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/resourcevar.h>
71 #include <sys/rwlock.h>
73 #include <sys/sysctl.h>
74 #include <sys/sysproto.h>
76 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/watchdog.h>
79 #include <geom/geom.h>
81 #include <vm/vm_param.h>
82 #include <vm/vm_kern.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_extern.h>
88 #include <vm/vm_map.h>
89 #include <vm/swap_pager.h>
91 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
93 struct bio_ops bioops; /* I/O operation notification */
95 struct buf_ops buf_ops_bio = {
96 .bop_name = "buf_ops_bio",
97 .bop_write = bufwrite,
98 .bop_strategy = bufstrategy,
100 .bop_bdflush = bufbdflush,
104 struct mtx_padalign bq_lock;
105 TAILQ_HEAD(, buf) bq_queue;
107 uint16_t bq_subqueue;
109 } __aligned(CACHE_LINE_SIZE);
111 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
112 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
113 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
114 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
117 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
118 struct bufqueue bd_dirtyq;
119 struct bufqueue *bd_cleanq;
120 struct mtx_padalign bd_run_lock;
125 long bd_bufspacethresh;
126 int bd_hifreebuffers;
127 int bd_lofreebuffers;
128 int bd_hidirtybuffers;
129 int bd_lodirtybuffers;
130 int bd_dirtybufthresh;
134 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
135 int __aligned(CACHE_LINE_SIZE) bd_running;
136 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
137 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
138 } __aligned(CACHE_LINE_SIZE);
140 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
141 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
142 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
143 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
144 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
145 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
146 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
147 #define BD_DOMAIN(bd) (bd - bdomain)
149 static struct buf *buf; /* buffer header pool */
150 extern struct buf *swbuf; /* Swap buffer header pool. */
151 caddr_t unmapped_buf;
153 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
154 struct proc *bufdaemonproc;
156 static int inmem(struct vnode *vp, daddr_t blkno);
157 static void vm_hold_free_pages(struct buf *bp, int newbsize);
158 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
160 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
161 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
163 static void vfs_clean_pages_dirty_buf(struct buf *bp);
164 static void vfs_setdirty_locked_object(struct buf *bp);
165 static void vfs_vmio_invalidate(struct buf *bp);
166 static void vfs_vmio_truncate(struct buf *bp, int npages);
167 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
168 static int vfs_bio_clcheck(struct vnode *vp, int size,
169 daddr_t lblkno, daddr_t blkno);
170 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
171 void (*)(struct buf *));
172 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
173 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
174 static void buf_daemon(void);
175 static __inline void bd_wakeup(void);
176 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
177 static void bufkva_reclaim(vmem_t *, int);
178 static void bufkva_free(struct buf *);
179 static int buf_import(void *, void **, int, int, int);
180 static void buf_release(void *, void **, int);
181 static void maxbcachebuf_adjust(void);
182 static inline struct bufdomain *bufdomain(struct buf *);
183 static void bq_remove(struct bufqueue *bq, struct buf *bp);
184 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
185 static int buf_recycle(struct bufdomain *, bool kva);
186 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
187 const char *lockname);
188 static void bd_init(struct bufdomain *bd);
189 static int bd_flushall(struct bufdomain *bd);
190 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
191 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
193 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
194 int vmiodirenable = TRUE;
195 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
196 "Use the VM system for directory writes");
197 long runningbufspace;
198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
199 "Amount of presently outstanding async buffer io");
200 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
201 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
202 static counter_u64_t bufkvaspace;
203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
204 "Kernel virtual memory used for buffers");
205 static long maxbufspace;
206 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
207 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
208 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
209 "Maximum allowed value of bufspace (including metadata)");
210 static long bufmallocspace;
211 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
212 "Amount of malloced memory for buffers");
213 static long maxbufmallocspace;
214 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
215 0, "Maximum amount of malloced memory for buffers");
216 static long lobufspace;
217 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
218 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
219 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
220 "Minimum amount of buffers we want to have");
222 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
223 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
224 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
225 "Maximum allowed value of bufspace (excluding metadata)");
227 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
228 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
229 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
230 "Bufspace consumed before waking the daemon to free some");
231 static counter_u64_t buffreekvacnt;
232 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
233 "Number of times we have freed the KVA space from some buffer");
234 static counter_u64_t bufdefragcnt;
235 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
236 "Number of times we have had to repeat buffer allocation to defragment");
237 static long lorunningspace;
238 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
239 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
240 "Minimum preferred space used for in-progress I/O");
241 static long hirunningspace;
242 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
243 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
244 "Maximum amount of space to use for in-progress I/O");
245 int dirtybufferflushes;
246 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
247 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
249 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
250 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
251 int altbufferflushes;
252 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
253 0, "Number of fsync flushes to limit dirty buffers");
254 static int recursiveflushes;
255 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
256 0, "Number of flushes skipped due to being recursive");
257 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
258 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
259 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
260 "Number of buffers that are dirty (has unwritten changes) at the moment");
261 static int lodirtybuffers;
262 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
263 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
264 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
265 "How many buffers we want to have free before bufdaemon can sleep");
266 static int hidirtybuffers;
267 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
268 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
269 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
270 "When the number of dirty buffers is considered severe");
272 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
273 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
274 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
275 "Number of bdwrite to bawrite conversions to clear dirty buffers");
276 static int numfreebuffers;
277 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
278 "Number of free buffers");
279 static int lofreebuffers;
280 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
281 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
282 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
283 "Target number of free buffers");
284 static int hifreebuffers;
285 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
286 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
287 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
288 "Threshold for clean buffer recycling");
289 static counter_u64_t getnewbufcalls;
290 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
291 &getnewbufcalls, "Number of calls to getnewbuf");
292 static counter_u64_t getnewbufrestarts;
293 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
295 "Number of times getnewbuf has had to restart a buffer acquisition");
296 static counter_u64_t mappingrestarts;
297 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
299 "Number of times getblk has had to restart a buffer mapping for "
301 static counter_u64_t numbufallocfails;
302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
303 &numbufallocfails, "Number of times buffer allocations failed");
304 static int flushbufqtarget = 100;
305 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
306 "Amount of work to do in flushbufqueues when helping bufdaemon");
307 static counter_u64_t notbufdflushes;
308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
309 "Number of dirty buffer flushes done by the bufdaemon helpers");
310 static long barrierwrites;
311 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
312 "Number of barrier writes");
313 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
314 &unmapped_buf_allowed, 0,
315 "Permit the use of the unmapped i/o");
316 int maxbcachebuf = MAXBCACHEBUF;
317 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
318 "Maximum size of a buffer cache block");
321 * This lock synchronizes access to bd_request.
323 static struct mtx_padalign __exclusive_cache_line bdlock;
326 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
327 * waitrunningbufspace().
329 static struct mtx_padalign __exclusive_cache_line rbreqlock;
332 * Lock that protects bdirtywait.
334 static struct mtx_padalign __exclusive_cache_line bdirtylock;
337 * Wakeup point for bufdaemon, as well as indicator of whether it is already
338 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
341 static int bd_request;
344 * Request for the buf daemon to write more buffers than is indicated by
345 * lodirtybuf. This may be necessary to push out excess dependencies or
346 * defragment the address space where a simple count of the number of dirty
347 * buffers is insufficient to characterize the demand for flushing them.
349 static int bd_speedupreq;
352 * Synchronization (sleep/wakeup) variable for active buffer space requests.
353 * Set when wait starts, cleared prior to wakeup().
354 * Used in runningbufwakeup() and waitrunningbufspace().
356 static int runningbufreq;
359 * Synchronization for bwillwrite() waiters.
361 static int bdirtywait;
364 * Definitions for the buffer free lists.
366 #define QUEUE_NONE 0 /* on no queue */
367 #define QUEUE_EMPTY 1 /* empty buffer headers */
368 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
369 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
370 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
372 /* Maximum number of buffer domains. */
373 #define BUF_DOMAINS 8
375 struct bufdomainset bdlodirty; /* Domains > lodirty */
376 struct bufdomainset bdhidirty; /* Domains > hidirty */
378 /* Configured number of clean queues. */
379 static int __read_mostly buf_domains;
381 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
382 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
383 struct bufqueue __exclusive_cache_line bqempty;
386 * per-cpu empty buffer cache.
391 * Single global constant for BUF_WMESG, to avoid getting multiple references.
392 * buf_wmesg is referred from macros.
394 const char *buf_wmesg = BUF_WMESG;
397 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
402 value = *(long *)arg1;
403 error = sysctl_handle_long(oidp, &value, 0, req);
404 if (error != 0 || req->newptr == NULL)
406 mtx_lock(&rbreqlock);
407 if (arg1 == &hirunningspace) {
408 if (value < lorunningspace)
411 hirunningspace = value;
413 KASSERT(arg1 == &lorunningspace,
414 ("%s: unknown arg1", __func__));
415 if (value > hirunningspace)
418 lorunningspace = value;
420 mtx_unlock(&rbreqlock);
425 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
431 value = *(int *)arg1;
432 error = sysctl_handle_int(oidp, &value, 0, req);
433 if (error != 0 || req->newptr == NULL)
435 *(int *)arg1 = value;
436 for (i = 0; i < buf_domains; i++)
437 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
444 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
450 value = *(long *)arg1;
451 error = sysctl_handle_long(oidp, &value, 0, req);
452 if (error != 0 || req->newptr == NULL)
454 *(long *)arg1 = value;
455 for (i = 0; i < buf_domains; i++)
456 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
462 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
463 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
465 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
472 for (i = 0; i < buf_domains; i++)
473 lvalue += bdomain[i].bd_bufspace;
474 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
475 return (sysctl_handle_long(oidp, &lvalue, 0, req));
476 if (lvalue > INT_MAX)
477 /* On overflow, still write out a long to trigger ENOMEM. */
478 return (sysctl_handle_long(oidp, &lvalue, 0, req));
480 return (sysctl_handle_int(oidp, &ivalue, 0, req));
484 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
490 for (i = 0; i < buf_domains; i++)
491 lvalue += bdomain[i].bd_bufspace;
492 return (sysctl_handle_long(oidp, &lvalue, 0, req));
497 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
503 for (i = 0; i < buf_domains; i++)
504 value += bdomain[i].bd_numdirtybuffers;
505 return (sysctl_handle_int(oidp, &value, 0, req));
511 * Wakeup any bwillwrite() waiters.
516 mtx_lock(&bdirtylock);
521 mtx_unlock(&bdirtylock);
527 * Clear a domain from the appropriate bitsets when dirtybuffers
531 bd_clear(struct bufdomain *bd)
534 mtx_lock(&bdirtylock);
535 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
536 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
537 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
538 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
539 mtx_unlock(&bdirtylock);
545 * Set a domain in the appropriate bitsets when dirtybuffers
549 bd_set(struct bufdomain *bd)
552 mtx_lock(&bdirtylock);
553 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
554 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
555 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
556 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
557 mtx_unlock(&bdirtylock);
563 * Decrement the numdirtybuffers count by one and wakeup any
564 * threads blocked in bwillwrite().
567 bdirtysub(struct buf *bp)
569 struct bufdomain *bd;
573 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
574 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
576 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
583 * Increment the numdirtybuffers count by one and wakeup the buf
587 bdirtyadd(struct buf *bp)
589 struct bufdomain *bd;
593 * Only do the wakeup once as we cross the boundary. The
594 * buf daemon will keep running until the condition clears.
597 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
598 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
600 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
605 * bufspace_daemon_wakeup:
607 * Wakeup the daemons responsible for freeing clean bufs.
610 bufspace_daemon_wakeup(struct bufdomain *bd)
614 * avoid the lock if the daemon is running.
616 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
618 atomic_store_int(&bd->bd_running, 1);
619 wakeup(&bd->bd_running);
625 * bufspace_daemon_wait:
627 * Sleep until the domain falls below a limit or one second passes.
630 bufspace_daemon_wait(struct bufdomain *bd)
633 * Re-check our limits and sleep. bd_running must be
634 * cleared prior to checking the limits to avoid missed
635 * wakeups. The waker will adjust one of bufspace or
636 * freebuffers prior to checking bd_running.
639 atomic_store_int(&bd->bd_running, 0);
640 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
641 bd->bd_freebuffers > bd->bd_lofreebuffers) {
642 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
645 /* Avoid spurious wakeups while running. */
646 atomic_store_int(&bd->bd_running, 1);
654 * Adjust the reported bufspace for a KVA managed buffer, possibly
655 * waking any waiters.
658 bufspace_adjust(struct buf *bp, int bufsize)
660 struct bufdomain *bd;
664 KASSERT((bp->b_flags & B_MALLOC) == 0,
665 ("bufspace_adjust: malloc buf %p", bp));
667 diff = bufsize - bp->b_bufsize;
669 atomic_subtract_long(&bd->bd_bufspace, -diff);
670 } else if (diff > 0) {
671 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
672 /* Wake up the daemon on the transition. */
673 if (space < bd->bd_bufspacethresh &&
674 space + diff >= bd->bd_bufspacethresh)
675 bufspace_daemon_wakeup(bd);
677 bp->b_bufsize = bufsize;
683 * Reserve bufspace before calling allocbuf(). metadata has a
684 * different space limit than data.
687 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
693 limit = bd->bd_maxbufspace;
695 limit = bd->bd_hibufspace;
696 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
699 atomic_subtract_long(&bd->bd_bufspace, size);
703 /* Wake up the daemon on the transition. */
704 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
705 bufspace_daemon_wakeup(bd);
713 * Release reserved bufspace after bufspace_adjust() has consumed it.
716 bufspace_release(struct bufdomain *bd, int size)
719 atomic_subtract_long(&bd->bd_bufspace, size);
725 * Wait for bufspace, acting as the buf daemon if a locked vnode is
726 * supplied. bd_wanted must be set prior to polling for space. The
727 * operation must be re-tried on return.
730 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
731 int slpflag, int slptimeo)
734 int error, fl, norunbuf;
736 if ((gbflags & GB_NOWAIT_BD) != 0)
741 while (bd->bd_wanted) {
742 if (vp != NULL && vp->v_type != VCHR &&
743 (td->td_pflags & TDP_BUFNEED) == 0) {
746 * getblk() is called with a vnode locked, and
747 * some majority of the dirty buffers may as
748 * well belong to the vnode. Flushing the
749 * buffers there would make a progress that
750 * cannot be achieved by the buf_daemon, that
751 * cannot lock the vnode.
753 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
754 (td->td_pflags & TDP_NORUNNINGBUF);
757 * Play bufdaemon. The getnewbuf() function
758 * may be called while the thread owns lock
759 * for another dirty buffer for the same
760 * vnode, which makes it impossible to use
761 * VOP_FSYNC() there, due to the buffer lock
764 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
765 fl = buf_flush(vp, bd, flushbufqtarget);
766 td->td_pflags &= norunbuf;
770 if (bd->bd_wanted == 0)
773 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
774 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
785 * buffer space management daemon. Tries to maintain some marginal
786 * amount of free buffer space so that requesting processes neither
787 * block nor work to reclaim buffers.
790 bufspace_daemon(void *arg)
792 struct bufdomain *bd;
794 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
795 SHUTDOWN_PRI_LAST + 100);
799 kthread_suspend_check();
802 * Free buffers from the clean queue until we meet our
805 * Theory of operation: The buffer cache is most efficient
806 * when some free buffer headers and space are always
807 * available to getnewbuf(). This daemon attempts to prevent
808 * the excessive blocking and synchronization associated
809 * with shortfall. It goes through three phases according
812 * 1) The daemon wakes up voluntarily once per-second
813 * during idle periods when the counters are below
814 * the wakeup thresholds (bufspacethresh, lofreebuffers).
816 * 2) The daemon wakes up as we cross the thresholds
817 * ahead of any potential blocking. This may bounce
818 * slightly according to the rate of consumption and
821 * 3) The daemon and consumers are starved for working
822 * clean buffers. This is the 'bufspace' sleep below
823 * which will inefficiently trade bufs with bqrelse
824 * until we return to condition 2.
826 while (bd->bd_bufspace > bd->bd_lobufspace ||
827 bd->bd_freebuffers < bd->bd_hifreebuffers) {
828 if (buf_recycle(bd, false) != 0) {
832 * Speedup dirty if we've run out of clean
833 * buffers. This is possible in particular
834 * because softdep may held many bufs locked
835 * pending writes to other bufs which are
836 * marked for delayed write, exhausting
837 * clean space until they are written.
842 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
843 PRIBIO|PDROP, "bufspace", hz/10);
849 bufspace_daemon_wait(bd);
856 * Adjust the reported bufspace for a malloc managed buffer, possibly
857 * waking any waiters.
860 bufmallocadjust(struct buf *bp, int bufsize)
864 KASSERT((bp->b_flags & B_MALLOC) != 0,
865 ("bufmallocadjust: non-malloc buf %p", bp));
866 diff = bufsize - bp->b_bufsize;
868 atomic_subtract_long(&bufmallocspace, -diff);
870 atomic_add_long(&bufmallocspace, diff);
871 bp->b_bufsize = bufsize;
877 * Wake up processes that are waiting on asynchronous writes to fall
878 * below lorunningspace.
884 mtx_lock(&rbreqlock);
887 wakeup(&runningbufreq);
889 mtx_unlock(&rbreqlock);
895 * Decrement the outstanding write count according.
898 runningbufwakeup(struct buf *bp)
902 bspace = bp->b_runningbufspace;
905 space = atomic_fetchadd_long(&runningbufspace, -bspace);
906 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
908 bp->b_runningbufspace = 0;
910 * Only acquire the lock and wakeup on the transition from exceeding
911 * the threshold to falling below it.
913 if (space < lorunningspace)
915 if (space - bspace > lorunningspace)
921 * waitrunningbufspace()
923 * runningbufspace is a measure of the amount of I/O currently
924 * running. This routine is used in async-write situations to
925 * prevent creating huge backups of pending writes to a device.
926 * Only asynchronous writes are governed by this function.
928 * This does NOT turn an async write into a sync write. It waits
929 * for earlier writes to complete and generally returns before the
930 * caller's write has reached the device.
933 waitrunningbufspace(void)
936 mtx_lock(&rbreqlock);
937 while (runningbufspace > hirunningspace) {
939 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
941 mtx_unlock(&rbreqlock);
946 * vfs_buf_test_cache:
948 * Called when a buffer is extended. This function clears the B_CACHE
949 * bit if the newly extended portion of the buffer does not contain
953 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
954 vm_offset_t size, vm_page_t m)
957 VM_OBJECT_ASSERT_LOCKED(m->object);
958 if (bp->b_flags & B_CACHE) {
959 int base = (foff + off) & PAGE_MASK;
960 if (vm_page_is_valid(m, base, size) == 0)
961 bp->b_flags &= ~B_CACHE;
965 /* Wake up the buffer daemon if necessary */
971 if (bd_request == 0) {
979 * Adjust the maxbcachbuf tunable.
982 maxbcachebuf_adjust(void)
987 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
990 while (i * 2 <= maxbcachebuf)
993 if (maxbcachebuf < MAXBSIZE)
994 maxbcachebuf = MAXBSIZE;
995 if (maxbcachebuf > MAXPHYS)
996 maxbcachebuf = MAXPHYS;
997 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
998 printf("maxbcachebuf=%d\n", maxbcachebuf);
1002 * bd_speedup - speedup the buffer cache flushing code
1011 if (bd_speedupreq == 0 || bd_request == 0)
1016 wakeup(&bd_request);
1017 mtx_unlock(&bdlock);
1021 #define TRANSIENT_DENOM 5
1023 #define TRANSIENT_DENOM 10
1027 * Calculating buffer cache scaling values and reserve space for buffer
1028 * headers. This is called during low level kernel initialization and
1029 * may be called more then once. We CANNOT write to the memory area
1030 * being reserved at this time.
1033 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1036 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1039 * physmem_est is in pages. Convert it to kilobytes (assumes
1040 * PAGE_SIZE is >= 1K)
1042 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1044 maxbcachebuf_adjust();
1046 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1047 * For the first 64MB of ram nominally allocate sufficient buffers to
1048 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1049 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1050 * the buffer cache we limit the eventual kva reservation to
1053 * factor represents the 1/4 x ram conversion.
1056 int factor = 4 * BKVASIZE / 1024;
1059 if (physmem_est > 4096)
1060 nbuf += min((physmem_est - 4096) / factor,
1062 if (physmem_est > 65536)
1063 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1064 32 * 1024 * 1024 / (factor * 5));
1066 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1067 nbuf = maxbcache / BKVASIZE;
1072 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1073 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1074 if (nbuf > maxbuf) {
1076 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1082 * Ideal allocation size for the transient bio submap is 10%
1083 * of the maximal space buffer map. This roughly corresponds
1084 * to the amount of the buffer mapped for typical UFS load.
1086 * Clip the buffer map to reserve space for the transient
1087 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1088 * maximum buffer map extent on the platform.
1090 * The fall-back to the maxbuf in case of maxbcache unset,
1091 * allows to not trim the buffer KVA for the architectures
1092 * with ample KVA space.
1094 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1095 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1096 buf_sz = (long)nbuf * BKVASIZE;
1097 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1098 (TRANSIENT_DENOM - 1)) {
1100 * There is more KVA than memory. Do not
1101 * adjust buffer map size, and assign the rest
1102 * of maxbuf to transient map.
1104 biotmap_sz = maxbuf_sz - buf_sz;
1107 * Buffer map spans all KVA we could afford on
1108 * this platform. Give 10% (20% on i386) of
1109 * the buffer map to the transient bio map.
1111 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1112 buf_sz -= biotmap_sz;
1114 if (biotmap_sz / INT_MAX > MAXPHYS)
1115 bio_transient_maxcnt = INT_MAX;
1117 bio_transient_maxcnt = biotmap_sz / MAXPHYS;
1119 * Artificially limit to 1024 simultaneous in-flight I/Os
1120 * using the transient mapping.
1122 if (bio_transient_maxcnt > 1024)
1123 bio_transient_maxcnt = 1024;
1125 nbuf = buf_sz / BKVASIZE;
1129 nswbuf = min(nbuf / 4, 256);
1130 if (nswbuf < NSWBUF_MIN)
1131 nswbuf = NSWBUF_MIN;
1135 * Reserve space for the buffer cache buffers
1138 v = (caddr_t)(buf + nbuf);
1143 /* Initialize the buffer subsystem. Called before use of any buffers. */
1150 KASSERT(maxbcachebuf >= MAXBSIZE,
1151 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1153 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1154 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1155 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1156 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1158 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
1160 /* finally, initialize each buffer header and stick on empty q */
1161 for (i = 0; i < nbuf; i++) {
1163 bzero(bp, sizeof *bp);
1164 bp->b_flags = B_INVAL;
1165 bp->b_rcred = NOCRED;
1166 bp->b_wcred = NOCRED;
1167 bp->b_qindex = QUEUE_NONE;
1169 bp->b_subqueue = mp_maxid + 1;
1171 bp->b_data = bp->b_kvabase = unmapped_buf;
1172 LIST_INIT(&bp->b_dep);
1174 bq_insert(&bqempty, bp, false);
1178 * maxbufspace is the absolute maximum amount of buffer space we are
1179 * allowed to reserve in KVM and in real terms. The absolute maximum
1180 * is nominally used by metadata. hibufspace is the nominal maximum
1181 * used by most other requests. The differential is required to
1182 * ensure that metadata deadlocks don't occur.
1184 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1185 * this may result in KVM fragmentation which is not handled optimally
1186 * by the system. XXX This is less true with vmem. We could use
1189 maxbufspace = (long)nbuf * BKVASIZE;
1190 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1191 lobufspace = (hibufspace / 20) * 19; /* 95% */
1192 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1195 * Note: The 16 MiB upper limit for hirunningspace was chosen
1196 * arbitrarily and may need further tuning. It corresponds to
1197 * 128 outstanding write IO requests (if IO size is 128 KiB),
1198 * which fits with many RAID controllers' tagged queuing limits.
1199 * The lower 1 MiB limit is the historical upper limit for
1202 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1203 16 * 1024 * 1024), 1024 * 1024);
1204 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1207 * Limit the amount of malloc memory since it is wired permanently into
1208 * the kernel space. Even though this is accounted for in the buffer
1209 * allocation, we don't want the malloced region to grow uncontrolled.
1210 * The malloc scheme improves memory utilization significantly on
1211 * average (small) directories.
1213 maxbufmallocspace = hibufspace / 20;
1216 * Reduce the chance of a deadlock occurring by limiting the number
1217 * of delayed-write dirty buffers we allow to stack up.
1219 hidirtybuffers = nbuf / 4 + 20;
1220 dirtybufthresh = hidirtybuffers * 9 / 10;
1222 * To support extreme low-memory systems, make sure hidirtybuffers
1223 * cannot eat up all available buffer space. This occurs when our
1224 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1225 * buffer space assuming BKVASIZE'd buffers.
1227 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1228 hidirtybuffers >>= 1;
1230 lodirtybuffers = hidirtybuffers / 2;
1233 * lofreebuffers should be sufficient to avoid stalling waiting on
1234 * buf headers under heavy utilization. The bufs in per-cpu caches
1235 * are counted as free but will be unavailable to threads executing
1238 * hifreebuffers is the free target for the bufspace daemon. This
1239 * should be set appropriately to limit work per-iteration.
1241 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1242 hifreebuffers = (3 * lofreebuffers) / 2;
1243 numfreebuffers = nbuf;
1245 /* Setup the kva and free list allocators. */
1246 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1247 buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf),
1248 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1251 * Size the clean queue according to the amount of buffer space.
1252 * One queue per-256mb up to the max. More queues gives better
1253 * concurrency but less accurate LRU.
1255 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1256 for (i = 0 ; i < buf_domains; i++) {
1257 struct bufdomain *bd;
1261 bd->bd_freebuffers = nbuf / buf_domains;
1262 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1263 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1264 bd->bd_bufspace = 0;
1265 bd->bd_maxbufspace = maxbufspace / buf_domains;
1266 bd->bd_hibufspace = hibufspace / buf_domains;
1267 bd->bd_lobufspace = lobufspace / buf_domains;
1268 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1269 bd->bd_numdirtybuffers = 0;
1270 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1271 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1272 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1273 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1274 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1276 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1277 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1278 mappingrestarts = counter_u64_alloc(M_WAITOK);
1279 numbufallocfails = counter_u64_alloc(M_WAITOK);
1280 notbufdflushes = counter_u64_alloc(M_WAITOK);
1281 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1282 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1283 bufkvaspace = counter_u64_alloc(M_WAITOK);
1288 vfs_buf_check_mapped(struct buf *bp)
1291 KASSERT(bp->b_kvabase != unmapped_buf,
1292 ("mapped buf: b_kvabase was not updated %p", bp));
1293 KASSERT(bp->b_data != unmapped_buf,
1294 ("mapped buf: b_data was not updated %p", bp));
1295 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1296 MAXPHYS, ("b_data + b_offset unmapped %p", bp));
1300 vfs_buf_check_unmapped(struct buf *bp)
1303 KASSERT(bp->b_data == unmapped_buf,
1304 ("unmapped buf: corrupted b_data %p", bp));
1307 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1308 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1310 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1311 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1315 isbufbusy(struct buf *bp)
1317 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1318 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1324 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1327 bufshutdown(int show_busybufs)
1329 static int first_buf_printf = 1;
1331 int iter, nbusy, pbusy;
1337 * Sync filesystems for shutdown
1339 wdog_kern_pat(WD_LASTVAL);
1340 sys_sync(curthread, NULL);
1343 * With soft updates, some buffers that are
1344 * written will be remarked as dirty until other
1345 * buffers are written.
1347 for (iter = pbusy = 0; iter < 20; iter++) {
1349 for (bp = &buf[nbuf]; --bp >= buf; )
1353 if (first_buf_printf)
1354 printf("All buffers synced.");
1357 if (first_buf_printf) {
1358 printf("Syncing disks, buffers remaining... ");
1359 first_buf_printf = 0;
1361 printf("%d ", nbusy);
1366 wdog_kern_pat(WD_LASTVAL);
1367 sys_sync(curthread, NULL);
1371 * Spin for a while to allow interrupt threads to run.
1373 DELAY(50000 * iter);
1376 * Context switch several times to allow interrupt
1379 for (subiter = 0; subiter < 50 * iter; subiter++) {
1380 thread_lock(curthread);
1381 mi_switch(SW_VOL, NULL);
1382 thread_unlock(curthread);
1389 * Count only busy local buffers to prevent forcing
1390 * a fsck if we're just a client of a wedged NFS server
1393 for (bp = &buf[nbuf]; --bp >= buf; ) {
1394 if (isbufbusy(bp)) {
1396 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1397 if (bp->b_dev == NULL) {
1398 TAILQ_REMOVE(&mountlist,
1399 bp->b_vp->v_mount, mnt_list);
1404 if (show_busybufs > 0) {
1406 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1407 nbusy, bp, bp->b_vp, bp->b_flags,
1408 (intmax_t)bp->b_blkno,
1409 (intmax_t)bp->b_lblkno);
1410 BUF_LOCKPRINTINFO(bp);
1411 if (show_busybufs > 1)
1419 * Failed to sync all blocks. Indicate this and don't
1420 * unmount filesystems (thus forcing an fsck on reboot).
1422 printf("Giving up on %d buffers\n", nbusy);
1423 DELAY(5000000); /* 5 seconds */
1425 if (!first_buf_printf)
1426 printf("Final sync complete\n");
1428 * Unmount filesystems
1430 if (panicstr == NULL)
1434 DELAY(100000); /* wait for console output to finish */
1438 bpmap_qenter(struct buf *bp)
1441 BUF_CHECK_MAPPED(bp);
1444 * bp->b_data is relative to bp->b_offset, but
1445 * bp->b_offset may be offset into the first page.
1447 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1448 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1449 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1450 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1453 static inline struct bufdomain *
1454 bufdomain(struct buf *bp)
1457 return (&bdomain[bp->b_domain]);
1460 static struct bufqueue *
1461 bufqueue(struct buf *bp)
1464 switch (bp->b_qindex) {
1467 case QUEUE_SENTINEL:
1472 return (&bufdomain(bp)->bd_dirtyq);
1474 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1478 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1482 * Return the locked bufqueue that bp is a member of.
1484 static struct bufqueue *
1485 bufqueue_acquire(struct buf *bp)
1487 struct bufqueue *bq, *nbq;
1490 * bp can be pushed from a per-cpu queue to the
1491 * cleanq while we're waiting on the lock. Retry
1492 * if the queues don't match.
1510 * Insert the buffer into the appropriate free list. Requires a
1511 * locked buffer on entry and buffer is unlocked before return.
1514 binsfree(struct buf *bp, int qindex)
1516 struct bufdomain *bd;
1517 struct bufqueue *bq;
1519 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1520 ("binsfree: Invalid qindex %d", qindex));
1521 BUF_ASSERT_XLOCKED(bp);
1524 * Handle delayed bremfree() processing.
1526 if (bp->b_flags & B_REMFREE) {
1527 if (bp->b_qindex == qindex) {
1528 bp->b_flags |= B_REUSE;
1529 bp->b_flags &= ~B_REMFREE;
1533 bq = bufqueue_acquire(bp);
1538 if (qindex == QUEUE_CLEAN) {
1539 if (bd->bd_lim != 0)
1540 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1544 bq = &bd->bd_dirtyq;
1545 bq_insert(bq, bp, true);
1551 * Free a buffer to the buf zone once it no longer has valid contents.
1554 buf_free(struct buf *bp)
1557 if (bp->b_flags & B_REMFREE)
1559 if (bp->b_vflags & BV_BKGRDINPROG)
1560 panic("losing buffer 1");
1561 if (bp->b_rcred != NOCRED) {
1562 crfree(bp->b_rcred);
1563 bp->b_rcred = NOCRED;
1565 if (bp->b_wcred != NOCRED) {
1566 crfree(bp->b_wcred);
1567 bp->b_wcred = NOCRED;
1569 if (!LIST_EMPTY(&bp->b_dep))
1572 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1574 uma_zfree(buf_zone, bp);
1580 * Import bufs into the uma cache from the buf list. The system still
1581 * expects a static array of bufs and much of the synchronization
1582 * around bufs assumes type stable storage. As a result, UMA is used
1583 * only as a per-cpu cache of bufs still maintained on a global list.
1586 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1592 for (i = 0; i < cnt; i++) {
1593 bp = TAILQ_FIRST(&bqempty.bq_queue);
1596 bq_remove(&bqempty, bp);
1599 BQ_UNLOCK(&bqempty);
1607 * Release bufs from the uma cache back to the buffer queues.
1610 buf_release(void *arg, void **store, int cnt)
1612 struct bufqueue *bq;
1618 for (i = 0; i < cnt; i++) {
1620 /* Inline bq_insert() to batch locking. */
1621 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1622 bp->b_flags &= ~(B_AGE | B_REUSE);
1624 bp->b_qindex = bq->bq_index;
1632 * Allocate an empty buffer header.
1635 buf_alloc(struct bufdomain *bd)
1641 * We can only run out of bufs in the buf zone if the average buf
1642 * is less than BKVASIZE. In this case the actual wait/block will
1643 * come from buf_reycle() failing to flush one of these small bufs.
1646 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1648 bp = uma_zalloc(buf_zone, M_NOWAIT);
1650 atomic_add_int(&bd->bd_freebuffers, 1);
1651 bufspace_daemon_wakeup(bd);
1652 counter_u64_add(numbufallocfails, 1);
1656 * Wake-up the bufspace daemon on transition below threshold.
1658 if (freebufs == bd->bd_lofreebuffers)
1659 bufspace_daemon_wakeup(bd);
1661 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1662 panic("getnewbuf_empty: Locked buf %p on free queue.", bp);
1664 KASSERT(bp->b_vp == NULL,
1665 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1666 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1667 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1668 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1669 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1670 KASSERT(bp->b_npages == 0,
1671 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1672 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1673 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1675 bp->b_domain = BD_DOMAIN(bd);
1681 bp->b_blkno = bp->b_lblkno = 0;
1682 bp->b_offset = NOOFFSET;
1688 bp->b_dirtyoff = bp->b_dirtyend = 0;
1689 bp->b_bufobj = NULL;
1690 bp->b_data = bp->b_kvabase = unmapped_buf;
1691 bp->b_fsprivate1 = NULL;
1692 bp->b_fsprivate2 = NULL;
1693 bp->b_fsprivate3 = NULL;
1694 LIST_INIT(&bp->b_dep);
1702 * Free a buffer from the given bufqueue. kva controls whether the
1703 * freed buf must own some kva resources. This is used for
1707 buf_recycle(struct bufdomain *bd, bool kva)
1709 struct bufqueue *bq;
1710 struct buf *bp, *nbp;
1713 counter_u64_add(bufdefragcnt, 1);
1717 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1718 ("buf_recycle: Locks don't match"));
1719 nbp = TAILQ_FIRST(&bq->bq_queue);
1722 * Run scan, possibly freeing data and/or kva mappings on the fly
1725 while ((bp = nbp) != NULL) {
1727 * Calculate next bp (we can only use it if we do not
1728 * release the bqlock).
1730 nbp = TAILQ_NEXT(bp, b_freelist);
1733 * If we are defragging then we need a buffer with
1734 * some kva to reclaim.
1736 if (kva && bp->b_kvasize == 0)
1739 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1743 * Implement a second chance algorithm for frequently
1746 if ((bp->b_flags & B_REUSE) != 0) {
1747 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1748 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1749 bp->b_flags &= ~B_REUSE;
1755 * Skip buffers with background writes in progress.
1757 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1762 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1763 ("buf_recycle: inconsistent queue %d bp %p",
1765 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1766 ("getnewbuf: queue domain %d doesn't match request %d",
1767 bp->b_domain, (int)BD_DOMAIN(bd)));
1769 * NOTE: nbp is now entirely invalid. We can only restart
1770 * the scan from this point on.
1776 * Requeue the background write buffer with error and
1779 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1782 nbp = TAILQ_FIRST(&bq->bq_queue);
1785 bp->b_flags |= B_INVAL;
1798 * Mark the buffer for removal from the appropriate free list.
1802 bremfree(struct buf *bp)
1805 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1806 KASSERT((bp->b_flags & B_REMFREE) == 0,
1807 ("bremfree: buffer %p already marked for delayed removal.", bp));
1808 KASSERT(bp->b_qindex != QUEUE_NONE,
1809 ("bremfree: buffer %p not on a queue.", bp));
1810 BUF_ASSERT_XLOCKED(bp);
1812 bp->b_flags |= B_REMFREE;
1818 * Force an immediate removal from a free list. Used only in nfs when
1819 * it abuses the b_freelist pointer.
1822 bremfreef(struct buf *bp)
1824 struct bufqueue *bq;
1826 bq = bufqueue_acquire(bp);
1832 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1835 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1836 TAILQ_INIT(&bq->bq_queue);
1838 bq->bq_index = qindex;
1839 bq->bq_subqueue = subqueue;
1843 bd_init(struct bufdomain *bd)
1847 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1848 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1849 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1850 for (i = 0; i <= mp_maxid; i++)
1851 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1852 "bufq clean subqueue lock");
1853 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1859 * Removes a buffer from the free list, must be called with the
1860 * correct qlock held.
1863 bq_remove(struct bufqueue *bq, struct buf *bp)
1866 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1867 bp, bp->b_vp, bp->b_flags);
1868 KASSERT(bp->b_qindex != QUEUE_NONE,
1869 ("bq_remove: buffer %p not on a queue.", bp));
1870 KASSERT(bufqueue(bp) == bq,
1871 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1873 BQ_ASSERT_LOCKED(bq);
1874 if (bp->b_qindex != QUEUE_EMPTY) {
1875 BUF_ASSERT_XLOCKED(bp);
1877 KASSERT(bq->bq_len >= 1,
1878 ("queue %d underflow", bp->b_qindex));
1879 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1881 bp->b_qindex = QUEUE_NONE;
1882 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1886 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1890 BQ_ASSERT_LOCKED(bq);
1891 if (bq != bd->bd_cleanq) {
1893 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1894 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1895 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1897 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1899 bd->bd_cleanq->bq_len += bq->bq_len;
1902 if (bd->bd_wanted) {
1904 wakeup(&bd->bd_wanted);
1906 if (bq != bd->bd_cleanq)
1911 bd_flushall(struct bufdomain *bd)
1913 struct bufqueue *bq;
1917 if (bd->bd_lim == 0)
1920 for (i = 0; i <= mp_maxid; i++) {
1921 bq = &bd->bd_subq[i];
1922 if (bq->bq_len == 0)
1934 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1936 struct bufdomain *bd;
1938 if (bp->b_qindex != QUEUE_NONE)
1939 panic("bq_insert: free buffer %p onto another queue?", bp);
1942 if (bp->b_flags & B_AGE) {
1943 /* Place this buf directly on the real queue. */
1944 if (bq->bq_index == QUEUE_CLEAN)
1947 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1950 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1952 bp->b_flags &= ~(B_AGE | B_REUSE);
1954 bp->b_qindex = bq->bq_index;
1955 bp->b_subqueue = bq->bq_subqueue;
1958 * Unlock before we notify so that we don't wakeup a waiter that
1959 * fails a trylock on the buf and sleeps again.
1964 if (bp->b_qindex == QUEUE_CLEAN) {
1966 * Flush the per-cpu queue and notify any waiters.
1968 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
1969 bq->bq_len >= bd->bd_lim))
1978 * Free the kva allocation for a buffer.
1982 bufkva_free(struct buf *bp)
1986 if (bp->b_kvasize == 0) {
1987 KASSERT(bp->b_kvabase == unmapped_buf &&
1988 bp->b_data == unmapped_buf,
1989 ("Leaked KVA space on %p", bp));
1990 } else if (buf_mapped(bp))
1991 BUF_CHECK_MAPPED(bp);
1993 BUF_CHECK_UNMAPPED(bp);
1995 if (bp->b_kvasize == 0)
1998 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
1999 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2000 counter_u64_add(buffreekvacnt, 1);
2001 bp->b_data = bp->b_kvabase = unmapped_buf;
2008 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2011 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2016 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2017 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2022 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2025 * Buffer map is too fragmented. Request the caller
2026 * to defragment the map.
2030 bp->b_kvabase = (caddr_t)addr;
2031 bp->b_kvasize = maxsize;
2032 counter_u64_add(bufkvaspace, bp->b_kvasize);
2033 if ((gbflags & GB_UNMAPPED) != 0) {
2034 bp->b_data = unmapped_buf;
2035 BUF_CHECK_UNMAPPED(bp);
2037 bp->b_data = bp->b_kvabase;
2038 BUF_CHECK_MAPPED(bp);
2046 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2047 * callback that fires to avoid returning failure.
2050 bufkva_reclaim(vmem_t *vmem, int flags)
2057 for (i = 0; i < 5; i++) {
2058 for (q = 0; q < buf_domains; q++)
2059 if (buf_recycle(&bdomain[q], true) != 0)
2068 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2069 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2070 * the buffer is valid and we do not have to do anything.
2073 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2074 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2082 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2083 if (inmem(vp, *rablkno))
2085 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2086 if ((rabp->b_flags & B_CACHE) != 0) {
2093 racct_add_buf(curproc, rabp, 0);
2094 PROC_UNLOCK(curproc);
2097 td->td_ru.ru_inblock++;
2098 rabp->b_flags |= B_ASYNC;
2099 rabp->b_flags &= ~B_INVAL;
2100 if ((flags & GB_CKHASH) != 0) {
2101 rabp->b_flags |= B_CKHASH;
2102 rabp->b_ckhashcalc = ckhashfunc;
2104 rabp->b_ioflags &= ~BIO_ERROR;
2105 rabp->b_iocmd = BIO_READ;
2106 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2107 rabp->b_rcred = crhold(cred);
2108 vfs_busy_pages(rabp, 0);
2110 rabp->b_iooffset = dbtob(rabp->b_blkno);
2116 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2118 * Get a buffer with the specified data. Look in the cache first. We
2119 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2120 * is set, the buffer is valid and we do not have to do anything, see
2121 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2123 * Always return a NULL buffer pointer (in bpp) when returning an error.
2126 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
2127 int *rabsize, int cnt, struct ucred *cred, int flags,
2128 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2132 int error, readwait, rv;
2134 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2137 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2140 error = getblkx(vp, blkno, size, 0, 0, flags, &bp);
2145 flags &= ~GB_NOSPARSE;
2149 * If not found in cache, do some I/O
2152 if ((bp->b_flags & B_CACHE) == 0) {
2155 PROC_LOCK(td->td_proc);
2156 racct_add_buf(td->td_proc, bp, 0);
2157 PROC_UNLOCK(td->td_proc);
2160 td->td_ru.ru_inblock++;
2161 bp->b_iocmd = BIO_READ;
2162 bp->b_flags &= ~B_INVAL;
2163 if ((flags & GB_CKHASH) != 0) {
2164 bp->b_flags |= B_CKHASH;
2165 bp->b_ckhashcalc = ckhashfunc;
2167 bp->b_ioflags &= ~BIO_ERROR;
2168 if (bp->b_rcred == NOCRED && cred != NOCRED)
2169 bp->b_rcred = crhold(cred);
2170 vfs_busy_pages(bp, 0);
2171 bp->b_iooffset = dbtob(bp->b_blkno);
2177 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2179 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2193 * Write, release buffer on completion. (Done by iodone
2194 * if async). Do not bother writing anything if the buffer
2197 * Note that we set B_CACHE here, indicating that buffer is
2198 * fully valid and thus cacheable. This is true even of NFS
2199 * now so we set it generally. This could be set either here
2200 * or in biodone() since the I/O is synchronous. We put it
2204 bufwrite(struct buf *bp)
2211 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2212 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2213 bp->b_flags |= B_INVAL | B_RELBUF;
2214 bp->b_flags &= ~B_CACHE;
2218 if (bp->b_flags & B_INVAL) {
2223 if (bp->b_flags & B_BARRIER)
2224 atomic_add_long(&barrierwrites, 1);
2226 oldflags = bp->b_flags;
2228 BUF_ASSERT_HELD(bp);
2230 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2231 ("FFS background buffer should not get here %p", bp));
2235 vp_md = vp->v_vflag & VV_MD;
2240 * Mark the buffer clean. Increment the bufobj write count
2241 * before bundirty() call, to prevent other thread from seeing
2242 * empty dirty list and zero counter for writes in progress,
2243 * falsely indicating that the bufobj is clean.
2245 bufobj_wref(bp->b_bufobj);
2248 bp->b_flags &= ~B_DONE;
2249 bp->b_ioflags &= ~BIO_ERROR;
2250 bp->b_flags |= B_CACHE;
2251 bp->b_iocmd = BIO_WRITE;
2253 vfs_busy_pages(bp, 1);
2256 * Normal bwrites pipeline writes
2258 bp->b_runningbufspace = bp->b_bufsize;
2259 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2264 racct_add_buf(curproc, bp, 1);
2265 PROC_UNLOCK(curproc);
2268 curthread->td_ru.ru_oublock++;
2269 if (oldflags & B_ASYNC)
2271 bp->b_iooffset = dbtob(bp->b_blkno);
2272 buf_track(bp, __func__);
2275 if ((oldflags & B_ASYNC) == 0) {
2276 int rtval = bufwait(bp);
2279 } else if (space > hirunningspace) {
2281 * don't allow the async write to saturate the I/O
2282 * system. We will not deadlock here because
2283 * we are blocking waiting for I/O that is already in-progress
2284 * to complete. We do not block here if it is the update
2285 * or syncer daemon trying to clean up as that can lead
2288 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2289 waitrunningbufspace();
2296 bufbdflush(struct bufobj *bo, struct buf *bp)
2300 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
2301 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2303 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
2306 * Try to find a buffer to flush.
2308 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2309 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2311 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2314 panic("bdwrite: found ourselves");
2316 /* Don't countdeps with the bo lock held. */
2317 if (buf_countdeps(nbp, 0)) {
2322 if (nbp->b_flags & B_CLUSTEROK) {
2323 vfs_bio_awrite(nbp);
2328 dirtybufferflushes++;
2337 * Delayed write. (Buffer is marked dirty). Do not bother writing
2338 * anything if the buffer is marked invalid.
2340 * Note that since the buffer must be completely valid, we can safely
2341 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2342 * biodone() in order to prevent getblk from writing the buffer
2343 * out synchronously.
2346 bdwrite(struct buf *bp)
2348 struct thread *td = curthread;
2352 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2353 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2354 KASSERT((bp->b_flags & B_BARRIER) == 0,
2355 ("Barrier request in delayed write %p", bp));
2356 BUF_ASSERT_HELD(bp);
2358 if (bp->b_flags & B_INVAL) {
2364 * If we have too many dirty buffers, don't create any more.
2365 * If we are wildly over our limit, then force a complete
2366 * cleanup. Otherwise, just keep the situation from getting
2367 * out of control. Note that we have to avoid a recursive
2368 * disaster and not try to clean up after our own cleanup!
2372 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2373 td->td_pflags |= TDP_INBDFLUSH;
2375 td->td_pflags &= ~TDP_INBDFLUSH;
2381 * Set B_CACHE, indicating that the buffer is fully valid. This is
2382 * true even of NFS now.
2384 bp->b_flags |= B_CACHE;
2387 * This bmap keeps the system from needing to do the bmap later,
2388 * perhaps when the system is attempting to do a sync. Since it
2389 * is likely that the indirect block -- or whatever other datastructure
2390 * that the filesystem needs is still in memory now, it is a good
2391 * thing to do this. Note also, that if the pageout daemon is
2392 * requesting a sync -- there might not be enough memory to do
2393 * the bmap then... So, this is important to do.
2395 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2396 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2399 buf_track(bp, __func__);
2402 * Set the *dirty* buffer range based upon the VM system dirty
2405 * Mark the buffer pages as clean. We need to do this here to
2406 * satisfy the vnode_pager and the pageout daemon, so that it
2407 * thinks that the pages have been "cleaned". Note that since
2408 * the pages are in a delayed write buffer -- the VFS layer
2409 * "will" see that the pages get written out on the next sync,
2410 * or perhaps the cluster will be completed.
2412 vfs_clean_pages_dirty_buf(bp);
2416 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2417 * due to the softdep code.
2424 * Turn buffer into delayed write request. We must clear BIO_READ and
2425 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2426 * itself to properly update it in the dirty/clean lists. We mark it
2427 * B_DONE to ensure that any asynchronization of the buffer properly
2428 * clears B_DONE ( else a panic will occur later ).
2430 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2431 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2432 * should only be called if the buffer is known-good.
2434 * Since the buffer is not on a queue, we do not update the numfreebuffers
2437 * The buffer must be on QUEUE_NONE.
2440 bdirty(struct buf *bp)
2443 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2444 bp, bp->b_vp, bp->b_flags);
2445 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2446 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2447 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2448 BUF_ASSERT_HELD(bp);
2449 bp->b_flags &= ~(B_RELBUF);
2450 bp->b_iocmd = BIO_WRITE;
2452 if ((bp->b_flags & B_DELWRI) == 0) {
2453 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2462 * Clear B_DELWRI for buffer.
2464 * Since the buffer is not on a queue, we do not update the numfreebuffers
2467 * The buffer must be on QUEUE_NONE.
2471 bundirty(struct buf *bp)
2474 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", 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 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2478 BUF_ASSERT_HELD(bp);
2480 if (bp->b_flags & B_DELWRI) {
2481 bp->b_flags &= ~B_DELWRI;
2486 * Since it is now being written, we can clear its deferred write flag.
2488 bp->b_flags &= ~B_DEFERRED;
2494 * Asynchronous write. Start output on a buffer, but do not wait for
2495 * it to complete. The buffer is released when the output completes.
2497 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2498 * B_INVAL buffers. Not us.
2501 bawrite(struct buf *bp)
2504 bp->b_flags |= B_ASYNC;
2511 * Asynchronous barrier write. Start output on a buffer, but do not
2512 * wait for it to complete. Place a write barrier after this write so
2513 * that this buffer and all buffers written before it are committed to
2514 * the disk before any buffers written after this write are committed
2515 * to the disk. The buffer is released when the output completes.
2518 babarrierwrite(struct buf *bp)
2521 bp->b_flags |= B_ASYNC | B_BARRIER;
2528 * Synchronous barrier write. Start output on a buffer and wait for
2529 * it to complete. Place a write barrier after this write so that
2530 * this buffer and all buffers written before it are committed to
2531 * the disk before any buffers written after this write are committed
2532 * to the disk. The buffer is released when the output completes.
2535 bbarrierwrite(struct buf *bp)
2538 bp->b_flags |= B_BARRIER;
2539 return (bwrite(bp));
2545 * Called prior to the locking of any vnodes when we are expecting to
2546 * write. We do not want to starve the buffer cache with too many
2547 * dirty buffers so we block here. By blocking prior to the locking
2548 * of any vnodes we attempt to avoid the situation where a locked vnode
2549 * prevents the various system daemons from flushing related buffers.
2555 if (buf_dirty_count_severe()) {
2556 mtx_lock(&bdirtylock);
2557 while (buf_dirty_count_severe()) {
2559 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2562 mtx_unlock(&bdirtylock);
2567 * Return true if we have too many dirty buffers.
2570 buf_dirty_count_severe(void)
2573 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2579 * Release a busy buffer and, if requested, free its resources. The
2580 * buffer will be stashed in the appropriate bufqueue[] allowing it
2581 * to be accessed later as a cache entity or reused for other purposes.
2584 brelse(struct buf *bp)
2586 struct mount *v_mnt;
2590 * Many functions erroneously call brelse with a NULL bp under rare
2591 * error conditions. Simply return when called with a NULL bp.
2595 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2596 bp, bp->b_vp, bp->b_flags);
2597 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2598 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2599 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2600 ("brelse: non-VMIO buffer marked NOREUSE"));
2602 if (BUF_LOCKRECURSED(bp)) {
2604 * Do not process, in particular, do not handle the
2605 * B_INVAL/B_RELBUF and do not release to free list.
2611 if (bp->b_flags & B_MANAGED) {
2616 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2617 BO_LOCK(bp->b_bufobj);
2618 bp->b_vflags &= ~BV_BKGRDERR;
2619 BO_UNLOCK(bp->b_bufobj);
2622 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2623 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2624 !(bp->b_flags & B_INVAL)) {
2626 * Failed write, redirty. All errors except ENXIO (which
2627 * means the device is gone) are treated as being
2630 * XXX Treating EIO as transient is not correct; the
2631 * contract with the local storage device drivers is that
2632 * they will only return EIO once the I/O is no longer
2633 * retriable. Network I/O also respects this through the
2634 * guarantees of TCP and/or the internal retries of NFS.
2635 * ENOMEM might be transient, but we also have no way of
2636 * knowing when its ok to retry/reschedule. In general,
2637 * this entire case should be made obsolete through better
2638 * error handling/recovery and resource scheduling.
2640 * Do this also for buffers that failed with ENXIO, but have
2641 * non-empty dependencies - the soft updates code might need
2642 * to access the buffer to untangle them.
2644 * Must clear BIO_ERROR to prevent pages from being scrapped.
2646 bp->b_ioflags &= ~BIO_ERROR;
2648 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2649 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2651 * Either a failed read I/O, or we were asked to free or not
2652 * cache the buffer, or we failed to write to a device that's
2653 * no longer present.
2655 bp->b_flags |= B_INVAL;
2656 if (!LIST_EMPTY(&bp->b_dep))
2658 if (bp->b_flags & B_DELWRI)
2660 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2661 if ((bp->b_flags & B_VMIO) == 0) {
2669 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2670 * is called with B_DELWRI set, the underlying pages may wind up
2671 * getting freed causing a previous write (bdwrite()) to get 'lost'
2672 * because pages associated with a B_DELWRI bp are marked clean.
2674 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2675 * if B_DELWRI is set.
2677 if (bp->b_flags & B_DELWRI)
2678 bp->b_flags &= ~B_RELBUF;
2681 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2682 * constituted, not even NFS buffers now. Two flags effect this. If
2683 * B_INVAL, the struct buf is invalidated but the VM object is kept
2684 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2686 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2687 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2688 * buffer is also B_INVAL because it hits the re-dirtying code above.
2690 * Normally we can do this whether a buffer is B_DELWRI or not. If
2691 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2692 * the commit state and we cannot afford to lose the buffer. If the
2693 * buffer has a background write in progress, we need to keep it
2694 * around to prevent it from being reconstituted and starting a second
2698 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2700 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2701 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2702 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2703 vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) {
2704 vfs_vmio_invalidate(bp);
2708 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2709 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2711 bp->b_flags &= ~B_NOREUSE;
2712 if (bp->b_vp != NULL)
2717 * If the buffer has junk contents signal it and eventually
2718 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2721 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2722 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2723 bp->b_flags |= B_INVAL;
2724 if (bp->b_flags & B_INVAL) {
2725 if (bp->b_flags & B_DELWRI)
2731 buf_track(bp, __func__);
2733 /* buffers with no memory */
2734 if (bp->b_bufsize == 0) {
2738 /* buffers with junk contents */
2739 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2740 (bp->b_ioflags & BIO_ERROR)) {
2741 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2742 if (bp->b_vflags & BV_BKGRDINPROG)
2743 panic("losing buffer 2");
2744 qindex = QUEUE_CLEAN;
2745 bp->b_flags |= B_AGE;
2746 /* remaining buffers */
2747 } else if (bp->b_flags & B_DELWRI)
2748 qindex = QUEUE_DIRTY;
2750 qindex = QUEUE_CLEAN;
2752 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2753 panic("brelse: not dirty");
2755 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2756 /* binsfree unlocks bp. */
2757 binsfree(bp, qindex);
2761 * Release a buffer back to the appropriate queue but do not try to free
2762 * it. The buffer is expected to be used again soon.
2764 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2765 * biodone() to requeue an async I/O on completion. It is also used when
2766 * known good buffers need to be requeued but we think we may need the data
2769 * XXX we should be able to leave the B_RELBUF hint set on completion.
2772 bqrelse(struct buf *bp)
2776 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2777 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2778 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2780 qindex = QUEUE_NONE;
2781 if (BUF_LOCKRECURSED(bp)) {
2782 /* do not release to free list */
2786 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2788 if (bp->b_flags & B_MANAGED) {
2789 if (bp->b_flags & B_REMFREE)
2794 /* buffers with stale but valid contents */
2795 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2796 BV_BKGRDERR)) == BV_BKGRDERR) {
2797 BO_LOCK(bp->b_bufobj);
2798 bp->b_vflags &= ~BV_BKGRDERR;
2799 BO_UNLOCK(bp->b_bufobj);
2800 qindex = QUEUE_DIRTY;
2802 if ((bp->b_flags & B_DELWRI) == 0 &&
2803 (bp->b_xflags & BX_VNDIRTY))
2804 panic("bqrelse: not dirty");
2805 if ((bp->b_flags & B_NOREUSE) != 0) {
2809 qindex = QUEUE_CLEAN;
2811 buf_track(bp, __func__);
2812 /* binsfree unlocks bp. */
2813 binsfree(bp, qindex);
2817 buf_track(bp, __func__);
2823 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2824 * restore bogus pages.
2827 vfs_vmio_iodone(struct buf *bp)
2832 struct vnode *vp __unused;
2833 int i, iosize, resid;
2836 obj = bp->b_bufobj->bo_object;
2837 KASSERT(obj->paging_in_progress >= bp->b_npages,
2838 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2839 obj->paging_in_progress, bp->b_npages));
2842 KASSERT(vp->v_holdcnt > 0,
2843 ("vfs_vmio_iodone: vnode %p has zero hold count", vp));
2844 KASSERT(vp->v_object != NULL,
2845 ("vfs_vmio_iodone: vnode %p has no vm_object", vp));
2847 foff = bp->b_offset;
2848 KASSERT(bp->b_offset != NOOFFSET,
2849 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2852 iosize = bp->b_bcount - bp->b_resid;
2853 VM_OBJECT_WLOCK(obj);
2854 for (i = 0; i < bp->b_npages; i++) {
2855 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2860 * cleanup bogus pages, restoring the originals
2863 if (m == bogus_page) {
2865 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2867 panic("biodone: page disappeared!");
2869 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2871 * In the write case, the valid and clean bits are
2872 * already changed correctly ( see bdwrite() ), so we
2873 * only need to do this here in the read case.
2875 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2876 resid)) == 0, ("vfs_vmio_iodone: page %p "
2877 "has unexpected dirty bits", m));
2878 vfs_page_set_valid(bp, foff, m);
2880 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2881 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2882 (intmax_t)foff, (uintmax_t)m->pindex));
2885 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2888 vm_object_pip_wakeupn(obj, bp->b_npages);
2889 VM_OBJECT_WUNLOCK(obj);
2890 if (bogus && buf_mapped(bp)) {
2891 BUF_CHECK_MAPPED(bp);
2892 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2893 bp->b_pages, bp->b_npages);
2898 * Perform page invalidation when a buffer is released. The fully invalid
2899 * pages will be reclaimed later in vfs_vmio_truncate().
2902 vfs_vmio_invalidate(struct buf *bp)
2906 int flags, i, resid, poffset, presid;
2908 if (buf_mapped(bp)) {
2909 BUF_CHECK_MAPPED(bp);
2910 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2912 BUF_CHECK_UNMAPPED(bp);
2914 * Get the base offset and length of the buffer. Note that
2915 * in the VMIO case if the buffer block size is not
2916 * page-aligned then b_data pointer may not be page-aligned.
2917 * But our b_pages[] array *IS* page aligned.
2919 * block sizes less then DEV_BSIZE (usually 512) are not
2920 * supported due to the page granularity bits (m->valid,
2921 * m->dirty, etc...).
2923 * See man buf(9) for more information
2925 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2926 obj = bp->b_bufobj->bo_object;
2927 resid = bp->b_bufsize;
2928 poffset = bp->b_offset & PAGE_MASK;
2929 VM_OBJECT_WLOCK(obj);
2930 for (i = 0; i < bp->b_npages; i++) {
2932 if (m == bogus_page)
2933 panic("vfs_vmio_invalidate: Unexpected bogus page.");
2934 bp->b_pages[i] = NULL;
2936 presid = resid > (PAGE_SIZE - poffset) ?
2937 (PAGE_SIZE - poffset) : resid;
2938 KASSERT(presid >= 0, ("brelse: extra page"));
2939 while (vm_page_xbusied(m)) {
2941 VM_OBJECT_WUNLOCK(obj);
2942 vm_page_busy_sleep(m, "mbncsh", true);
2943 VM_OBJECT_WLOCK(obj);
2945 if (pmap_page_wired_mappings(m) == 0)
2946 vm_page_set_invalid(m, poffset, presid);
2947 vm_page_release_locked(m, flags);
2951 VM_OBJECT_WUNLOCK(obj);
2956 * Page-granular truncation of an existing VMIO buffer.
2959 vfs_vmio_truncate(struct buf *bp, int desiredpages)
2965 if (bp->b_npages == desiredpages)
2968 if (buf_mapped(bp)) {
2969 BUF_CHECK_MAPPED(bp);
2970 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
2971 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
2973 BUF_CHECK_UNMAPPED(bp);
2976 * The object lock is needed only if we will attempt to free pages.
2978 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2979 if ((bp->b_flags & B_DIRECT) != 0) {
2980 flags |= VPR_TRYFREE;
2981 obj = bp->b_bufobj->bo_object;
2982 VM_OBJECT_WLOCK(obj);
2986 for (i = desiredpages; i < bp->b_npages; i++) {
2988 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
2989 bp->b_pages[i] = NULL;
2991 vm_page_release_locked(m, flags);
2993 vm_page_release(m, flags);
2996 VM_OBJECT_WUNLOCK(obj);
2997 bp->b_npages = desiredpages;
3001 * Byte granular extension of VMIO buffers.
3004 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3007 * We are growing the buffer, possibly in a
3008 * byte-granular fashion.
3016 * Step 1, bring in the VM pages from the object, allocating
3017 * them if necessary. We must clear B_CACHE if these pages
3018 * are not valid for the range covered by the buffer.
3020 obj = bp->b_bufobj->bo_object;
3021 VM_OBJECT_WLOCK(obj);
3022 if (bp->b_npages < desiredpages) {
3024 * We must allocate system pages since blocking
3025 * here could interfere with paging I/O, no
3026 * matter which process we are.
3028 * Only exclusive busy can be tested here.
3029 * Blocking on shared busy might lead to
3030 * deadlocks once allocbuf() is called after
3031 * pages are vfs_busy_pages().
3033 (void)vm_page_grab_pages(obj,
3034 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3035 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3036 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3037 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3038 bp->b_npages = desiredpages;
3042 * Step 2. We've loaded the pages into the buffer,
3043 * we have to figure out if we can still have B_CACHE
3044 * set. Note that B_CACHE is set according to the
3045 * byte-granular range ( bcount and size ), not the
3046 * aligned range ( newbsize ).
3048 * The VM test is against m->valid, which is DEV_BSIZE
3049 * aligned. Needless to say, the validity of the data
3050 * needs to also be DEV_BSIZE aligned. Note that this
3051 * fails with NFS if the server or some other client
3052 * extends the file's EOF. If our buffer is resized,
3053 * B_CACHE may remain set! XXX
3055 toff = bp->b_bcount;
3056 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3057 while ((bp->b_flags & B_CACHE) && toff < size) {
3060 if (tinc > (size - toff))
3062 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3063 m = bp->b_pages[pi];
3064 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3068 VM_OBJECT_WUNLOCK(obj);
3071 * Step 3, fixup the KVA pmap.
3076 BUF_CHECK_UNMAPPED(bp);
3080 * Check to see if a block at a particular lbn is available for a clustered
3084 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3091 /* If the buf isn't in core skip it */
3092 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3095 /* If the buf is busy we don't want to wait for it */
3096 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3099 /* Only cluster with valid clusterable delayed write buffers */
3100 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3101 (B_DELWRI | B_CLUSTEROK))
3104 if (bpa->b_bufsize != size)
3108 * Check to see if it is in the expected place on disk and that the
3109 * block has been mapped.
3111 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3121 * Implement clustered async writes for clearing out B_DELWRI buffers.
3122 * This is much better then the old way of writing only one buffer at
3123 * a time. Note that we may not be presented with the buffers in the
3124 * correct order, so we search for the cluster in both directions.
3127 vfs_bio_awrite(struct buf *bp)
3132 daddr_t lblkno = bp->b_lblkno;
3133 struct vnode *vp = bp->b_vp;
3141 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3143 * right now we support clustered writing only to regular files. If
3144 * we find a clusterable block we could be in the middle of a cluster
3145 * rather then at the beginning.
3147 if ((vp->v_type == VREG) &&
3148 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3149 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3151 size = vp->v_mount->mnt_stat.f_iosize;
3152 maxcl = MAXPHYS / size;
3155 for (i = 1; i < maxcl; i++)
3156 if (vfs_bio_clcheck(vp, size, lblkno + i,
3157 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3160 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3161 if (vfs_bio_clcheck(vp, size, lblkno - j,
3162 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3168 * this is a possible cluster write
3172 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3178 bp->b_flags |= B_ASYNC;
3180 * default (old) behavior, writing out only one block
3182 * XXX returns b_bufsize instead of b_bcount for nwritten?
3184 nwritten = bp->b_bufsize;
3193 * Allocate KVA for an empty buf header according to gbflags.
3196 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3199 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3201 * In order to keep fragmentation sane we only allocate kva
3202 * in BKVASIZE chunks. XXX with vmem we can do page size.
3204 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3206 if (maxsize != bp->b_kvasize &&
3207 bufkva_alloc(bp, maxsize, gbflags))
3216 * Find and initialize a new buffer header, freeing up existing buffers
3217 * in the bufqueues as necessary. The new buffer is returned locked.
3220 * We have insufficient buffer headers
3221 * We have insufficient buffer space
3222 * buffer_arena is too fragmented ( space reservation fails )
3223 * If we have to flush dirty buffers ( but we try to avoid this )
3225 * The caller is responsible for releasing the reserved bufspace after
3226 * allocbuf() is called.
3229 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3231 struct bufdomain *bd;
3233 bool metadata, reserved;
3236 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3237 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3238 if (!unmapped_buf_allowed)
3239 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3241 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3249 bd = &bdomain[vp->v_bufobj.bo_domain];
3251 counter_u64_add(getnewbufcalls, 1);
3254 if (reserved == false &&
3255 bufspace_reserve(bd, maxsize, metadata) != 0) {
3256 counter_u64_add(getnewbufrestarts, 1);
3260 if ((bp = buf_alloc(bd)) == NULL) {
3261 counter_u64_add(getnewbufrestarts, 1);
3264 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3267 } while (buf_recycle(bd, false) == 0);
3270 bufspace_release(bd, maxsize);
3272 bp->b_flags |= B_INVAL;
3275 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3283 * buffer flushing daemon. Buffers are normally flushed by the
3284 * update daemon but if it cannot keep up this process starts to
3285 * take the load in an attempt to prevent getnewbuf() from blocking.
3287 static struct kproc_desc buf_kp = {
3292 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3295 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3299 flushed = flushbufqueues(vp, bd, target, 0);
3302 * Could not find any buffers without rollback
3303 * dependencies, so just write the first one
3304 * in the hopes of eventually making progress.
3306 if (vp != NULL && target > 2)
3308 flushbufqueues(vp, bd, target, 1);
3316 struct bufdomain *bd;
3322 * This process needs to be suspended prior to shutdown sync.
3324 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3325 SHUTDOWN_PRI_LAST + 100);
3328 * Start the buf clean daemons as children threads.
3330 for (i = 0 ; i < buf_domains; i++) {
3333 error = kthread_add((void (*)(void *))bufspace_daemon,
3334 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3336 panic("error %d spawning bufspace daemon", error);
3340 * This process is allowed to take the buffer cache to the limit
3342 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3346 mtx_unlock(&bdlock);
3348 kthread_suspend_check();
3351 * Save speedupreq for this pass and reset to capture new
3354 speedupreq = bd_speedupreq;
3358 * Flush each domain sequentially according to its level and
3359 * the speedup request.
3361 for (i = 0; i < buf_domains; i++) {
3364 lodirty = bd->bd_numdirtybuffers / 2;
3366 lodirty = bd->bd_lodirtybuffers;
3367 while (bd->bd_numdirtybuffers > lodirty) {
3368 if (buf_flush(NULL, bd,
3369 bd->bd_numdirtybuffers - lodirty) == 0)
3371 kern_yield(PRI_USER);
3376 * Only clear bd_request if we have reached our low water
3377 * mark. The buf_daemon normally waits 1 second and
3378 * then incrementally flushes any dirty buffers that have
3379 * built up, within reason.
3381 * If we were unable to hit our low water mark and couldn't
3382 * find any flushable buffers, we sleep for a short period
3383 * to avoid endless loops on unlockable buffers.
3386 if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3388 * We reached our low water mark, reset the
3389 * request and sleep until we are needed again.
3390 * The sleep is just so the suspend code works.
3394 * Do an extra wakeup in case dirty threshold
3395 * changed via sysctl and the explicit transition
3396 * out of shortfall was missed.
3399 if (runningbufspace <= lorunningspace)
3401 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3404 * We couldn't find any flushable dirty buffers but
3405 * still have too many dirty buffers, we
3406 * have to sleep and try again. (rare)
3408 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3416 * Try to flush a buffer in the dirty queue. We must be careful to
3417 * free up B_INVAL buffers instead of write them, which NFS is
3418 * particularly sensitive to.
3420 static int flushwithdeps = 0;
3421 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
3422 0, "Number of buffers flushed with dependecies that require rollbacks");
3425 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3428 struct bufqueue *bq;
3429 struct buf *sentinel;
3439 bq = &bd->bd_dirtyq;
3441 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3442 sentinel->b_qindex = QUEUE_SENTINEL;
3444 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3446 while (flushed != target) {
3449 bp = TAILQ_NEXT(sentinel, b_freelist);
3451 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3452 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3459 * Skip sentinels inserted by other invocations of the
3460 * flushbufqueues(), taking care to not reorder them.
3462 * Only flush the buffers that belong to the
3463 * vnode locked by the curthread.
3465 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3470 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3476 * BKGRDINPROG can only be set with the buf and bufobj
3477 * locks both held. We tolerate a race to clear it here.
3479 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3480 (bp->b_flags & B_DELWRI) == 0) {
3484 if (bp->b_flags & B_INVAL) {
3491 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3492 if (flushdeps == 0) {
3500 * We must hold the lock on a vnode before writing
3501 * one of its buffers. Otherwise we may confuse, or
3502 * in the case of a snapshot vnode, deadlock the
3505 * The lock order here is the reverse of the normal
3506 * of vnode followed by buf lock. This is ok because
3507 * the NOWAIT will prevent deadlock.
3510 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3516 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3518 ASSERT_VOP_LOCKED(vp, "getbuf");
3520 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3521 vn_lock(vp, LK_TRYUPGRADE);
3524 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3525 bp, bp->b_vp, bp->b_flags);
3526 if (curproc == bufdaemonproc) {
3531 counter_u64_add(notbufdflushes, 1);
3533 vn_finished_write(mp);
3536 flushwithdeps += hasdeps;
3540 * Sleeping on runningbufspace while holding
3541 * vnode lock leads to deadlock.
3543 if (curproc == bufdaemonproc &&
3544 runningbufspace > hirunningspace)
3545 waitrunningbufspace();
3548 vn_finished_write(mp);
3552 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3554 free(sentinel, M_TEMP);
3559 * Check to see if a block is currently memory resident.
3562 incore(struct bufobj *bo, daddr_t blkno)
3567 bp = gbincore(bo, blkno);
3573 * Returns true if no I/O is needed to access the
3574 * associated VM object. This is like incore except
3575 * it also hunts around in the VM system for the data.
3579 inmem(struct vnode * vp, daddr_t blkno)
3582 vm_offset_t toff, tinc, size;
3586 ASSERT_VOP_LOCKED(vp, "inmem");
3588 if (incore(&vp->v_bufobj, blkno))
3590 if (vp->v_mount == NULL)
3597 if (size > vp->v_mount->mnt_stat.f_iosize)
3598 size = vp->v_mount->mnt_stat.f_iosize;
3599 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3601 VM_OBJECT_RLOCK(obj);
3602 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3603 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
3607 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3608 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3609 if (vm_page_is_valid(m,
3610 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
3613 VM_OBJECT_RUNLOCK(obj);
3617 VM_OBJECT_RUNLOCK(obj);
3622 * Set the dirty range for a buffer based on the status of the dirty
3623 * bits in the pages comprising the buffer. The range is limited
3624 * to the size of the buffer.
3626 * Tell the VM system that the pages associated with this buffer
3627 * are clean. This is used for delayed writes where the data is
3628 * going to go to disk eventually without additional VM intevention.
3630 * Note that while we only really need to clean through to b_bcount, we
3631 * just go ahead and clean through to b_bufsize.
3634 vfs_clean_pages_dirty_buf(struct buf *bp)
3636 vm_ooffset_t foff, noff, eoff;
3640 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3643 foff = bp->b_offset;
3644 KASSERT(bp->b_offset != NOOFFSET,
3645 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3647 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3648 vfs_drain_busy_pages(bp);
3649 vfs_setdirty_locked_object(bp);
3650 for (i = 0; i < bp->b_npages; i++) {
3651 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3653 if (eoff > bp->b_offset + bp->b_bufsize)
3654 eoff = bp->b_offset + bp->b_bufsize;
3656 vfs_page_set_validclean(bp, foff, m);
3657 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3660 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3664 vfs_setdirty_locked_object(struct buf *bp)
3669 object = bp->b_bufobj->bo_object;
3670 VM_OBJECT_ASSERT_WLOCKED(object);
3673 * We qualify the scan for modified pages on whether the
3674 * object has been flushed yet.
3676 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
3677 vm_offset_t boffset;
3678 vm_offset_t eoffset;
3681 * test the pages to see if they have been modified directly
3682 * by users through the VM system.
3684 for (i = 0; i < bp->b_npages; i++)
3685 vm_page_test_dirty(bp->b_pages[i]);
3688 * Calculate the encompassing dirty range, boffset and eoffset,
3689 * (eoffset - boffset) bytes.
3692 for (i = 0; i < bp->b_npages; i++) {
3693 if (bp->b_pages[i]->dirty)
3696 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3698 for (i = bp->b_npages - 1; i >= 0; --i) {
3699 if (bp->b_pages[i]->dirty) {
3703 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3706 * Fit it to the buffer.
3709 if (eoffset > bp->b_bcount)
3710 eoffset = bp->b_bcount;
3713 * If we have a good dirty range, merge with the existing
3717 if (boffset < eoffset) {
3718 if (bp->b_dirtyoff > boffset)
3719 bp->b_dirtyoff = boffset;
3720 if (bp->b_dirtyend < eoffset)
3721 bp->b_dirtyend = eoffset;
3727 * Allocate the KVA mapping for an existing buffer.
3728 * If an unmapped buffer is provided but a mapped buffer is requested, take
3729 * also care to properly setup mappings between pages and KVA.
3732 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3734 int bsize, maxsize, need_mapping, need_kva;
3737 need_mapping = bp->b_data == unmapped_buf &&
3738 (gbflags & GB_UNMAPPED) == 0;
3739 need_kva = bp->b_kvabase == unmapped_buf &&
3740 bp->b_data == unmapped_buf &&
3741 (gbflags & GB_KVAALLOC) != 0;
3742 if (!need_mapping && !need_kva)
3745 BUF_CHECK_UNMAPPED(bp);
3747 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3749 * Buffer is not mapped, but the KVA was already
3750 * reserved at the time of the instantiation. Use the
3757 * Calculate the amount of the address space we would reserve
3758 * if the buffer was mapped.
3760 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3761 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3762 offset = blkno * bsize;
3763 maxsize = size + (offset & PAGE_MASK);
3764 maxsize = imax(maxsize, bsize);
3766 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3767 if ((gbflags & GB_NOWAIT_BD) != 0) {
3769 * XXXKIB: defragmentation cannot
3770 * succeed, not sure what else to do.
3772 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3774 counter_u64_add(mappingrestarts, 1);
3775 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3779 /* b_offset is handled by bpmap_qenter. */
3780 bp->b_data = bp->b_kvabase;
3781 BUF_CHECK_MAPPED(bp);
3787 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3793 error = getblkx(vp, blkno, size, slpflag, slptimeo, flags, &bp);
3802 * Get a block given a specified block and offset into a file/device.
3803 * The buffers B_DONE bit will be cleared on return, making it almost
3804 * ready for an I/O initiation. B_INVAL may or may not be set on
3805 * return. The caller should clear B_INVAL prior to initiating a
3808 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3809 * an existing buffer.
3811 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3812 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3813 * and then cleared based on the backing VM. If the previous buffer is
3814 * non-0-sized but invalid, B_CACHE will be cleared.
3816 * If getblk() must create a new buffer, the new buffer is returned with
3817 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3818 * case it is returned with B_INVAL clear and B_CACHE set based on the
3821 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
3822 * B_CACHE bit is clear.
3824 * What this means, basically, is that the caller should use B_CACHE to
3825 * determine whether the buffer is fully valid or not and should clear
3826 * B_INVAL prior to issuing a read. If the caller intends to validate
3827 * the buffer by loading its data area with something, the caller needs
3828 * to clear B_INVAL. If the caller does this without issuing an I/O,
3829 * the caller should set B_CACHE ( as an optimization ), else the caller
3830 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3831 * a write attempt or if it was a successful read. If the caller
3832 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3833 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3836 getblkx(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3837 int flags, struct buf **bpp)
3842 int bsize, error, maxsize, vmio;
3845 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3846 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3847 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3848 ASSERT_VOP_LOCKED(vp, "getblk");
3849 if (size > maxbcachebuf)
3850 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3852 if (!unmapped_buf_allowed)
3853 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3859 bp = gbincore(bo, blkno);
3863 * Buffer is in-core. If the buffer is not busy nor managed,
3864 * it must be on a queue.
3866 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3868 if ((flags & GB_LOCK_NOWAIT) != 0)
3869 lockflags |= LK_NOWAIT;
3871 error = BUF_TIMELOCK(bp, lockflags,
3872 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3875 * If we slept and got the lock we have to restart in case
3876 * the buffer changed identities.
3878 if (error == ENOLCK)
3880 /* We timed out or were interrupted. */
3881 else if (error != 0)
3883 /* If recursed, assume caller knows the rules. */
3884 else if (BUF_LOCKRECURSED(bp))
3888 * The buffer is locked. B_CACHE is cleared if the buffer is
3889 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3890 * and for a VMIO buffer B_CACHE is adjusted according to the
3893 if (bp->b_flags & B_INVAL)
3894 bp->b_flags &= ~B_CACHE;
3895 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3896 bp->b_flags |= B_CACHE;
3897 if (bp->b_flags & B_MANAGED)
3898 MPASS(bp->b_qindex == QUEUE_NONE);
3903 * check for size inconsistencies for non-VMIO case.
3905 if (bp->b_bcount != size) {
3906 if ((bp->b_flags & B_VMIO) == 0 ||
3907 (size > bp->b_kvasize)) {
3908 if (bp->b_flags & B_DELWRI) {
3909 bp->b_flags |= B_NOCACHE;
3912 if (LIST_EMPTY(&bp->b_dep)) {
3913 bp->b_flags |= B_RELBUF;
3916 bp->b_flags |= B_NOCACHE;
3925 * Handle the case of unmapped buffer which should
3926 * become mapped, or the buffer for which KVA
3927 * reservation is requested.
3929 bp_unmapped_get_kva(bp, blkno, size, flags);
3932 * If the size is inconsistent in the VMIO case, we can resize
3933 * the buffer. This might lead to B_CACHE getting set or
3934 * cleared. If the size has not changed, B_CACHE remains
3935 * unchanged from its previous state.
3939 KASSERT(bp->b_offset != NOOFFSET,
3940 ("getblk: no buffer offset"));
3943 * A buffer with B_DELWRI set and B_CACHE clear must
3944 * be committed before we can return the buffer in
3945 * order to prevent the caller from issuing a read
3946 * ( due to B_CACHE not being set ) and overwriting
3949 * Most callers, including NFS and FFS, need this to
3950 * operate properly either because they assume they
3951 * can issue a read if B_CACHE is not set, or because
3952 * ( for example ) an uncached B_DELWRI might loop due
3953 * to softupdates re-dirtying the buffer. In the latter
3954 * case, B_CACHE is set after the first write completes,
3955 * preventing further loops.
3956 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3957 * above while extending the buffer, we cannot allow the
3958 * buffer to remain with B_CACHE set after the write
3959 * completes or it will represent a corrupt state. To
3960 * deal with this we set B_NOCACHE to scrap the buffer
3963 * We might be able to do something fancy, like setting
3964 * B_CACHE in bwrite() except if B_DELWRI is already set,
3965 * so the below call doesn't set B_CACHE, but that gets real
3966 * confusing. This is much easier.
3969 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3970 bp->b_flags |= B_NOCACHE;
3974 bp->b_flags &= ~B_DONE;
3977 * Buffer is not in-core, create new buffer. The buffer
3978 * returned by getnewbuf() is locked. Note that the returned
3979 * buffer is also considered valid (not marked B_INVAL).
3983 * If the user does not want us to create the buffer, bail out
3986 if (flags & GB_NOCREAT)
3989 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3990 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3991 offset = blkno * bsize;
3992 vmio = vp->v_object != NULL;
3994 maxsize = size + (offset & PAGE_MASK);
3997 /* Do not allow non-VMIO notmapped buffers. */
3998 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4000 maxsize = imax(maxsize, bsize);
4001 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4002 !vn_isdisk(vp, NULL)) {
4003 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4004 KASSERT(error != EOPNOTSUPP,
4005 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4010 return (EJUSTRETURN);
4013 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4015 if (slpflag || slptimeo)
4018 * XXX This is here until the sleep path is diagnosed
4019 * enough to work under very low memory conditions.
4021 * There's an issue on low memory, 4BSD+non-preempt
4022 * systems (eg MIPS routers with 32MB RAM) where buffer
4023 * exhaustion occurs without sleeping for buffer
4024 * reclaimation. This just sticks in a loop and
4025 * constantly attempts to allocate a buffer, which
4026 * hits exhaustion and tries to wakeup bufdaemon.
4027 * This never happens because we never yield.
4029 * The real solution is to identify and fix these cases
4030 * so we aren't effectively busy-waiting in a loop
4031 * until the reclaimation path has cycles to run.
4033 kern_yield(PRI_USER);
4038 * This code is used to make sure that a buffer is not
4039 * created while the getnewbuf routine is blocked.
4040 * This can be a problem whether the vnode is locked or not.
4041 * If the buffer is created out from under us, we have to
4042 * throw away the one we just created.
4044 * Note: this must occur before we associate the buffer
4045 * with the vp especially considering limitations in
4046 * the splay tree implementation when dealing with duplicate
4050 if (gbincore(bo, blkno)) {
4052 bp->b_flags |= B_INVAL;
4053 bufspace_release(bufdomain(bp), maxsize);
4059 * Insert the buffer into the hash, so that it can
4060 * be found by incore.
4062 bp->b_lblkno = blkno;
4063 bp->b_blkno = d_blkno;
4064 bp->b_offset = offset;
4069 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4070 * buffer size starts out as 0, B_CACHE will be set by
4071 * allocbuf() for the VMIO case prior to it testing the
4072 * backing store for validity.
4076 bp->b_flags |= B_VMIO;
4077 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4078 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4079 bp, vp->v_object, bp->b_bufobj->bo_object));
4081 bp->b_flags &= ~B_VMIO;
4082 KASSERT(bp->b_bufobj->bo_object == NULL,
4083 ("ARGH! has b_bufobj->bo_object %p %p\n",
4084 bp, bp->b_bufobj->bo_object));
4085 BUF_CHECK_MAPPED(bp);
4089 bufspace_release(bufdomain(bp), maxsize);
4090 bp->b_flags &= ~B_DONE;
4092 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4093 BUF_ASSERT_HELD(bp);
4095 buf_track(bp, __func__);
4096 KASSERT(bp->b_bufobj == bo,
4097 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4103 * Get an empty, disassociated buffer of given size. The buffer is initially
4107 geteblk(int size, int flags)
4112 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4113 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4114 if ((flags & GB_NOWAIT_BD) &&
4115 (curthread->td_pflags & TDP_BUFNEED) != 0)
4119 bufspace_release(bufdomain(bp), maxsize);
4120 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4121 BUF_ASSERT_HELD(bp);
4126 * Truncate the backing store for a non-vmio buffer.
4129 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4132 if (bp->b_flags & B_MALLOC) {
4134 * malloced buffers are not shrunk
4136 if (newbsize == 0) {
4137 bufmallocadjust(bp, 0);
4138 free(bp->b_data, M_BIOBUF);
4139 bp->b_data = bp->b_kvabase;
4140 bp->b_flags &= ~B_MALLOC;
4144 vm_hold_free_pages(bp, newbsize);
4145 bufspace_adjust(bp, newbsize);
4149 * Extend the backing for a non-VMIO buffer.
4152 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4158 * We only use malloced memory on the first allocation.
4159 * and revert to page-allocated memory when the buffer
4162 * There is a potential smp race here that could lead
4163 * to bufmallocspace slightly passing the max. It
4164 * is probably extremely rare and not worth worrying
4167 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4168 bufmallocspace < maxbufmallocspace) {
4169 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4170 bp->b_flags |= B_MALLOC;
4171 bufmallocadjust(bp, newbsize);
4176 * If the buffer is growing on its other-than-first
4177 * allocation then we revert to the page-allocation
4182 if (bp->b_flags & B_MALLOC) {
4183 origbuf = bp->b_data;
4184 origbufsize = bp->b_bufsize;
4185 bp->b_data = bp->b_kvabase;
4186 bufmallocadjust(bp, 0);
4187 bp->b_flags &= ~B_MALLOC;
4188 newbsize = round_page(newbsize);
4190 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4191 (vm_offset_t) bp->b_data + newbsize);
4192 if (origbuf != NULL) {
4193 bcopy(origbuf, bp->b_data, origbufsize);
4194 free(origbuf, M_BIOBUF);
4196 bufspace_adjust(bp, newbsize);
4200 * This code constitutes the buffer memory from either anonymous system
4201 * memory (in the case of non-VMIO operations) or from an associated
4202 * VM object (in the case of VMIO operations). This code is able to
4203 * resize a buffer up or down.
4205 * Note that this code is tricky, and has many complications to resolve
4206 * deadlock or inconsistent data situations. Tread lightly!!!
4207 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4208 * the caller. Calling this code willy nilly can result in the loss of data.
4210 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4211 * B_CACHE for the non-VMIO case.
4214 allocbuf(struct buf *bp, int size)
4218 BUF_ASSERT_HELD(bp);
4220 if (bp->b_bcount == size)
4223 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4224 panic("allocbuf: buffer too small");
4226 newbsize = roundup2(size, DEV_BSIZE);
4227 if ((bp->b_flags & B_VMIO) == 0) {
4228 if ((bp->b_flags & B_MALLOC) == 0)
4229 newbsize = round_page(newbsize);
4231 * Just get anonymous memory from the kernel. Don't
4232 * mess with B_CACHE.
4234 if (newbsize < bp->b_bufsize)
4235 vfs_nonvmio_truncate(bp, newbsize);
4236 else if (newbsize > bp->b_bufsize)
4237 vfs_nonvmio_extend(bp, newbsize);
4241 desiredpages = (size == 0) ? 0 :
4242 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4244 if (bp->b_flags & B_MALLOC)
4245 panic("allocbuf: VMIO buffer can't be malloced");
4247 * Set B_CACHE initially if buffer is 0 length or will become
4250 if (size == 0 || bp->b_bufsize == 0)
4251 bp->b_flags |= B_CACHE;
4253 if (newbsize < bp->b_bufsize)
4254 vfs_vmio_truncate(bp, desiredpages);
4255 /* XXX This looks as if it should be newbsize > b_bufsize */
4256 else if (size > bp->b_bcount)
4257 vfs_vmio_extend(bp, desiredpages, size);
4258 bufspace_adjust(bp, newbsize);
4260 bp->b_bcount = size; /* requested buffer size. */
4264 extern int inflight_transient_maps;
4266 static struct bio_queue nondump_bios;
4269 biodone(struct bio *bp)
4272 void (*done)(struct bio *);
4273 vm_offset_t start, end;
4275 biotrack(bp, __func__);
4278 * Avoid completing I/O when dumping after a panic since that may
4279 * result in a deadlock in the filesystem or pager code. Note that
4280 * this doesn't affect dumps that were started manually since we aim
4281 * to keep the system usable after it has been resumed.
4283 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4284 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4287 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4288 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4289 bp->bio_flags |= BIO_UNMAPPED;
4290 start = trunc_page((vm_offset_t)bp->bio_data);
4291 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4292 bp->bio_data = unmapped_buf;
4293 pmap_qremove(start, atop(end - start));
4294 vmem_free(transient_arena, start, end - start);
4295 atomic_add_int(&inflight_transient_maps, -1);
4297 done = bp->bio_done;
4299 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4301 bp->bio_flags |= BIO_DONE;
4309 * Wait for a BIO to finish.
4312 biowait(struct bio *bp, const char *wchan)
4316 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4318 while ((bp->bio_flags & BIO_DONE) == 0)
4319 msleep(bp, mtxp, PRIBIO, wchan, 0);
4321 if (bp->bio_error != 0)
4322 return (bp->bio_error);
4323 if (!(bp->bio_flags & BIO_ERROR))
4329 biofinish(struct bio *bp, struct devstat *stat, int error)
4333 bp->bio_error = error;
4334 bp->bio_flags |= BIO_ERROR;
4337 devstat_end_transaction_bio(stat, bp);
4341 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4343 biotrack_buf(struct bio *bp, const char *location)
4346 buf_track(bp->bio_track_bp, location);
4353 * Wait for buffer I/O completion, returning error status. The buffer
4354 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4355 * error and cleared.
4358 bufwait(struct buf *bp)
4360 if (bp->b_iocmd == BIO_READ)
4361 bwait(bp, PRIBIO, "biord");
4363 bwait(bp, PRIBIO, "biowr");
4364 if (bp->b_flags & B_EINTR) {
4365 bp->b_flags &= ~B_EINTR;
4368 if (bp->b_ioflags & BIO_ERROR) {
4369 return (bp->b_error ? bp->b_error : EIO);
4378 * Finish I/O on a buffer, optionally calling a completion function.
4379 * This is usually called from an interrupt so process blocking is
4382 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4383 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4384 * assuming B_INVAL is clear.
4386 * For the VMIO case, we set B_CACHE if the op was a read and no
4387 * read error occurred, or if the op was a write. B_CACHE is never
4388 * set if the buffer is invalid or otherwise uncacheable.
4390 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4391 * initiator to leave B_INVAL set to brelse the buffer out of existence
4392 * in the biodone routine.
4395 bufdone(struct buf *bp)
4397 struct bufobj *dropobj;
4398 void (*biodone)(struct buf *);
4400 buf_track(bp, __func__);
4401 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4404 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4405 BUF_ASSERT_HELD(bp);
4407 runningbufwakeup(bp);
4408 if (bp->b_iocmd == BIO_WRITE)
4409 dropobj = bp->b_bufobj;
4410 /* call optional completion function if requested */
4411 if (bp->b_iodone != NULL) {
4412 biodone = bp->b_iodone;
4413 bp->b_iodone = NULL;
4416 bufobj_wdrop(dropobj);
4419 if (bp->b_flags & B_VMIO) {
4421 * Set B_CACHE if the op was a normal read and no error
4422 * occurred. B_CACHE is set for writes in the b*write()
4425 if (bp->b_iocmd == BIO_READ &&
4426 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4427 !(bp->b_ioflags & BIO_ERROR))
4428 bp->b_flags |= B_CACHE;
4429 vfs_vmio_iodone(bp);
4431 if (!LIST_EMPTY(&bp->b_dep))
4433 if ((bp->b_flags & B_CKHASH) != 0) {
4434 KASSERT(bp->b_iocmd == BIO_READ,
4435 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4436 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4437 (*bp->b_ckhashcalc)(bp);
4440 * For asynchronous completions, release the buffer now. The brelse
4441 * will do a wakeup there if necessary - so no need to do a wakeup
4442 * here in the async case. The sync case always needs to do a wakeup.
4444 if (bp->b_flags & B_ASYNC) {
4445 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4446 (bp->b_ioflags & BIO_ERROR))
4453 bufobj_wdrop(dropobj);
4457 * This routine is called in lieu of iodone in the case of
4458 * incomplete I/O. This keeps the busy status for pages
4462 vfs_unbusy_pages(struct buf *bp)
4468 runningbufwakeup(bp);
4469 if (!(bp->b_flags & B_VMIO))
4472 obj = bp->b_bufobj->bo_object;
4473 VM_OBJECT_WLOCK(obj);
4474 for (i = 0; i < bp->b_npages; i++) {
4476 if (m == bogus_page) {
4477 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4479 panic("vfs_unbusy_pages: page missing\n");
4481 if (buf_mapped(bp)) {
4482 BUF_CHECK_MAPPED(bp);
4483 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4484 bp->b_pages, bp->b_npages);
4486 BUF_CHECK_UNMAPPED(bp);
4490 vm_object_pip_wakeupn(obj, bp->b_npages);
4491 VM_OBJECT_WUNLOCK(obj);
4495 * vfs_page_set_valid:
4497 * Set the valid bits in a page based on the supplied offset. The
4498 * range is restricted to the buffer's size.
4500 * This routine is typically called after a read completes.
4503 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4508 * Compute the end offset, eoff, such that [off, eoff) does not span a
4509 * page boundary and eoff is not greater than the end of the buffer.
4510 * The end of the buffer, in this case, is our file EOF, not the
4511 * allocation size of the buffer.
4513 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4514 if (eoff > bp->b_offset + bp->b_bcount)
4515 eoff = bp->b_offset + bp->b_bcount;
4518 * Set valid range. This is typically the entire buffer and thus the
4522 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4526 * vfs_page_set_validclean:
4528 * Set the valid bits and clear the dirty bits in a page based on the
4529 * supplied offset. The range is restricted to the buffer's size.
4532 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4534 vm_ooffset_t soff, eoff;
4537 * Start and end offsets in buffer. eoff - soff may not cross a
4538 * page boundary or cross the end of the buffer. The end of the
4539 * buffer, in this case, is our file EOF, not the allocation size
4543 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4544 if (eoff > bp->b_offset + bp->b_bcount)
4545 eoff = bp->b_offset + bp->b_bcount;
4548 * Set valid range. This is typically the entire buffer and thus the
4552 vm_page_set_validclean(
4554 (vm_offset_t) (soff & PAGE_MASK),
4555 (vm_offset_t) (eoff - soff)
4561 * Ensure that all buffer pages are not exclusive busied. If any page is
4562 * exclusive busy, drain it.
4565 vfs_drain_busy_pages(struct buf *bp)
4570 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4572 for (i = 0; i < bp->b_npages; i++) {
4574 if (vm_page_xbusied(m)) {
4575 for (; last_busied < i; last_busied++)
4576 vm_page_sbusy(bp->b_pages[last_busied]);
4577 while (vm_page_xbusied(m)) {
4579 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4580 vm_page_busy_sleep(m, "vbpage", true);
4581 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4585 for (i = 0; i < last_busied; i++)
4586 vm_page_sunbusy(bp->b_pages[i]);
4590 * This routine is called before a device strategy routine.
4591 * It is used to tell the VM system that paging I/O is in
4592 * progress, and treat the pages associated with the buffer
4593 * almost as being exclusive busy. Also the object paging_in_progress
4594 * flag is handled to make sure that the object doesn't become
4597 * Since I/O has not been initiated yet, certain buffer flags
4598 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4599 * and should be ignored.
4602 vfs_busy_pages(struct buf *bp, int clear_modify)
4610 if (!(bp->b_flags & B_VMIO))
4613 obj = bp->b_bufobj->bo_object;
4614 foff = bp->b_offset;
4615 KASSERT(bp->b_offset != NOOFFSET,
4616 ("vfs_busy_pages: no buffer offset"));
4617 VM_OBJECT_WLOCK(obj);
4618 vfs_drain_busy_pages(bp);
4619 if (bp->b_bufsize != 0)
4620 vfs_setdirty_locked_object(bp);
4622 for (i = 0; i < bp->b_npages; i++) {
4625 if ((bp->b_flags & B_CLUSTER) == 0) {
4626 vm_object_pip_add(obj, 1);
4630 * When readying a buffer for a read ( i.e
4631 * clear_modify == 0 ), it is important to do
4632 * bogus_page replacement for valid pages in
4633 * partially instantiated buffers. Partially
4634 * instantiated buffers can, in turn, occur when
4635 * reconstituting a buffer from its VM backing store
4636 * base. We only have to do this if B_CACHE is
4637 * clear ( which causes the I/O to occur in the
4638 * first place ). The replacement prevents the read
4639 * I/O from overwriting potentially dirty VM-backed
4640 * pages. XXX bogus page replacement is, uh, bogus.
4641 * It may not work properly with small-block devices.
4642 * We need to find a better way.
4645 pmap_remove_write(m);
4646 vfs_page_set_validclean(bp, foff, m);
4647 } else if (m->valid == VM_PAGE_BITS_ALL &&
4648 (bp->b_flags & B_CACHE) == 0) {
4649 bp->b_pages[i] = bogus_page;
4652 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4654 VM_OBJECT_WUNLOCK(obj);
4655 if (bogus && buf_mapped(bp)) {
4656 BUF_CHECK_MAPPED(bp);
4657 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4658 bp->b_pages, bp->b_npages);
4663 * vfs_bio_set_valid:
4665 * Set the range within the buffer to valid. The range is
4666 * relative to the beginning of the buffer, b_offset. Note that
4667 * b_offset itself may be offset from the beginning of the first
4671 vfs_bio_set_valid(struct buf *bp, int base, int size)
4676 if (!(bp->b_flags & B_VMIO))
4680 * Fixup base to be relative to beginning of first page.
4681 * Set initial n to be the maximum number of bytes in the
4682 * first page that can be validated.
4684 base += (bp->b_offset & PAGE_MASK);
4685 n = PAGE_SIZE - (base & PAGE_MASK);
4687 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4688 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4692 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4697 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4703 * If the specified buffer is a non-VMIO buffer, clear the entire
4704 * buffer. If the specified buffer is a VMIO buffer, clear and
4705 * validate only the previously invalid portions of the buffer.
4706 * This routine essentially fakes an I/O, so we need to clear
4707 * BIO_ERROR and B_INVAL.
4709 * Note that while we only theoretically need to clear through b_bcount,
4710 * we go ahead and clear through b_bufsize.
4713 vfs_bio_clrbuf(struct buf *bp)
4715 int i, j, mask, sa, ea, slide;
4717 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4721 bp->b_flags &= ~B_INVAL;
4722 bp->b_ioflags &= ~BIO_ERROR;
4723 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4724 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4725 (bp->b_offset & PAGE_MASK) == 0) {
4726 if (bp->b_pages[0] == bogus_page)
4728 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4729 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4730 if ((bp->b_pages[0]->valid & mask) == mask)
4732 if ((bp->b_pages[0]->valid & mask) == 0) {
4733 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4734 bp->b_pages[0]->valid |= mask;
4738 sa = bp->b_offset & PAGE_MASK;
4740 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4741 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4742 ea = slide & PAGE_MASK;
4745 if (bp->b_pages[i] == bogus_page)
4748 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4749 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4750 if ((bp->b_pages[i]->valid & mask) == mask)
4752 if ((bp->b_pages[i]->valid & mask) == 0)
4753 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4755 for (; sa < ea; sa += DEV_BSIZE, j++) {
4756 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4757 pmap_zero_page_area(bp->b_pages[i],
4762 bp->b_pages[i]->valid |= mask;
4765 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4770 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4775 if (buf_mapped(bp)) {
4776 BUF_CHECK_MAPPED(bp);
4777 bzero(bp->b_data + base, size);
4779 BUF_CHECK_UNMAPPED(bp);
4780 n = PAGE_SIZE - (base & PAGE_MASK);
4781 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4785 pmap_zero_page_area(m, base & PAGE_MASK, n);
4794 * Update buffer flags based on I/O request parameters, optionally releasing the
4795 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4796 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4797 * I/O). Otherwise the buffer is released to the cache.
4800 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4803 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4804 ("buf %p non-VMIO noreuse", bp));
4806 if ((ioflag & IO_DIRECT) != 0)
4807 bp->b_flags |= B_DIRECT;
4808 if ((ioflag & IO_EXT) != 0)
4809 bp->b_xflags |= BX_ALTDATA;
4810 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4811 bp->b_flags |= B_RELBUF;
4812 if ((ioflag & IO_NOREUSE) != 0)
4813 bp->b_flags |= B_NOREUSE;
4821 vfs_bio_brelse(struct buf *bp, int ioflag)
4824 b_io_dismiss(bp, ioflag, true);
4828 vfs_bio_set_flags(struct buf *bp, int ioflag)
4831 b_io_dismiss(bp, ioflag, false);
4835 * vm_hold_load_pages and vm_hold_free_pages get pages into
4836 * a buffers address space. The pages are anonymous and are
4837 * not associated with a file object.
4840 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4846 BUF_CHECK_MAPPED(bp);
4848 to = round_page(to);
4849 from = round_page(from);
4850 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4852 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4854 * note: must allocate system pages since blocking here
4855 * could interfere with paging I/O, no matter which
4858 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4859 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4861 pmap_qenter(pg, &p, 1);
4862 bp->b_pages[index] = p;
4864 bp->b_npages = index;
4867 /* Return pages associated with this buf to the vm system */
4869 vm_hold_free_pages(struct buf *bp, int newbsize)
4873 int index, newnpages;
4875 BUF_CHECK_MAPPED(bp);
4877 from = round_page((vm_offset_t)bp->b_data + newbsize);
4878 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4879 if (bp->b_npages > newnpages)
4880 pmap_qremove(from, bp->b_npages - newnpages);
4881 for (index = newnpages; index < bp->b_npages; index++) {
4882 p = bp->b_pages[index];
4883 bp->b_pages[index] = NULL;
4887 vm_wire_sub(bp->b_npages - newnpages);
4888 bp->b_npages = newnpages;
4892 * Map an IO request into kernel virtual address space.
4894 * All requests are (re)mapped into kernel VA space.
4895 * Notice that we use b_bufsize for the size of the buffer
4896 * to be mapped. b_bcount might be modified by the driver.
4898 * Note that even if the caller determines that the address space should
4899 * be valid, a race or a smaller-file mapped into a larger space may
4900 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4901 * check the return value.
4903 * This function only works with pager buffers.
4906 vmapbuf(struct buf *bp, int mapbuf)
4911 if (bp->b_bufsize < 0)
4913 prot = VM_PROT_READ;
4914 if (bp->b_iocmd == BIO_READ)
4915 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4916 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4917 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4918 btoc(MAXPHYS))) < 0)
4920 bp->b_npages = pidx;
4921 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4922 if (mapbuf || !unmapped_buf_allowed) {
4923 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4924 bp->b_data = bp->b_kvabase + bp->b_offset;
4926 bp->b_data = unmapped_buf;
4931 * Free the io map PTEs associated with this IO operation.
4932 * We also invalidate the TLB entries and restore the original b_addr.
4934 * This function only works with pager buffers.
4937 vunmapbuf(struct buf *bp)
4941 npages = bp->b_npages;
4943 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4944 vm_page_unhold_pages(bp->b_pages, npages);
4946 bp->b_data = unmapped_buf;
4950 bdone(struct buf *bp)
4954 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4956 bp->b_flags |= B_DONE;
4962 bwait(struct buf *bp, u_char pri, const char *wchan)
4966 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4968 while ((bp->b_flags & B_DONE) == 0)
4969 msleep(bp, mtxp, pri, wchan, 0);
4974 bufsync(struct bufobj *bo, int waitfor)
4977 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
4981 bufstrategy(struct bufobj *bo, struct buf *bp)
4987 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4988 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4989 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4990 i = VOP_STRATEGY(vp, bp);
4991 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4995 * Initialize a struct bufobj before use. Memory is assumed zero filled.
4998 bufobj_init(struct bufobj *bo, void *private)
5000 static volatile int bufobj_cleanq;
5003 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5004 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5005 bo->bo_private = private;
5006 TAILQ_INIT(&bo->bo_clean.bv_hd);
5007 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5011 bufobj_wrefl(struct bufobj *bo)
5014 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5015 ASSERT_BO_WLOCKED(bo);
5020 bufobj_wref(struct bufobj *bo)
5023 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5030 bufobj_wdrop(struct bufobj *bo)
5033 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5035 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5036 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5037 bo->bo_flag &= ~BO_WWAIT;
5038 wakeup(&bo->bo_numoutput);
5044 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5048 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5049 ASSERT_BO_WLOCKED(bo);
5051 while (bo->bo_numoutput) {
5052 bo->bo_flag |= BO_WWAIT;
5053 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5054 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5062 * Set bio_data or bio_ma for struct bio from the struct buf.
5065 bdata2bio(struct buf *bp, struct bio *bip)
5068 if (!buf_mapped(bp)) {
5069 KASSERT(unmapped_buf_allowed, ("unmapped"));
5070 bip->bio_ma = bp->b_pages;
5071 bip->bio_ma_n = bp->b_npages;
5072 bip->bio_data = unmapped_buf;
5073 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5074 bip->bio_flags |= BIO_UNMAPPED;
5075 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5076 PAGE_SIZE == bp->b_npages,
5077 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5078 (long long)bip->bio_length, bip->bio_ma_n));
5080 bip->bio_data = bp->b_data;
5086 * The MIPS pmap code currently doesn't handle aliased pages.
5087 * The VIPT caches may not handle page aliasing themselves, leading
5088 * to data corruption.
5090 * As such, this code makes a system extremely unhappy if said
5091 * system doesn't support unaliasing the above situation in hardware.
5092 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5093 * this feature at build time, so it has to be handled in software.
5095 * Once the MIPS pmap/cache code grows to support this function on
5096 * earlier chips, it should be flipped back off.
5099 static int buf_pager_relbuf = 1;
5101 static int buf_pager_relbuf = 0;
5103 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5104 &buf_pager_relbuf, 0,
5105 "Make buffer pager release buffers after reading");
5108 * The buffer pager. It uses buffer reads to validate pages.
5110 * In contrast to the generic local pager from vm/vnode_pager.c, this
5111 * pager correctly and easily handles volumes where the underlying
5112 * device block size is greater than the machine page size. The
5113 * buffer cache transparently extends the requested page run to be
5114 * aligned at the block boundary, and does the necessary bogus page
5115 * replacements in the addends to avoid obliterating already valid
5118 * The only non-trivial issue is that the exclusive busy state for
5119 * pages, which is assumed by the vm_pager_getpages() interface, is
5120 * incompatible with the VMIO buffer cache's desire to share-busy the
5121 * pages. This function performs a trivial downgrade of the pages'
5122 * state before reading buffers, and a less trivial upgrade from the
5123 * shared-busy to excl-busy state after the read.
5126 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5127 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5128 vbg_get_blksize_t get_blksize)
5135 vm_ooffset_t la, lb, poff, poffe;
5137 int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5140 object = vp->v_object;
5143 la = IDX_TO_OFF(ma[count - 1]->pindex);
5144 if (la >= object->un_pager.vnp.vnp_size)
5145 return (VM_PAGER_BAD);
5148 * Change the meaning of la from where the last requested page starts
5149 * to where it ends, because that's the end of the requested region
5150 * and the start of the potential read-ahead region.
5153 lpart = la > object->un_pager.vnp.vnp_size;
5154 bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)));
5157 * Calculate read-ahead, behind and total pages.
5160 lb = IDX_TO_OFF(ma[0]->pindex);
5161 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5163 if (rbehind != NULL)
5165 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5166 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5167 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5172 VM_CNT_INC(v_vnodein);
5173 VM_CNT_ADD(v_vnodepgsin, pgsin);
5175 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5176 != 0) ? GB_UNMAPPED : 0;
5177 VM_OBJECT_WLOCK(object);
5179 for (i = 0; i < count; i++)
5180 vm_page_busy_downgrade(ma[i]);
5181 VM_OBJECT_WUNLOCK(object);
5184 for (i = 0; i < count; i++) {
5188 * Pages are shared busy and the object lock is not
5189 * owned, which together allow for the pages'
5190 * invalidation. The racy test for validity avoids
5191 * useless creation of the buffer for the most typical
5192 * case when invalidation is not used in redo or for
5193 * parallel read. The shared->excl upgrade loop at
5194 * the end of the function catches the race in a
5195 * reliable way (protected by the object lock).
5197 if (m->valid == VM_PAGE_BITS_ALL)
5200 poff = IDX_TO_OFF(m->pindex);
5201 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5202 for (; poff < poffe; poff += bsize) {
5203 lbn = get_lblkno(vp, poff);
5208 bsize = get_blksize(vp, lbn);
5209 error = bread_gb(vp, lbn, bsize, curthread->td_ucred,
5213 if (LIST_EMPTY(&bp->b_dep)) {
5215 * Invalidation clears m->valid, but
5216 * may leave B_CACHE flag if the
5217 * buffer existed at the invalidation
5218 * time. In this case, recycle the
5219 * buffer to do real read on next
5220 * bread() after redo.
5222 * Otherwise B_RELBUF is not strictly
5223 * necessary, enable to reduce buf
5226 if (buf_pager_relbuf ||
5227 m->valid != VM_PAGE_BITS_ALL)
5228 bp->b_flags |= B_RELBUF;
5230 bp->b_flags &= ~B_NOCACHE;
5236 KASSERT(1 /* racy, enable for debugging */ ||
5237 m->valid == VM_PAGE_BITS_ALL || i == count - 1,
5238 ("buf %d %p invalid", i, m));
5239 if (i == count - 1 && lpart) {
5240 VM_OBJECT_WLOCK(object);
5241 if (m->valid != 0 &&
5242 m->valid != VM_PAGE_BITS_ALL)
5243 vm_page_zero_invalid(m, TRUE);
5244 VM_OBJECT_WUNLOCK(object);
5250 VM_OBJECT_WLOCK(object);
5252 for (i = 0; i < count; i++) {
5253 vm_page_sunbusy(ma[i]);
5254 ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL);
5257 * Since the pages were only sbusy while neither the
5258 * buffer nor the object lock was held by us, or
5259 * reallocated while vm_page_grab() slept for busy
5260 * relinguish, they could have been invalidated.
5261 * Recheck the valid bits and re-read as needed.
5263 * Note that the last page is made fully valid in the
5264 * read loop, and partial validity for the page at
5265 * index count - 1 could mean that the page was
5266 * invalidated or removed, so we must restart for
5269 if (ma[i]->valid != VM_PAGE_BITS_ALL)
5272 if (redo && error == 0)
5274 VM_OBJECT_WUNLOCK(object);
5275 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5278 #include "opt_ddb.h"
5280 #include <ddb/ddb.h>
5282 /* DDB command to show buffer data */
5283 DB_SHOW_COMMAND(buffer, db_show_buffer)
5286 struct buf *bp = (struct buf *)addr;
5287 #ifdef FULL_BUF_TRACKING
5292 db_printf("usage: show buffer <addr>\n");
5296 db_printf("buf at %p\n", bp);
5297 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5298 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5299 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5300 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5301 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5302 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5304 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5305 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5306 "b_vp = %p, b_dep = %p\n",
5307 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5308 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5309 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5310 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5311 bp->b_kvabase, bp->b_kvasize);
5314 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5315 for (i = 0; i < bp->b_npages; i++) {
5319 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5321 (u_long)VM_PAGE_TO_PHYS(m));
5323 db_printf("( ??? )");
5324 if ((i + 1) < bp->b_npages)
5329 BUF_LOCKPRINTINFO(bp);
5330 #if defined(FULL_BUF_TRACKING)
5331 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5333 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5334 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5335 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5337 db_printf(" %2u: %s\n", j,
5338 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5340 #elif defined(BUF_TRACKING)
5341 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5346 DB_SHOW_COMMAND(bufqueues, bufqueues)
5348 struct bufdomain *bd;
5353 db_printf("bqempty: %d\n", bqempty.bq_len);
5355 for (i = 0; i < buf_domains; i++) {
5357 db_printf("Buf domain %d\n", i);
5358 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5359 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5360 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5362 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5363 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5364 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5365 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5366 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5368 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5369 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5370 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5371 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5374 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5375 total += bp->b_bufsize;
5376 db_printf("\tcleanq count\t%d (%ld)\n",
5377 bd->bd_cleanq->bq_len, total);
5379 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5380 total += bp->b_bufsize;
5381 db_printf("\tdirtyq count\t%d (%ld)\n",
5382 bd->bd_dirtyq.bq_len, total);
5383 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5384 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5385 db_printf("\tCPU ");
5386 for (j = 0; j <= mp_maxid; j++)
5387 db_printf("%d, ", bd->bd_subq[j].bq_len);
5391 for (j = 0; j < nbuf; j++)
5392 if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) {
5394 total += buf[j].b_bufsize;
5396 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5399 for (j = 0; j < nbuf; j++)
5400 if (buf[j].b_domain == i) {
5402 total += buf[j].b_bufsize;
5404 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5408 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5413 for (i = 0; i < nbuf; i++) {
5415 if (BUF_ISLOCKED(bp)) {
5416 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5424 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5430 db_printf("usage: show vnodebufs <addr>\n");
5433 vp = (struct vnode *)addr;
5434 db_printf("Clean buffers:\n");
5435 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5436 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5439 db_printf("Dirty buffers:\n");
5440 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5441 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5446 DB_COMMAND(countfreebufs, db_coundfreebufs)
5449 int i, used = 0, nfree = 0;
5452 db_printf("usage: countfreebufs\n");
5456 for (i = 0; i < nbuf; i++) {
5458 if (bp->b_qindex == QUEUE_EMPTY)
5464 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5466 db_printf("numfreebuffers is %d\n", numfreebuffers);