228 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
229 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
230 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
231 "Maximum allowed value of bufspace (excluding metadata)");
233 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
234 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
235 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
236 "Bufspace consumed before waking the daemon to free some");
237 static counter_u64_t buffreekvacnt;
238 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
239 "Number of times we have freed the KVA space from some buffer");
240 static counter_u64_t bufdefragcnt;
241 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
242 "Number of times we have had to repeat buffer allocation to defragment");
243 static long lorunningspace;
244 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
245 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
246 "Minimum preferred space used for in-progress I/O");
247 static long hirunningspace;
248 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
249 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
250 "Maximum amount of space to use for in-progress I/O");
251 int dirtybufferflushes;
252 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
253 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
255 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
256 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
257 int altbufferflushes;
258 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
259 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
260 static int recursiveflushes;
261 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
262 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
263 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
264 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
265 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
266 "Number of buffers that are dirty (has unwritten changes) at the moment");
267 static int lodirtybuffers;
268 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
269 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
270 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
271 "How many buffers we want to have free before bufdaemon can sleep");
272 static int hidirtybuffers;
273 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
274 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
275 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
276 "When the number of dirty buffers is considered severe");
278 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
279 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
280 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
281 "Number of bdwrite to bawrite conversions to clear dirty buffers");
282 static int numfreebuffers;
283 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
284 "Number of free buffers");
285 static int lofreebuffers;
286 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
287 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
288 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
289 "Target number of free buffers");
290 static int hifreebuffers;
291 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
292 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
293 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
294 "Threshold for clean buffer recycling");
295 static counter_u64_t getnewbufcalls;
296 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
297 &getnewbufcalls, "Number of calls to getnewbuf");
298 static counter_u64_t getnewbufrestarts;
299 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
301 "Number of times getnewbuf has had to restart a buffer acquisition");
302 static counter_u64_t mappingrestarts;
303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
305 "Number of times getblk has had to restart a buffer mapping for "
307 static counter_u64_t numbufallocfails;
308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
309 &numbufallocfails, "Number of times buffer allocations failed");
310 static int flushbufqtarget = 100;
311 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
312 "Amount of work to do in flushbufqueues when helping bufdaemon");
313 static counter_u64_t notbufdflushes;
314 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
315 "Number of dirty buffer flushes done by the bufdaemon helpers");
316 static long barrierwrites;
317 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
318 &barrierwrites, 0, "Number of barrier writes");
319 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
320 &unmapped_buf_allowed, 0,
321 "Permit the use of the unmapped i/o");
322 int maxbcachebuf = MAXBCACHEBUF;
323 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
324 "Maximum size of a buffer cache block");
327 * This lock synchronizes access to bd_request.
329 static struct mtx_padalign __exclusive_cache_line bdlock;
332 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
333 * waitrunningbufspace().
335 static struct mtx_padalign __exclusive_cache_line rbreqlock;
338 * Lock that protects bdirtywait.
340 static struct mtx_padalign __exclusive_cache_line bdirtylock;
343 * Wakeup point for bufdaemon, as well as indicator of whether it is already
344 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
347 static int bd_request;
350 * Request for the buf daemon to write more buffers than is indicated by
351 * lodirtybuf. This may be necessary to push out excess dependencies or
352 * defragment the address space where a simple count of the number of dirty
353 * buffers is insufficient to characterize the demand for flushing them.
355 static int bd_speedupreq;
358 * Synchronization (sleep/wakeup) variable for active buffer space requests.
359 * Set when wait starts, cleared prior to wakeup().
360 * Used in runningbufwakeup() and waitrunningbufspace().
362 static int runningbufreq;
365 * Synchronization for bwillwrite() waiters.
367 static int bdirtywait;
370 * Definitions for the buffer free lists.
372 #define QUEUE_NONE 0 /* on no queue */
373 #define QUEUE_EMPTY 1 /* empty buffer headers */
374 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
375 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
376 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
378 /* Maximum number of buffer domains. */
379 #define BUF_DOMAINS 8
381 struct bufdomainset bdlodirty; /* Domains > lodirty */
382 struct bufdomainset bdhidirty; /* Domains > hidirty */
384 /* Configured number of clean queues. */
385 static int __read_mostly buf_domains;
387 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
388 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
389 struct bufqueue __exclusive_cache_line bqempty;
392 * per-cpu empty buffer cache.
397 * Single global constant for BUF_WMESG, to avoid getting multiple references.
398 * buf_wmesg is referred from macros.
400 const char *buf_wmesg = BUF_WMESG;
403 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
408 value = *(long *)arg1;
409 error = sysctl_handle_long(oidp, &value, 0, req);
410 if (error != 0 || req->newptr == NULL)
412 mtx_lock(&rbreqlock);
413 if (arg1 == &hirunningspace) {
414 if (value < lorunningspace)
417 hirunningspace = value;
419 KASSERT(arg1 == &lorunningspace,
420 ("%s: unknown arg1", __func__));
421 if (value > hirunningspace)
424 lorunningspace = value;
426 mtx_unlock(&rbreqlock);
431 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
437 value = *(int *)arg1;
438 error = sysctl_handle_int(oidp, &value, 0, req);
439 if (error != 0 || req->newptr == NULL)
441 *(int *)arg1 = value;
442 for (i = 0; i < buf_domains; i++)
443 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
450 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
456 value = *(long *)arg1;
457 error = sysctl_handle_long(oidp, &value, 0, req);
458 if (error != 0 || req->newptr == NULL)
460 *(long *)arg1 = value;
461 for (i = 0; i < buf_domains; i++)
462 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
468 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
469 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
471 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
478 for (i = 0; i < buf_domains; i++)
479 lvalue += bdomain[i].bd_bufspace;
480 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
481 return (sysctl_handle_long(oidp, &lvalue, 0, req));
482 if (lvalue > INT_MAX)
483 /* On overflow, still write out a long to trigger ENOMEM. */
484 return (sysctl_handle_long(oidp, &lvalue, 0, req));
486 return (sysctl_handle_int(oidp, &ivalue, 0, req));
490 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
496 for (i = 0; i < buf_domains; i++)
497 lvalue += bdomain[i].bd_bufspace;
498 return (sysctl_handle_long(oidp, &lvalue, 0, req));
503 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
509 for (i = 0; i < buf_domains; i++)
510 value += bdomain[i].bd_numdirtybuffers;
511 return (sysctl_handle_int(oidp, &value, 0, req));
517 * Wakeup any bwillwrite() waiters.
522 mtx_lock(&bdirtylock);
527 mtx_unlock(&bdirtylock);
533 * Clear a domain from the appropriate bitsets when dirtybuffers
537 bd_clear(struct bufdomain *bd)
540 mtx_lock(&bdirtylock);
541 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
542 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
543 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
544 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
545 mtx_unlock(&bdirtylock);
551 * Set a domain in the appropriate bitsets when dirtybuffers
555 bd_set(struct bufdomain *bd)
558 mtx_lock(&bdirtylock);
559 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
560 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
561 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
562 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
563 mtx_unlock(&bdirtylock);
569 * Decrement the numdirtybuffers count by one and wakeup any
570 * threads blocked in bwillwrite().
573 bdirtysub(struct buf *bp)
575 struct bufdomain *bd;
579 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
580 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
582 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
589 * Increment the numdirtybuffers count by one and wakeup the buf
593 bdirtyadd(struct buf *bp)
595 struct bufdomain *bd;
599 * Only do the wakeup once as we cross the boundary. The
600 * buf daemon will keep running until the condition clears.
603 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
604 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
606 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
611 * bufspace_daemon_wakeup:
613 * Wakeup the daemons responsible for freeing clean bufs.
616 bufspace_daemon_wakeup(struct bufdomain *bd)
620 * avoid the lock if the daemon is running.
622 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
624 atomic_store_int(&bd->bd_running, 1);
625 wakeup(&bd->bd_running);
631 * bufspace_daemon_wait:
633 * Sleep until the domain falls below a limit or one second passes.
636 bufspace_daemon_wait(struct bufdomain *bd)
639 * Re-check our limits and sleep. bd_running must be
640 * cleared prior to checking the limits to avoid missed
641 * wakeups. The waker will adjust one of bufspace or
642 * freebuffers prior to checking bd_running.
645 atomic_store_int(&bd->bd_running, 0);
646 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
647 bd->bd_freebuffers > bd->bd_lofreebuffers) {
648 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
651 /* Avoid spurious wakeups while running. */
652 atomic_store_int(&bd->bd_running, 1);
660 * Adjust the reported bufspace for a KVA managed buffer, possibly
661 * waking any waiters.
664 bufspace_adjust(struct buf *bp, int bufsize)
666 struct bufdomain *bd;
670 KASSERT((bp->b_flags & B_MALLOC) == 0,
671 ("bufspace_adjust: malloc buf %p", bp));
673 diff = bufsize - bp->b_bufsize;
675 atomic_subtract_long(&bd->bd_bufspace, -diff);
676 } else if (diff > 0) {
677 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
678 /* Wake up the daemon on the transition. */
679 if (space < bd->bd_bufspacethresh &&
680 space + diff >= bd->bd_bufspacethresh)
681 bufspace_daemon_wakeup(bd);
683 bp->b_bufsize = bufsize;
689 * Reserve bufspace before calling allocbuf(). metadata has a
690 * different space limit than data.
693 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
699 limit = bd->bd_maxbufspace;
701 limit = bd->bd_hibufspace;
702 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
705 atomic_subtract_long(&bd->bd_bufspace, size);
709 /* Wake up the daemon on the transition. */
710 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
711 bufspace_daemon_wakeup(bd);
719 * Release reserved bufspace after bufspace_adjust() has consumed it.
722 bufspace_release(struct bufdomain *bd, int size)
725 atomic_subtract_long(&bd->bd_bufspace, size);
731 * Wait for bufspace, acting as the buf daemon if a locked vnode is
732 * supplied. bd_wanted must be set prior to polling for space. The
733 * operation must be re-tried on return.
736 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
737 int slpflag, int slptimeo)
740 int error, fl, norunbuf;
742 if ((gbflags & GB_NOWAIT_BD) != 0)
747 while (bd->bd_wanted) {
748 if (vp != NULL && vp->v_type != VCHR &&
749 (td->td_pflags & TDP_BUFNEED) == 0) {
752 * getblk() is called with a vnode locked, and
753 * some majority of the dirty buffers may as
754 * well belong to the vnode. Flushing the
755 * buffers there would make a progress that
756 * cannot be achieved by the buf_daemon, that
757 * cannot lock the vnode.
759 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
760 (td->td_pflags & TDP_NORUNNINGBUF);
763 * Play bufdaemon. The getnewbuf() function
764 * may be called while the thread owns lock
765 * for another dirty buffer for the same
766 * vnode, which makes it impossible to use
767 * VOP_FSYNC() there, due to the buffer lock
770 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
771 fl = buf_flush(vp, bd, flushbufqtarget);
772 td->td_pflags &= norunbuf;
776 if (bd->bd_wanted == 0)
779 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
780 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
790 * buffer space management daemon. Tries to maintain some marginal
791 * amount of free buffer space so that requesting processes neither
792 * block nor work to reclaim buffers.
795 bufspace_daemon(void *arg)
797 struct bufdomain *bd;
799 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
800 SHUTDOWN_PRI_LAST + 100);
804 kthread_suspend_check();
807 * Free buffers from the clean queue until we meet our
810 * Theory of operation: The buffer cache is most efficient
811 * when some free buffer headers and space are always
812 * available to getnewbuf(). This daemon attempts to prevent
813 * the excessive blocking and synchronization associated
814 * with shortfall. It goes through three phases according
817 * 1) The daemon wakes up voluntarily once per-second
818 * during idle periods when the counters are below
819 * the wakeup thresholds (bufspacethresh, lofreebuffers).
821 * 2) The daemon wakes up as we cross the thresholds
822 * ahead of any potential blocking. This may bounce
823 * slightly according to the rate of consumption and
826 * 3) The daemon and consumers are starved for working
827 * clean buffers. This is the 'bufspace' sleep below
828 * which will inefficiently trade bufs with bqrelse
829 * until we return to condition 2.
831 while (bd->bd_bufspace > bd->bd_lobufspace ||
832 bd->bd_freebuffers < bd->bd_hifreebuffers) {
833 if (buf_recycle(bd, false) != 0) {
837 * Speedup dirty if we've run out of clean
838 * buffers. This is possible in particular
839 * because softdep may held many bufs locked
840 * pending writes to other bufs which are
841 * marked for delayed write, exhausting
842 * clean space until they are written.
847 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
848 PRIBIO|PDROP, "bufspace", hz/10);
854 bufspace_daemon_wait(bd);
861 * Adjust the reported bufspace for a malloc managed buffer, possibly
862 * waking any waiters.
865 bufmallocadjust(struct buf *bp, int bufsize)
869 KASSERT((bp->b_flags & B_MALLOC) != 0,
870 ("bufmallocadjust: non-malloc buf %p", bp));
871 diff = bufsize - bp->b_bufsize;
873 atomic_subtract_long(&bufmallocspace, -diff);
875 atomic_add_long(&bufmallocspace, diff);
876 bp->b_bufsize = bufsize;
882 * Wake up processes that are waiting on asynchronous writes to fall
883 * below lorunningspace.
889 mtx_lock(&rbreqlock);
892 wakeup(&runningbufreq);
894 mtx_unlock(&rbreqlock);
900 * Decrement the outstanding write count according.
903 runningbufwakeup(struct buf *bp)
907 bspace = bp->b_runningbufspace;
910 space = atomic_fetchadd_long(&runningbufspace, -bspace);
911 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
913 bp->b_runningbufspace = 0;
915 * Only acquire the lock and wakeup on the transition from exceeding
916 * the threshold to falling below it.
918 if (space < lorunningspace)
920 if (space - bspace > lorunningspace)
926 * waitrunningbufspace()
928 * runningbufspace is a measure of the amount of I/O currently
929 * running. This routine is used in async-write situations to
930 * prevent creating huge backups of pending writes to a device.
931 * Only asynchronous writes are governed by this function.
933 * This does NOT turn an async write into a sync write. It waits
934 * for earlier writes to complete and generally returns before the
935 * caller's write has reached the device.
938 waitrunningbufspace(void)
941 mtx_lock(&rbreqlock);
942 while (runningbufspace > hirunningspace) {
944 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
946 mtx_unlock(&rbreqlock);
950 * vfs_buf_test_cache:
952 * Called when a buffer is extended. This function clears the B_CACHE
953 * bit if the newly extended portion of the buffer does not contain
957 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
958 vm_offset_t size, vm_page_t m)
962 * This function and its results are protected by higher level
963 * synchronization requiring vnode and buf locks to page in and
966 if (bp->b_flags & B_CACHE) {
967 int base = (foff + off) & PAGE_MASK;
968 if (vm_page_is_valid(m, base, size) == 0)
969 bp->b_flags &= ~B_CACHE;
973 /* Wake up the buffer daemon if necessary */
979 if (bd_request == 0) {
987 * Adjust the maxbcachbuf tunable.
990 maxbcachebuf_adjust(void)
995 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
998 while (i * 2 <= maxbcachebuf)
1001 if (maxbcachebuf < MAXBSIZE)
1002 maxbcachebuf = MAXBSIZE;
1003 if (maxbcachebuf > maxphys)
1004 maxbcachebuf = maxphys;
1005 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1006 printf("maxbcachebuf=%d\n", maxbcachebuf);
1010 * bd_speedup - speedup the buffer cache flushing code
1019 if (bd_speedupreq == 0 || bd_request == 0)
1024 wakeup(&bd_request);
1025 mtx_unlock(&bdlock);
1029 #define TRANSIENT_DENOM 5
1031 #define TRANSIENT_DENOM 10
1035 * Calculating buffer cache scaling values and reserve space for buffer
1036 * headers. This is called during low level kernel initialization and
1037 * may be called more then once. We CANNOT write to the memory area
1038 * being reserved at this time.
1041 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1044 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1047 * physmem_est is in pages. Convert it to kilobytes (assumes
1048 * PAGE_SIZE is >= 1K)
1050 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1052 maxbcachebuf_adjust();
1054 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1055 * For the first 64MB of ram nominally allocate sufficient buffers to
1056 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1057 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1058 * the buffer cache we limit the eventual kva reservation to
1061 * factor represents the 1/4 x ram conversion.
1064 int factor = 4 * BKVASIZE / 1024;
1067 if (physmem_est > 4096)
1068 nbuf += min((physmem_est - 4096) / factor,
1070 if (physmem_est > 65536)
1071 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1072 32 * 1024 * 1024 / (factor * 5));
1074 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1075 nbuf = maxbcache / BKVASIZE;
1080 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1081 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1082 if (nbuf > maxbuf) {
1084 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1090 * Ideal allocation size for the transient bio submap is 10%
1091 * of the maximal space buffer map. This roughly corresponds
1092 * to the amount of the buffer mapped for typical UFS load.
1094 * Clip the buffer map to reserve space for the transient
1095 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1096 * maximum buffer map extent on the platform.
1098 * The fall-back to the maxbuf in case of maxbcache unset,
1099 * allows to not trim the buffer KVA for the architectures
1100 * with ample KVA space.
1102 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1103 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1104 buf_sz = (long)nbuf * BKVASIZE;
1105 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1106 (TRANSIENT_DENOM - 1)) {
1108 * There is more KVA than memory. Do not
1109 * adjust buffer map size, and assign the rest
1110 * of maxbuf to transient map.
1112 biotmap_sz = maxbuf_sz - buf_sz;
1115 * Buffer map spans all KVA we could afford on
1116 * this platform. Give 10% (20% on i386) of
1117 * the buffer map to the transient bio map.
1119 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1120 buf_sz -= biotmap_sz;
1122 if (biotmap_sz / INT_MAX > maxphys)
1123 bio_transient_maxcnt = INT_MAX;
1125 bio_transient_maxcnt = biotmap_sz / maxphys;
1127 * Artificially limit to 1024 simultaneous in-flight I/Os
1128 * using the transient mapping.
1130 if (bio_transient_maxcnt > 1024)
1131 bio_transient_maxcnt = 1024;
1133 nbuf = buf_sz / BKVASIZE;
1137 nswbuf = min(nbuf / 4, 256);
1138 if (nswbuf < NSWBUF_MIN)
1139 nswbuf = NSWBUF_MIN;
1143 * Reserve space for the buffer cache buffers
1146 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1147 atop(maxbcachebuf)) * nbuf;
1152 /* Initialize the buffer subsystem. Called before use of any buffers. */
1159 KASSERT(maxbcachebuf >= MAXBSIZE,
1160 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1162 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1163 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1164 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1165 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1167 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1169 /* finally, initialize each buffer header and stick on empty q */
1170 for (i = 0; i < nbuf; i++) {
1172 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1173 bp->b_flags = B_INVAL;
1174 bp->b_rcred = NOCRED;
1175 bp->b_wcred = NOCRED;
1176 bp->b_qindex = QUEUE_NONE;
1178 bp->b_subqueue = mp_maxid + 1;
1180 bp->b_data = bp->b_kvabase = unmapped_buf;
1181 LIST_INIT(&bp->b_dep);
1183 bq_insert(&bqempty, bp, false);
1187 * maxbufspace is the absolute maximum amount of buffer space we are
1188 * allowed to reserve in KVM and in real terms. The absolute maximum
1189 * is nominally used by metadata. hibufspace is the nominal maximum
1190 * used by most other requests. The differential is required to
1191 * ensure that metadata deadlocks don't occur.
1193 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1194 * this may result in KVM fragmentation which is not handled optimally
1195 * by the system. XXX This is less true with vmem. We could use
1198 maxbufspace = (long)nbuf * BKVASIZE;
1199 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1200 lobufspace = (hibufspace / 20) * 19; /* 95% */
1201 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1204 * Note: The 16 MiB upper limit for hirunningspace was chosen
1205 * arbitrarily and may need further tuning. It corresponds to
1206 * 128 outstanding write IO requests (if IO size is 128 KiB),
1207 * which fits with many RAID controllers' tagged queuing limits.
1208 * The lower 1 MiB limit is the historical upper limit for
1211 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1212 16 * 1024 * 1024), 1024 * 1024);
1213 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1216 * Limit the amount of malloc memory since it is wired permanently into
1217 * the kernel space. Even though this is accounted for in the buffer
1218 * allocation, we don't want the malloced region to grow uncontrolled.
1219 * The malloc scheme improves memory utilization significantly on
1220 * average (small) directories.
1222 maxbufmallocspace = hibufspace / 20;
1225 * Reduce the chance of a deadlock occurring by limiting the number
1226 * of delayed-write dirty buffers we allow to stack up.
1228 hidirtybuffers = nbuf / 4 + 20;
1229 dirtybufthresh = hidirtybuffers * 9 / 10;
1231 * To support extreme low-memory systems, make sure hidirtybuffers
1232 * cannot eat up all available buffer space. This occurs when our
1233 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1234 * buffer space assuming BKVASIZE'd buffers.
1236 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1237 hidirtybuffers >>= 1;
1239 lodirtybuffers = hidirtybuffers / 2;
1242 * lofreebuffers should be sufficient to avoid stalling waiting on
1243 * buf headers under heavy utilization. The bufs in per-cpu caches
1244 * are counted as free but will be unavailable to threads executing
1247 * hifreebuffers is the free target for the bufspace daemon. This
1248 * should be set appropriately to limit work per-iteration.
1250 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1251 hifreebuffers = (3 * lofreebuffers) / 2;
1252 numfreebuffers = nbuf;
1254 /* Setup the kva and free list allocators. */
1255 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1256 buf_zone = uma_zcache_create("buf free cache",
1257 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1258 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1261 * Size the clean queue according to the amount of buffer space.
1262 * One queue per-256mb up to the max. More queues gives better
1263 * concurrency but less accurate LRU.
1265 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1266 for (i = 0 ; i < buf_domains; i++) {
1267 struct bufdomain *bd;
1271 bd->bd_freebuffers = nbuf / buf_domains;
1272 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1273 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1274 bd->bd_bufspace = 0;
1275 bd->bd_maxbufspace = maxbufspace / buf_domains;
1276 bd->bd_hibufspace = hibufspace / buf_domains;
1277 bd->bd_lobufspace = lobufspace / buf_domains;
1278 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1279 bd->bd_numdirtybuffers = 0;
1280 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1281 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1282 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1283 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1284 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1286 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1287 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1288 mappingrestarts = counter_u64_alloc(M_WAITOK);
1289 numbufallocfails = counter_u64_alloc(M_WAITOK);
1290 notbufdflushes = counter_u64_alloc(M_WAITOK);
1291 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1292 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1293 bufkvaspace = counter_u64_alloc(M_WAITOK);
1298 vfs_buf_check_mapped(struct buf *bp)
1301 KASSERT(bp->b_kvabase != unmapped_buf,
1302 ("mapped buf: b_kvabase was not updated %p", bp));
1303 KASSERT(bp->b_data != unmapped_buf,
1304 ("mapped buf: b_data was not updated %p", bp));
1305 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1306 maxphys, ("b_data + b_offset unmapped %p", bp));
1310 vfs_buf_check_unmapped(struct buf *bp)
1313 KASSERT(bp->b_data == unmapped_buf,
1314 ("unmapped buf: corrupted b_data %p", bp));
1317 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1318 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1320 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1321 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1325 isbufbusy(struct buf *bp)
1327 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1328 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1334 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1337 bufshutdown(int show_busybufs)
1339 static int first_buf_printf = 1;
1341 int i, iter, nbusy, pbusy;
1347 * Sync filesystems for shutdown
1349 wdog_kern_pat(WD_LASTVAL);
1350 kern_sync(curthread);
1353 * With soft updates, some buffers that are
1354 * written will be remarked as dirty until other
1355 * buffers are written.
1357 for (iter = pbusy = 0; iter < 20; iter++) {
1359 for (i = nbuf - 1; i >= 0; i--) {
1365 if (first_buf_printf)
1366 printf("All buffers synced.");
1369 if (first_buf_printf) {
1370 printf("Syncing disks, buffers remaining... ");
1371 first_buf_printf = 0;
1373 printf("%d ", nbusy);
1378 wdog_kern_pat(WD_LASTVAL);
1379 kern_sync(curthread);
1383 * Spin for a while to allow interrupt threads to run.
1385 DELAY(50000 * iter);
1388 * Context switch several times to allow interrupt
1391 for (subiter = 0; subiter < 50 * iter; subiter++) {
1392 thread_lock(curthread);
1400 * Count only busy local buffers to prevent forcing
1401 * a fsck if we're just a client of a wedged NFS server
1404 for (i = nbuf - 1; i >= 0; i--) {
1406 if (isbufbusy(bp)) {
1408 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1409 if (bp->b_dev == NULL) {
1410 TAILQ_REMOVE(&mountlist,
1411 bp->b_vp->v_mount, mnt_list);
1416 if (show_busybufs > 0) {
1418 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1419 nbusy, bp, bp->b_vp, bp->b_flags,
1420 (intmax_t)bp->b_blkno,
1421 (intmax_t)bp->b_lblkno);
1422 BUF_LOCKPRINTINFO(bp);
1423 if (show_busybufs > 1)
1431 * Failed to sync all blocks. Indicate this and don't
1432 * unmount filesystems (thus forcing an fsck on reboot).
1434 printf("Giving up on %d buffers\n", nbusy);
1435 DELAY(5000000); /* 5 seconds */
1437 if (!first_buf_printf)
1438 printf("Final sync complete\n");
1440 * Unmount filesystems
1442 if (!KERNEL_PANICKED())
1446 DELAY(100000); /* wait for console output to finish */
1450 bpmap_qenter(struct buf *bp)
1453 BUF_CHECK_MAPPED(bp);
1456 * bp->b_data is relative to bp->b_offset, but
1457 * bp->b_offset may be offset into the first page.
1459 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1460 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1461 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1462 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1465 static inline struct bufdomain *
1466 bufdomain(struct buf *bp)
1469 return (&bdomain[bp->b_domain]);
1472 static struct bufqueue *
1473 bufqueue(struct buf *bp)
1476 switch (bp->b_qindex) {
1479 case QUEUE_SENTINEL:
1484 return (&bufdomain(bp)->bd_dirtyq);
1486 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1490 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1494 * Return the locked bufqueue that bp is a member of.
1496 static struct bufqueue *
1497 bufqueue_acquire(struct buf *bp)
1499 struct bufqueue *bq, *nbq;
1502 * bp can be pushed from a per-cpu queue to the
1503 * cleanq while we're waiting on the lock. Retry
1504 * if the queues don't match.
1522 * Insert the buffer into the appropriate free list. Requires a
1523 * locked buffer on entry and buffer is unlocked before return.
1526 binsfree(struct buf *bp, int qindex)
1528 struct bufdomain *bd;
1529 struct bufqueue *bq;
1531 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1532 ("binsfree: Invalid qindex %d", qindex));
1533 BUF_ASSERT_XLOCKED(bp);
1536 * Handle delayed bremfree() processing.
1538 if (bp->b_flags & B_REMFREE) {
1539 if (bp->b_qindex == qindex) {
1540 bp->b_flags |= B_REUSE;
1541 bp->b_flags &= ~B_REMFREE;
1545 bq = bufqueue_acquire(bp);
1550 if (qindex == QUEUE_CLEAN) {
1551 if (bd->bd_lim != 0)
1552 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1556 bq = &bd->bd_dirtyq;
1557 bq_insert(bq, bp, true);
1563 * Free a buffer to the buf zone once it no longer has valid contents.
1566 buf_free(struct buf *bp)
1569 if (bp->b_flags & B_REMFREE)
1571 if (bp->b_vflags & BV_BKGRDINPROG)
1572 panic("losing buffer 1");
1573 if (bp->b_rcred != NOCRED) {
1574 crfree(bp->b_rcred);
1575 bp->b_rcred = NOCRED;
1577 if (bp->b_wcred != NOCRED) {
1578 crfree(bp->b_wcred);
1579 bp->b_wcred = NOCRED;
1581 if (!LIST_EMPTY(&bp->b_dep))
1584 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1585 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1587 uma_zfree(buf_zone, bp);
1593 * Import bufs into the uma cache from the buf list. The system still
1594 * expects a static array of bufs and much of the synchronization
1595 * around bufs assumes type stable storage. As a result, UMA is used
1596 * only as a per-cpu cache of bufs still maintained on a global list.
1599 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1605 for (i = 0; i < cnt; i++) {
1606 bp = TAILQ_FIRST(&bqempty.bq_queue);
1609 bq_remove(&bqempty, bp);
1612 BQ_UNLOCK(&bqempty);
1620 * Release bufs from the uma cache back to the buffer queues.
1623 buf_release(void *arg, void **store, int cnt)
1625 struct bufqueue *bq;
1631 for (i = 0; i < cnt; i++) {
1633 /* Inline bq_insert() to batch locking. */
1634 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1635 bp->b_flags &= ~(B_AGE | B_REUSE);
1637 bp->b_qindex = bq->bq_index;
1645 * Allocate an empty buffer header.
1648 buf_alloc(struct bufdomain *bd)
1651 int freebufs, error;
1654 * We can only run out of bufs in the buf zone if the average buf
1655 * is less than BKVASIZE. In this case the actual wait/block will
1656 * come from buf_reycle() failing to flush one of these small bufs.
1659 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1661 bp = uma_zalloc(buf_zone, M_NOWAIT);
1663 atomic_add_int(&bd->bd_freebuffers, 1);
1664 bufspace_daemon_wakeup(bd);
1665 counter_u64_add(numbufallocfails, 1);
1669 * Wake-up the bufspace daemon on transition below threshold.
1671 if (freebufs == bd->bd_lofreebuffers)
1672 bufspace_daemon_wakeup(bd);
1674 error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1675 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1679 KASSERT(bp->b_vp == NULL,
1680 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1681 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1682 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1683 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1684 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1685 KASSERT(bp->b_npages == 0,
1686 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1687 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1688 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1689 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1691 bp->b_domain = BD_DOMAIN(bd);
1697 bp->b_blkno = bp->b_lblkno = 0;
1698 bp->b_offset = NOOFFSET;
1704 bp->b_dirtyoff = bp->b_dirtyend = 0;
1705 bp->b_bufobj = NULL;
1706 bp->b_data = bp->b_kvabase = unmapped_buf;
1707 bp->b_fsprivate1 = NULL;
1708 bp->b_fsprivate2 = NULL;
1709 bp->b_fsprivate3 = NULL;
1710 LIST_INIT(&bp->b_dep);
1718 * Free a buffer from the given bufqueue. kva controls whether the
1719 * freed buf must own some kva resources. This is used for
1723 buf_recycle(struct bufdomain *bd, bool kva)
1725 struct bufqueue *bq;
1726 struct buf *bp, *nbp;
1729 counter_u64_add(bufdefragcnt, 1);
1733 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1734 ("buf_recycle: Locks don't match"));
1735 nbp = TAILQ_FIRST(&bq->bq_queue);
1738 * Run scan, possibly freeing data and/or kva mappings on the fly
1741 while ((bp = nbp) != NULL) {
1743 * Calculate next bp (we can only use it if we do not
1744 * release the bqlock).
1746 nbp = TAILQ_NEXT(bp, b_freelist);
1749 * If we are defragging then we need a buffer with
1750 * some kva to reclaim.
1752 if (kva && bp->b_kvasize == 0)
1755 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1759 * Implement a second chance algorithm for frequently
1762 if ((bp->b_flags & B_REUSE) != 0) {
1763 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1764 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1765 bp->b_flags &= ~B_REUSE;
1771 * Skip buffers with background writes in progress.
1773 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1778 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1779 ("buf_recycle: inconsistent queue %d bp %p",
1781 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1782 ("getnewbuf: queue domain %d doesn't match request %d",
1783 bp->b_domain, (int)BD_DOMAIN(bd)));
1785 * NOTE: nbp is now entirely invalid. We can only restart
1786 * the scan from this point on.
1792 * Requeue the background write buffer with error and
1795 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1798 nbp = TAILQ_FIRST(&bq->bq_queue);
1801 bp->b_flags |= B_INVAL;
1814 * Mark the buffer for removal from the appropriate free list.
1818 bremfree(struct buf *bp)
1821 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1822 KASSERT((bp->b_flags & B_REMFREE) == 0,
1823 ("bremfree: buffer %p already marked for delayed removal.", bp));
1824 KASSERT(bp->b_qindex != QUEUE_NONE,
1825 ("bremfree: buffer %p not on a queue.", bp));
1826 BUF_ASSERT_XLOCKED(bp);
1828 bp->b_flags |= B_REMFREE;
1834 * Force an immediate removal from a free list. Used only in nfs when
1835 * it abuses the b_freelist pointer.
1838 bremfreef(struct buf *bp)
1840 struct bufqueue *bq;
1842 bq = bufqueue_acquire(bp);
1848 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1851 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1852 TAILQ_INIT(&bq->bq_queue);
1854 bq->bq_index = qindex;
1855 bq->bq_subqueue = subqueue;
1859 bd_init(struct bufdomain *bd)
1863 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1864 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1865 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1866 for (i = 0; i <= mp_maxid; i++)
1867 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1868 "bufq clean subqueue lock");
1869 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1875 * Removes a buffer from the free list, must be called with the
1876 * correct qlock held.
1879 bq_remove(struct bufqueue *bq, struct buf *bp)
1882 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1883 bp, bp->b_vp, bp->b_flags);
1884 KASSERT(bp->b_qindex != QUEUE_NONE,
1885 ("bq_remove: buffer %p not on a queue.", bp));
1886 KASSERT(bufqueue(bp) == bq,
1887 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1889 BQ_ASSERT_LOCKED(bq);
1890 if (bp->b_qindex != QUEUE_EMPTY) {
1891 BUF_ASSERT_XLOCKED(bp);
1893 KASSERT(bq->bq_len >= 1,
1894 ("queue %d underflow", bp->b_qindex));
1895 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1897 bp->b_qindex = QUEUE_NONE;
1898 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1902 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1906 BQ_ASSERT_LOCKED(bq);
1907 if (bq != bd->bd_cleanq) {
1909 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1910 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1911 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1913 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1915 bd->bd_cleanq->bq_len += bq->bq_len;
1918 if (bd->bd_wanted) {
1920 wakeup(&bd->bd_wanted);
1922 if (bq != bd->bd_cleanq)
1927 bd_flushall(struct bufdomain *bd)
1929 struct bufqueue *bq;
1933 if (bd->bd_lim == 0)
1936 for (i = 0; i <= mp_maxid; i++) {
1937 bq = &bd->bd_subq[i];
1938 if (bq->bq_len == 0)
1950 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1952 struct bufdomain *bd;
1954 if (bp->b_qindex != QUEUE_NONE)
1955 panic("bq_insert: free buffer %p onto another queue?", bp);
1958 if (bp->b_flags & B_AGE) {
1959 /* Place this buf directly on the real queue. */
1960 if (bq->bq_index == QUEUE_CLEAN)
1963 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1966 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1968 bp->b_flags &= ~(B_AGE | B_REUSE);
1970 bp->b_qindex = bq->bq_index;
1971 bp->b_subqueue = bq->bq_subqueue;
1974 * Unlock before we notify so that we don't wakeup a waiter that
1975 * fails a trylock on the buf and sleeps again.
1980 if (bp->b_qindex == QUEUE_CLEAN) {
1982 * Flush the per-cpu queue and notify any waiters.
1984 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1985 bq->bq_len >= bd->bd_lim))
1994 * Free the kva allocation for a buffer.
1998 bufkva_free(struct buf *bp)
2002 if (bp->b_kvasize == 0) {
2003 KASSERT(bp->b_kvabase == unmapped_buf &&
2004 bp->b_data == unmapped_buf,
2005 ("Leaked KVA space on %p", bp));
2006 } else if (buf_mapped(bp))
2007 BUF_CHECK_MAPPED(bp);
2009 BUF_CHECK_UNMAPPED(bp);
2011 if (bp->b_kvasize == 0)
2014 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2015 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2016 counter_u64_add(buffreekvacnt, 1);
2017 bp->b_data = bp->b_kvabase = unmapped_buf;
2024 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2027 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2032 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2033 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2034 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2035 KASSERT(maxsize <= maxbcachebuf,
2036 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2041 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2044 * Buffer map is too fragmented. Request the caller
2045 * to defragment the map.
2049 bp->b_kvabase = (caddr_t)addr;
2050 bp->b_kvasize = maxsize;
2051 counter_u64_add(bufkvaspace, bp->b_kvasize);
2052 if ((gbflags & GB_UNMAPPED) != 0) {
2053 bp->b_data = unmapped_buf;
2054 BUF_CHECK_UNMAPPED(bp);
2056 bp->b_data = bp->b_kvabase;
2057 BUF_CHECK_MAPPED(bp);
2065 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2066 * callback that fires to avoid returning failure.
2069 bufkva_reclaim(vmem_t *vmem, int flags)
2076 for (i = 0; i < 5; i++) {
2077 for (q = 0; q < buf_domains; q++)
2078 if (buf_recycle(&bdomain[q], true) != 0)
2087 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2088 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2089 * the buffer is valid and we do not have to do anything.
2092 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2093 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2101 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2102 if (inmem(vp, *rablkno))
2104 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2105 if ((rabp->b_flags & B_CACHE) != 0) {
2112 racct_add_buf(curproc, rabp, 0);
2113 PROC_UNLOCK(curproc);
2116 td->td_ru.ru_inblock++;
2117 rabp->b_flags |= B_ASYNC;
2118 rabp->b_flags &= ~B_INVAL;
2119 if ((flags & GB_CKHASH) != 0) {
2120 rabp->b_flags |= B_CKHASH;
2121 rabp->b_ckhashcalc = ckhashfunc;
2123 rabp->b_ioflags &= ~BIO_ERROR;
2124 rabp->b_iocmd = BIO_READ;
2125 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2126 rabp->b_rcred = crhold(cred);
2127 vfs_busy_pages(rabp, 0);
2129 rabp->b_iooffset = dbtob(rabp->b_blkno);
2135 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2137 * Get a buffer with the specified data. Look in the cache first. We
2138 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2139 * is set, the buffer is valid and we do not have to do anything, see
2140 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2142 * Always return a NULL buffer pointer (in bpp) when returning an error.
2144 * The blkno parameter is the logical block being requested. Normally
2145 * the mapping of logical block number to disk block address is done
2146 * by calling VOP_BMAP(). However, if the mapping is already known, the
2147 * disk block address can be passed using the dblkno parameter. If the
2148 * disk block address is not known, then the same value should be passed
2149 * for blkno and dblkno.
2152 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2153 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2154 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2158 int error, readwait, rv;
2160 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2163 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2166 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2171 KASSERT(blkno == bp->b_lblkno,
2172 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2173 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2174 flags &= ~GB_NOSPARSE;
2178 * If not found in cache, do some I/O
2181 if ((bp->b_flags & B_CACHE) == 0) {
2184 PROC_LOCK(td->td_proc);
2185 racct_add_buf(td->td_proc, bp, 0);
2186 PROC_UNLOCK(td->td_proc);
2189 td->td_ru.ru_inblock++;
2190 bp->b_iocmd = BIO_READ;
2191 bp->b_flags &= ~B_INVAL;
2192 if ((flags & GB_CKHASH) != 0) {
2193 bp->b_flags |= B_CKHASH;
2194 bp->b_ckhashcalc = ckhashfunc;
2196 if ((flags & GB_CVTENXIO) != 0)
2197 bp->b_xflags |= BX_CVTENXIO;
2198 bp->b_ioflags &= ~BIO_ERROR;
2199 if (bp->b_rcred == NOCRED && cred != NOCRED)
2200 bp->b_rcred = crhold(cred);
2201 vfs_busy_pages(bp, 0);
2202 bp->b_iooffset = dbtob(bp->b_blkno);
2208 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2210 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2224 * Write, release buffer on completion. (Done by iodone
2225 * if async). Do not bother writing anything if the buffer
2228 * Note that we set B_CACHE here, indicating that buffer is
2229 * fully valid and thus cacheable. This is true even of NFS
2230 * now so we set it generally. This could be set either here
2231 * or in biodone() since the I/O is synchronous. We put it
2235 bufwrite(struct buf *bp)
2242 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2243 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2244 bp->b_flags |= B_INVAL | B_RELBUF;
2245 bp->b_flags &= ~B_CACHE;
2249 if (bp->b_flags & B_INVAL) {
2254 if (bp->b_flags & B_BARRIER)
2255 atomic_add_long(&barrierwrites, 1);
2257 oldflags = bp->b_flags;
2259 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2260 ("FFS background buffer should not get here %p", bp));
2264 vp_md = vp->v_vflag & VV_MD;
2269 * Mark the buffer clean. Increment the bufobj write count
2270 * before bundirty() call, to prevent other thread from seeing
2271 * empty dirty list and zero counter for writes in progress,
2272 * falsely indicating that the bufobj is clean.
2274 bufobj_wref(bp->b_bufobj);
2277 bp->b_flags &= ~B_DONE;
2278 bp->b_ioflags &= ~BIO_ERROR;
2279 bp->b_flags |= B_CACHE;
2280 bp->b_iocmd = BIO_WRITE;
2282 vfs_busy_pages(bp, 1);
2285 * Normal bwrites pipeline writes
2287 bp->b_runningbufspace = bp->b_bufsize;
2288 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2293 racct_add_buf(curproc, bp, 1);
2294 PROC_UNLOCK(curproc);
2297 curthread->td_ru.ru_oublock++;
2298 if (oldflags & B_ASYNC)
2300 bp->b_iooffset = dbtob(bp->b_blkno);
2301 buf_track(bp, __func__);
2304 if ((oldflags & B_ASYNC) == 0) {
2305 int rtval = bufwait(bp);
2308 } else if (space > hirunningspace) {
2310 * don't allow the async write to saturate the I/O
2311 * system. We will not deadlock here because
2312 * we are blocking waiting for I/O that is already in-progress
2313 * to complete. We do not block here if it is the update
2314 * or syncer daemon trying to clean up as that can lead
2317 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2318 waitrunningbufspace();
2325 bufbdflush(struct bufobj *bo, struct buf *bp)
2328 struct bufdomain *bd;
2330 bd = &bdomain[bo->bo_domain];
2331 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2332 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2334 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2337 * Try to find a buffer to flush.
2339 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2340 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2342 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2345 panic("bdwrite: found ourselves");
2347 /* Don't countdeps with the bo lock held. */
2348 if (buf_countdeps(nbp, 0)) {
2353 if (nbp->b_flags & B_CLUSTEROK) {
2354 vfs_bio_awrite(nbp);
2359 dirtybufferflushes++;
2368 * Delayed write. (Buffer is marked dirty). Do not bother writing
2369 * anything if the buffer is marked invalid.
2371 * Note that since the buffer must be completely valid, we can safely
2372 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2373 * biodone() in order to prevent getblk from writing the buffer
2374 * out synchronously.
2377 bdwrite(struct buf *bp)
2379 struct thread *td = curthread;
2383 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2384 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2385 KASSERT((bp->b_flags & B_BARRIER) == 0,
2386 ("Barrier request in delayed write %p", bp));
2388 if (bp->b_flags & B_INVAL) {
2394 * If we have too many dirty buffers, don't create any more.
2395 * If we are wildly over our limit, then force a complete
2396 * cleanup. Otherwise, just keep the situation from getting
2397 * out of control. Note that we have to avoid a recursive
2398 * disaster and not try to clean up after our own cleanup!
2402 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2403 td->td_pflags |= TDP_INBDFLUSH;
2405 td->td_pflags &= ~TDP_INBDFLUSH;
2411 * Set B_CACHE, indicating that the buffer is fully valid. This is
2412 * true even of NFS now.
2414 bp->b_flags |= B_CACHE;
2417 * This bmap keeps the system from needing to do the bmap later,
2418 * perhaps when the system is attempting to do a sync. Since it
2419 * is likely that the indirect block -- or whatever other datastructure
2420 * that the filesystem needs is still in memory now, it is a good
2421 * thing to do this. Note also, that if the pageout daemon is
2422 * requesting a sync -- there might not be enough memory to do
2423 * the bmap then... So, this is important to do.
2425 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2426 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2429 buf_track(bp, __func__);
2432 * Set the *dirty* buffer range based upon the VM system dirty
2435 * Mark the buffer pages as clean. We need to do this here to
2436 * satisfy the vnode_pager and the pageout daemon, so that it
2437 * thinks that the pages have been "cleaned". Note that since
2438 * the pages are in a delayed write buffer -- the VFS layer
2439 * "will" see that the pages get written out on the next sync,
2440 * or perhaps the cluster will be completed.
2442 vfs_clean_pages_dirty_buf(bp);
2446 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2447 * due to the softdep code.
2454 * Turn buffer into delayed write request. We must clear BIO_READ and
2455 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2456 * itself to properly update it in the dirty/clean lists. We mark it
2457 * B_DONE to ensure that any asynchronization of the buffer properly
2458 * clears B_DONE ( else a panic will occur later ).
2460 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2461 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2462 * should only be called if the buffer is known-good.
2464 * Since the buffer is not on a queue, we do not update the numfreebuffers
2467 * The buffer must be on QUEUE_NONE.
2470 bdirty(struct buf *bp)
2473 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2474 bp, bp->b_vp, bp->b_flags);
2475 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2476 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2477 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2478 bp->b_flags &= ~(B_RELBUF);
2479 bp->b_iocmd = BIO_WRITE;
2481 if ((bp->b_flags & B_DELWRI) == 0) {
2482 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2491 * Clear B_DELWRI for buffer.
2493 * Since the buffer is not on a queue, we do not update the numfreebuffers
2496 * The buffer must be on QUEUE_NONE.
2500 bundirty(struct buf *bp)
2503 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2504 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2505 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2506 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2508 if (bp->b_flags & B_DELWRI) {
2509 bp->b_flags &= ~B_DELWRI;
2514 * Since it is now being written, we can clear its deferred write flag.
2516 bp->b_flags &= ~B_DEFERRED;
2522 * Asynchronous write. Start output on a buffer, but do not wait for
2523 * it to complete. The buffer is released when the output completes.
2525 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2526 * B_INVAL buffers. Not us.
2529 bawrite(struct buf *bp)
2532 bp->b_flags |= B_ASYNC;
2539 * Asynchronous barrier write. Start output on a buffer, but do not
2540 * wait for it to complete. Place a write barrier after this write so
2541 * that this buffer and all buffers written before it are committed to
2542 * the disk before any buffers written after this write are committed
2543 * to the disk. The buffer is released when the output completes.
2546 babarrierwrite(struct buf *bp)
2549 bp->b_flags |= B_ASYNC | B_BARRIER;
2556 * Synchronous barrier write. Start output on a buffer and wait for
2557 * it to complete. Place a write barrier after this write so that
2558 * this buffer and all buffers written before it are committed to
2559 * the disk before any buffers written after this write are committed
2560 * to the disk. The buffer is released when the output completes.
2563 bbarrierwrite(struct buf *bp)
2566 bp->b_flags |= B_BARRIER;
2567 return (bwrite(bp));
2573 * Called prior to the locking of any vnodes when we are expecting to
2574 * write. We do not want to starve the buffer cache with too many
2575 * dirty buffers so we block here. By blocking prior to the locking
2576 * of any vnodes we attempt to avoid the situation where a locked vnode
2577 * prevents the various system daemons from flushing related buffers.
2583 if (buf_dirty_count_severe()) {
2584 mtx_lock(&bdirtylock);
2585 while (buf_dirty_count_severe()) {
2587 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2590 mtx_unlock(&bdirtylock);
2595 * Return true if we have too many dirty buffers.
2598 buf_dirty_count_severe(void)
2601 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2607 * Release a busy buffer and, if requested, free its resources. The
2608 * buffer will be stashed in the appropriate bufqueue[] allowing it
2609 * to be accessed later as a cache entity or reused for other purposes.
2612 brelse(struct buf *bp)
2614 struct mount *v_mnt;
2618 * Many functions erroneously call brelse with a NULL bp under rare
2619 * error conditions. Simply return when called with a NULL bp.
2623 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2624 bp, bp->b_vp, bp->b_flags);
2625 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2626 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2627 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2628 ("brelse: non-VMIO buffer marked NOREUSE"));
2630 if (BUF_LOCKRECURSED(bp)) {
2632 * Do not process, in particular, do not handle the
2633 * B_INVAL/B_RELBUF and do not release to free list.
2639 if (bp->b_flags & B_MANAGED) {
2644 if (LIST_EMPTY(&bp->b_dep)) {
2645 bp->b_flags &= ~B_IOSTARTED;
2647 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2648 ("brelse: SU io not finished bp %p", bp));
2651 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2652 BO_LOCK(bp->b_bufobj);
2653 bp->b_vflags &= ~BV_BKGRDERR;
2654 BO_UNLOCK(bp->b_bufobj);
2658 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2659 (bp->b_flags & B_INVALONERR)) {
2661 * Forced invalidation of dirty buffer contents, to be used
2662 * after a failed write in the rare case that the loss of the
2663 * contents is acceptable. The buffer is invalidated and
2666 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2667 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2670 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2671 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2672 !(bp->b_flags & B_INVAL)) {
2674 * Failed write, redirty. All errors except ENXIO (which
2675 * means the device is gone) are treated as being
2678 * XXX Treating EIO as transient is not correct; the
2679 * contract with the local storage device drivers is that
2680 * they will only return EIO once the I/O is no longer
2681 * retriable. Network I/O also respects this through the
2682 * guarantees of TCP and/or the internal retries of NFS.
2683 * ENOMEM might be transient, but we also have no way of
2684 * knowing when its ok to retry/reschedule. In general,
2685 * this entire case should be made obsolete through better
2686 * error handling/recovery and resource scheduling.
2688 * Do this also for buffers that failed with ENXIO, but have
2689 * non-empty dependencies - the soft updates code might need
2690 * to access the buffer to untangle them.
2692 * Must clear BIO_ERROR to prevent pages from being scrapped.
2694 bp->b_ioflags &= ~BIO_ERROR;
2696 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2697 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2699 * Either a failed read I/O, or we were asked to free or not
2700 * cache the buffer, or we failed to write to a device that's
2701 * no longer present.
2703 bp->b_flags |= B_INVAL;
2704 if (!LIST_EMPTY(&bp->b_dep))
2706 if (bp->b_flags & B_DELWRI)
2708 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2709 if ((bp->b_flags & B_VMIO) == 0) {
2717 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2718 * is called with B_DELWRI set, the underlying pages may wind up
2719 * getting freed causing a previous write (bdwrite()) to get 'lost'
2720 * because pages associated with a B_DELWRI bp are marked clean.
2722 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2723 * if B_DELWRI is set.
2725 if (bp->b_flags & B_DELWRI)
2726 bp->b_flags &= ~B_RELBUF;
2729 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2730 * constituted, not even NFS buffers now. Two flags effect this. If
2731 * B_INVAL, the struct buf is invalidated but the VM object is kept
2732 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2734 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2735 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2736 * buffer is also B_INVAL because it hits the re-dirtying code above.
2738 * Normally we can do this whether a buffer is B_DELWRI or not. If
2739 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2740 * the commit state and we cannot afford to lose the buffer. If the
2741 * buffer has a background write in progress, we need to keep it
2742 * around to prevent it from being reconstituted and starting a second
2746 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2748 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2749 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2750 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2751 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2752 vfs_vmio_invalidate(bp);
2756 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2757 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2759 bp->b_flags &= ~B_NOREUSE;
2760 if (bp->b_vp != NULL)
2765 * If the buffer has junk contents signal it and eventually
2766 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2769 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2770 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2771 bp->b_flags |= B_INVAL;
2772 if (bp->b_flags & B_INVAL) {
2773 if (bp->b_flags & B_DELWRI)
2779 buf_track(bp, __func__);
2781 /* buffers with no memory */
2782 if (bp->b_bufsize == 0) {
2786 /* buffers with junk contents */
2787 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2788 (bp->b_ioflags & BIO_ERROR)) {
2789 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2790 if (bp->b_vflags & BV_BKGRDINPROG)
2791 panic("losing buffer 2");
2792 qindex = QUEUE_CLEAN;
2793 bp->b_flags |= B_AGE;
2794 /* remaining buffers */
2795 } else if (bp->b_flags & B_DELWRI)
2796 qindex = QUEUE_DIRTY;
2798 qindex = QUEUE_CLEAN;
2800 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2801 panic("brelse: not dirty");
2803 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2804 bp->b_xflags &= ~(BX_CVTENXIO);
2805 /* binsfree unlocks bp. */
2806 binsfree(bp, qindex);
2810 * Release a buffer back to the appropriate queue but do not try to free
2811 * it. The buffer is expected to be used again soon.
2813 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2814 * biodone() to requeue an async I/O on completion. It is also used when
2815 * known good buffers need to be requeued but we think we may need the data
2818 * XXX we should be able to leave the B_RELBUF hint set on completion.
2821 bqrelse(struct buf *bp)
2825 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2826 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2827 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2829 qindex = QUEUE_NONE;
2830 if (BUF_LOCKRECURSED(bp)) {
2831 /* do not release to free list */
2835 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2836 bp->b_xflags &= ~(BX_CVTENXIO);
2838 if (LIST_EMPTY(&bp->b_dep)) {
2839 bp->b_flags &= ~B_IOSTARTED;
2841 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2842 ("bqrelse: SU io not finished bp %p", bp));
2845 if (bp->b_flags & B_MANAGED) {
2846 if (bp->b_flags & B_REMFREE)
2851 /* buffers with stale but valid contents */
2852 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2853 BV_BKGRDERR)) == BV_BKGRDERR) {
2854 BO_LOCK(bp->b_bufobj);
2855 bp->b_vflags &= ~BV_BKGRDERR;
2856 BO_UNLOCK(bp->b_bufobj);
2857 qindex = QUEUE_DIRTY;
2859 if ((bp->b_flags & B_DELWRI) == 0 &&
2860 (bp->b_xflags & BX_VNDIRTY))
2861 panic("bqrelse: not dirty");
2862 if ((bp->b_flags & B_NOREUSE) != 0) {
2866 qindex = QUEUE_CLEAN;
2868 buf_track(bp, __func__);
2869 /* binsfree unlocks bp. */
2870 binsfree(bp, qindex);
2874 buf_track(bp, __func__);
2880 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2881 * restore bogus pages.
2884 vfs_vmio_iodone(struct buf *bp)
2889 struct vnode *vp __unused;
2890 int i, iosize, resid;
2893 obj = bp->b_bufobj->bo_object;
2894 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2895 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2896 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2899 VNPASS(vp->v_holdcnt > 0, vp);
2900 VNPASS(vp->v_object != NULL, vp);
2902 foff = bp->b_offset;
2903 KASSERT(bp->b_offset != NOOFFSET,
2904 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2907 iosize = bp->b_bcount - bp->b_resid;
2908 for (i = 0; i < bp->b_npages; i++) {
2909 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2914 * cleanup bogus pages, restoring the originals
2917 if (m == bogus_page) {
2919 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2921 panic("biodone: page disappeared!");
2923 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2925 * In the write case, the valid and clean bits are
2926 * already changed correctly ( see bdwrite() ), so we
2927 * only need to do this here in the read case.
2929 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2930 resid)) == 0, ("vfs_vmio_iodone: page %p "
2931 "has unexpected dirty bits", m));
2932 vfs_page_set_valid(bp, foff, m);
2934 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2935 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2936 (intmax_t)foff, (uintmax_t)m->pindex));
2939 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2942 vm_object_pip_wakeupn(obj, bp->b_npages);
2943 if (bogus && buf_mapped(bp)) {
2944 BUF_CHECK_MAPPED(bp);
2945 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2946 bp->b_pages, bp->b_npages);
2951 * Perform page invalidation when a buffer is released. The fully invalid
2952 * pages will be reclaimed later in vfs_vmio_truncate().
2955 vfs_vmio_invalidate(struct buf *bp)
2959 int flags, i, resid, poffset, presid;
2961 if (buf_mapped(bp)) {
2962 BUF_CHECK_MAPPED(bp);
2963 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2965 BUF_CHECK_UNMAPPED(bp);
2967 * Get the base offset and length of the buffer. Note that
2968 * in the VMIO case if the buffer block size is not
2969 * page-aligned then b_data pointer may not be page-aligned.
2970 * But our b_pages[] array *IS* page aligned.
2972 * block sizes less then DEV_BSIZE (usually 512) are not
2973 * supported due to the page granularity bits (m->valid,
2974 * m->dirty, etc...).
2976 * See man buf(9) for more information
2978 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2979 obj = bp->b_bufobj->bo_object;
2980 resid = bp->b_bufsize;
2981 poffset = bp->b_offset & PAGE_MASK;
2982 VM_OBJECT_WLOCK(obj);
2983 for (i = 0; i < bp->b_npages; i++) {
2985 if (m == bogus_page)
2986 panic("vfs_vmio_invalidate: Unexpected bogus page.");
2987 bp->b_pages[i] = NULL;
2989 presid = resid > (PAGE_SIZE - poffset) ?
2990 (PAGE_SIZE - poffset) : resid;
2991 KASSERT(presid >= 0, ("brelse: extra page"));
2992 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
2993 if (pmap_page_wired_mappings(m) == 0)
2994 vm_page_set_invalid(m, poffset, presid);
2996 vm_page_release_locked(m, flags);
3000 VM_OBJECT_WUNLOCK(obj);
3005 * Page-granular truncation of an existing VMIO buffer.
3008 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3014 if (bp->b_npages == desiredpages)
3017 if (buf_mapped(bp)) {
3018 BUF_CHECK_MAPPED(bp);
3019 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3020 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3022 BUF_CHECK_UNMAPPED(bp);
3025 * The object lock is needed only if we will attempt to free pages.
3027 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3028 if ((bp->b_flags & B_DIRECT) != 0) {
3029 flags |= VPR_TRYFREE;
3030 obj = bp->b_bufobj->bo_object;
3031 VM_OBJECT_WLOCK(obj);
3035 for (i = desiredpages; i < bp->b_npages; i++) {
3037 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3038 bp->b_pages[i] = NULL;
3040 vm_page_release_locked(m, flags);
3042 vm_page_release(m, flags);
3045 VM_OBJECT_WUNLOCK(obj);
3046 bp->b_npages = desiredpages;
3050 * Byte granular extension of VMIO buffers.
3053 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3056 * We are growing the buffer, possibly in a
3057 * byte-granular fashion.
3065 * Step 1, bring in the VM pages from the object, allocating
3066 * them if necessary. We must clear B_CACHE if these pages
3067 * are not valid for the range covered by the buffer.
3069 obj = bp->b_bufobj->bo_object;
3070 if (bp->b_npages < desiredpages) {
3071 KASSERT(desiredpages <= atop(maxbcachebuf),
3072 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3073 bp, desiredpages, maxbcachebuf));
3076 * We must allocate system pages since blocking
3077 * here could interfere with paging I/O, no
3078 * matter which process we are.
3080 * Only exclusive busy can be tested here.
3081 * Blocking on shared busy might lead to
3082 * deadlocks once allocbuf() is called after
3083 * pages are vfs_busy_pages().
3085 (void)vm_page_grab_pages_unlocked(obj,
3086 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3087 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3088 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3089 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3090 bp->b_npages = desiredpages;
3094 * Step 2. We've loaded the pages into the buffer,
3095 * we have to figure out if we can still have B_CACHE
3096 * set. Note that B_CACHE is set according to the
3097 * byte-granular range ( bcount and size ), not the
3098 * aligned range ( newbsize ).
3100 * The VM test is against m->valid, which is DEV_BSIZE
3101 * aligned. Needless to say, the validity of the data
3102 * needs to also be DEV_BSIZE aligned. Note that this
3103 * fails with NFS if the server or some other client
3104 * extends the file's EOF. If our buffer is resized,
3105 * B_CACHE may remain set! XXX
3107 toff = bp->b_bcount;
3108 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3109 while ((bp->b_flags & B_CACHE) && toff < size) {
3112 if (tinc > (size - toff))
3114 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3115 m = bp->b_pages[pi];
3116 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3122 * Step 3, fixup the KVA pmap.
3127 BUF_CHECK_UNMAPPED(bp);
3131 * Check to see if a block at a particular lbn is available for a clustered
3135 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3142 /* If the buf isn't in core skip it */
3143 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3146 /* If the buf is busy we don't want to wait for it */
3147 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3150 /* Only cluster with valid clusterable delayed write buffers */
3151 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3152 (B_DELWRI | B_CLUSTEROK))
3155 if (bpa->b_bufsize != size)
3159 * Check to see if it is in the expected place on disk and that the
3160 * block has been mapped.
3162 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3172 * Implement clustered async writes for clearing out B_DELWRI buffers.
3173 * This is much better then the old way of writing only one buffer at
3174 * a time. Note that we may not be presented with the buffers in the
3175 * correct order, so we search for the cluster in both directions.
3178 vfs_bio_awrite(struct buf *bp)
3183 daddr_t lblkno = bp->b_lblkno;
3184 struct vnode *vp = bp->b_vp;
3192 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3194 * right now we support clustered writing only to regular files. If
3195 * we find a clusterable block we could be in the middle of a cluster
3196 * rather then at the beginning.
3198 if ((vp->v_type == VREG) &&
3199 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3200 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3201 size = vp->v_mount->mnt_stat.f_iosize;
3202 maxcl = maxphys / size;
3205 for (i = 1; i < maxcl; i++)
3206 if (vfs_bio_clcheck(vp, size, lblkno + i,
3207 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3210 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3211 if (vfs_bio_clcheck(vp, size, lblkno - j,
3212 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3218 * this is a possible cluster write
3222 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3228 bp->b_flags |= B_ASYNC;
3230 * default (old) behavior, writing out only one block
3232 * XXX returns b_bufsize instead of b_bcount for nwritten?
3234 nwritten = bp->b_bufsize;
3243 * Allocate KVA for an empty buf header according to gbflags.
3246 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3249 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3251 * In order to keep fragmentation sane we only allocate kva
3252 * in BKVASIZE chunks. XXX with vmem we can do page size.
3254 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3256 if (maxsize != bp->b_kvasize &&
3257 bufkva_alloc(bp, maxsize, gbflags))
3266 * Find and initialize a new buffer header, freeing up existing buffers
3267 * in the bufqueues as necessary. The new buffer is returned locked.
3270 * We have insufficient buffer headers
3271 * We have insufficient buffer space
3272 * buffer_arena is too fragmented ( space reservation fails )
3273 * If we have to flush dirty buffers ( but we try to avoid this )
3275 * The caller is responsible for releasing the reserved bufspace after
3276 * allocbuf() is called.
3279 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3281 struct bufdomain *bd;
3283 bool metadata, reserved;
3286 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3287 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3288 if (!unmapped_buf_allowed)
3289 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3291 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3299 bd = &bdomain[vp->v_bufobj.bo_domain];
3301 counter_u64_add(getnewbufcalls, 1);
3304 if (reserved == false &&
3305 bufspace_reserve(bd, maxsize, metadata) != 0) {
3306 counter_u64_add(getnewbufrestarts, 1);
3310 if ((bp = buf_alloc(bd)) == NULL) {
3311 counter_u64_add(getnewbufrestarts, 1);
3314 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3317 } while (buf_recycle(bd, false) == 0);
3320 bufspace_release(bd, maxsize);
3322 bp->b_flags |= B_INVAL;
3325 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3333 * buffer flushing daemon. Buffers are normally flushed by the
3334 * update daemon but if it cannot keep up this process starts to
3335 * take the load in an attempt to prevent getnewbuf() from blocking.
3337 static struct kproc_desc buf_kp = {
3342 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3345 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3349 flushed = flushbufqueues(vp, bd, target, 0);
3352 * Could not find any buffers without rollback
3353 * dependencies, so just write the first one
3354 * in the hopes of eventually making progress.
3356 if (vp != NULL && target > 2)
3358 flushbufqueues(vp, bd, target, 1);
3366 struct bufdomain *bd;
3372 * This process needs to be suspended prior to shutdown sync.
3374 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3375 SHUTDOWN_PRI_LAST + 100);
3378 * Start the buf clean daemons as children threads.
3380 for (i = 0 ; i < buf_domains; i++) {
3383 error = kthread_add((void (*)(void *))bufspace_daemon,
3384 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3386 panic("error %d spawning bufspace daemon", error);
3390 * This process is allowed to take the buffer cache to the limit
3392 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3396 mtx_unlock(&bdlock);
3398 kthread_suspend_check();
3401 * Save speedupreq for this pass and reset to capture new
3404 speedupreq = bd_speedupreq;
3408 * Flush each domain sequentially according to its level and
3409 * the speedup request.
3411 for (i = 0; i < buf_domains; i++) {
3414 lodirty = bd->bd_numdirtybuffers / 2;
3416 lodirty = bd->bd_lodirtybuffers;
3417 while (bd->bd_numdirtybuffers > lodirty) {
3418 if (buf_flush(NULL, bd,
3419 bd->bd_numdirtybuffers - lodirty) == 0)
3421 kern_yield(PRI_USER);
3426 * Only clear bd_request if we have reached our low water
3427 * mark. The buf_daemon normally waits 1 second and
3428 * then incrementally flushes any dirty buffers that have
3429 * built up, within reason.
3431 * If we were unable to hit our low water mark and couldn't
3432 * find any flushable buffers, we sleep for a short period
3433 * to avoid endless loops on unlockable buffers.
3436 if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3438 * We reached our low water mark, reset the
3439 * request and sleep until we are needed again.
3440 * The sleep is just so the suspend code works.
3444 * Do an extra wakeup in case dirty threshold
3445 * changed via sysctl and the explicit transition
3446 * out of shortfall was missed.
3449 if (runningbufspace <= lorunningspace)
3451 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3454 * We couldn't find any flushable dirty buffers but
3455 * still have too many dirty buffers, we
3456 * have to sleep and try again. (rare)
3458 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3466 * Try to flush a buffer in the dirty queue. We must be careful to
3467 * free up B_INVAL buffers instead of write them, which NFS is
3468 * particularly sensitive to.
3470 static int flushwithdeps = 0;
3471 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3473 "Number of buffers flushed with dependecies that require rollbacks");
3476 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3479 struct bufqueue *bq;
3480 struct buf *sentinel;
3490 bq = &bd->bd_dirtyq;
3492 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3493 sentinel->b_qindex = QUEUE_SENTINEL;
3495 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3497 while (flushed != target) {
3500 bp = TAILQ_NEXT(sentinel, b_freelist);
3502 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3503 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3510 * Skip sentinels inserted by other invocations of the
3511 * flushbufqueues(), taking care to not reorder them.
3513 * Only flush the buffers that belong to the
3514 * vnode locked by the curthread.
3516 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3521 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3527 * BKGRDINPROG can only be set with the buf and bufobj
3528 * locks both held. We tolerate a race to clear it here.
3530 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3531 (bp->b_flags & B_DELWRI) == 0) {
3535 if (bp->b_flags & B_INVAL) {
3542 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3543 if (flushdeps == 0) {
3551 * We must hold the lock on a vnode before writing
3552 * one of its buffers. Otherwise we may confuse, or
3553 * in the case of a snapshot vnode, deadlock the
3556 * The lock order here is the reverse of the normal
3557 * of vnode followed by buf lock. This is ok because
3558 * the NOWAIT will prevent deadlock.
3561 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3567 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3569 ASSERT_VOP_LOCKED(vp, "getbuf");
3571 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3572 vn_lock(vp, LK_TRYUPGRADE);
3575 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3576 bp, bp->b_vp, bp->b_flags);
3577 if (curproc == bufdaemonproc) {
3582 counter_u64_add(notbufdflushes, 1);
3584 vn_finished_write(mp);
3587 flushwithdeps += hasdeps;
3591 * Sleeping on runningbufspace while holding
3592 * vnode lock leads to deadlock.
3594 if (curproc == bufdaemonproc &&
3595 runningbufspace > hirunningspace)
3596 waitrunningbufspace();
3599 vn_finished_write(mp);
3603 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3605 free(sentinel, M_TEMP);
3610 * Check to see if a block is currently memory resident.
3613 incore(struct bufobj *bo, daddr_t blkno)
3615 return (gbincore_unlocked(bo, blkno));
3619 * Returns true if no I/O is needed to access the
3620 * associated VM object. This is like incore except
3621 * it also hunts around in the VM system for the data.
3624 inmem(struct vnode * vp, daddr_t blkno)
3627 vm_offset_t toff, tinc, size;
3632 ASSERT_VOP_LOCKED(vp, "inmem");
3634 if (incore(&vp->v_bufobj, blkno))
3636 if (vp->v_mount == NULL)
3643 if (size > vp->v_mount->mnt_stat.f_iosize)
3644 size = vp->v_mount->mnt_stat.f_iosize;
3645 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3647 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3648 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3654 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3655 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3657 * Consider page validity only if page mapping didn't change
3660 valid = vm_page_is_valid(m,
3661 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3662 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3674 * Set the dirty range for a buffer based on the status of the dirty
3675 * bits in the pages comprising the buffer. The range is limited
3676 * to the size of the buffer.
3678 * Tell the VM system that the pages associated with this buffer
3679 * are clean. This is used for delayed writes where the data is
3680 * going to go to disk eventually without additional VM intevention.
3682 * Note that while we only really need to clean through to b_bcount, we
3683 * just go ahead and clean through to b_bufsize.
3686 vfs_clean_pages_dirty_buf(struct buf *bp)
3688 vm_ooffset_t foff, noff, eoff;
3692 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3695 foff = bp->b_offset;
3696 KASSERT(bp->b_offset != NOOFFSET,
3697 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3699 vfs_busy_pages_acquire(bp);
3700 vfs_setdirty_range(bp);
3701 for (i = 0; i < bp->b_npages; i++) {
3702 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3704 if (eoff > bp->b_offset + bp->b_bufsize)
3705 eoff = bp->b_offset + bp->b_bufsize;
3707 vfs_page_set_validclean(bp, foff, m);
3708 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3711 vfs_busy_pages_release(bp);
3715 vfs_setdirty_range(struct buf *bp)
3717 vm_offset_t boffset;
3718 vm_offset_t eoffset;
3722 * test the pages to see if they have been modified directly
3723 * by users through the VM system.
3725 for (i = 0; i < bp->b_npages; i++)
3726 vm_page_test_dirty(bp->b_pages[i]);
3729 * Calculate the encompassing dirty range, boffset and eoffset,
3730 * (eoffset - boffset) bytes.
3733 for (i = 0; i < bp->b_npages; i++) {
3734 if (bp->b_pages[i]->dirty)
3737 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3739 for (i = bp->b_npages - 1; i >= 0; --i) {
3740 if (bp->b_pages[i]->dirty) {
3744 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3747 * Fit it to the buffer.
3750 if (eoffset > bp->b_bcount)
3751 eoffset = bp->b_bcount;
3754 * If we have a good dirty range, merge with the existing
3758 if (boffset < eoffset) {
3759 if (bp->b_dirtyoff > boffset)
3760 bp->b_dirtyoff = boffset;
3761 if (bp->b_dirtyend < eoffset)
3762 bp->b_dirtyend = eoffset;
3767 * Allocate the KVA mapping for an existing buffer.
3768 * If an unmapped buffer is provided but a mapped buffer is requested, take
3769 * also care to properly setup mappings between pages and KVA.
3772 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3774 int bsize, maxsize, need_mapping, need_kva;
3777 need_mapping = bp->b_data == unmapped_buf &&
3778 (gbflags & GB_UNMAPPED) == 0;
3779 need_kva = bp->b_kvabase == unmapped_buf &&
3780 bp->b_data == unmapped_buf &&
3781 (gbflags & GB_KVAALLOC) != 0;
3782 if (!need_mapping && !need_kva)
3785 BUF_CHECK_UNMAPPED(bp);
3787 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3789 * Buffer is not mapped, but the KVA was already
3790 * reserved at the time of the instantiation. Use the
3797 * Calculate the amount of the address space we would reserve
3798 * if the buffer was mapped.
3800 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3801 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3802 offset = blkno * bsize;
3803 maxsize = size + (offset & PAGE_MASK);
3804 maxsize = imax(maxsize, bsize);
3806 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3807 if ((gbflags & GB_NOWAIT_BD) != 0) {
3809 * XXXKIB: defragmentation cannot
3810 * succeed, not sure what else to do.
3812 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3814 counter_u64_add(mappingrestarts, 1);
3815 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3819 /* b_offset is handled by bpmap_qenter. */
3820 bp->b_data = bp->b_kvabase;
3821 BUF_CHECK_MAPPED(bp);
3827 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3833 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3842 * Get a block given a specified block and offset into a file/device.
3843 * The buffers B_DONE bit will be cleared on return, making it almost
3844 * ready for an I/O initiation. B_INVAL may or may not be set on
3845 * return. The caller should clear B_INVAL prior to initiating a
3848 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3849 * an existing buffer.
3851 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3852 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3853 * and then cleared based on the backing VM. If the previous buffer is
3854 * non-0-sized but invalid, B_CACHE will be cleared.
3856 * If getblk() must create a new buffer, the new buffer is returned with
3857 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3858 * case it is returned with B_INVAL clear and B_CACHE set based on the
3861 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3862 * B_CACHE bit is clear.
3864 * What this means, basically, is that the caller should use B_CACHE to
3865 * determine whether the buffer is fully valid or not and should clear
3866 * B_INVAL prior to issuing a read. If the caller intends to validate
3867 * the buffer by loading its data area with something, the caller needs
3868 * to clear B_INVAL. If the caller does this without issuing an I/O,
3869 * the caller should set B_CACHE ( as an optimization ), else the caller
3870 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3871 * a write attempt or if it was a successful read. If the caller
3872 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3873 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3875 * The blkno parameter is the logical block being requested. Normally
3876 * the mapping of logical block number to disk block address is done
3877 * by calling VOP_BMAP(). However, if the mapping is already known, the
3878 * disk block address can be passed using the dblkno parameter. If the
3879 * disk block address is not known, then the same value should be passed
3880 * for blkno and dblkno.
3883 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3884 int slptimeo, int flags, struct buf **bpp)
3889 int bsize, error, maxsize, vmio;
3892 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3893 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3894 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3895 ASSERT_VOP_LOCKED(vp, "getblk");
3896 if (size > maxbcachebuf)
3897 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3899 if (!unmapped_buf_allowed)
3900 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3905 /* Attempt lockless lookup first. */
3906 bp = gbincore_unlocked(bo, blkno);
3908 goto newbuf_unlocked;
3910 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3915 /* Verify buf identify has not changed since lookup. */
3916 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3917 goto foundbuf_fastpath;
3919 /* It changed, fallback to locked lookup. */
3924 bp = gbincore(bo, blkno);
3929 * Buffer is in-core. If the buffer is not busy nor managed,
3930 * it must be on a queue.
3932 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3933 ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3935 error = BUF_TIMELOCK(bp, lockflags,
3936 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3939 * If we slept and got the lock we have to restart in case
3940 * the buffer changed identities.
3942 if (error == ENOLCK)
3944 /* We timed out or were interrupted. */
3945 else if (error != 0)
3949 /* If recursed, assume caller knows the rules. */
3950 if (BUF_LOCKRECURSED(bp))
3954 * The buffer is locked. B_CACHE is cleared if the buffer is
3955 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3956 * and for a VMIO buffer B_CACHE is adjusted according to the
3959 if (bp->b_flags & B_INVAL)
3960 bp->b_flags &= ~B_CACHE;
3961 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3962 bp->b_flags |= B_CACHE;
3963 if (bp->b_flags & B_MANAGED)
3964 MPASS(bp->b_qindex == QUEUE_NONE);
3969 * check for size inconsistencies for non-VMIO case.
3971 if (bp->b_bcount != size) {
3972 if ((bp->b_flags & B_VMIO) == 0 ||
3973 (size > bp->b_kvasize)) {
3974 if (bp->b_flags & B_DELWRI) {
3975 bp->b_flags |= B_NOCACHE;
3978 if (LIST_EMPTY(&bp->b_dep)) {
3979 bp->b_flags |= B_RELBUF;
3982 bp->b_flags |= B_NOCACHE;
3991 * Handle the case of unmapped buffer which should
3992 * become mapped, or the buffer for which KVA
3993 * reservation is requested.
3995 bp_unmapped_get_kva(bp, blkno, size, flags);
3998 * If the size is inconsistent in the VMIO case, we can resize
3999 * the buffer. This might lead to B_CACHE getting set or
4000 * cleared. If the size has not changed, B_CACHE remains
4001 * unchanged from its previous state.
4005 KASSERT(bp->b_offset != NOOFFSET,
4006 ("getblk: no buffer offset"));
4009 * A buffer with B_DELWRI set and B_CACHE clear must
4010 * be committed before we can return the buffer in
4011 * order to prevent the caller from issuing a read
4012 * ( due to B_CACHE not being set ) and overwriting
4015 * Most callers, including NFS and FFS, need this to
4016 * operate properly either because they assume they
4017 * can issue a read if B_CACHE is not set, or because
4018 * ( for example ) an uncached B_DELWRI might loop due
4019 * to softupdates re-dirtying the buffer. In the latter
4020 * case, B_CACHE is set after the first write completes,
4021 * preventing further loops.
4022 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4023 * above while extending the buffer, we cannot allow the
4024 * buffer to remain with B_CACHE set after the write
4025 * completes or it will represent a corrupt state. To
4026 * deal with this we set B_NOCACHE to scrap the buffer
4029 * We might be able to do something fancy, like setting
4030 * B_CACHE in bwrite() except if B_DELWRI is already set,
4031 * so the below call doesn't set B_CACHE, but that gets real
4032 * confusing. This is much easier.
4035 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4036 bp->b_flags |= B_NOCACHE;
4040 bp->b_flags &= ~B_DONE;
4043 * Buffer is not in-core, create new buffer. The buffer
4044 * returned by getnewbuf() is locked. Note that the returned
4045 * buffer is also considered valid (not marked B_INVAL).
4050 * If the user does not want us to create the buffer, bail out
4053 if (flags & GB_NOCREAT)
4056 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4057 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4058 offset = blkno * bsize;
4059 vmio = vp->v_object != NULL;
4061 maxsize = size + (offset & PAGE_MASK);
4064 /* Do not allow non-VMIO notmapped buffers. */
4065 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4067 maxsize = imax(maxsize, bsize);
4068 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4070 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4071 KASSERT(error != EOPNOTSUPP,
4072 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4077 return (EJUSTRETURN);
4080 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4082 if (slpflag || slptimeo)
4085 * XXX This is here until the sleep path is diagnosed
4086 * enough to work under very low memory conditions.
4088 * There's an issue on low memory, 4BSD+non-preempt
4089 * systems (eg MIPS routers with 32MB RAM) where buffer
4090 * exhaustion occurs without sleeping for buffer
4091 * reclaimation. This just sticks in a loop and
4092 * constantly attempts to allocate a buffer, which
4093 * hits exhaustion and tries to wakeup bufdaemon.
4094 * This never happens because we never yield.
4096 * The real solution is to identify and fix these cases
4097 * so we aren't effectively busy-waiting in a loop
4098 * until the reclaimation path has cycles to run.
4100 kern_yield(PRI_USER);
4105 * This code is used to make sure that a buffer is not
4106 * created while the getnewbuf routine is blocked.
4107 * This can be a problem whether the vnode is locked or not.
4108 * If the buffer is created out from under us, we have to
4109 * throw away the one we just created.
4111 * Note: this must occur before we associate the buffer
4112 * with the vp especially considering limitations in
4113 * the splay tree implementation when dealing with duplicate
4117 if (gbincore(bo, blkno)) {
4119 bp->b_flags |= B_INVAL;
4120 bufspace_release(bufdomain(bp), maxsize);
4126 * Insert the buffer into the hash, so that it can
4127 * be found by incore.
4129 bp->b_lblkno = blkno;
4130 bp->b_blkno = d_blkno;
4131 bp->b_offset = offset;
4136 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4137 * buffer size starts out as 0, B_CACHE will be set by
4138 * allocbuf() for the VMIO case prior to it testing the
4139 * backing store for validity.
4143 bp->b_flags |= B_VMIO;
4144 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4145 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4146 bp, vp->v_object, bp->b_bufobj->bo_object));
4148 bp->b_flags &= ~B_VMIO;
4149 KASSERT(bp->b_bufobj->bo_object == NULL,
4150 ("ARGH! has b_bufobj->bo_object %p %p\n",
4151 bp, bp->b_bufobj->bo_object));
4152 BUF_CHECK_MAPPED(bp);
4156 bufspace_release(bufdomain(bp), maxsize);
4157 bp->b_flags &= ~B_DONE;
4159 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4161 buf_track(bp, __func__);
4162 KASSERT(bp->b_bufobj == bo,
4163 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4169 * Get an empty, disassociated buffer of given size. The buffer is initially
4173 geteblk(int size, int flags)
4178 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4179 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4180 if ((flags & GB_NOWAIT_BD) &&
4181 (curthread->td_pflags & TDP_BUFNEED) != 0)
4185 bufspace_release(bufdomain(bp), maxsize);
4186 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4191 * Truncate the backing store for a non-vmio buffer.
4194 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4197 if (bp->b_flags & B_MALLOC) {
4199 * malloced buffers are not shrunk
4201 if (newbsize == 0) {
4202 bufmallocadjust(bp, 0);
4203 free(bp->b_data, M_BIOBUF);
4204 bp->b_data = bp->b_kvabase;
4205 bp->b_flags &= ~B_MALLOC;
4209 vm_hold_free_pages(bp, newbsize);
4210 bufspace_adjust(bp, newbsize);
4214 * Extend the backing for a non-VMIO buffer.
4217 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4223 * We only use malloced memory on the first allocation.
4224 * and revert to page-allocated memory when the buffer
4227 * There is a potential smp race here that could lead
4228 * to bufmallocspace slightly passing the max. It
4229 * is probably extremely rare and not worth worrying
4232 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4233 bufmallocspace < maxbufmallocspace) {
4234 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4235 bp->b_flags |= B_MALLOC;
4236 bufmallocadjust(bp, newbsize);
4241 * If the buffer is growing on its other-than-first
4242 * allocation then we revert to the page-allocation
4247 if (bp->b_flags & B_MALLOC) {
4248 origbuf = bp->b_data;
4249 origbufsize = bp->b_bufsize;
4250 bp->b_data = bp->b_kvabase;
4251 bufmallocadjust(bp, 0);
4252 bp->b_flags &= ~B_MALLOC;
4253 newbsize = round_page(newbsize);
4255 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4256 (vm_offset_t) bp->b_data + newbsize);
4257 if (origbuf != NULL) {
4258 bcopy(origbuf, bp->b_data, origbufsize);
4259 free(origbuf, M_BIOBUF);
4261 bufspace_adjust(bp, newbsize);
4265 * This code constitutes the buffer memory from either anonymous system
4266 * memory (in the case of non-VMIO operations) or from an associated
4267 * VM object (in the case of VMIO operations). This code is able to
4268 * resize a buffer up or down.
4270 * Note that this code is tricky, and has many complications to resolve
4271 * deadlock or inconsistent data situations. Tread lightly!!!
4272 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4273 * the caller. Calling this code willy nilly can result in the loss of data.
4275 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4276 * B_CACHE for the non-VMIO case.
4279 allocbuf(struct buf *bp, int size)
4283 if (bp->b_bcount == size)
4286 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4287 panic("allocbuf: buffer too small");
4289 newbsize = roundup2(size, DEV_BSIZE);
4290 if ((bp->b_flags & B_VMIO) == 0) {
4291 if ((bp->b_flags & B_MALLOC) == 0)
4292 newbsize = round_page(newbsize);
4294 * Just get anonymous memory from the kernel. Don't
4295 * mess with B_CACHE.
4297 if (newbsize < bp->b_bufsize)
4298 vfs_nonvmio_truncate(bp, newbsize);
4299 else if (newbsize > bp->b_bufsize)
4300 vfs_nonvmio_extend(bp, newbsize);
4304 desiredpages = (size == 0) ? 0 :
4305 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4307 if (bp->b_flags & B_MALLOC)
4308 panic("allocbuf: VMIO buffer can't be malloced");
4310 * Set B_CACHE initially if buffer is 0 length or will become
4313 if (size == 0 || bp->b_bufsize == 0)
4314 bp->b_flags |= B_CACHE;
4316 if (newbsize < bp->b_bufsize)
4317 vfs_vmio_truncate(bp, desiredpages);
4318 /* XXX This looks as if it should be newbsize > b_bufsize */
4319 else if (size > bp->b_bcount)
4320 vfs_vmio_extend(bp, desiredpages, size);
4321 bufspace_adjust(bp, newbsize);
4323 bp->b_bcount = size; /* requested buffer size. */
4327 extern int inflight_transient_maps;
4329 static struct bio_queue nondump_bios;
4332 biodone(struct bio *bp)
4335 void (*done)(struct bio *);
4336 vm_offset_t start, end;
4338 biotrack(bp, __func__);
4341 * Avoid completing I/O when dumping after a panic since that may
4342 * result in a deadlock in the filesystem or pager code. Note that
4343 * this doesn't affect dumps that were started manually since we aim
4344 * to keep the system usable after it has been resumed.
4346 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4347 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4350 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4351 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4352 bp->bio_flags |= BIO_UNMAPPED;
4353 start = trunc_page((vm_offset_t)bp->bio_data);
4354 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4355 bp->bio_data = unmapped_buf;
4356 pmap_qremove(start, atop(end - start));
4357 vmem_free(transient_arena, start, end - start);
4358 atomic_add_int(&inflight_transient_maps, -1);
4360 done = bp->bio_done;
4362 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4364 bp->bio_flags |= BIO_DONE;
4372 * Wait for a BIO to finish.
4375 biowait(struct bio *bp, const char *wchan)
4379 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4381 while ((bp->bio_flags & BIO_DONE) == 0)
4382 msleep(bp, mtxp, PRIBIO, wchan, 0);
4384 if (bp->bio_error != 0)
4385 return (bp->bio_error);
4386 if (!(bp->bio_flags & BIO_ERROR))
4392 biofinish(struct bio *bp, struct devstat *stat, int error)
4396 bp->bio_error = error;
4397 bp->bio_flags |= BIO_ERROR;
4400 devstat_end_transaction_bio(stat, bp);
4404 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4406 biotrack_buf(struct bio *bp, const char *location)
4409 buf_track(bp->bio_track_bp, location);
4416 * Wait for buffer I/O completion, returning error status. The buffer
4417 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4418 * error and cleared.
4421 bufwait(struct buf *bp)
4423 if (bp->b_iocmd == BIO_READ)
4424 bwait(bp, PRIBIO, "biord");
4426 bwait(bp, PRIBIO, "biowr");
4427 if (bp->b_flags & B_EINTR) {
4428 bp->b_flags &= ~B_EINTR;
4431 if (bp->b_ioflags & BIO_ERROR) {
4432 return (bp->b_error ? bp->b_error : EIO);
4441 * Finish I/O on a buffer, optionally calling a completion function.
4442 * This is usually called from an interrupt so process blocking is
4445 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4446 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4447 * assuming B_INVAL is clear.
4449 * For the VMIO case, we set B_CACHE if the op was a read and no
4450 * read error occurred, or if the op was a write. B_CACHE is never
4451 * set if the buffer is invalid or otherwise uncacheable.
4453 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4454 * initiator to leave B_INVAL set to brelse the buffer out of existence
4455 * in the biodone routine.
4458 bufdone(struct buf *bp)
4460 struct bufobj *dropobj;
4461 void (*biodone)(struct buf *);
4463 buf_track(bp, __func__);
4464 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4467 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4469 runningbufwakeup(bp);
4470 if (bp->b_iocmd == BIO_WRITE)
4471 dropobj = bp->b_bufobj;
4472 /* call optional completion function if requested */
4473 if (bp->b_iodone != NULL) {
4474 biodone = bp->b_iodone;
4475 bp->b_iodone = NULL;
4478 bufobj_wdrop(dropobj);
4481 if (bp->b_flags & B_VMIO) {
4483 * Set B_CACHE if the op was a normal read and no error
4484 * occurred. B_CACHE is set for writes in the b*write()
4487 if (bp->b_iocmd == BIO_READ &&
4488 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4489 !(bp->b_ioflags & BIO_ERROR))
4490 bp->b_flags |= B_CACHE;
4491 vfs_vmio_iodone(bp);
4493 if (!LIST_EMPTY(&bp->b_dep))
4495 if ((bp->b_flags & B_CKHASH) != 0) {
4496 KASSERT(bp->b_iocmd == BIO_READ,
4497 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4498 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4499 (*bp->b_ckhashcalc)(bp);
4502 * For asynchronous completions, release the buffer now. The brelse
4503 * will do a wakeup there if necessary - so no need to do a wakeup
4504 * here in the async case. The sync case always needs to do a wakeup.
4506 if (bp->b_flags & B_ASYNC) {
4507 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4508 (bp->b_ioflags & BIO_ERROR))
4515 bufobj_wdrop(dropobj);
4519 * This routine is called in lieu of iodone in the case of
4520 * incomplete I/O. This keeps the busy status for pages
4524 vfs_unbusy_pages(struct buf *bp)
4530 runningbufwakeup(bp);
4531 if (!(bp->b_flags & B_VMIO))
4534 obj = bp->b_bufobj->bo_object;
4535 for (i = 0; i < bp->b_npages; i++) {
4537 if (m == bogus_page) {
4538 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4540 panic("vfs_unbusy_pages: page missing\n");
4542 if (buf_mapped(bp)) {
4543 BUF_CHECK_MAPPED(bp);
4544 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4545 bp->b_pages, bp->b_npages);
4547 BUF_CHECK_UNMAPPED(bp);
4551 vm_object_pip_wakeupn(obj, bp->b_npages);
4555 * vfs_page_set_valid:
4557 * Set the valid bits in a page based on the supplied offset. The
4558 * range is restricted to the buffer's size.
4560 * This routine is typically called after a read completes.
4563 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4568 * Compute the end offset, eoff, such that [off, eoff) does not span a
4569 * page boundary and eoff is not greater than the end of the buffer.
4570 * The end of the buffer, in this case, is our file EOF, not the
4571 * allocation size of the buffer.
4573 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4574 if (eoff > bp->b_offset + bp->b_bcount)
4575 eoff = bp->b_offset + bp->b_bcount;
4578 * Set valid range. This is typically the entire buffer and thus the
4582 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4586 * vfs_page_set_validclean:
4588 * Set the valid bits and clear the dirty bits in a page based on the
4589 * supplied offset. The range is restricted to the buffer's size.
4592 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4594 vm_ooffset_t soff, eoff;
4597 * Start and end offsets in buffer. eoff - soff may not cross a
4598 * page boundary or cross the end of the buffer. The end of the
4599 * buffer, in this case, is our file EOF, not the allocation size
4603 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4604 if (eoff > bp->b_offset + bp->b_bcount)
4605 eoff = bp->b_offset + bp->b_bcount;
4608 * Set valid range. This is typically the entire buffer and thus the
4612 vm_page_set_validclean(
4614 (vm_offset_t) (soff & PAGE_MASK),
4615 (vm_offset_t) (eoff - soff)
4621 * Acquire a shared busy on all pages in the buf.
4624 vfs_busy_pages_acquire(struct buf *bp)
4628 for (i = 0; i < bp->b_npages; i++)
4629 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4633 vfs_busy_pages_release(struct buf *bp)
4637 for (i = 0; i < bp->b_npages; i++)
4638 vm_page_sunbusy(bp->b_pages[i]);
4642 * This routine is called before a device strategy routine.
4643 * It is used to tell the VM system that paging I/O is in
4644 * progress, and treat the pages associated with the buffer
4645 * almost as being exclusive busy. Also the object paging_in_progress
4646 * flag is handled to make sure that the object doesn't become
4649 * Since I/O has not been initiated yet, certain buffer flags
4650 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4651 * and should be ignored.
4654 vfs_busy_pages(struct buf *bp, int clear_modify)
4662 if (!(bp->b_flags & B_VMIO))
4665 obj = bp->b_bufobj->bo_object;
4666 foff = bp->b_offset;
4667 KASSERT(bp->b_offset != NOOFFSET,
4668 ("vfs_busy_pages: no buffer offset"));
4669 if ((bp->b_flags & B_CLUSTER) == 0) {
4670 vm_object_pip_add(obj, bp->b_npages);
4671 vfs_busy_pages_acquire(bp);
4673 if (bp->b_bufsize != 0)
4674 vfs_setdirty_range(bp);
4676 for (i = 0; i < bp->b_npages; i++) {
4678 vm_page_assert_sbusied(m);
4681 * When readying a buffer for a read ( i.e
4682 * clear_modify == 0 ), it is important to do
4683 * bogus_page replacement for valid pages in
4684 * partially instantiated buffers. Partially
4685 * instantiated buffers can, in turn, occur when
4686 * reconstituting a buffer from its VM backing store
4687 * base. We only have to do this if B_CACHE is
4688 * clear ( which causes the I/O to occur in the
4689 * first place ). The replacement prevents the read
4690 * I/O from overwriting potentially dirty VM-backed
4691 * pages. XXX bogus page replacement is, uh, bogus.
4692 * It may not work properly with small-block devices.
4693 * We need to find a better way.
4696 pmap_remove_write(m);
4697 vfs_page_set_validclean(bp, foff, m);
4698 } else if (vm_page_all_valid(m) &&
4699 (bp->b_flags & B_CACHE) == 0) {
4700 bp->b_pages[i] = bogus_page;
4703 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4705 if (bogus && buf_mapped(bp)) {
4706 BUF_CHECK_MAPPED(bp);
4707 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4708 bp->b_pages, bp->b_npages);
4713 * vfs_bio_set_valid:
4715 * Set the range within the buffer to valid. The range is
4716 * relative to the beginning of the buffer, b_offset. Note that
4717 * b_offset itself may be offset from the beginning of the first
4721 vfs_bio_set_valid(struct buf *bp, int base, int size)
4726 if (!(bp->b_flags & B_VMIO))
4730 * Fixup base to be relative to beginning of first page.
4731 * Set initial n to be the maximum number of bytes in the
4732 * first page that can be validated.
4734 base += (bp->b_offset & PAGE_MASK);
4735 n = PAGE_SIZE - (base & PAGE_MASK);
4738 * Busy may not be strictly necessary here because the pages are
4739 * unlikely to be fully valid and the vnode lock will synchronize
4740 * their access via getpages. It is grabbed for consistency with
4741 * other page validation.
4743 vfs_busy_pages_acquire(bp);
4744 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4748 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4753 vfs_busy_pages_release(bp);
4759 * If the specified buffer is a non-VMIO buffer, clear the entire
4760 * buffer. If the specified buffer is a VMIO buffer, clear and
4761 * validate only the previously invalid portions of the buffer.
4762 * This routine essentially fakes an I/O, so we need to clear
4763 * BIO_ERROR and B_INVAL.
4765 * Note that while we only theoretically need to clear through b_bcount,
4766 * we go ahead and clear through b_bufsize.
4769 vfs_bio_clrbuf(struct buf *bp)
4771 int i, j, mask, sa, ea, slide;
4773 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4777 bp->b_flags &= ~B_INVAL;
4778 bp->b_ioflags &= ~BIO_ERROR;
4779 vfs_busy_pages_acquire(bp);
4780 sa = bp->b_offset & PAGE_MASK;
4782 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4783 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4784 ea = slide & PAGE_MASK;
4787 if (bp->b_pages[i] == bogus_page)
4790 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4791 if ((bp->b_pages[i]->valid & mask) == mask)
4793 if ((bp->b_pages[i]->valid & mask) == 0)
4794 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4796 for (; sa < ea; sa += DEV_BSIZE, j++) {
4797 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4798 pmap_zero_page_area(bp->b_pages[i],
4803 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4804 roundup2(ea - sa, DEV_BSIZE));
4806 vfs_busy_pages_release(bp);
4811 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4816 if (buf_mapped(bp)) {
4817 BUF_CHECK_MAPPED(bp);
4818 bzero(bp->b_data + base, size);
4820 BUF_CHECK_UNMAPPED(bp);
4821 n = PAGE_SIZE - (base & PAGE_MASK);
4822 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4826 pmap_zero_page_area(m, base & PAGE_MASK, n);
4835 * Update buffer flags based on I/O request parameters, optionally releasing the
4836 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4837 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4838 * I/O). Otherwise the buffer is released to the cache.
4841 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4844 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4845 ("buf %p non-VMIO noreuse", bp));
4847 if ((ioflag & IO_DIRECT) != 0)
4848 bp->b_flags |= B_DIRECT;
4849 if ((ioflag & IO_EXT) != 0)
4850 bp->b_xflags |= BX_ALTDATA;
4851 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4852 bp->b_flags |= B_RELBUF;
4853 if ((ioflag & IO_NOREUSE) != 0)
4854 bp->b_flags |= B_NOREUSE;
4862 vfs_bio_brelse(struct buf *bp, int ioflag)
4865 b_io_dismiss(bp, ioflag, true);
4869 vfs_bio_set_flags(struct buf *bp, int ioflag)
4872 b_io_dismiss(bp, ioflag, false);
4876 * vm_hold_load_pages and vm_hold_free_pages get pages into
4877 * a buffers address space. The pages are anonymous and are
4878 * not associated with a file object.
4881 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4887 BUF_CHECK_MAPPED(bp);
4889 to = round_page(to);
4890 from = round_page(from);
4891 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4892 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4893 KASSERT(to - from <= maxbcachebuf,
4894 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4895 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4897 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4899 * note: must allocate system pages since blocking here
4900 * could interfere with paging I/O, no matter which
4903 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4904 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4906 pmap_qenter(pg, &p, 1);
4907 bp->b_pages[index] = p;
4909 bp->b_npages = index;
4912 /* Return pages associated with this buf to the vm system */
4914 vm_hold_free_pages(struct buf *bp, int newbsize)
4918 int index, newnpages;
4920 BUF_CHECK_MAPPED(bp);
4922 from = round_page((vm_offset_t)bp->b_data + newbsize);
4923 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4924 if (bp->b_npages > newnpages)
4925 pmap_qremove(from, bp->b_npages - newnpages);
4926 for (index = newnpages; index < bp->b_npages; index++) {
4927 p = bp->b_pages[index];
4928 bp->b_pages[index] = NULL;
4929 vm_page_unwire_noq(p);
4932 bp->b_npages = newnpages;
4936 * Map an IO request into kernel virtual address space.
4938 * All requests are (re)mapped into kernel VA space.
4939 * Notice that we use b_bufsize for the size of the buffer
4940 * to be mapped. b_bcount might be modified by the driver.
4942 * Note that even if the caller determines that the address space should
4943 * be valid, a race or a smaller-file mapped into a larger space may
4944 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4945 * check the return value.
4947 * This function only works with pager buffers.
4950 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4955 MPASS((bp->b_flags & B_MAXPHYS) != 0);
4956 prot = VM_PROT_READ;
4957 if (bp->b_iocmd == BIO_READ)
4958 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4959 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4960 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
4963 bp->b_bufsize = len;
4964 bp->b_npages = pidx;
4965 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
4966 if (mapbuf || !unmapped_buf_allowed) {
4967 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4968 bp->b_data = bp->b_kvabase + bp->b_offset;
4970 bp->b_data = unmapped_buf;
4975 * Free the io map PTEs associated with this IO operation.
4976 * We also invalidate the TLB entries and restore the original b_addr.
4978 * This function only works with pager buffers.
4981 vunmapbuf(struct buf *bp)
4985 npages = bp->b_npages;
4987 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4988 vm_page_unhold_pages(bp->b_pages, npages);
4990 bp->b_data = unmapped_buf;
4994 bdone(struct buf *bp)
4998 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5000 bp->b_flags |= B_DONE;
5006 bwait(struct buf *bp, u_char pri, const char *wchan)
5010 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5012 while ((bp->b_flags & B_DONE) == 0)
5013 msleep(bp, mtxp, pri, wchan, 0);
5018 bufsync(struct bufobj *bo, int waitfor)
5021 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5025 bufstrategy(struct bufobj *bo, struct buf *bp)
5031 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5032 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5033 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5034 i = VOP_STRATEGY(vp, bp);
5035 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5039 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5042 bufobj_init(struct bufobj *bo, void *private)
5044 static volatile int bufobj_cleanq;
5047 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5048 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5049 bo->bo_private = private;
5050 TAILQ_INIT(&bo->bo_clean.bv_hd);
5051 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5055 bufobj_wrefl(struct bufobj *bo)
5058 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5059 ASSERT_BO_WLOCKED(bo);
5064 bufobj_wref(struct bufobj *bo)
5067 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5074 bufobj_wdrop(struct bufobj *bo)
5077 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5079 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5080 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5081 bo->bo_flag &= ~BO_WWAIT;
5082 wakeup(&bo->bo_numoutput);
5088 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5092 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5093 ASSERT_BO_WLOCKED(bo);
5095 while (bo->bo_numoutput) {
5096 bo->bo_flag |= BO_WWAIT;
5097 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5098 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5106 * Set bio_data or bio_ma for struct bio from the struct buf.
5109 bdata2bio(struct buf *bp, struct bio *bip)
5112 if (!buf_mapped(bp)) {
5113 KASSERT(unmapped_buf_allowed, ("unmapped"));
5114 bip->bio_ma = bp->b_pages;
5115 bip->bio_ma_n = bp->b_npages;
5116 bip->bio_data = unmapped_buf;
5117 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5118 bip->bio_flags |= BIO_UNMAPPED;
5119 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5120 PAGE_SIZE == bp->b_npages,
5121 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5122 (long long)bip->bio_length, bip->bio_ma_n));
5124 bip->bio_data = bp->b_data;
5130 * The MIPS pmap code currently doesn't handle aliased pages.
5131 * The VIPT caches may not handle page aliasing themselves, leading
5132 * to data corruption.
5134 * As such, this code makes a system extremely unhappy if said
5135 * system doesn't support unaliasing the above situation in hardware.
5136 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5137 * this feature at build time, so it has to be handled in software.
5139 * Once the MIPS pmap/cache code grows to support this function on
5140 * earlier chips, it should be flipped back off.
5143 static int buf_pager_relbuf = 1;
5145 static int buf_pager_relbuf = 0;
5147 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5148 &buf_pager_relbuf, 0,
5149 "Make buffer pager release buffers after reading");
5152 * The buffer pager. It uses buffer reads to validate pages.
5154 * In contrast to the generic local pager from vm/vnode_pager.c, this
5155 * pager correctly and easily handles volumes where the underlying
5156 * device block size is greater than the machine page size. The
5157 * buffer cache transparently extends the requested page run to be
5158 * aligned at the block boundary, and does the necessary bogus page
5159 * replacements in the addends to avoid obliterating already valid
5162 * The only non-trivial issue is that the exclusive busy state for
5163 * pages, which is assumed by the vm_pager_getpages() interface, is
5164 * incompatible with the VMIO buffer cache's desire to share-busy the
5165 * pages. This function performs a trivial downgrade of the pages'
5166 * state before reading buffers, and a less trivial upgrade from the
5167 * shared-busy to excl-busy state after the read.
5170 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5171 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5172 vbg_get_blksize_t get_blksize)
5179 vm_ooffset_t la, lb, poff, poffe;
5181 int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5184 object = vp->v_object;
5187 la = IDX_TO_OFF(ma[count - 1]->pindex);
5188 if (la >= object->un_pager.vnp.vnp_size)
5189 return (VM_PAGER_BAD);
5192 * Change the meaning of la from where the last requested page starts
5193 * to where it ends, because that's the end of the requested region
5194 * and the start of the potential read-ahead region.
5197 lpart = la > object->un_pager.vnp.vnp_size;
5198 bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5201 * Calculate read-ahead, behind and total pages.
5204 lb = IDX_TO_OFF(ma[0]->pindex);
5205 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5207 if (rbehind != NULL)
5209 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5210 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5211 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5216 VM_CNT_INC(v_vnodein);
5217 VM_CNT_ADD(v_vnodepgsin, pgsin);
5219 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5220 != 0) ? GB_UNMAPPED : 0;
5222 for (i = 0; i < count; i++) {
5223 if (ma[i] != bogus_page)
5224 vm_page_busy_downgrade(ma[i]);
5228 for (i = 0; i < count; i++) {
5230 if (m == bogus_page)
5234 * Pages are shared busy and the object lock is not
5235 * owned, which together allow for the pages'
5236 * invalidation. The racy test for validity avoids
5237 * useless creation of the buffer for the most typical
5238 * case when invalidation is not used in redo or for
5239 * parallel read. The shared->excl upgrade loop at
5240 * the end of the function catches the race in a
5241 * reliable way (protected by the object lock).
5243 if (vm_page_all_valid(m))
5246 poff = IDX_TO_OFF(m->pindex);
5247 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5248 for (; poff < poffe; poff += bsize) {
5249 lbn = get_lblkno(vp, poff);
5254 bsize = get_blksize(vp, lbn);
5255 error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5259 if (bp->b_rcred == curthread->td_ucred) {
5260 crfree(bp->b_rcred);
5261 bp->b_rcred = NOCRED;
5263 if (LIST_EMPTY(&bp->b_dep)) {
5265 * Invalidation clears m->valid, but
5266 * may leave B_CACHE flag if the
5267 * buffer existed at the invalidation
5268 * time. In this case, recycle the
5269 * buffer to do real read on next
5270 * bread() after redo.
5272 * Otherwise B_RELBUF is not strictly
5273 * necessary, enable to reduce buf
5276 if (buf_pager_relbuf ||
5277 !vm_page_all_valid(m))
5278 bp->b_flags |= B_RELBUF;
5280 bp->b_flags &= ~B_NOCACHE;
5286 KASSERT(1 /* racy, enable for debugging */ ||
5287 vm_page_all_valid(m) || i == count - 1,
5288 ("buf %d %p invalid", i, m));
5289 if (i == count - 1 && lpart) {
5290 if (!vm_page_none_valid(m) &&
5291 !vm_page_all_valid(m))
5292 vm_page_zero_invalid(m, TRUE);
5299 for (i = 0; i < count; i++) {
5300 if (ma[i] == bogus_page)
5302 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5303 vm_page_sunbusy(ma[i]);
5304 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5309 * Since the pages were only sbusy while neither the
5310 * buffer nor the object lock was held by us, or
5311 * reallocated while vm_page_grab() slept for busy
5312 * relinguish, they could have been invalidated.
5313 * Recheck the valid bits and re-read as needed.
5315 * Note that the last page is made fully valid in the
5316 * read loop, and partial validity for the page at
5317 * index count - 1 could mean that the page was
5318 * invalidated or removed, so we must restart for
5321 if (!vm_page_all_valid(ma[i]))
5324 if (redo && error == 0)
5326 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5329 #include "opt_ddb.h"
5331 #include <ddb/ddb.h>
5333 /* DDB command to show buffer data */
5334 DB_SHOW_COMMAND(buffer, db_show_buffer)
5337 struct buf *bp = (struct buf *)addr;
5338 #ifdef FULL_BUF_TRACKING
5343 db_printf("usage: show buffer <addr>\n");
5347 db_printf("buf at %p\n", bp);
5348 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5349 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5350 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5351 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5352 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5353 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5355 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5356 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5357 "b_vp = %p, b_dep = %p\n",
5358 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5359 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5360 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5361 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5362 bp->b_kvabase, bp->b_kvasize);
5365 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5366 for (i = 0; i < bp->b_npages; i++) {
5370 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5372 (u_long)VM_PAGE_TO_PHYS(m));
5374 db_printf("( ??? )");
5375 if ((i + 1) < bp->b_npages)
5380 BUF_LOCKPRINTINFO(bp);
5381 #if defined(FULL_BUF_TRACKING)
5382 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5384 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5385 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5386 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5388 db_printf(" %2u: %s\n", j,
5389 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5391 #elif defined(BUF_TRACKING)
5392 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5397 DB_SHOW_COMMAND(bufqueues, bufqueues)
5399 struct bufdomain *bd;
5404 db_printf("bqempty: %d\n", bqempty.bq_len);
5406 for (i = 0; i < buf_domains; i++) {
5408 db_printf("Buf domain %d\n", i);
5409 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5410 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5411 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5413 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5414 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5415 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5416 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5417 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5419 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5420 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5421 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5422 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5425 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5426 total += bp->b_bufsize;
5427 db_printf("\tcleanq count\t%d (%ld)\n",
5428 bd->bd_cleanq->bq_len, total);
5430 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5431 total += bp->b_bufsize;
5432 db_printf("\tdirtyq count\t%d (%ld)\n",
5433 bd->bd_dirtyq.bq_len, total);
5434 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5435 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5436 db_printf("\tCPU ");
5437 for (j = 0; j <= mp_maxid; j++)
5438 db_printf("%d, ", bd->bd_subq[j].bq_len);
5442 for (j = 0; j < nbuf; j++) {
5444 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5446 total += bp->b_bufsize;
5449 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5452 for (j = 0; j < nbuf; j++) {
5454 if (bp->b_domain == i) {
5456 total += bp->b_bufsize;
5459 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5463 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5468 for (i = 0; i < nbuf; i++) {
5470 if (BUF_ISLOCKED(bp)) {
5471 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5479 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5485 db_printf("usage: show vnodebufs <addr>\n");
5488 vp = (struct vnode *)addr;
5489 db_printf("Clean buffers:\n");
5490 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5491 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5494 db_printf("Dirty buffers:\n");
5495 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5496 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5501 DB_COMMAND(countfreebufs, db_coundfreebufs)
5504 int i, used = 0, nfree = 0;
5507 db_printf("usage: countfreebufs\n");
5511 for (i = 0; i < nbuf; i++) {
5513 if (bp->b_qindex == QUEUE_EMPTY)
5519 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5521 db_printf("numfreebuffers is %d\n", numfreebuffers);