2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2004 Poul-Henning Kamp
5 * Copyright (c) 1994,1997 John S. Dyson
6 * Copyright (c) 2013 The FreeBSD Foundation
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * this file contains a new buffer I/O scheme implementing a coherent
36 * VM object and buffer cache scheme. Pains have been taken to make
37 * sure that the performance degradation associated with schemes such
38 * as this is not realized.
40 * Author: John S. Dyson
41 * Significant help during the development and debugging phases
42 * had been provided by David Greenman, also of the FreeBSD core team.
44 * see man buf(9) for more info.
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/bitset.h>
56 #include <sys/counter.h>
58 #include <sys/devicestat.h>
59 #include <sys/eventhandler.h>
62 #include <sys/limits.h>
64 #include <sys/malloc.h>
65 #include <sys/mount.h>
66 #include <sys/mutex.h>
67 #include <sys/kernel.h>
68 #include <sys/kthread.h>
70 #include <sys/racct.h>
71 #include <sys/refcount.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
75 #include <sys/sysctl.h>
76 #include <sys/syscallsubr.h>
78 #include <sys/vmmeter.h>
79 #include <sys/vnode.h>
80 #include <sys/watchdog.h>
81 #include <geom/geom.h>
83 #include <vm/vm_param.h>
84 #include <vm/vm_kern.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_map.h>
91 #include <vm/swap_pager.h>
93 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
95 struct bio_ops bioops; /* I/O operation notification */
97 struct buf_ops buf_ops_bio = {
98 .bop_name = "buf_ops_bio",
99 .bop_write = bufwrite,
100 .bop_strategy = bufstrategy,
102 .bop_bdflush = bufbdflush,
106 struct mtx_padalign bq_lock;
107 TAILQ_HEAD(, buf) bq_queue;
109 uint16_t bq_subqueue;
111 } __aligned(CACHE_LINE_SIZE);
113 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
114 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
115 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
116 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
119 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
120 struct bufqueue bd_dirtyq;
121 struct bufqueue *bd_cleanq;
122 struct mtx_padalign bd_run_lock;
127 long bd_bufspacethresh;
128 int bd_hifreebuffers;
129 int bd_lofreebuffers;
130 int bd_hidirtybuffers;
131 int bd_lodirtybuffers;
132 int bd_dirtybufthresh;
136 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
137 int __aligned(CACHE_LINE_SIZE) bd_running;
138 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
139 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
140 } __aligned(CACHE_LINE_SIZE);
142 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
143 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
144 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
145 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
146 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
147 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
148 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
149 #define BD_DOMAIN(bd) (bd - bdomain)
151 static char *buf; /* buffer header pool */
155 return ((struct buf *)(buf + (sizeof(struct buf) +
156 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
159 caddr_t __read_mostly unmapped_buf;
161 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
162 struct proc *bufdaemonproc;
164 static void vm_hold_free_pages(struct buf *bp, int newbsize);
165 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
167 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
168 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
170 static void vfs_clean_pages_dirty_buf(struct buf *bp);
171 static void vfs_setdirty_range(struct buf *bp);
172 static void vfs_vmio_invalidate(struct buf *bp);
173 static void vfs_vmio_truncate(struct buf *bp, int npages);
174 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
175 static int vfs_bio_clcheck(struct vnode *vp, int size,
176 daddr_t lblkno, daddr_t blkno);
177 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
178 void (*)(struct buf *));
179 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
180 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
181 static void buf_daemon(void);
182 static __inline void bd_wakeup(void);
183 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
184 static void bufkva_reclaim(vmem_t *, int);
185 static void bufkva_free(struct buf *);
186 static int buf_import(void *, void **, int, int, int);
187 static void buf_release(void *, void **, int);
188 static void maxbcachebuf_adjust(void);
189 static inline struct bufdomain *bufdomain(struct buf *);
190 static void bq_remove(struct bufqueue *bq, struct buf *bp);
191 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
192 static int buf_recycle(struct bufdomain *, bool kva);
193 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
194 const char *lockname);
195 static void bd_init(struct bufdomain *bd);
196 static int bd_flushall(struct bufdomain *bd);
197 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
198 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
200 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
201 int vmiodirenable = TRUE;
202 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
203 "Use the VM system for directory writes");
204 long runningbufspace;
205 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
206 "Amount of presently outstanding async buffer io");
207 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
208 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
209 static counter_u64_t bufkvaspace;
210 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
211 "Kernel virtual memory used for buffers");
212 static long maxbufspace;
213 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
214 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
215 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
216 "Maximum allowed value of bufspace (including metadata)");
217 static long bufmallocspace;
218 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
219 "Amount of malloced memory for buffers");
220 static long maxbufmallocspace;
221 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
222 0, "Maximum amount of malloced memory for buffers");
223 static long lobufspace;
224 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
225 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
226 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
227 "Minimum amount of buffers we want to have");
229 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
230 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
231 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
232 "Maximum allowed value of bufspace (excluding metadata)");
234 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
235 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
236 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
237 "Bufspace consumed before waking the daemon to free some");
238 static counter_u64_t buffreekvacnt;
239 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
240 "Number of times we have freed the KVA space from some buffer");
241 static counter_u64_t bufdefragcnt;
242 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
243 "Number of times we have had to repeat buffer allocation to defragment");
244 static long lorunningspace;
245 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
246 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
247 "Minimum preferred space used for in-progress I/O");
248 static long hirunningspace;
249 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
250 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
251 "Maximum amount of space to use for in-progress I/O");
252 int dirtybufferflushes;
253 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
254 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
256 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
257 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
258 int altbufferflushes;
259 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
260 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
261 static int recursiveflushes;
262 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
263 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
264 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
265 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
266 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
267 "Number of buffers that are dirty (has unwritten changes) at the moment");
268 static int lodirtybuffers;
269 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
270 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
271 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
272 "How many buffers we want to have free before bufdaemon can sleep");
273 static int hidirtybuffers;
274 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
275 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
276 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
277 "When the number of dirty buffers is considered severe");
279 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
280 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
281 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
282 "Number of bdwrite to bawrite conversions to clear dirty buffers");
283 static int numfreebuffers;
284 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
285 "Number of free buffers");
286 static int lofreebuffers;
287 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
288 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
289 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
290 "Target number of free buffers");
291 static int hifreebuffers;
292 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
293 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
294 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
295 "Threshold for clean buffer recycling");
296 static counter_u64_t getnewbufcalls;
297 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
298 &getnewbufcalls, "Number of calls to getnewbuf");
299 static counter_u64_t getnewbufrestarts;
300 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
302 "Number of times getnewbuf has had to restart a buffer acquisition");
303 static counter_u64_t mappingrestarts;
304 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
306 "Number of times getblk has had to restart a buffer mapping for "
308 static counter_u64_t numbufallocfails;
309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
310 &numbufallocfails, "Number of times buffer allocations failed");
311 static int flushbufqtarget = 100;
312 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
313 "Amount of work to do in flushbufqueues when helping bufdaemon");
314 static counter_u64_t notbufdflushes;
315 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
316 "Number of dirty buffer flushes done by the bufdaemon helpers");
317 static long barrierwrites;
318 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
319 &barrierwrites, 0, "Number of barrier writes");
320 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
321 &unmapped_buf_allowed, 0,
322 "Permit the use of the unmapped i/o");
323 int maxbcachebuf = MAXBCACHEBUF;
324 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
325 "Maximum size of a buffer cache block");
328 * This lock synchronizes access to bd_request.
330 static struct mtx_padalign __exclusive_cache_line bdlock;
333 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
334 * waitrunningbufspace().
336 static struct mtx_padalign __exclusive_cache_line rbreqlock;
339 * Lock that protects bdirtywait.
341 static struct mtx_padalign __exclusive_cache_line bdirtylock;
344 * Wakeup point for bufdaemon, as well as indicator of whether it is already
345 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
348 static int bd_request;
351 * Request for the buf daemon to write more buffers than is indicated by
352 * lodirtybuf. This may be necessary to push out excess dependencies or
353 * defragment the address space where a simple count of the number of dirty
354 * buffers is insufficient to characterize the demand for flushing them.
356 static int bd_speedupreq;
359 * Synchronization (sleep/wakeup) variable for active buffer space requests.
360 * Set when wait starts, cleared prior to wakeup().
361 * Used in runningbufwakeup() and waitrunningbufspace().
363 static int runningbufreq;
366 * Synchronization for bwillwrite() waiters.
368 static int bdirtywait;
371 * Definitions for the buffer free lists.
373 #define QUEUE_NONE 0 /* on no queue */
374 #define QUEUE_EMPTY 1 /* empty buffer headers */
375 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
376 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
377 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
379 /* Maximum number of buffer domains. */
380 #define BUF_DOMAINS 8
382 struct bufdomainset bdlodirty; /* Domains > lodirty */
383 struct bufdomainset bdhidirty; /* Domains > hidirty */
385 /* Configured number of clean queues. */
386 static int __read_mostly buf_domains;
388 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
389 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
390 struct bufqueue __exclusive_cache_line bqempty;
393 * per-cpu empty buffer cache.
398 * Single global constant for BUF_WMESG, to avoid getting multiple references.
399 * buf_wmesg is referred from macros.
401 const char *buf_wmesg = BUF_WMESG;
404 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
409 value = *(long *)arg1;
410 error = sysctl_handle_long(oidp, &value, 0, req);
411 if (error != 0 || req->newptr == NULL)
413 mtx_lock(&rbreqlock);
414 if (arg1 == &hirunningspace) {
415 if (value < lorunningspace)
418 hirunningspace = value;
420 KASSERT(arg1 == &lorunningspace,
421 ("%s: unknown arg1", __func__));
422 if (value > hirunningspace)
425 lorunningspace = value;
427 mtx_unlock(&rbreqlock);
432 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
438 value = *(int *)arg1;
439 error = sysctl_handle_int(oidp, &value, 0, req);
440 if (error != 0 || req->newptr == NULL)
442 *(int *)arg1 = value;
443 for (i = 0; i < buf_domains; i++)
444 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
451 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
457 value = *(long *)arg1;
458 error = sysctl_handle_long(oidp, &value, 0, req);
459 if (error != 0 || req->newptr == NULL)
461 *(long *)arg1 = value;
462 for (i = 0; i < buf_domains; i++)
463 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
469 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
470 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
472 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
479 for (i = 0; i < buf_domains; i++)
480 lvalue += bdomain[i].bd_bufspace;
481 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
482 return (sysctl_handle_long(oidp, &lvalue, 0, req));
483 if (lvalue > INT_MAX)
484 /* On overflow, still write out a long to trigger ENOMEM. */
485 return (sysctl_handle_long(oidp, &lvalue, 0, req));
487 return (sysctl_handle_int(oidp, &ivalue, 0, req));
491 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
497 for (i = 0; i < buf_domains; i++)
498 lvalue += bdomain[i].bd_bufspace;
499 return (sysctl_handle_long(oidp, &lvalue, 0, req));
504 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
510 for (i = 0; i < buf_domains; i++)
511 value += bdomain[i].bd_numdirtybuffers;
512 return (sysctl_handle_int(oidp, &value, 0, req));
518 * Wakeup any bwillwrite() waiters.
523 mtx_lock(&bdirtylock);
528 mtx_unlock(&bdirtylock);
534 * Clear a domain from the appropriate bitsets when dirtybuffers
538 bd_clear(struct bufdomain *bd)
541 mtx_lock(&bdirtylock);
542 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
543 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
544 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
545 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
546 mtx_unlock(&bdirtylock);
552 * Set a domain in the appropriate bitsets when dirtybuffers
556 bd_set(struct bufdomain *bd)
559 mtx_lock(&bdirtylock);
560 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
561 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
562 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
563 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
564 mtx_unlock(&bdirtylock);
570 * Decrement the numdirtybuffers count by one and wakeup any
571 * threads blocked in bwillwrite().
574 bdirtysub(struct buf *bp)
576 struct bufdomain *bd;
580 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
581 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
583 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
590 * Increment the numdirtybuffers count by one and wakeup the buf
594 bdirtyadd(struct buf *bp)
596 struct bufdomain *bd;
600 * Only do the wakeup once as we cross the boundary. The
601 * buf daemon will keep running until the condition clears.
604 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
605 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
607 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
612 * bufspace_daemon_wakeup:
614 * Wakeup the daemons responsible for freeing clean bufs.
617 bufspace_daemon_wakeup(struct bufdomain *bd)
621 * avoid the lock if the daemon is running.
623 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
625 atomic_store_int(&bd->bd_running, 1);
626 wakeup(&bd->bd_running);
632 * bufspace_daemon_wait:
634 * Sleep until the domain falls below a limit or one second passes.
637 bufspace_daemon_wait(struct bufdomain *bd)
640 * Re-check our limits and sleep. bd_running must be
641 * cleared prior to checking the limits to avoid missed
642 * wakeups. The waker will adjust one of bufspace or
643 * freebuffers prior to checking bd_running.
646 atomic_store_int(&bd->bd_running, 0);
647 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
648 bd->bd_freebuffers > bd->bd_lofreebuffers) {
649 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP,
652 /* Avoid spurious wakeups while running. */
653 atomic_store_int(&bd->bd_running, 1);
661 * Adjust the reported bufspace for a KVA managed buffer, possibly
662 * waking any waiters.
665 bufspace_adjust(struct buf *bp, int bufsize)
667 struct bufdomain *bd;
671 KASSERT((bp->b_flags & B_MALLOC) == 0,
672 ("bufspace_adjust: malloc buf %p", bp));
674 diff = bufsize - bp->b_bufsize;
676 atomic_subtract_long(&bd->bd_bufspace, -diff);
677 } else if (diff > 0) {
678 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
679 /* Wake up the daemon on the transition. */
680 if (space < bd->bd_bufspacethresh &&
681 space + diff >= bd->bd_bufspacethresh)
682 bufspace_daemon_wakeup(bd);
684 bp->b_bufsize = bufsize;
690 * Reserve bufspace before calling allocbuf(). metadata has a
691 * different space limit than data.
694 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
700 limit = bd->bd_maxbufspace;
702 limit = bd->bd_hibufspace;
703 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
706 atomic_subtract_long(&bd->bd_bufspace, size);
710 /* Wake up the daemon on the transition. */
711 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
712 bufspace_daemon_wakeup(bd);
720 * Release reserved bufspace after bufspace_adjust() has consumed it.
723 bufspace_release(struct bufdomain *bd, int size)
726 atomic_subtract_long(&bd->bd_bufspace, size);
732 * Wait for bufspace, acting as the buf daemon if a locked vnode is
733 * supplied. bd_wanted must be set prior to polling for space. The
734 * operation must be re-tried on return.
737 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
738 int slpflag, int slptimeo)
741 int error, fl, norunbuf;
743 if ((gbflags & GB_NOWAIT_BD) != 0)
748 while (bd->bd_wanted) {
749 if (vp != NULL && vp->v_type != VCHR &&
750 (td->td_pflags & TDP_BUFNEED) == 0) {
753 * getblk() is called with a vnode locked, and
754 * some majority of the dirty buffers may as
755 * well belong to the vnode. Flushing the
756 * buffers there would make a progress that
757 * cannot be achieved by the buf_daemon, that
758 * cannot lock the vnode.
760 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
761 (td->td_pflags & TDP_NORUNNINGBUF);
764 * Play bufdaemon. The getnewbuf() function
765 * may be called while the thread owns lock
766 * for another dirty buffer for the same
767 * vnode, which makes it impossible to use
768 * VOP_FSYNC() there, due to the buffer lock
771 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
772 fl = buf_flush(vp, bd, flushbufqtarget);
773 td->td_pflags &= norunbuf;
777 if (bd->bd_wanted == 0)
780 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
781 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
791 * buffer space management daemon. Tries to maintain some marginal
792 * amount of free buffer space so that requesting processes neither
793 * block nor work to reclaim buffers.
796 bufspace_daemon(void *arg)
798 struct bufdomain *bd;
800 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
801 SHUTDOWN_PRI_LAST + 100);
805 kthread_suspend_check();
808 * Free buffers from the clean queue until we meet our
811 * Theory of operation: The buffer cache is most efficient
812 * when some free buffer headers and space are always
813 * available to getnewbuf(). This daemon attempts to prevent
814 * the excessive blocking and synchronization associated
815 * with shortfall. It goes through three phases according
818 * 1) The daemon wakes up voluntarily once per-second
819 * during idle periods when the counters are below
820 * the wakeup thresholds (bufspacethresh, lofreebuffers).
822 * 2) The daemon wakes up as we cross the thresholds
823 * ahead of any potential blocking. This may bounce
824 * slightly according to the rate of consumption and
827 * 3) The daemon and consumers are starved for working
828 * clean buffers. This is the 'bufspace' sleep below
829 * which will inefficiently trade bufs with bqrelse
830 * until we return to condition 2.
832 while (bd->bd_bufspace > bd->bd_lobufspace ||
833 bd->bd_freebuffers < bd->bd_hifreebuffers) {
834 if (buf_recycle(bd, false) != 0) {
838 * Speedup dirty if we've run out of clean
839 * buffers. This is possible in particular
840 * because softdep may held many bufs locked
841 * pending writes to other bufs which are
842 * marked for delayed write, exhausting
843 * clean space until they are written.
848 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
849 PRIBIO|PDROP, "bufspace", hz/10);
855 bufspace_daemon_wait(bd);
862 * Adjust the reported bufspace for a malloc managed buffer, possibly
863 * waking any waiters.
866 bufmallocadjust(struct buf *bp, int bufsize)
870 KASSERT((bp->b_flags & B_MALLOC) != 0,
871 ("bufmallocadjust: non-malloc buf %p", bp));
872 diff = bufsize - bp->b_bufsize;
874 atomic_subtract_long(&bufmallocspace, -diff);
876 atomic_add_long(&bufmallocspace, diff);
877 bp->b_bufsize = bufsize;
883 * Wake up processes that are waiting on asynchronous writes to fall
884 * below lorunningspace.
890 mtx_lock(&rbreqlock);
893 wakeup(&runningbufreq);
895 mtx_unlock(&rbreqlock);
901 * Decrement the outstanding write count according.
904 runningbufwakeup(struct buf *bp)
908 bspace = bp->b_runningbufspace;
911 space = atomic_fetchadd_long(&runningbufspace, -bspace);
912 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
914 bp->b_runningbufspace = 0;
916 * Only acquire the lock and wakeup on the transition from exceeding
917 * the threshold to falling below it.
919 if (space < lorunningspace)
921 if (space - bspace > lorunningspace)
927 * waitrunningbufspace()
929 * runningbufspace is a measure of the amount of I/O currently
930 * running. This routine is used in async-write situations to
931 * prevent creating huge backups of pending writes to a device.
932 * Only asynchronous writes are governed by this function.
934 * This does NOT turn an async write into a sync write. It waits
935 * for earlier writes to complete and generally returns before the
936 * caller's write has reached the device.
939 waitrunningbufspace(void)
942 mtx_lock(&rbreqlock);
943 while (runningbufspace > hirunningspace) {
945 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
947 mtx_unlock(&rbreqlock);
951 * vfs_buf_test_cache:
953 * Called when a buffer is extended. This function clears the B_CACHE
954 * bit if the newly extended portion of the buffer does not contain
958 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
959 vm_offset_t size, vm_page_t m)
963 * This function and its results are protected by higher level
964 * synchronization requiring vnode and buf locks to page in and
967 if (bp->b_flags & B_CACHE) {
968 int base = (foff + off) & PAGE_MASK;
969 if (vm_page_is_valid(m, base, size) == 0)
970 bp->b_flags &= ~B_CACHE;
974 /* Wake up the buffer daemon if necessary */
980 if (bd_request == 0) {
988 * Adjust the maxbcachbuf tunable.
991 maxbcachebuf_adjust(void)
996 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
999 while (i * 2 <= maxbcachebuf)
1002 if (maxbcachebuf < MAXBSIZE)
1003 maxbcachebuf = MAXBSIZE;
1004 if (maxbcachebuf > maxphys)
1005 maxbcachebuf = maxphys;
1006 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1007 printf("maxbcachebuf=%d\n", maxbcachebuf);
1011 * bd_speedup - speedup the buffer cache flushing code
1020 if (bd_speedupreq == 0 || bd_request == 0)
1025 wakeup(&bd_request);
1026 mtx_unlock(&bdlock);
1030 #define TRANSIENT_DENOM 5
1032 #define TRANSIENT_DENOM 10
1036 * Calculating buffer cache scaling values and reserve space for buffer
1037 * headers. This is called during low level kernel initialization and
1038 * may be called more then once. We CANNOT write to the memory area
1039 * being reserved at this time.
1042 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1045 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1049 * With KASAN enabled, the kernel map is shadowed. Account for this
1050 * when sizing maps based on the amount of physical memory available.
1052 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1053 (KASAN_SHADOW_SCALE + 1);
1057 * physmem_est is in pages. Convert it to kilobytes (assumes
1058 * PAGE_SIZE is >= 1K)
1060 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1062 maxbcachebuf_adjust();
1064 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1065 * For the first 64MB of ram nominally allocate sufficient buffers to
1066 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1067 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1068 * the buffer cache we limit the eventual kva reservation to
1071 * factor represents the 1/4 x ram conversion.
1074 int factor = 4 * BKVASIZE / 1024;
1077 if (physmem_est > 4096)
1078 nbuf += min((physmem_est - 4096) / factor,
1080 if (physmem_est > 65536)
1081 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1082 32 * 1024 * 1024 / (factor * 5));
1084 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1085 nbuf = maxbcache / BKVASIZE;
1090 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1091 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1092 if (nbuf > maxbuf) {
1094 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1100 * Ideal allocation size for the transient bio submap is 10%
1101 * of the maximal space buffer map. This roughly corresponds
1102 * to the amount of the buffer mapped for typical UFS load.
1104 * Clip the buffer map to reserve space for the transient
1105 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1106 * maximum buffer map extent on the platform.
1108 * The fall-back to the maxbuf in case of maxbcache unset,
1109 * allows to not trim the buffer KVA for the architectures
1110 * with ample KVA space.
1112 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1113 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1114 buf_sz = (long)nbuf * BKVASIZE;
1115 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1116 (TRANSIENT_DENOM - 1)) {
1118 * There is more KVA than memory. Do not
1119 * adjust buffer map size, and assign the rest
1120 * of maxbuf to transient map.
1122 biotmap_sz = maxbuf_sz - buf_sz;
1125 * Buffer map spans all KVA we could afford on
1126 * this platform. Give 10% (20% on i386) of
1127 * the buffer map to the transient bio map.
1129 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1130 buf_sz -= biotmap_sz;
1132 if (biotmap_sz / INT_MAX > maxphys)
1133 bio_transient_maxcnt = INT_MAX;
1135 bio_transient_maxcnt = biotmap_sz / maxphys;
1137 * Artificially limit to 1024 simultaneous in-flight I/Os
1138 * using the transient mapping.
1140 if (bio_transient_maxcnt > 1024)
1141 bio_transient_maxcnt = 1024;
1143 nbuf = buf_sz / BKVASIZE;
1147 nswbuf = min(nbuf / 4, 256);
1148 if (nswbuf < NSWBUF_MIN)
1149 nswbuf = NSWBUF_MIN;
1153 * Reserve space for the buffer cache buffers
1156 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1157 atop(maxbcachebuf)) * nbuf;
1162 /* Initialize the buffer subsystem. Called before use of any buffers. */
1169 KASSERT(maxbcachebuf >= MAXBSIZE,
1170 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1172 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1173 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1174 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1175 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1177 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1179 /* finally, initialize each buffer header and stick on empty q */
1180 for (i = 0; i < nbuf; i++) {
1182 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1183 bp->b_flags = B_INVAL;
1184 bp->b_rcred = NOCRED;
1185 bp->b_wcred = NOCRED;
1186 bp->b_qindex = QUEUE_NONE;
1188 bp->b_subqueue = mp_maxid + 1;
1190 bp->b_data = bp->b_kvabase = unmapped_buf;
1191 LIST_INIT(&bp->b_dep);
1193 bq_insert(&bqempty, bp, false);
1197 * maxbufspace is the absolute maximum amount of buffer space we are
1198 * allowed to reserve in KVM and in real terms. The absolute maximum
1199 * is nominally used by metadata. hibufspace is the nominal maximum
1200 * used by most other requests. The differential is required to
1201 * ensure that metadata deadlocks don't occur.
1203 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1204 * this may result in KVM fragmentation which is not handled optimally
1205 * by the system. XXX This is less true with vmem. We could use
1208 maxbufspace = (long)nbuf * BKVASIZE;
1209 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1210 lobufspace = (hibufspace / 20) * 19; /* 95% */
1211 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1214 * Note: The 16 MiB upper limit for hirunningspace was chosen
1215 * arbitrarily and may need further tuning. It corresponds to
1216 * 128 outstanding write IO requests (if IO size is 128 KiB),
1217 * which fits with many RAID controllers' tagged queuing limits.
1218 * The lower 1 MiB limit is the historical upper limit for
1221 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1222 16 * 1024 * 1024), 1024 * 1024);
1223 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1226 * Limit the amount of malloc memory since it is wired permanently into
1227 * the kernel space. Even though this is accounted for in the buffer
1228 * allocation, we don't want the malloced region to grow uncontrolled.
1229 * The malloc scheme improves memory utilization significantly on
1230 * average (small) directories.
1232 maxbufmallocspace = hibufspace / 20;
1235 * Reduce the chance of a deadlock occurring by limiting the number
1236 * of delayed-write dirty buffers we allow to stack up.
1238 hidirtybuffers = nbuf / 4 + 20;
1239 dirtybufthresh = hidirtybuffers * 9 / 10;
1241 * To support extreme low-memory systems, make sure hidirtybuffers
1242 * cannot eat up all available buffer space. This occurs when our
1243 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1244 * buffer space assuming BKVASIZE'd buffers.
1246 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1247 hidirtybuffers >>= 1;
1249 lodirtybuffers = hidirtybuffers / 2;
1252 * lofreebuffers should be sufficient to avoid stalling waiting on
1253 * buf headers under heavy utilization. The bufs in per-cpu caches
1254 * are counted as free but will be unavailable to threads executing
1257 * hifreebuffers is the free target for the bufspace daemon. This
1258 * should be set appropriately to limit work per-iteration.
1260 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1261 hifreebuffers = (3 * lofreebuffers) / 2;
1262 numfreebuffers = nbuf;
1264 /* Setup the kva and free list allocators. */
1265 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1266 buf_zone = uma_zcache_create("buf free cache",
1267 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1268 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1271 * Size the clean queue according to the amount of buffer space.
1272 * One queue per-256mb up to the max. More queues gives better
1273 * concurrency but less accurate LRU.
1275 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1276 for (i = 0 ; i < buf_domains; i++) {
1277 struct bufdomain *bd;
1281 bd->bd_freebuffers = nbuf / buf_domains;
1282 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1283 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1284 bd->bd_bufspace = 0;
1285 bd->bd_maxbufspace = maxbufspace / buf_domains;
1286 bd->bd_hibufspace = hibufspace / buf_domains;
1287 bd->bd_lobufspace = lobufspace / buf_domains;
1288 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1289 bd->bd_numdirtybuffers = 0;
1290 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1291 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1292 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1293 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1294 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1296 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1297 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1298 mappingrestarts = counter_u64_alloc(M_WAITOK);
1299 numbufallocfails = counter_u64_alloc(M_WAITOK);
1300 notbufdflushes = counter_u64_alloc(M_WAITOK);
1301 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1302 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1303 bufkvaspace = counter_u64_alloc(M_WAITOK);
1308 vfs_buf_check_mapped(struct buf *bp)
1311 KASSERT(bp->b_kvabase != unmapped_buf,
1312 ("mapped buf: b_kvabase was not updated %p", bp));
1313 KASSERT(bp->b_data != unmapped_buf,
1314 ("mapped buf: b_data was not updated %p", bp));
1315 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1316 maxphys, ("b_data + b_offset unmapped %p", bp));
1320 vfs_buf_check_unmapped(struct buf *bp)
1323 KASSERT(bp->b_data == unmapped_buf,
1324 ("unmapped buf: corrupted b_data %p", bp));
1327 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1328 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1330 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1331 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1335 isbufbusy(struct buf *bp)
1337 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1338 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1344 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1347 bufshutdown(int show_busybufs)
1349 static int first_buf_printf = 1;
1351 int i, iter, nbusy, pbusy;
1357 * Sync filesystems for shutdown
1359 wdog_kern_pat(WD_LASTVAL);
1360 kern_sync(curthread);
1363 * With soft updates, some buffers that are
1364 * written will be remarked as dirty until other
1365 * buffers are written.
1367 for (iter = pbusy = 0; iter < 20; iter++) {
1369 for (i = nbuf - 1; i >= 0; i--) {
1375 if (first_buf_printf)
1376 printf("All buffers synced.");
1379 if (first_buf_printf) {
1380 printf("Syncing disks, buffers remaining... ");
1381 first_buf_printf = 0;
1383 printf("%d ", nbusy);
1388 wdog_kern_pat(WD_LASTVAL);
1389 kern_sync(curthread);
1393 * Spin for a while to allow interrupt threads to run.
1395 DELAY(50000 * iter);
1398 * Context switch several times to allow interrupt
1401 for (subiter = 0; subiter < 50 * iter; subiter++) {
1402 thread_lock(curthread);
1410 * Count only busy local buffers to prevent forcing
1411 * a fsck if we're just a client of a wedged NFS server
1414 for (i = nbuf - 1; i >= 0; i--) {
1416 if (isbufbusy(bp)) {
1418 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1419 if (bp->b_dev == NULL) {
1420 TAILQ_REMOVE(&mountlist,
1421 bp->b_vp->v_mount, mnt_list);
1426 if (show_busybufs > 0) {
1428 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1429 nbusy, bp, bp->b_vp, bp->b_flags,
1430 (intmax_t)bp->b_blkno,
1431 (intmax_t)bp->b_lblkno);
1432 BUF_LOCKPRINTINFO(bp);
1433 if (show_busybufs > 1)
1441 * Failed to sync all blocks. Indicate this and don't
1442 * unmount filesystems (thus forcing an fsck on reboot).
1444 printf("Giving up on %d buffers\n", nbusy);
1445 DELAY(5000000); /* 5 seconds */
1448 if (!first_buf_printf)
1449 printf("Final sync complete\n");
1452 * Unmount filesystems and perform swapoff, to quiesce
1453 * the system as much as possible. In particular, no
1454 * I/O should be initiated from top levels since it
1455 * might be abruptly terminated by reset, or otherwise
1456 * erronously handled because other parts of the
1457 * system are disabled.
1459 * Swapoff before unmount, because file-backed swap is
1460 * non-operational after unmount of the underlying
1463 if (!KERNEL_PANICKED()) {
1468 DELAY(100000); /* wait for console output to finish */
1472 bpmap_qenter(struct buf *bp)
1475 BUF_CHECK_MAPPED(bp);
1478 * bp->b_data is relative to bp->b_offset, but
1479 * bp->b_offset may be offset into the first page.
1481 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1482 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1483 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1484 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1487 static inline struct bufdomain *
1488 bufdomain(struct buf *bp)
1491 return (&bdomain[bp->b_domain]);
1494 static struct bufqueue *
1495 bufqueue(struct buf *bp)
1498 switch (bp->b_qindex) {
1501 case QUEUE_SENTINEL:
1506 return (&bufdomain(bp)->bd_dirtyq);
1508 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1512 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1516 * Return the locked bufqueue that bp is a member of.
1518 static struct bufqueue *
1519 bufqueue_acquire(struct buf *bp)
1521 struct bufqueue *bq, *nbq;
1524 * bp can be pushed from a per-cpu queue to the
1525 * cleanq while we're waiting on the lock. Retry
1526 * if the queues don't match.
1544 * Insert the buffer into the appropriate free list. Requires a
1545 * locked buffer on entry and buffer is unlocked before return.
1548 binsfree(struct buf *bp, int qindex)
1550 struct bufdomain *bd;
1551 struct bufqueue *bq;
1553 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1554 ("binsfree: Invalid qindex %d", qindex));
1555 BUF_ASSERT_XLOCKED(bp);
1558 * Handle delayed bremfree() processing.
1560 if (bp->b_flags & B_REMFREE) {
1561 if (bp->b_qindex == qindex) {
1562 bp->b_flags |= B_REUSE;
1563 bp->b_flags &= ~B_REMFREE;
1567 bq = bufqueue_acquire(bp);
1572 if (qindex == QUEUE_CLEAN) {
1573 if (bd->bd_lim != 0)
1574 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1578 bq = &bd->bd_dirtyq;
1579 bq_insert(bq, bp, true);
1585 * Free a buffer to the buf zone once it no longer has valid contents.
1588 buf_free(struct buf *bp)
1591 if (bp->b_flags & B_REMFREE)
1593 if (bp->b_vflags & BV_BKGRDINPROG)
1594 panic("losing buffer 1");
1595 if (bp->b_rcred != NOCRED) {
1596 crfree(bp->b_rcred);
1597 bp->b_rcred = NOCRED;
1599 if (bp->b_wcred != NOCRED) {
1600 crfree(bp->b_wcred);
1601 bp->b_wcred = NOCRED;
1603 if (!LIST_EMPTY(&bp->b_dep))
1606 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1607 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1609 uma_zfree(buf_zone, bp);
1615 * Import bufs into the uma cache from the buf list. The system still
1616 * expects a static array of bufs and much of the synchronization
1617 * around bufs assumes type stable storage. As a result, UMA is used
1618 * only as a per-cpu cache of bufs still maintained on a global list.
1621 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1627 for (i = 0; i < cnt; i++) {
1628 bp = TAILQ_FIRST(&bqempty.bq_queue);
1631 bq_remove(&bqempty, bp);
1634 BQ_UNLOCK(&bqempty);
1642 * Release bufs from the uma cache back to the buffer queues.
1645 buf_release(void *arg, void **store, int cnt)
1647 struct bufqueue *bq;
1653 for (i = 0; i < cnt; i++) {
1655 /* Inline bq_insert() to batch locking. */
1656 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1657 bp->b_flags &= ~(B_AGE | B_REUSE);
1659 bp->b_qindex = bq->bq_index;
1667 * Allocate an empty buffer header.
1670 buf_alloc(struct bufdomain *bd)
1673 int freebufs, error;
1676 * We can only run out of bufs in the buf zone if the average buf
1677 * is less than BKVASIZE. In this case the actual wait/block will
1678 * come from buf_reycle() failing to flush one of these small bufs.
1681 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1683 bp = uma_zalloc(buf_zone, M_NOWAIT);
1685 atomic_add_int(&bd->bd_freebuffers, 1);
1686 bufspace_daemon_wakeup(bd);
1687 counter_u64_add(numbufallocfails, 1);
1691 * Wake-up the bufspace daemon on transition below threshold.
1693 if (freebufs == bd->bd_lofreebuffers)
1694 bufspace_daemon_wakeup(bd);
1696 error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1697 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1701 KASSERT(bp->b_vp == NULL,
1702 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1703 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1704 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1705 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1706 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1707 KASSERT(bp->b_npages == 0,
1708 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1709 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1710 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1711 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1713 bp->b_domain = BD_DOMAIN(bd);
1719 bp->b_blkno = bp->b_lblkno = 0;
1720 bp->b_offset = NOOFFSET;
1726 bp->b_dirtyoff = bp->b_dirtyend = 0;
1727 bp->b_bufobj = NULL;
1728 bp->b_data = bp->b_kvabase = unmapped_buf;
1729 bp->b_fsprivate1 = NULL;
1730 bp->b_fsprivate2 = NULL;
1731 bp->b_fsprivate3 = NULL;
1732 LIST_INIT(&bp->b_dep);
1740 * Free a buffer from the given bufqueue. kva controls whether the
1741 * freed buf must own some kva resources. This is used for
1745 buf_recycle(struct bufdomain *bd, bool kva)
1747 struct bufqueue *bq;
1748 struct buf *bp, *nbp;
1751 counter_u64_add(bufdefragcnt, 1);
1755 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1756 ("buf_recycle: Locks don't match"));
1757 nbp = TAILQ_FIRST(&bq->bq_queue);
1760 * Run scan, possibly freeing data and/or kva mappings on the fly
1763 while ((bp = nbp) != NULL) {
1765 * Calculate next bp (we can only use it if we do not
1766 * release the bqlock).
1768 nbp = TAILQ_NEXT(bp, b_freelist);
1771 * If we are defragging then we need a buffer with
1772 * some kva to reclaim.
1774 if (kva && bp->b_kvasize == 0)
1777 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1781 * Implement a second chance algorithm for frequently
1784 if ((bp->b_flags & B_REUSE) != 0) {
1785 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1786 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1787 bp->b_flags &= ~B_REUSE;
1793 * Skip buffers with background writes in progress.
1795 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1800 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1801 ("buf_recycle: inconsistent queue %d bp %p",
1803 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1804 ("getnewbuf: queue domain %d doesn't match request %d",
1805 bp->b_domain, (int)BD_DOMAIN(bd)));
1807 * NOTE: nbp is now entirely invalid. We can only restart
1808 * the scan from this point on.
1814 * Requeue the background write buffer with error and
1817 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1820 nbp = TAILQ_FIRST(&bq->bq_queue);
1823 bp->b_flags |= B_INVAL;
1836 * Mark the buffer for removal from the appropriate free list.
1840 bremfree(struct buf *bp)
1843 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1844 KASSERT((bp->b_flags & B_REMFREE) == 0,
1845 ("bremfree: buffer %p already marked for delayed removal.", bp));
1846 KASSERT(bp->b_qindex != QUEUE_NONE,
1847 ("bremfree: buffer %p not on a queue.", bp));
1848 BUF_ASSERT_XLOCKED(bp);
1850 bp->b_flags |= B_REMFREE;
1856 * Force an immediate removal from a free list. Used only in nfs when
1857 * it abuses the b_freelist pointer.
1860 bremfreef(struct buf *bp)
1862 struct bufqueue *bq;
1864 bq = bufqueue_acquire(bp);
1870 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1873 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1874 TAILQ_INIT(&bq->bq_queue);
1876 bq->bq_index = qindex;
1877 bq->bq_subqueue = subqueue;
1881 bd_init(struct bufdomain *bd)
1885 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1886 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1887 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1888 for (i = 0; i <= mp_maxid; i++)
1889 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1890 "bufq clean subqueue lock");
1891 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1897 * Removes a buffer from the free list, must be called with the
1898 * correct qlock held.
1901 bq_remove(struct bufqueue *bq, struct buf *bp)
1904 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1905 bp, bp->b_vp, bp->b_flags);
1906 KASSERT(bp->b_qindex != QUEUE_NONE,
1907 ("bq_remove: buffer %p not on a queue.", bp));
1908 KASSERT(bufqueue(bp) == bq,
1909 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1911 BQ_ASSERT_LOCKED(bq);
1912 if (bp->b_qindex != QUEUE_EMPTY) {
1913 BUF_ASSERT_XLOCKED(bp);
1915 KASSERT(bq->bq_len >= 1,
1916 ("queue %d underflow", bp->b_qindex));
1917 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1919 bp->b_qindex = QUEUE_NONE;
1920 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1924 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1928 BQ_ASSERT_LOCKED(bq);
1929 if (bq != bd->bd_cleanq) {
1931 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1932 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1933 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1935 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1937 bd->bd_cleanq->bq_len += bq->bq_len;
1940 if (bd->bd_wanted) {
1942 wakeup(&bd->bd_wanted);
1944 if (bq != bd->bd_cleanq)
1949 bd_flushall(struct bufdomain *bd)
1951 struct bufqueue *bq;
1955 if (bd->bd_lim == 0)
1958 for (i = 0; i <= mp_maxid; i++) {
1959 bq = &bd->bd_subq[i];
1960 if (bq->bq_len == 0)
1972 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1974 struct bufdomain *bd;
1976 if (bp->b_qindex != QUEUE_NONE)
1977 panic("bq_insert: free buffer %p onto another queue?", bp);
1980 if (bp->b_flags & B_AGE) {
1981 /* Place this buf directly on the real queue. */
1982 if (bq->bq_index == QUEUE_CLEAN)
1985 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1988 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1990 bp->b_flags &= ~(B_AGE | B_REUSE);
1992 bp->b_qindex = bq->bq_index;
1993 bp->b_subqueue = bq->bq_subqueue;
1996 * Unlock before we notify so that we don't wakeup a waiter that
1997 * fails a trylock on the buf and sleeps again.
2002 if (bp->b_qindex == QUEUE_CLEAN) {
2004 * Flush the per-cpu queue and notify any waiters.
2006 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2007 bq->bq_len >= bd->bd_lim))
2016 * Free the kva allocation for a buffer.
2020 bufkva_free(struct buf *bp)
2024 if (bp->b_kvasize == 0) {
2025 KASSERT(bp->b_kvabase == unmapped_buf &&
2026 bp->b_data == unmapped_buf,
2027 ("Leaked KVA space on %p", bp));
2028 } else if (buf_mapped(bp))
2029 BUF_CHECK_MAPPED(bp);
2031 BUF_CHECK_UNMAPPED(bp);
2033 if (bp->b_kvasize == 0)
2036 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2037 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2038 counter_u64_add(buffreekvacnt, 1);
2039 bp->b_data = bp->b_kvabase = unmapped_buf;
2046 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2049 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2054 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2055 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2056 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2057 KASSERT(maxsize <= maxbcachebuf,
2058 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2063 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2066 * Buffer map is too fragmented. Request the caller
2067 * to defragment the map.
2071 bp->b_kvabase = (caddr_t)addr;
2072 bp->b_kvasize = maxsize;
2073 counter_u64_add(bufkvaspace, bp->b_kvasize);
2074 if ((gbflags & GB_UNMAPPED) != 0) {
2075 bp->b_data = unmapped_buf;
2076 BUF_CHECK_UNMAPPED(bp);
2078 bp->b_data = bp->b_kvabase;
2079 BUF_CHECK_MAPPED(bp);
2087 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2088 * callback that fires to avoid returning failure.
2091 bufkva_reclaim(vmem_t *vmem, int flags)
2098 for (i = 0; i < 5; i++) {
2099 for (q = 0; q < buf_domains; q++)
2100 if (buf_recycle(&bdomain[q], true) != 0)
2109 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2110 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2111 * the buffer is valid and we do not have to do anything.
2114 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2115 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2123 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2124 if (inmem(vp, *rablkno))
2126 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2127 if ((rabp->b_flags & B_CACHE) != 0) {
2134 racct_add_buf(curproc, rabp, 0);
2135 PROC_UNLOCK(curproc);
2138 td->td_ru.ru_inblock++;
2139 rabp->b_flags |= B_ASYNC;
2140 rabp->b_flags &= ~B_INVAL;
2141 if ((flags & GB_CKHASH) != 0) {
2142 rabp->b_flags |= B_CKHASH;
2143 rabp->b_ckhashcalc = ckhashfunc;
2145 rabp->b_ioflags &= ~BIO_ERROR;
2146 rabp->b_iocmd = BIO_READ;
2147 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2148 rabp->b_rcred = crhold(cred);
2149 vfs_busy_pages(rabp, 0);
2151 rabp->b_iooffset = dbtob(rabp->b_blkno);
2157 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2159 * Get a buffer with the specified data. Look in the cache first. We
2160 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2161 * is set, the buffer is valid and we do not have to do anything, see
2162 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2164 * Always return a NULL buffer pointer (in bpp) when returning an error.
2166 * The blkno parameter is the logical block being requested. Normally
2167 * the mapping of logical block number to disk block address is done
2168 * by calling VOP_BMAP(). However, if the mapping is already known, the
2169 * disk block address can be passed using the dblkno parameter. If the
2170 * disk block address is not known, then the same value should be passed
2171 * for blkno and dblkno.
2174 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2175 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2176 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2180 int error, readwait, rv;
2182 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2185 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2188 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2193 KASSERT(blkno == bp->b_lblkno,
2194 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2195 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2196 flags &= ~GB_NOSPARSE;
2200 * If not found in cache, do some I/O
2203 if ((bp->b_flags & B_CACHE) == 0) {
2206 PROC_LOCK(td->td_proc);
2207 racct_add_buf(td->td_proc, bp, 0);
2208 PROC_UNLOCK(td->td_proc);
2211 td->td_ru.ru_inblock++;
2212 bp->b_iocmd = BIO_READ;
2213 bp->b_flags &= ~B_INVAL;
2214 if ((flags & GB_CKHASH) != 0) {
2215 bp->b_flags |= B_CKHASH;
2216 bp->b_ckhashcalc = ckhashfunc;
2218 if ((flags & GB_CVTENXIO) != 0)
2219 bp->b_xflags |= BX_CVTENXIO;
2220 bp->b_ioflags &= ~BIO_ERROR;
2221 if (bp->b_rcred == NOCRED && cred != NOCRED)
2222 bp->b_rcred = crhold(cred);
2223 vfs_busy_pages(bp, 0);
2224 bp->b_iooffset = dbtob(bp->b_blkno);
2230 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2232 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2246 * Write, release buffer on completion. (Done by iodone
2247 * if async). Do not bother writing anything if the buffer
2250 * Note that we set B_CACHE here, indicating that buffer is
2251 * fully valid and thus cacheable. This is true even of NFS
2252 * now so we set it generally. This could be set either here
2253 * or in biodone() since the I/O is synchronous. We put it
2257 bufwrite(struct buf *bp)
2264 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2265 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2266 bp->b_flags |= B_INVAL | B_RELBUF;
2267 bp->b_flags &= ~B_CACHE;
2271 if (bp->b_flags & B_INVAL) {
2276 if (bp->b_flags & B_BARRIER)
2277 atomic_add_long(&barrierwrites, 1);
2279 oldflags = bp->b_flags;
2281 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2282 ("FFS background buffer should not get here %p", bp));
2286 vp_md = vp->v_vflag & VV_MD;
2291 * Mark the buffer clean. Increment the bufobj write count
2292 * before bundirty() call, to prevent other thread from seeing
2293 * empty dirty list and zero counter for writes in progress,
2294 * falsely indicating that the bufobj is clean.
2296 bufobj_wref(bp->b_bufobj);
2299 bp->b_flags &= ~B_DONE;
2300 bp->b_ioflags &= ~BIO_ERROR;
2301 bp->b_flags |= B_CACHE;
2302 bp->b_iocmd = BIO_WRITE;
2304 vfs_busy_pages(bp, 1);
2307 * Normal bwrites pipeline writes
2309 bp->b_runningbufspace = bp->b_bufsize;
2310 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2315 racct_add_buf(curproc, bp, 1);
2316 PROC_UNLOCK(curproc);
2319 curthread->td_ru.ru_oublock++;
2320 if (oldflags & B_ASYNC)
2322 bp->b_iooffset = dbtob(bp->b_blkno);
2323 buf_track(bp, __func__);
2326 if ((oldflags & B_ASYNC) == 0) {
2327 int rtval = bufwait(bp);
2330 } else if (space > hirunningspace) {
2332 * don't allow the async write to saturate the I/O
2333 * system. We will not deadlock here because
2334 * we are blocking waiting for I/O that is already in-progress
2335 * to complete. We do not block here if it is the update
2336 * or syncer daemon trying to clean up as that can lead
2339 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2340 waitrunningbufspace();
2347 bufbdflush(struct bufobj *bo, struct buf *bp)
2350 struct bufdomain *bd;
2352 bd = &bdomain[bo->bo_domain];
2353 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2354 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2356 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2359 * Try to find a buffer to flush.
2361 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2362 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2364 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2367 panic("bdwrite: found ourselves");
2369 /* Don't countdeps with the bo lock held. */
2370 if (buf_countdeps(nbp, 0)) {
2375 if (nbp->b_flags & B_CLUSTEROK) {
2376 vfs_bio_awrite(nbp);
2381 dirtybufferflushes++;
2390 * Delayed write. (Buffer is marked dirty). Do not bother writing
2391 * anything if the buffer is marked invalid.
2393 * Note that since the buffer must be completely valid, we can safely
2394 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2395 * biodone() in order to prevent getblk from writing the buffer
2396 * out synchronously.
2399 bdwrite(struct buf *bp)
2401 struct thread *td = curthread;
2405 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2406 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2407 KASSERT((bp->b_flags & B_BARRIER) == 0,
2408 ("Barrier request in delayed write %p", bp));
2410 if (bp->b_flags & B_INVAL) {
2416 * If we have too many dirty buffers, don't create any more.
2417 * If we are wildly over our limit, then force a complete
2418 * cleanup. Otherwise, just keep the situation from getting
2419 * out of control. Note that we have to avoid a recursive
2420 * disaster and not try to clean up after our own cleanup!
2424 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2425 td->td_pflags |= TDP_INBDFLUSH;
2427 td->td_pflags &= ~TDP_INBDFLUSH;
2433 * Set B_CACHE, indicating that the buffer is fully valid. This is
2434 * true even of NFS now.
2436 bp->b_flags |= B_CACHE;
2439 * This bmap keeps the system from needing to do the bmap later,
2440 * perhaps when the system is attempting to do a sync. Since it
2441 * is likely that the indirect block -- or whatever other datastructure
2442 * that the filesystem needs is still in memory now, it is a good
2443 * thing to do this. Note also, that if the pageout daemon is
2444 * requesting a sync -- there might not be enough memory to do
2445 * the bmap then... So, this is important to do.
2447 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2448 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2451 buf_track(bp, __func__);
2454 * Set the *dirty* buffer range based upon the VM system dirty
2457 * Mark the buffer pages as clean. We need to do this here to
2458 * satisfy the vnode_pager and the pageout daemon, so that it
2459 * thinks that the pages have been "cleaned". Note that since
2460 * the pages are in a delayed write buffer -- the VFS layer
2461 * "will" see that the pages get written out on the next sync,
2462 * or perhaps the cluster will be completed.
2464 vfs_clean_pages_dirty_buf(bp);
2468 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2469 * due to the softdep code.
2476 * Turn buffer into delayed write request. We must clear BIO_READ and
2477 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2478 * itself to properly update it in the dirty/clean lists. We mark it
2479 * B_DONE to ensure that any asynchronization of the buffer properly
2480 * clears B_DONE ( else a panic will occur later ).
2482 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2483 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2484 * should only be called if the buffer is known-good.
2486 * Since the buffer is not on a queue, we do not update the numfreebuffers
2489 * The buffer must be on QUEUE_NONE.
2492 bdirty(struct buf *bp)
2495 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2496 bp, bp->b_vp, bp->b_flags);
2497 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2498 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2499 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2500 bp->b_flags &= ~(B_RELBUF);
2501 bp->b_iocmd = BIO_WRITE;
2503 if ((bp->b_flags & B_DELWRI) == 0) {
2504 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2513 * Clear B_DELWRI for buffer.
2515 * Since the buffer is not on a queue, we do not update the numfreebuffers
2518 * The buffer must be on QUEUE_NONE.
2522 bundirty(struct buf *bp)
2525 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2526 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2527 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2528 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2530 if (bp->b_flags & B_DELWRI) {
2531 bp->b_flags &= ~B_DELWRI;
2536 * Since it is now being written, we can clear its deferred write flag.
2538 bp->b_flags &= ~B_DEFERRED;
2544 * Asynchronous write. Start output on a buffer, but do not wait for
2545 * it to complete. The buffer is released when the output completes.
2547 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2548 * B_INVAL buffers. Not us.
2551 bawrite(struct buf *bp)
2554 bp->b_flags |= B_ASYNC;
2561 * Asynchronous barrier write. Start output on a buffer, but do not
2562 * wait for it to complete. Place a write barrier after this write so
2563 * that this buffer and all buffers written before it are committed to
2564 * the disk before any buffers written after this write are committed
2565 * to the disk. The buffer is released when the output completes.
2568 babarrierwrite(struct buf *bp)
2571 bp->b_flags |= B_ASYNC | B_BARRIER;
2578 * Synchronous barrier write. Start output on a buffer and wait for
2579 * it to complete. Place a write barrier after this write so that
2580 * this buffer and all buffers written before it are committed to
2581 * the disk before any buffers written after this write are committed
2582 * to the disk. The buffer is released when the output completes.
2585 bbarrierwrite(struct buf *bp)
2588 bp->b_flags |= B_BARRIER;
2589 return (bwrite(bp));
2595 * Called prior to the locking of any vnodes when we are expecting to
2596 * write. We do not want to starve the buffer cache with too many
2597 * dirty buffers so we block here. By blocking prior to the locking
2598 * of any vnodes we attempt to avoid the situation where a locked vnode
2599 * prevents the various system daemons from flushing related buffers.
2605 if (buf_dirty_count_severe()) {
2606 mtx_lock(&bdirtylock);
2607 while (buf_dirty_count_severe()) {
2609 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2612 mtx_unlock(&bdirtylock);
2617 * Return true if we have too many dirty buffers.
2620 buf_dirty_count_severe(void)
2623 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2629 * Release a busy buffer and, if requested, free its resources. The
2630 * buffer will be stashed in the appropriate bufqueue[] allowing it
2631 * to be accessed later as a cache entity or reused for other purposes.
2634 brelse(struct buf *bp)
2636 struct mount *v_mnt;
2640 * Many functions erroneously call brelse with a NULL bp under rare
2641 * error conditions. Simply return when called with a NULL bp.
2645 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2646 bp, bp->b_vp, bp->b_flags);
2647 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2648 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2649 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2650 ("brelse: non-VMIO buffer marked NOREUSE"));
2652 if (BUF_LOCKRECURSED(bp)) {
2654 * Do not process, in particular, do not handle the
2655 * B_INVAL/B_RELBUF and do not release to free list.
2661 if (bp->b_flags & B_MANAGED) {
2666 if (LIST_EMPTY(&bp->b_dep)) {
2667 bp->b_flags &= ~B_IOSTARTED;
2669 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2670 ("brelse: SU io not finished bp %p", bp));
2673 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2674 BO_LOCK(bp->b_bufobj);
2675 bp->b_vflags &= ~BV_BKGRDERR;
2676 BO_UNLOCK(bp->b_bufobj);
2680 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2681 (bp->b_flags & B_INVALONERR)) {
2683 * Forced invalidation of dirty buffer contents, to be used
2684 * after a failed write in the rare case that the loss of the
2685 * contents is acceptable. The buffer is invalidated and
2688 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2689 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2692 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2693 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2694 !(bp->b_flags & B_INVAL)) {
2696 * Failed write, redirty. All errors except ENXIO (which
2697 * means the device is gone) are treated as being
2700 * XXX Treating EIO as transient is not correct; the
2701 * contract with the local storage device drivers is that
2702 * they will only return EIO once the I/O is no longer
2703 * retriable. Network I/O also respects this through the
2704 * guarantees of TCP and/or the internal retries of NFS.
2705 * ENOMEM might be transient, but we also have no way of
2706 * knowing when its ok to retry/reschedule. In general,
2707 * this entire case should be made obsolete through better
2708 * error handling/recovery and resource scheduling.
2710 * Do this also for buffers that failed with ENXIO, but have
2711 * non-empty dependencies - the soft updates code might need
2712 * to access the buffer to untangle them.
2714 * Must clear BIO_ERROR to prevent pages from being scrapped.
2716 bp->b_ioflags &= ~BIO_ERROR;
2718 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2719 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2721 * Either a failed read I/O, or we were asked to free or not
2722 * cache the buffer, or we failed to write to a device that's
2723 * no longer present.
2725 bp->b_flags |= B_INVAL;
2726 if (!LIST_EMPTY(&bp->b_dep))
2728 if (bp->b_flags & B_DELWRI)
2730 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2731 if ((bp->b_flags & B_VMIO) == 0) {
2739 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2740 * is called with B_DELWRI set, the underlying pages may wind up
2741 * getting freed causing a previous write (bdwrite()) to get 'lost'
2742 * because pages associated with a B_DELWRI bp are marked clean.
2744 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2745 * if B_DELWRI is set.
2747 if (bp->b_flags & B_DELWRI)
2748 bp->b_flags &= ~B_RELBUF;
2751 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2752 * constituted, not even NFS buffers now. Two flags effect this. If
2753 * B_INVAL, the struct buf is invalidated but the VM object is kept
2754 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2756 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2757 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2758 * buffer is also B_INVAL because it hits the re-dirtying code above.
2760 * Normally we can do this whether a buffer is B_DELWRI or not. If
2761 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2762 * the commit state and we cannot afford to lose the buffer. If the
2763 * buffer has a background write in progress, we need to keep it
2764 * around to prevent it from being reconstituted and starting a second
2768 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2770 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2771 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2772 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2773 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2774 vfs_vmio_invalidate(bp);
2778 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2779 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2781 bp->b_flags &= ~B_NOREUSE;
2782 if (bp->b_vp != NULL)
2787 * If the buffer has junk contents signal it and eventually
2788 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2791 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2792 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2793 bp->b_flags |= B_INVAL;
2794 if (bp->b_flags & B_INVAL) {
2795 if (bp->b_flags & B_DELWRI)
2801 buf_track(bp, __func__);
2803 /* buffers with no memory */
2804 if (bp->b_bufsize == 0) {
2808 /* buffers with junk contents */
2809 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2810 (bp->b_ioflags & BIO_ERROR)) {
2811 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2812 if (bp->b_vflags & BV_BKGRDINPROG)
2813 panic("losing buffer 2");
2814 qindex = QUEUE_CLEAN;
2815 bp->b_flags |= B_AGE;
2816 /* remaining buffers */
2817 } else if (bp->b_flags & B_DELWRI)
2818 qindex = QUEUE_DIRTY;
2820 qindex = QUEUE_CLEAN;
2822 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2823 panic("brelse: not dirty");
2825 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2826 bp->b_xflags &= ~(BX_CVTENXIO);
2827 /* binsfree unlocks bp. */
2828 binsfree(bp, qindex);
2832 * Release a buffer back to the appropriate queue but do not try to free
2833 * it. The buffer is expected to be used again soon.
2835 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2836 * biodone() to requeue an async I/O on completion. It is also used when
2837 * known good buffers need to be requeued but we think we may need the data
2840 * XXX we should be able to leave the B_RELBUF hint set on completion.
2843 bqrelse(struct buf *bp)
2847 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2848 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2849 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2851 qindex = QUEUE_NONE;
2852 if (BUF_LOCKRECURSED(bp)) {
2853 /* do not release to free list */
2857 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2858 bp->b_xflags &= ~(BX_CVTENXIO);
2860 if (LIST_EMPTY(&bp->b_dep)) {
2861 bp->b_flags &= ~B_IOSTARTED;
2863 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2864 ("bqrelse: SU io not finished bp %p", bp));
2867 if (bp->b_flags & B_MANAGED) {
2868 if (bp->b_flags & B_REMFREE)
2873 /* buffers with stale but valid contents */
2874 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2875 BV_BKGRDERR)) == BV_BKGRDERR) {
2876 BO_LOCK(bp->b_bufobj);
2877 bp->b_vflags &= ~BV_BKGRDERR;
2878 BO_UNLOCK(bp->b_bufobj);
2879 qindex = QUEUE_DIRTY;
2881 if ((bp->b_flags & B_DELWRI) == 0 &&
2882 (bp->b_xflags & BX_VNDIRTY))
2883 panic("bqrelse: not dirty");
2884 if ((bp->b_flags & B_NOREUSE) != 0) {
2888 qindex = QUEUE_CLEAN;
2890 buf_track(bp, __func__);
2891 /* binsfree unlocks bp. */
2892 binsfree(bp, qindex);
2896 buf_track(bp, __func__);
2902 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2903 * restore bogus pages.
2906 vfs_vmio_iodone(struct buf *bp)
2911 struct vnode *vp __unused;
2912 int i, iosize, resid;
2915 obj = bp->b_bufobj->bo_object;
2916 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2917 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2918 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2921 VNPASS(vp->v_holdcnt > 0, vp);
2922 VNPASS(vp->v_object != NULL, vp);
2924 foff = bp->b_offset;
2925 KASSERT(bp->b_offset != NOOFFSET,
2926 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2929 iosize = bp->b_bcount - bp->b_resid;
2930 for (i = 0; i < bp->b_npages; i++) {
2931 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2936 * cleanup bogus pages, restoring the originals
2939 if (m == bogus_page) {
2941 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2943 panic("biodone: page disappeared!");
2945 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2947 * In the write case, the valid and clean bits are
2948 * already changed correctly ( see bdwrite() ), so we
2949 * only need to do this here in the read case.
2951 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2952 resid)) == 0, ("vfs_vmio_iodone: page %p "
2953 "has unexpected dirty bits", m));
2954 vfs_page_set_valid(bp, foff, m);
2956 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2957 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2958 (intmax_t)foff, (uintmax_t)m->pindex));
2961 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2964 vm_object_pip_wakeupn(obj, bp->b_npages);
2965 if (bogus && buf_mapped(bp)) {
2966 BUF_CHECK_MAPPED(bp);
2967 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2968 bp->b_pages, bp->b_npages);
2973 * Perform page invalidation when a buffer is released. The fully invalid
2974 * pages will be reclaimed later in vfs_vmio_truncate().
2977 vfs_vmio_invalidate(struct buf *bp)
2981 int flags, i, resid, poffset, presid;
2983 if (buf_mapped(bp)) {
2984 BUF_CHECK_MAPPED(bp);
2985 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2987 BUF_CHECK_UNMAPPED(bp);
2989 * Get the base offset and length of the buffer. Note that
2990 * in the VMIO case if the buffer block size is not
2991 * page-aligned then b_data pointer may not be page-aligned.
2992 * But our b_pages[] array *IS* page aligned.
2994 * block sizes less then DEV_BSIZE (usually 512) are not
2995 * supported due to the page granularity bits (m->valid,
2996 * m->dirty, etc...).
2998 * See man buf(9) for more information
3000 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3001 obj = bp->b_bufobj->bo_object;
3002 resid = bp->b_bufsize;
3003 poffset = bp->b_offset & PAGE_MASK;
3004 VM_OBJECT_WLOCK(obj);
3005 for (i = 0; i < bp->b_npages; i++) {
3007 if (m == bogus_page)
3008 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3009 bp->b_pages[i] = NULL;
3011 presid = resid > (PAGE_SIZE - poffset) ?
3012 (PAGE_SIZE - poffset) : resid;
3013 KASSERT(presid >= 0, ("brelse: extra page"));
3014 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3015 if (pmap_page_wired_mappings(m) == 0)
3016 vm_page_set_invalid(m, poffset, presid);
3018 vm_page_release_locked(m, flags);
3022 VM_OBJECT_WUNLOCK(obj);
3027 * Page-granular truncation of an existing VMIO buffer.
3030 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3036 if (bp->b_npages == desiredpages)
3039 if (buf_mapped(bp)) {
3040 BUF_CHECK_MAPPED(bp);
3041 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3042 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3044 BUF_CHECK_UNMAPPED(bp);
3047 * The object lock is needed only if we will attempt to free pages.
3049 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3050 if ((bp->b_flags & B_DIRECT) != 0) {
3051 flags |= VPR_TRYFREE;
3052 obj = bp->b_bufobj->bo_object;
3053 VM_OBJECT_WLOCK(obj);
3057 for (i = desiredpages; i < bp->b_npages; i++) {
3059 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3060 bp->b_pages[i] = NULL;
3062 vm_page_release_locked(m, flags);
3064 vm_page_release(m, flags);
3067 VM_OBJECT_WUNLOCK(obj);
3068 bp->b_npages = desiredpages;
3072 * Byte granular extension of VMIO buffers.
3075 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3078 * We are growing the buffer, possibly in a
3079 * byte-granular fashion.
3087 * Step 1, bring in the VM pages from the object, allocating
3088 * them if necessary. We must clear B_CACHE if these pages
3089 * are not valid for the range covered by the buffer.
3091 obj = bp->b_bufobj->bo_object;
3092 if (bp->b_npages < desiredpages) {
3093 KASSERT(desiredpages <= atop(maxbcachebuf),
3094 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3095 bp, desiredpages, maxbcachebuf));
3098 * We must allocate system pages since blocking
3099 * here could interfere with paging I/O, no
3100 * matter which process we are.
3102 * Only exclusive busy can be tested here.
3103 * Blocking on shared busy might lead to
3104 * deadlocks once allocbuf() is called after
3105 * pages are vfs_busy_pages().
3107 (void)vm_page_grab_pages_unlocked(obj,
3108 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3109 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3110 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3111 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3112 bp->b_npages = desiredpages;
3116 * Step 2. We've loaded the pages into the buffer,
3117 * we have to figure out if we can still have B_CACHE
3118 * set. Note that B_CACHE is set according to the
3119 * byte-granular range ( bcount and size ), not the
3120 * aligned range ( newbsize ).
3122 * The VM test is against m->valid, which is DEV_BSIZE
3123 * aligned. Needless to say, the validity of the data
3124 * needs to also be DEV_BSIZE aligned. Note that this
3125 * fails with NFS if the server or some other client
3126 * extends the file's EOF. If our buffer is resized,
3127 * B_CACHE may remain set! XXX
3129 toff = bp->b_bcount;
3130 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3131 while ((bp->b_flags & B_CACHE) && toff < size) {
3134 if (tinc > (size - toff))
3136 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3137 m = bp->b_pages[pi];
3138 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3144 * Step 3, fixup the KVA pmap.
3149 BUF_CHECK_UNMAPPED(bp);
3153 * Check to see if a block at a particular lbn is available for a clustered
3157 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3164 /* If the buf isn't in core skip it */
3165 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3168 /* If the buf is busy we don't want to wait for it */
3169 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3172 /* Only cluster with valid clusterable delayed write buffers */
3173 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3174 (B_DELWRI | B_CLUSTEROK))
3177 if (bpa->b_bufsize != size)
3181 * Check to see if it is in the expected place on disk and that the
3182 * block has been mapped.
3184 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3194 * Implement clustered async writes for clearing out B_DELWRI buffers.
3195 * This is much better then the old way of writing only one buffer at
3196 * a time. Note that we may not be presented with the buffers in the
3197 * correct order, so we search for the cluster in both directions.
3200 vfs_bio_awrite(struct buf *bp)
3205 daddr_t lblkno = bp->b_lblkno;
3206 struct vnode *vp = bp->b_vp;
3214 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3216 * right now we support clustered writing only to regular files. If
3217 * we find a clusterable block we could be in the middle of a cluster
3218 * rather then at the beginning.
3220 if ((vp->v_type == VREG) &&
3221 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3222 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3223 size = vp->v_mount->mnt_stat.f_iosize;
3224 maxcl = maxphys / size;
3227 for (i = 1; i < maxcl; i++)
3228 if (vfs_bio_clcheck(vp, size, lblkno + i,
3229 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3232 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3233 if (vfs_bio_clcheck(vp, size, lblkno - j,
3234 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3240 * this is a possible cluster write
3244 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3250 bp->b_flags |= B_ASYNC;
3252 * default (old) behavior, writing out only one block
3254 * XXX returns b_bufsize instead of b_bcount for nwritten?
3256 nwritten = bp->b_bufsize;
3265 * Allocate KVA for an empty buf header according to gbflags.
3268 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3271 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3273 * In order to keep fragmentation sane we only allocate kva
3274 * in BKVASIZE chunks. XXX with vmem we can do page size.
3276 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3278 if (maxsize != bp->b_kvasize &&
3279 bufkva_alloc(bp, maxsize, gbflags))
3288 * Find and initialize a new buffer header, freeing up existing buffers
3289 * in the bufqueues as necessary. The new buffer is returned locked.
3292 * We have insufficient buffer headers
3293 * We have insufficient buffer space
3294 * buffer_arena is too fragmented ( space reservation fails )
3295 * If we have to flush dirty buffers ( but we try to avoid this )
3297 * The caller is responsible for releasing the reserved bufspace after
3298 * allocbuf() is called.
3301 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3303 struct bufdomain *bd;
3305 bool metadata, reserved;
3308 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3309 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3310 if (!unmapped_buf_allowed)
3311 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3313 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3321 bd = &bdomain[vp->v_bufobj.bo_domain];
3323 counter_u64_add(getnewbufcalls, 1);
3326 if (reserved == false &&
3327 bufspace_reserve(bd, maxsize, metadata) != 0) {
3328 counter_u64_add(getnewbufrestarts, 1);
3332 if ((bp = buf_alloc(bd)) == NULL) {
3333 counter_u64_add(getnewbufrestarts, 1);
3336 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3339 } while (buf_recycle(bd, false) == 0);
3342 bufspace_release(bd, maxsize);
3344 bp->b_flags |= B_INVAL;
3347 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3355 * buffer flushing daemon. Buffers are normally flushed by the
3356 * update daemon but if it cannot keep up this process starts to
3357 * take the load in an attempt to prevent getnewbuf() from blocking.
3359 static struct kproc_desc buf_kp = {
3364 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3367 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3371 flushed = flushbufqueues(vp, bd, target, 0);
3374 * Could not find any buffers without rollback
3375 * dependencies, so just write the first one
3376 * in the hopes of eventually making progress.
3378 if (vp != NULL && target > 2)
3380 flushbufqueues(vp, bd, target, 1);
3388 struct bufdomain *bd;
3394 * This process needs to be suspended prior to shutdown sync.
3396 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3397 SHUTDOWN_PRI_LAST + 100);
3400 * Start the buf clean daemons as children threads.
3402 for (i = 0 ; i < buf_domains; i++) {
3405 error = kthread_add((void (*)(void *))bufspace_daemon,
3406 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3408 panic("error %d spawning bufspace daemon", error);
3412 * This process is allowed to take the buffer cache to the limit
3414 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3418 mtx_unlock(&bdlock);
3420 kthread_suspend_check();
3423 * Save speedupreq for this pass and reset to capture new
3426 speedupreq = bd_speedupreq;
3430 * Flush each domain sequentially according to its level and
3431 * the speedup request.
3433 for (i = 0; i < buf_domains; i++) {
3436 lodirty = bd->bd_numdirtybuffers / 2;
3438 lodirty = bd->bd_lodirtybuffers;
3439 while (bd->bd_numdirtybuffers > lodirty) {
3440 if (buf_flush(NULL, bd,
3441 bd->bd_numdirtybuffers - lodirty) == 0)
3443 kern_yield(PRI_USER);
3448 * Only clear bd_request if we have reached our low water
3449 * mark. The buf_daemon normally waits 1 second and
3450 * then incrementally flushes any dirty buffers that have
3451 * built up, within reason.
3453 * If we were unable to hit our low water mark and couldn't
3454 * find any flushable buffers, we sleep for a short period
3455 * to avoid endless loops on unlockable buffers.
3458 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3460 * We reached our low water mark, reset the
3461 * request and sleep until we are needed again.
3462 * The sleep is just so the suspend code works.
3466 * Do an extra wakeup in case dirty threshold
3467 * changed via sysctl and the explicit transition
3468 * out of shortfall was missed.
3471 if (runningbufspace <= lorunningspace)
3473 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3476 * We couldn't find any flushable dirty buffers but
3477 * still have too many dirty buffers, we
3478 * have to sleep and try again. (rare)
3480 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3488 * Try to flush a buffer in the dirty queue. We must be careful to
3489 * free up B_INVAL buffers instead of write them, which NFS is
3490 * particularly sensitive to.
3492 static int flushwithdeps = 0;
3493 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3495 "Number of buffers flushed with dependencies that require rollbacks");
3498 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3501 struct bufqueue *bq;
3502 struct buf *sentinel;
3512 bq = &bd->bd_dirtyq;
3514 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3515 sentinel->b_qindex = QUEUE_SENTINEL;
3517 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3519 while (flushed != target) {
3522 bp = TAILQ_NEXT(sentinel, b_freelist);
3524 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3525 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3532 * Skip sentinels inserted by other invocations of the
3533 * flushbufqueues(), taking care to not reorder them.
3535 * Only flush the buffers that belong to the
3536 * vnode locked by the curthread.
3538 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3543 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3549 * BKGRDINPROG can only be set with the buf and bufobj
3550 * locks both held. We tolerate a race to clear it here.
3552 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3553 (bp->b_flags & B_DELWRI) == 0) {
3557 if (bp->b_flags & B_INVAL) {
3564 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3565 if (flushdeps == 0) {
3573 * We must hold the lock on a vnode before writing
3574 * one of its buffers. Otherwise we may confuse, or
3575 * in the case of a snapshot vnode, deadlock the
3578 * The lock order here is the reverse of the normal
3579 * of vnode followed by buf lock. This is ok because
3580 * the NOWAIT will prevent deadlock.
3583 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3589 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3591 ASSERT_VOP_LOCKED(vp, "getbuf");
3593 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3594 vn_lock(vp, LK_TRYUPGRADE);
3597 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3598 bp, bp->b_vp, bp->b_flags);
3599 if (curproc == bufdaemonproc) {
3604 counter_u64_add(notbufdflushes, 1);
3606 vn_finished_write(mp);
3609 flushwithdeps += hasdeps;
3613 * Sleeping on runningbufspace while holding
3614 * vnode lock leads to deadlock.
3616 if (curproc == bufdaemonproc &&
3617 runningbufspace > hirunningspace)
3618 waitrunningbufspace();
3621 vn_finished_write(mp);
3625 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3627 free(sentinel, M_TEMP);
3632 * Check to see if a block is currently memory resident.
3635 incore(struct bufobj *bo, daddr_t blkno)
3637 return (gbincore_unlocked(bo, blkno));
3641 * Returns true if no I/O is needed to access the
3642 * associated VM object. This is like incore except
3643 * it also hunts around in the VM system for the data.
3646 inmem(struct vnode * vp, daddr_t blkno)
3649 vm_offset_t toff, tinc, size;
3654 ASSERT_VOP_LOCKED(vp, "inmem");
3656 if (incore(&vp->v_bufobj, blkno))
3658 if (vp->v_mount == NULL)
3665 if (size > vp->v_mount->mnt_stat.f_iosize)
3666 size = vp->v_mount->mnt_stat.f_iosize;
3667 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3669 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3670 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3676 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3677 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3679 * Consider page validity only if page mapping didn't change
3682 valid = vm_page_is_valid(m,
3683 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3684 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3696 * Set the dirty range for a buffer based on the status of the dirty
3697 * bits in the pages comprising the buffer. The range is limited
3698 * to the size of the buffer.
3700 * Tell the VM system that the pages associated with this buffer
3701 * are clean. This is used for delayed writes where the data is
3702 * going to go to disk eventually without additional VM intevention.
3704 * Note that while we only really need to clean through to b_bcount, we
3705 * just go ahead and clean through to b_bufsize.
3708 vfs_clean_pages_dirty_buf(struct buf *bp)
3710 vm_ooffset_t foff, noff, eoff;
3714 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3717 foff = bp->b_offset;
3718 KASSERT(bp->b_offset != NOOFFSET,
3719 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3721 vfs_busy_pages_acquire(bp);
3722 vfs_setdirty_range(bp);
3723 for (i = 0; i < bp->b_npages; i++) {
3724 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3726 if (eoff > bp->b_offset + bp->b_bufsize)
3727 eoff = bp->b_offset + bp->b_bufsize;
3729 vfs_page_set_validclean(bp, foff, m);
3730 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3733 vfs_busy_pages_release(bp);
3737 vfs_setdirty_range(struct buf *bp)
3739 vm_offset_t boffset;
3740 vm_offset_t eoffset;
3744 * test the pages to see if they have been modified directly
3745 * by users through the VM system.
3747 for (i = 0; i < bp->b_npages; i++)
3748 vm_page_test_dirty(bp->b_pages[i]);
3751 * Calculate the encompassing dirty range, boffset and eoffset,
3752 * (eoffset - boffset) bytes.
3755 for (i = 0; i < bp->b_npages; i++) {
3756 if (bp->b_pages[i]->dirty)
3759 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3761 for (i = bp->b_npages - 1; i >= 0; --i) {
3762 if (bp->b_pages[i]->dirty) {
3766 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3769 * Fit it to the buffer.
3772 if (eoffset > bp->b_bcount)
3773 eoffset = bp->b_bcount;
3776 * If we have a good dirty range, merge with the existing
3780 if (boffset < eoffset) {
3781 if (bp->b_dirtyoff > boffset)
3782 bp->b_dirtyoff = boffset;
3783 if (bp->b_dirtyend < eoffset)
3784 bp->b_dirtyend = eoffset;
3789 * Allocate the KVA mapping for an existing buffer.
3790 * If an unmapped buffer is provided but a mapped buffer is requested, take
3791 * also care to properly setup mappings between pages and KVA.
3794 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3796 int bsize, maxsize, need_mapping, need_kva;
3799 need_mapping = bp->b_data == unmapped_buf &&
3800 (gbflags & GB_UNMAPPED) == 0;
3801 need_kva = bp->b_kvabase == unmapped_buf &&
3802 bp->b_data == unmapped_buf &&
3803 (gbflags & GB_KVAALLOC) != 0;
3804 if (!need_mapping && !need_kva)
3807 BUF_CHECK_UNMAPPED(bp);
3809 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3811 * Buffer is not mapped, but the KVA was already
3812 * reserved at the time of the instantiation. Use the
3819 * Calculate the amount of the address space we would reserve
3820 * if the buffer was mapped.
3822 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3823 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3824 offset = blkno * bsize;
3825 maxsize = size + (offset & PAGE_MASK);
3826 maxsize = imax(maxsize, bsize);
3828 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3829 if ((gbflags & GB_NOWAIT_BD) != 0) {
3831 * XXXKIB: defragmentation cannot
3832 * succeed, not sure what else to do.
3834 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3836 counter_u64_add(mappingrestarts, 1);
3837 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3841 /* b_offset is handled by bpmap_qenter. */
3842 bp->b_data = bp->b_kvabase;
3843 BUF_CHECK_MAPPED(bp);
3849 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3855 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3864 * Get a block given a specified block and offset into a file/device.
3865 * The buffers B_DONE bit will be cleared on return, making it almost
3866 * ready for an I/O initiation. B_INVAL may or may not be set on
3867 * return. The caller should clear B_INVAL prior to initiating a
3870 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3871 * an existing buffer.
3873 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3874 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3875 * and then cleared based on the backing VM. If the previous buffer is
3876 * non-0-sized but invalid, B_CACHE will be cleared.
3878 * If getblk() must create a new buffer, the new buffer is returned with
3879 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3880 * case it is returned with B_INVAL clear and B_CACHE set based on the
3883 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3884 * B_CACHE bit is clear.
3886 * What this means, basically, is that the caller should use B_CACHE to
3887 * determine whether the buffer is fully valid or not and should clear
3888 * B_INVAL prior to issuing a read. If the caller intends to validate
3889 * the buffer by loading its data area with something, the caller needs
3890 * to clear B_INVAL. If the caller does this without issuing an I/O,
3891 * the caller should set B_CACHE ( as an optimization ), else the caller
3892 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3893 * a write attempt or if it was a successful read. If the caller
3894 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3895 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3897 * The blkno parameter is the logical block being requested. Normally
3898 * the mapping of logical block number to disk block address is done
3899 * by calling VOP_BMAP(). However, if the mapping is already known, the
3900 * disk block address can be passed using the dblkno parameter. If the
3901 * disk block address is not known, then the same value should be passed
3902 * for blkno and dblkno.
3905 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3906 int slptimeo, int flags, struct buf **bpp)
3911 int bsize, error, maxsize, vmio;
3914 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3915 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3916 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3917 if (vp->v_type != VCHR)
3918 ASSERT_VOP_LOCKED(vp, "getblk");
3919 if (size > maxbcachebuf)
3920 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3922 if (!unmapped_buf_allowed)
3923 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3928 /* Attempt lockless lookup first. */
3929 bp = gbincore_unlocked(bo, blkno);
3931 goto newbuf_unlocked;
3933 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3938 /* Verify buf identify has not changed since lookup. */
3939 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3940 goto foundbuf_fastpath;
3942 /* It changed, fallback to locked lookup. */
3947 bp = gbincore(bo, blkno);
3952 * Buffer is in-core. If the buffer is not busy nor managed,
3953 * it must be on a queue.
3955 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3956 ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3958 error = BUF_TIMELOCK(bp, lockflags,
3959 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3962 * If we slept and got the lock we have to restart in case
3963 * the buffer changed identities.
3965 if (error == ENOLCK)
3967 /* We timed out or were interrupted. */
3968 else if (error != 0)
3972 /* If recursed, assume caller knows the rules. */
3973 if (BUF_LOCKRECURSED(bp))
3977 * The buffer is locked. B_CACHE is cleared if the buffer is
3978 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3979 * and for a VMIO buffer B_CACHE is adjusted according to the
3982 if (bp->b_flags & B_INVAL)
3983 bp->b_flags &= ~B_CACHE;
3984 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3985 bp->b_flags |= B_CACHE;
3986 if (bp->b_flags & B_MANAGED)
3987 MPASS(bp->b_qindex == QUEUE_NONE);
3992 * check for size inconsistencies for non-VMIO case.
3994 if (bp->b_bcount != size) {
3995 if ((bp->b_flags & B_VMIO) == 0 ||
3996 (size > bp->b_kvasize)) {
3997 if (bp->b_flags & B_DELWRI) {
3998 bp->b_flags |= B_NOCACHE;
4001 if (LIST_EMPTY(&bp->b_dep)) {
4002 bp->b_flags |= B_RELBUF;
4005 bp->b_flags |= B_NOCACHE;
4014 * Handle the case of unmapped buffer which should
4015 * become mapped, or the buffer for which KVA
4016 * reservation is requested.
4018 bp_unmapped_get_kva(bp, blkno, size, flags);
4021 * If the size is inconsistent in the VMIO case, we can resize
4022 * the buffer. This might lead to B_CACHE getting set or
4023 * cleared. If the size has not changed, B_CACHE remains
4024 * unchanged from its previous state.
4028 KASSERT(bp->b_offset != NOOFFSET,
4029 ("getblk: no buffer offset"));
4032 * A buffer with B_DELWRI set and B_CACHE clear must
4033 * be committed before we can return the buffer in
4034 * order to prevent the caller from issuing a read
4035 * ( due to B_CACHE not being set ) and overwriting
4038 * Most callers, including NFS and FFS, need this to
4039 * operate properly either because they assume they
4040 * can issue a read if B_CACHE is not set, or because
4041 * ( for example ) an uncached B_DELWRI might loop due
4042 * to softupdates re-dirtying the buffer. In the latter
4043 * case, B_CACHE is set after the first write completes,
4044 * preventing further loops.
4045 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4046 * above while extending the buffer, we cannot allow the
4047 * buffer to remain with B_CACHE set after the write
4048 * completes or it will represent a corrupt state. To
4049 * deal with this we set B_NOCACHE to scrap the buffer
4052 * We might be able to do something fancy, like setting
4053 * B_CACHE in bwrite() except if B_DELWRI is already set,
4054 * so the below call doesn't set B_CACHE, but that gets real
4055 * confusing. This is much easier.
4058 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4059 bp->b_flags |= B_NOCACHE;
4063 bp->b_flags &= ~B_DONE;
4066 * Buffer is not in-core, create new buffer. The buffer
4067 * returned by getnewbuf() is locked. Note that the returned
4068 * buffer is also considered valid (not marked B_INVAL).
4073 * If the user does not want us to create the buffer, bail out
4076 if (flags & GB_NOCREAT)
4079 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4080 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4081 offset = blkno * bsize;
4082 vmio = vp->v_object != NULL;
4084 maxsize = size + (offset & PAGE_MASK);
4087 /* Do not allow non-VMIO notmapped buffers. */
4088 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4090 maxsize = imax(maxsize, bsize);
4091 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4093 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4094 KASSERT(error != EOPNOTSUPP,
4095 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4100 return (EJUSTRETURN);
4103 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4105 if (slpflag || slptimeo)
4108 * XXX This is here until the sleep path is diagnosed
4109 * enough to work under very low memory conditions.
4111 * There's an issue on low memory, 4BSD+non-preempt
4112 * systems (eg MIPS routers with 32MB RAM) where buffer
4113 * exhaustion occurs without sleeping for buffer
4114 * reclaimation. This just sticks in a loop and
4115 * constantly attempts to allocate a buffer, which
4116 * hits exhaustion and tries to wakeup bufdaemon.
4117 * This never happens because we never yield.
4119 * The real solution is to identify and fix these cases
4120 * so we aren't effectively busy-waiting in a loop
4121 * until the reclaimation path has cycles to run.
4123 kern_yield(PRI_USER);
4128 * This code is used to make sure that a buffer is not
4129 * created while the getnewbuf routine is blocked.
4130 * This can be a problem whether the vnode is locked or not.
4131 * If the buffer is created out from under us, we have to
4132 * throw away the one we just created.
4134 * Note: this must occur before we associate the buffer
4135 * with the vp especially considering limitations in
4136 * the splay tree implementation when dealing with duplicate
4140 if (gbincore(bo, blkno)) {
4142 bp->b_flags |= B_INVAL;
4143 bufspace_release(bufdomain(bp), maxsize);
4149 * Insert the buffer into the hash, so that it can
4150 * be found by incore.
4152 bp->b_lblkno = blkno;
4153 bp->b_blkno = d_blkno;
4154 bp->b_offset = offset;
4159 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4160 * buffer size starts out as 0, B_CACHE will be set by
4161 * allocbuf() for the VMIO case prior to it testing the
4162 * backing store for validity.
4166 bp->b_flags |= B_VMIO;
4167 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4168 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4169 bp, vp->v_object, bp->b_bufobj->bo_object));
4171 bp->b_flags &= ~B_VMIO;
4172 KASSERT(bp->b_bufobj->bo_object == NULL,
4173 ("ARGH! has b_bufobj->bo_object %p %p\n",
4174 bp, bp->b_bufobj->bo_object));
4175 BUF_CHECK_MAPPED(bp);
4179 bufspace_release(bufdomain(bp), maxsize);
4180 bp->b_flags &= ~B_DONE;
4182 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4184 buf_track(bp, __func__);
4185 KASSERT(bp->b_bufobj == bo,
4186 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4192 * Get an empty, disassociated buffer of given size. The buffer is initially
4196 geteblk(int size, int flags)
4201 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4202 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4203 if ((flags & GB_NOWAIT_BD) &&
4204 (curthread->td_pflags & TDP_BUFNEED) != 0)
4208 bufspace_release(bufdomain(bp), maxsize);
4209 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4214 * Truncate the backing store for a non-vmio buffer.
4217 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4220 if (bp->b_flags & B_MALLOC) {
4222 * malloced buffers are not shrunk
4224 if (newbsize == 0) {
4225 bufmallocadjust(bp, 0);
4226 free(bp->b_data, M_BIOBUF);
4227 bp->b_data = bp->b_kvabase;
4228 bp->b_flags &= ~B_MALLOC;
4232 vm_hold_free_pages(bp, newbsize);
4233 bufspace_adjust(bp, newbsize);
4237 * Extend the backing for a non-VMIO buffer.
4240 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4246 * We only use malloced memory on the first allocation.
4247 * and revert to page-allocated memory when the buffer
4250 * There is a potential smp race here that could lead
4251 * to bufmallocspace slightly passing the max. It
4252 * is probably extremely rare and not worth worrying
4255 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4256 bufmallocspace < maxbufmallocspace) {
4257 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4258 bp->b_flags |= B_MALLOC;
4259 bufmallocadjust(bp, newbsize);
4264 * If the buffer is growing on its other-than-first
4265 * allocation then we revert to the page-allocation
4270 if (bp->b_flags & B_MALLOC) {
4271 origbuf = bp->b_data;
4272 origbufsize = bp->b_bufsize;
4273 bp->b_data = bp->b_kvabase;
4274 bufmallocadjust(bp, 0);
4275 bp->b_flags &= ~B_MALLOC;
4276 newbsize = round_page(newbsize);
4278 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4279 (vm_offset_t) bp->b_data + newbsize);
4280 if (origbuf != NULL) {
4281 bcopy(origbuf, bp->b_data, origbufsize);
4282 free(origbuf, M_BIOBUF);
4284 bufspace_adjust(bp, newbsize);
4288 * This code constitutes the buffer memory from either anonymous system
4289 * memory (in the case of non-VMIO operations) or from an associated
4290 * VM object (in the case of VMIO operations). This code is able to
4291 * resize a buffer up or down.
4293 * Note that this code is tricky, and has many complications to resolve
4294 * deadlock or inconsistent data situations. Tread lightly!!!
4295 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4296 * the caller. Calling this code willy nilly can result in the loss of data.
4298 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4299 * B_CACHE for the non-VMIO case.
4302 allocbuf(struct buf *bp, int size)
4306 if (bp->b_bcount == size)
4309 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4310 panic("allocbuf: buffer too small");
4312 newbsize = roundup2(size, DEV_BSIZE);
4313 if ((bp->b_flags & B_VMIO) == 0) {
4314 if ((bp->b_flags & B_MALLOC) == 0)
4315 newbsize = round_page(newbsize);
4317 * Just get anonymous memory from the kernel. Don't
4318 * mess with B_CACHE.
4320 if (newbsize < bp->b_bufsize)
4321 vfs_nonvmio_truncate(bp, newbsize);
4322 else if (newbsize > bp->b_bufsize)
4323 vfs_nonvmio_extend(bp, newbsize);
4327 desiredpages = (size == 0) ? 0 :
4328 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4330 if (bp->b_flags & B_MALLOC)
4331 panic("allocbuf: VMIO buffer can't be malloced");
4333 * Set B_CACHE initially if buffer is 0 length or will become
4336 if (size == 0 || bp->b_bufsize == 0)
4337 bp->b_flags |= B_CACHE;
4339 if (newbsize < bp->b_bufsize)
4340 vfs_vmio_truncate(bp, desiredpages);
4341 /* XXX This looks as if it should be newbsize > b_bufsize */
4342 else if (size > bp->b_bcount)
4343 vfs_vmio_extend(bp, desiredpages, size);
4344 bufspace_adjust(bp, newbsize);
4346 bp->b_bcount = size; /* requested buffer size. */
4350 extern int inflight_transient_maps;
4352 static struct bio_queue nondump_bios;
4355 biodone(struct bio *bp)
4358 void (*done)(struct bio *);
4359 vm_offset_t start, end;
4361 biotrack(bp, __func__);
4364 * Avoid completing I/O when dumping after a panic since that may
4365 * result in a deadlock in the filesystem or pager code. Note that
4366 * this doesn't affect dumps that were started manually since we aim
4367 * to keep the system usable after it has been resumed.
4369 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4370 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4373 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4374 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4375 bp->bio_flags |= BIO_UNMAPPED;
4376 start = trunc_page((vm_offset_t)bp->bio_data);
4377 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4378 bp->bio_data = unmapped_buf;
4379 pmap_qremove(start, atop(end - start));
4380 vmem_free(transient_arena, start, end - start);
4381 atomic_add_int(&inflight_transient_maps, -1);
4383 done = bp->bio_done;
4385 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4387 bp->bio_flags |= BIO_DONE;
4395 * Wait for a BIO to finish.
4398 biowait(struct bio *bp, const char *wmesg)
4402 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4404 while ((bp->bio_flags & BIO_DONE) == 0)
4405 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4407 if (bp->bio_error != 0)
4408 return (bp->bio_error);
4409 if (!(bp->bio_flags & BIO_ERROR))
4415 biofinish(struct bio *bp, struct devstat *stat, int error)
4419 bp->bio_error = error;
4420 bp->bio_flags |= BIO_ERROR;
4423 devstat_end_transaction_bio(stat, bp);
4427 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4429 biotrack_buf(struct bio *bp, const char *location)
4432 buf_track(bp->bio_track_bp, location);
4439 * Wait for buffer I/O completion, returning error status. The buffer
4440 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4441 * error and cleared.
4444 bufwait(struct buf *bp)
4446 if (bp->b_iocmd == BIO_READ)
4447 bwait(bp, PRIBIO, "biord");
4449 bwait(bp, PRIBIO, "biowr");
4450 if (bp->b_flags & B_EINTR) {
4451 bp->b_flags &= ~B_EINTR;
4454 if (bp->b_ioflags & BIO_ERROR) {
4455 return (bp->b_error ? bp->b_error : EIO);
4464 * Finish I/O on a buffer, optionally calling a completion function.
4465 * This is usually called from an interrupt so process blocking is
4468 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4469 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4470 * assuming B_INVAL is clear.
4472 * For the VMIO case, we set B_CACHE if the op was a read and no
4473 * read error occurred, or if the op was a write. B_CACHE is never
4474 * set if the buffer is invalid or otherwise uncacheable.
4476 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4477 * initiator to leave B_INVAL set to brelse the buffer out of existence
4478 * in the biodone routine.
4481 bufdone(struct buf *bp)
4483 struct bufobj *dropobj;
4484 void (*biodone)(struct buf *);
4486 buf_track(bp, __func__);
4487 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4490 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4492 runningbufwakeup(bp);
4493 if (bp->b_iocmd == BIO_WRITE)
4494 dropobj = bp->b_bufobj;
4495 /* call optional completion function if requested */
4496 if (bp->b_iodone != NULL) {
4497 biodone = bp->b_iodone;
4498 bp->b_iodone = NULL;
4501 bufobj_wdrop(dropobj);
4504 if (bp->b_flags & B_VMIO) {
4506 * Set B_CACHE if the op was a normal read and no error
4507 * occurred. B_CACHE is set for writes in the b*write()
4510 if (bp->b_iocmd == BIO_READ &&
4511 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4512 !(bp->b_ioflags & BIO_ERROR))
4513 bp->b_flags |= B_CACHE;
4514 vfs_vmio_iodone(bp);
4516 if (!LIST_EMPTY(&bp->b_dep))
4518 if ((bp->b_flags & B_CKHASH) != 0) {
4519 KASSERT(bp->b_iocmd == BIO_READ,
4520 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4521 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4522 (*bp->b_ckhashcalc)(bp);
4525 * For asynchronous completions, release the buffer now. The brelse
4526 * will do a wakeup there if necessary - so no need to do a wakeup
4527 * here in the async case. The sync case always needs to do a wakeup.
4529 if (bp->b_flags & B_ASYNC) {
4530 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4531 (bp->b_ioflags & BIO_ERROR))
4538 bufobj_wdrop(dropobj);
4542 * This routine is called in lieu of iodone in the case of
4543 * incomplete I/O. This keeps the busy status for pages
4547 vfs_unbusy_pages(struct buf *bp)
4553 runningbufwakeup(bp);
4554 if (!(bp->b_flags & B_VMIO))
4557 obj = bp->b_bufobj->bo_object;
4558 for (i = 0; i < bp->b_npages; i++) {
4560 if (m == bogus_page) {
4561 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4563 panic("vfs_unbusy_pages: page missing\n");
4565 if (buf_mapped(bp)) {
4566 BUF_CHECK_MAPPED(bp);
4567 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4568 bp->b_pages, bp->b_npages);
4570 BUF_CHECK_UNMAPPED(bp);
4574 vm_object_pip_wakeupn(obj, bp->b_npages);
4578 * vfs_page_set_valid:
4580 * Set the valid bits in a page based on the supplied offset. The
4581 * range is restricted to the buffer's size.
4583 * This routine is typically called after a read completes.
4586 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4591 * Compute the end offset, eoff, such that [off, eoff) does not span a
4592 * page boundary and eoff is not greater than the end of the buffer.
4593 * The end of the buffer, in this case, is our file EOF, not the
4594 * allocation size of the buffer.
4596 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4597 if (eoff > bp->b_offset + bp->b_bcount)
4598 eoff = bp->b_offset + bp->b_bcount;
4601 * Set valid range. This is typically the entire buffer and thus the
4605 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4609 * vfs_page_set_validclean:
4611 * Set the valid bits and clear the dirty bits in a page based on the
4612 * supplied offset. The range is restricted to the buffer's size.
4615 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4617 vm_ooffset_t soff, eoff;
4620 * Start and end offsets in buffer. eoff - soff may not cross a
4621 * page boundary or cross the end of the buffer. The end of the
4622 * buffer, in this case, is our file EOF, not the allocation size
4626 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4627 if (eoff > bp->b_offset + bp->b_bcount)
4628 eoff = bp->b_offset + bp->b_bcount;
4631 * Set valid range. This is typically the entire buffer and thus the
4635 vm_page_set_validclean(
4637 (vm_offset_t) (soff & PAGE_MASK),
4638 (vm_offset_t) (eoff - soff)
4644 * Acquire a shared busy on all pages in the buf.
4647 vfs_busy_pages_acquire(struct buf *bp)
4651 for (i = 0; i < bp->b_npages; i++)
4652 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4656 vfs_busy_pages_release(struct buf *bp)
4660 for (i = 0; i < bp->b_npages; i++)
4661 vm_page_sunbusy(bp->b_pages[i]);
4665 * This routine is called before a device strategy routine.
4666 * It is used to tell the VM system that paging I/O is in
4667 * progress, and treat the pages associated with the buffer
4668 * almost as being exclusive busy. Also the object paging_in_progress
4669 * flag is handled to make sure that the object doesn't become
4672 * Since I/O has not been initiated yet, certain buffer flags
4673 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4674 * and should be ignored.
4677 vfs_busy_pages(struct buf *bp, int clear_modify)
4685 if (!(bp->b_flags & B_VMIO))
4688 obj = bp->b_bufobj->bo_object;
4689 foff = bp->b_offset;
4690 KASSERT(bp->b_offset != NOOFFSET,
4691 ("vfs_busy_pages: no buffer offset"));
4692 if ((bp->b_flags & B_CLUSTER) == 0) {
4693 vm_object_pip_add(obj, bp->b_npages);
4694 vfs_busy_pages_acquire(bp);
4696 if (bp->b_bufsize != 0)
4697 vfs_setdirty_range(bp);
4699 for (i = 0; i < bp->b_npages; i++) {
4701 vm_page_assert_sbusied(m);
4704 * When readying a buffer for a read ( i.e
4705 * clear_modify == 0 ), it is important to do
4706 * bogus_page replacement for valid pages in
4707 * partially instantiated buffers. Partially
4708 * instantiated buffers can, in turn, occur when
4709 * reconstituting a buffer from its VM backing store
4710 * base. We only have to do this if B_CACHE is
4711 * clear ( which causes the I/O to occur in the
4712 * first place ). The replacement prevents the read
4713 * I/O from overwriting potentially dirty VM-backed
4714 * pages. XXX bogus page replacement is, uh, bogus.
4715 * It may not work properly with small-block devices.
4716 * We need to find a better way.
4719 pmap_remove_write(m);
4720 vfs_page_set_validclean(bp, foff, m);
4721 } else if (vm_page_all_valid(m) &&
4722 (bp->b_flags & B_CACHE) == 0) {
4723 bp->b_pages[i] = bogus_page;
4726 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4728 if (bogus && buf_mapped(bp)) {
4729 BUF_CHECK_MAPPED(bp);
4730 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4731 bp->b_pages, bp->b_npages);
4736 * vfs_bio_set_valid:
4738 * Set the range within the buffer to valid. The range is
4739 * relative to the beginning of the buffer, b_offset. Note that
4740 * b_offset itself may be offset from the beginning of the first
4744 vfs_bio_set_valid(struct buf *bp, int base, int size)
4749 if (!(bp->b_flags & B_VMIO))
4753 * Fixup base to be relative to beginning of first page.
4754 * Set initial n to be the maximum number of bytes in the
4755 * first page that can be validated.
4757 base += (bp->b_offset & PAGE_MASK);
4758 n = PAGE_SIZE - (base & PAGE_MASK);
4761 * Busy may not be strictly necessary here because the pages are
4762 * unlikely to be fully valid and the vnode lock will synchronize
4763 * their access via getpages. It is grabbed for consistency with
4764 * other page validation.
4766 vfs_busy_pages_acquire(bp);
4767 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4771 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4776 vfs_busy_pages_release(bp);
4782 * If the specified buffer is a non-VMIO buffer, clear the entire
4783 * buffer. If the specified buffer is a VMIO buffer, clear and
4784 * validate only the previously invalid portions of the buffer.
4785 * This routine essentially fakes an I/O, so we need to clear
4786 * BIO_ERROR and B_INVAL.
4788 * Note that while we only theoretically need to clear through b_bcount,
4789 * we go ahead and clear through b_bufsize.
4792 vfs_bio_clrbuf(struct buf *bp)
4794 int i, j, mask, sa, ea, slide;
4796 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4800 bp->b_flags &= ~B_INVAL;
4801 bp->b_ioflags &= ~BIO_ERROR;
4802 vfs_busy_pages_acquire(bp);
4803 sa = bp->b_offset & PAGE_MASK;
4805 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4806 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4807 ea = slide & PAGE_MASK;
4810 if (bp->b_pages[i] == bogus_page)
4813 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4814 if ((bp->b_pages[i]->valid & mask) == mask)
4816 if ((bp->b_pages[i]->valid & mask) == 0)
4817 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4819 for (; sa < ea; sa += DEV_BSIZE, j++) {
4820 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4821 pmap_zero_page_area(bp->b_pages[i],
4826 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4827 roundup2(ea - sa, DEV_BSIZE));
4829 vfs_busy_pages_release(bp);
4834 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4839 if (buf_mapped(bp)) {
4840 BUF_CHECK_MAPPED(bp);
4841 bzero(bp->b_data + base, size);
4843 BUF_CHECK_UNMAPPED(bp);
4844 n = PAGE_SIZE - (base & PAGE_MASK);
4845 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4849 pmap_zero_page_area(m, base & PAGE_MASK, n);
4858 * Update buffer flags based on I/O request parameters, optionally releasing the
4859 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4860 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4861 * I/O). Otherwise the buffer is released to the cache.
4864 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4867 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4868 ("buf %p non-VMIO noreuse", bp));
4870 if ((ioflag & IO_DIRECT) != 0)
4871 bp->b_flags |= B_DIRECT;
4872 if ((ioflag & IO_EXT) != 0)
4873 bp->b_xflags |= BX_ALTDATA;
4874 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4875 bp->b_flags |= B_RELBUF;
4876 if ((ioflag & IO_NOREUSE) != 0)
4877 bp->b_flags |= B_NOREUSE;
4885 vfs_bio_brelse(struct buf *bp, int ioflag)
4888 b_io_dismiss(bp, ioflag, true);
4892 vfs_bio_set_flags(struct buf *bp, int ioflag)
4895 b_io_dismiss(bp, ioflag, false);
4899 * vm_hold_load_pages and vm_hold_free_pages get pages into
4900 * a buffers address space. The pages are anonymous and are
4901 * not associated with a file object.
4904 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4910 BUF_CHECK_MAPPED(bp);
4912 to = round_page(to);
4913 from = round_page(from);
4914 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4915 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4916 KASSERT(to - from <= maxbcachebuf,
4917 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4918 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4920 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4922 * note: must allocate system pages since blocking here
4923 * could interfere with paging I/O, no matter which
4926 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4927 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4928 pmap_qenter(pg, &p, 1);
4929 bp->b_pages[index] = p;
4931 bp->b_npages = index;
4934 /* Return pages associated with this buf to the vm system */
4936 vm_hold_free_pages(struct buf *bp, int newbsize)
4940 int index, newnpages;
4942 BUF_CHECK_MAPPED(bp);
4944 from = round_page((vm_offset_t)bp->b_data + newbsize);
4945 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4946 if (bp->b_npages > newnpages)
4947 pmap_qremove(from, bp->b_npages - newnpages);
4948 for (index = newnpages; index < bp->b_npages; index++) {
4949 p = bp->b_pages[index];
4950 bp->b_pages[index] = NULL;
4951 vm_page_unwire_noq(p);
4954 bp->b_npages = newnpages;
4958 * Map an IO request into kernel virtual address space.
4960 * All requests are (re)mapped into kernel VA space.
4961 * Notice that we use b_bufsize for the size of the buffer
4962 * to be mapped. b_bcount might be modified by the driver.
4964 * Note that even if the caller determines that the address space should
4965 * be valid, a race or a smaller-file mapped into a larger space may
4966 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4967 * check the return value.
4969 * This function only works with pager buffers.
4972 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4977 MPASS((bp->b_flags & B_MAXPHYS) != 0);
4978 prot = VM_PROT_READ;
4979 if (bp->b_iocmd == BIO_READ)
4980 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4981 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4982 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
4985 bp->b_bufsize = len;
4986 bp->b_npages = pidx;
4987 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
4988 if (mapbuf || !unmapped_buf_allowed) {
4989 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4990 bp->b_data = bp->b_kvabase + bp->b_offset;
4992 bp->b_data = unmapped_buf;
4997 * Free the io map PTEs associated with this IO operation.
4998 * We also invalidate the TLB entries and restore the original b_addr.
5000 * This function only works with pager buffers.
5003 vunmapbuf(struct buf *bp)
5007 npages = bp->b_npages;
5009 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5010 vm_page_unhold_pages(bp->b_pages, npages);
5012 bp->b_data = unmapped_buf;
5016 bdone(struct buf *bp)
5020 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5022 bp->b_flags |= B_DONE;
5028 bwait(struct buf *bp, u_char pri, const char *wchan)
5032 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5034 while ((bp->b_flags & B_DONE) == 0)
5035 msleep(bp, mtxp, pri, wchan, 0);
5040 bufsync(struct bufobj *bo, int waitfor)
5043 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5047 bufstrategy(struct bufobj *bo, struct buf *bp)
5053 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5054 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5055 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5056 i = VOP_STRATEGY(vp, bp);
5057 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5061 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5064 bufobj_init(struct bufobj *bo, void *private)
5066 static volatile int bufobj_cleanq;
5069 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5070 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5071 bo->bo_private = private;
5072 TAILQ_INIT(&bo->bo_clean.bv_hd);
5073 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5077 bufobj_wrefl(struct bufobj *bo)
5080 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5081 ASSERT_BO_WLOCKED(bo);
5086 bufobj_wref(struct bufobj *bo)
5089 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5096 bufobj_wdrop(struct bufobj *bo)
5099 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5101 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5102 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5103 bo->bo_flag &= ~BO_WWAIT;
5104 wakeup(&bo->bo_numoutput);
5110 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5114 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5115 ASSERT_BO_WLOCKED(bo);
5117 while (bo->bo_numoutput) {
5118 bo->bo_flag |= BO_WWAIT;
5119 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5120 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5128 * Set bio_data or bio_ma for struct bio from the struct buf.
5131 bdata2bio(struct buf *bp, struct bio *bip)
5134 if (!buf_mapped(bp)) {
5135 KASSERT(unmapped_buf_allowed, ("unmapped"));
5136 bip->bio_ma = bp->b_pages;
5137 bip->bio_ma_n = bp->b_npages;
5138 bip->bio_data = unmapped_buf;
5139 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5140 bip->bio_flags |= BIO_UNMAPPED;
5141 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5142 PAGE_SIZE == bp->b_npages,
5143 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5144 (long long)bip->bio_length, bip->bio_ma_n));
5146 bip->bio_data = bp->b_data;
5152 * The MIPS pmap code currently doesn't handle aliased pages.
5153 * The VIPT caches may not handle page aliasing themselves, leading
5154 * to data corruption.
5156 * As such, this code makes a system extremely unhappy if said
5157 * system doesn't support unaliasing the above situation in hardware.
5158 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5159 * this feature at build time, so it has to be handled in software.
5161 * Once the MIPS pmap/cache code grows to support this function on
5162 * earlier chips, it should be flipped back off.
5165 static int buf_pager_relbuf = 1;
5167 static int buf_pager_relbuf = 0;
5169 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5170 &buf_pager_relbuf, 0,
5171 "Make buffer pager release buffers after reading");
5174 * The buffer pager. It uses buffer reads to validate pages.
5176 * In contrast to the generic local pager from vm/vnode_pager.c, this
5177 * pager correctly and easily handles volumes where the underlying
5178 * device block size is greater than the machine page size. The
5179 * buffer cache transparently extends the requested page run to be
5180 * aligned at the block boundary, and does the necessary bogus page
5181 * replacements in the addends to avoid obliterating already valid
5184 * The only non-trivial issue is that the exclusive busy state for
5185 * pages, which is assumed by the vm_pager_getpages() interface, is
5186 * incompatible with the VMIO buffer cache's desire to share-busy the
5187 * pages. This function performs a trivial downgrade of the pages'
5188 * state before reading buffers, and a less trivial upgrade from the
5189 * shared-busy to excl-busy state after the read.
5192 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5193 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5194 vbg_get_blksize_t get_blksize)
5201 vm_ooffset_t la, lb, poff, poffe;
5203 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5206 object = vp->v_object;
5209 la = IDX_TO_OFF(ma[count - 1]->pindex);
5210 if (la >= object->un_pager.vnp.vnp_size)
5211 return (VM_PAGER_BAD);
5214 * Change the meaning of la from where the last requested page starts
5215 * to where it ends, because that's the end of the requested region
5216 * and the start of the potential read-ahead region.
5219 lpart = la > object->un_pager.vnp.vnp_size;
5220 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5223 return (VM_PAGER_ERROR);
5226 * Calculate read-ahead, behind and total pages.
5229 lb = IDX_TO_OFF(ma[0]->pindex);
5230 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5232 if (rbehind != NULL)
5234 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5235 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5236 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5241 VM_CNT_INC(v_vnodein);
5242 VM_CNT_ADD(v_vnodepgsin, pgsin);
5244 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5245 != 0) ? GB_UNMAPPED : 0;
5247 for (i = 0; i < count; i++) {
5248 if (ma[i] != bogus_page)
5249 vm_page_busy_downgrade(ma[i]);
5253 for (i = 0; i < count; i++) {
5255 if (m == bogus_page)
5259 * Pages are shared busy and the object lock is not
5260 * owned, which together allow for the pages'
5261 * invalidation. The racy test for validity avoids
5262 * useless creation of the buffer for the most typical
5263 * case when invalidation is not used in redo or for
5264 * parallel read. The shared->excl upgrade loop at
5265 * the end of the function catches the race in a
5266 * reliable way (protected by the object lock).
5268 if (vm_page_all_valid(m))
5271 poff = IDX_TO_OFF(m->pindex);
5272 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5273 for (; poff < poffe; poff += bsize) {
5274 lbn = get_lblkno(vp, poff);
5279 error = get_blksize(vp, lbn, &bsize);
5281 error = bread_gb(vp, lbn, bsize,
5282 curthread->td_ucred, br_flags, &bp);
5285 if (bp->b_rcred == curthread->td_ucred) {
5286 crfree(bp->b_rcred);
5287 bp->b_rcred = NOCRED;
5289 if (LIST_EMPTY(&bp->b_dep)) {
5291 * Invalidation clears m->valid, but
5292 * may leave B_CACHE flag if the
5293 * buffer existed at the invalidation
5294 * time. In this case, recycle the
5295 * buffer to do real read on next
5296 * bread() after redo.
5298 * Otherwise B_RELBUF is not strictly
5299 * necessary, enable to reduce buf
5302 if (buf_pager_relbuf ||
5303 !vm_page_all_valid(m))
5304 bp->b_flags |= B_RELBUF;
5306 bp->b_flags &= ~B_NOCACHE;
5312 KASSERT(1 /* racy, enable for debugging */ ||
5313 vm_page_all_valid(m) || i == count - 1,
5314 ("buf %d %p invalid", i, m));
5315 if (i == count - 1 && lpart) {
5316 if (!vm_page_none_valid(m) &&
5317 !vm_page_all_valid(m))
5318 vm_page_zero_invalid(m, TRUE);
5325 for (i = 0; i < count; i++) {
5326 if (ma[i] == bogus_page)
5328 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5329 vm_page_sunbusy(ma[i]);
5330 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5335 * Since the pages were only sbusy while neither the
5336 * buffer nor the object lock was held by us, or
5337 * reallocated while vm_page_grab() slept for busy
5338 * relinguish, they could have been invalidated.
5339 * Recheck the valid bits and re-read as needed.
5341 * Note that the last page is made fully valid in the
5342 * read loop, and partial validity for the page at
5343 * index count - 1 could mean that the page was
5344 * invalidated or removed, so we must restart for
5347 if (!vm_page_all_valid(ma[i]))
5350 if (redo && error == 0)
5352 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5355 #include "opt_ddb.h"
5357 #include <ddb/ddb.h>
5359 /* DDB command to show buffer data */
5360 DB_SHOW_COMMAND(buffer, db_show_buffer)
5363 struct buf *bp = (struct buf *)addr;
5364 #ifdef FULL_BUF_TRACKING
5369 db_printf("usage: show buffer <addr>\n");
5373 db_printf("buf at %p\n", bp);
5374 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5375 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5376 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5377 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5378 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5379 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5381 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5382 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5383 "b_vp = %p, b_dep = %p\n",
5384 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5385 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5386 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5387 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5388 bp->b_kvabase, bp->b_kvasize);
5391 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5392 for (i = 0; i < bp->b_npages; i++) {
5396 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5398 (u_long)VM_PAGE_TO_PHYS(m));
5400 db_printf("( ??? )");
5401 if ((i + 1) < bp->b_npages)
5406 BUF_LOCKPRINTINFO(bp);
5407 #if defined(FULL_BUF_TRACKING)
5408 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5410 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5411 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5412 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5414 db_printf(" %2u: %s\n", j,
5415 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5417 #elif defined(BUF_TRACKING)
5418 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5423 DB_SHOW_COMMAND(bufqueues, bufqueues)
5425 struct bufdomain *bd;
5430 db_printf("bqempty: %d\n", bqempty.bq_len);
5432 for (i = 0; i < buf_domains; i++) {
5434 db_printf("Buf domain %d\n", i);
5435 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5436 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5437 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5439 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5440 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5441 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5442 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5443 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5445 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5446 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5447 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5448 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5451 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5452 total += bp->b_bufsize;
5453 db_printf("\tcleanq count\t%d (%ld)\n",
5454 bd->bd_cleanq->bq_len, total);
5456 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5457 total += bp->b_bufsize;
5458 db_printf("\tdirtyq count\t%d (%ld)\n",
5459 bd->bd_dirtyq.bq_len, total);
5460 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5461 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5462 db_printf("\tCPU ");
5463 for (j = 0; j <= mp_maxid; j++)
5464 db_printf("%d, ", bd->bd_subq[j].bq_len);
5468 for (j = 0; j < nbuf; j++) {
5470 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5472 total += bp->b_bufsize;
5475 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5478 for (j = 0; j < nbuf; j++) {
5480 if (bp->b_domain == i) {
5482 total += bp->b_bufsize;
5485 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5489 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5494 for (i = 0; i < nbuf; i++) {
5496 if (BUF_ISLOCKED(bp)) {
5497 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5505 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5511 db_printf("usage: show vnodebufs <addr>\n");
5514 vp = (struct vnode *)addr;
5515 db_printf("Clean buffers:\n");
5516 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5517 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5520 db_printf("Dirty buffers:\n");
5521 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5522 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5527 DB_COMMAND(countfreebufs, db_coundfreebufs)
5530 int i, used = 0, nfree = 0;
5533 db_printf("usage: countfreebufs\n");
5537 for (i = 0; i < nbuf; i++) {
5539 if (bp->b_qindex == QUEUE_EMPTY)
5545 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5547 db_printf("numfreebuffers is %d\n", numfreebuffers);