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;
1048 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for
1049 * this when sizing maps based on the amount of physical memory
1053 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1054 (KASAN_SHADOW_SCALE + 1);
1055 #elif defined(KMSAN)
1059 * KMSAN cannot reliably determine whether buffer data is initialized
1060 * unless it is updated through a KVA mapping.
1062 unmapped_buf_allowed = 0;
1066 * physmem_est is in pages. Convert it to kilobytes (assumes
1067 * PAGE_SIZE is >= 1K)
1069 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1071 maxbcachebuf_adjust();
1073 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1074 * For the first 64MB of ram nominally allocate sufficient buffers to
1075 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1076 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1077 * the buffer cache we limit the eventual kva reservation to
1080 * factor represents the 1/4 x ram conversion.
1083 int factor = 4 * BKVASIZE / 1024;
1086 if (physmem_est > 4096)
1087 nbuf += min((physmem_est - 4096) / factor,
1089 if (physmem_est > 65536)
1090 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1091 32 * 1024 * 1024 / (factor * 5));
1093 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1094 nbuf = maxbcache / BKVASIZE;
1099 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1100 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1101 if (nbuf > maxbuf) {
1103 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1109 * Ideal allocation size for the transient bio submap is 10%
1110 * of the maximal space buffer map. This roughly corresponds
1111 * to the amount of the buffer mapped for typical UFS load.
1113 * Clip the buffer map to reserve space for the transient
1114 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1115 * maximum buffer map extent on the platform.
1117 * The fall-back to the maxbuf in case of maxbcache unset,
1118 * allows to not trim the buffer KVA for the architectures
1119 * with ample KVA space.
1121 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1122 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1123 buf_sz = (long)nbuf * BKVASIZE;
1124 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1125 (TRANSIENT_DENOM - 1)) {
1127 * There is more KVA than memory. Do not
1128 * adjust buffer map size, and assign the rest
1129 * of maxbuf to transient map.
1131 biotmap_sz = maxbuf_sz - buf_sz;
1134 * Buffer map spans all KVA we could afford on
1135 * this platform. Give 10% (20% on i386) of
1136 * the buffer map to the transient bio map.
1138 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1139 buf_sz -= biotmap_sz;
1141 if (biotmap_sz / INT_MAX > maxphys)
1142 bio_transient_maxcnt = INT_MAX;
1144 bio_transient_maxcnt = biotmap_sz / maxphys;
1146 * Artificially limit to 1024 simultaneous in-flight I/Os
1147 * using the transient mapping.
1149 if (bio_transient_maxcnt > 1024)
1150 bio_transient_maxcnt = 1024;
1152 nbuf = buf_sz / BKVASIZE;
1156 nswbuf = min(nbuf / 4, 256);
1157 if (nswbuf < NSWBUF_MIN)
1158 nswbuf = NSWBUF_MIN;
1162 * Reserve space for the buffer cache buffers
1165 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1166 atop(maxbcachebuf)) * nbuf;
1171 /* Initialize the buffer subsystem. Called before use of any buffers. */
1178 KASSERT(maxbcachebuf >= MAXBSIZE,
1179 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1181 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1182 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1183 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1184 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1186 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1188 /* finally, initialize each buffer header and stick on empty q */
1189 for (i = 0; i < nbuf; i++) {
1191 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1192 bp->b_flags = B_INVAL;
1193 bp->b_rcred = NOCRED;
1194 bp->b_wcred = NOCRED;
1195 bp->b_qindex = QUEUE_NONE;
1197 bp->b_subqueue = mp_maxid + 1;
1199 bp->b_data = bp->b_kvabase = unmapped_buf;
1200 LIST_INIT(&bp->b_dep);
1202 bq_insert(&bqempty, bp, false);
1206 * maxbufspace is the absolute maximum amount of buffer space we are
1207 * allowed to reserve in KVM and in real terms. The absolute maximum
1208 * is nominally used by metadata. hibufspace is the nominal maximum
1209 * used by most other requests. The differential is required to
1210 * ensure that metadata deadlocks don't occur.
1212 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1213 * this may result in KVM fragmentation which is not handled optimally
1214 * by the system. XXX This is less true with vmem. We could use
1217 maxbufspace = (long)nbuf * BKVASIZE;
1218 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1219 lobufspace = (hibufspace / 20) * 19; /* 95% */
1220 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1223 * Note: The 16 MiB upper limit for hirunningspace was chosen
1224 * arbitrarily and may need further tuning. It corresponds to
1225 * 128 outstanding write IO requests (if IO size is 128 KiB),
1226 * which fits with many RAID controllers' tagged queuing limits.
1227 * The lower 1 MiB limit is the historical upper limit for
1230 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1231 16 * 1024 * 1024), 1024 * 1024);
1232 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1235 * Limit the amount of malloc memory since it is wired permanently into
1236 * the kernel space. Even though this is accounted for in the buffer
1237 * allocation, we don't want the malloced region to grow uncontrolled.
1238 * The malloc scheme improves memory utilization significantly on
1239 * average (small) directories.
1241 maxbufmallocspace = hibufspace / 20;
1244 * Reduce the chance of a deadlock occurring by limiting the number
1245 * of delayed-write dirty buffers we allow to stack up.
1247 hidirtybuffers = nbuf / 4 + 20;
1248 dirtybufthresh = hidirtybuffers * 9 / 10;
1250 * To support extreme low-memory systems, make sure hidirtybuffers
1251 * cannot eat up all available buffer space. This occurs when our
1252 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1253 * buffer space assuming BKVASIZE'd buffers.
1255 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1256 hidirtybuffers >>= 1;
1258 lodirtybuffers = hidirtybuffers / 2;
1261 * lofreebuffers should be sufficient to avoid stalling waiting on
1262 * buf headers under heavy utilization. The bufs in per-cpu caches
1263 * are counted as free but will be unavailable to threads executing
1266 * hifreebuffers is the free target for the bufspace daemon. This
1267 * should be set appropriately to limit work per-iteration.
1269 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1270 hifreebuffers = (3 * lofreebuffers) / 2;
1271 numfreebuffers = nbuf;
1273 /* Setup the kva and free list allocators. */
1274 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1275 buf_zone = uma_zcache_create("buf free cache",
1276 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1277 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1280 * Size the clean queue according to the amount of buffer space.
1281 * One queue per-256mb up to the max. More queues gives better
1282 * concurrency but less accurate LRU.
1284 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1285 for (i = 0 ; i < buf_domains; i++) {
1286 struct bufdomain *bd;
1290 bd->bd_freebuffers = nbuf / buf_domains;
1291 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1292 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1293 bd->bd_bufspace = 0;
1294 bd->bd_maxbufspace = maxbufspace / buf_domains;
1295 bd->bd_hibufspace = hibufspace / buf_domains;
1296 bd->bd_lobufspace = lobufspace / buf_domains;
1297 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1298 bd->bd_numdirtybuffers = 0;
1299 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1300 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1301 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1302 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1303 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1305 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1306 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1307 mappingrestarts = counter_u64_alloc(M_WAITOK);
1308 numbufallocfails = counter_u64_alloc(M_WAITOK);
1309 notbufdflushes = counter_u64_alloc(M_WAITOK);
1310 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1311 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1312 bufkvaspace = counter_u64_alloc(M_WAITOK);
1317 vfs_buf_check_mapped(struct buf *bp)
1320 KASSERT(bp->b_kvabase != unmapped_buf,
1321 ("mapped buf: b_kvabase was not updated %p", bp));
1322 KASSERT(bp->b_data != unmapped_buf,
1323 ("mapped buf: b_data was not updated %p", bp));
1324 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1325 maxphys, ("b_data + b_offset unmapped %p", bp));
1329 vfs_buf_check_unmapped(struct buf *bp)
1332 KASSERT(bp->b_data == unmapped_buf,
1333 ("unmapped buf: corrupted b_data %p", bp));
1336 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1337 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1339 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1340 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1344 isbufbusy(struct buf *bp)
1346 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1347 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1353 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1356 bufshutdown(int show_busybufs)
1358 static int first_buf_printf = 1;
1360 int i, iter, nbusy, pbusy;
1366 * Sync filesystems for shutdown
1368 wdog_kern_pat(WD_LASTVAL);
1369 kern_sync(curthread);
1372 * With soft updates, some buffers that are
1373 * written will be remarked as dirty until other
1374 * buffers are written.
1376 for (iter = pbusy = 0; iter < 20; iter++) {
1378 for (i = nbuf - 1; i >= 0; i--) {
1384 if (first_buf_printf)
1385 printf("All buffers synced.");
1388 if (first_buf_printf) {
1389 printf("Syncing disks, buffers remaining... ");
1390 first_buf_printf = 0;
1392 printf("%d ", nbusy);
1397 wdog_kern_pat(WD_LASTVAL);
1398 kern_sync(curthread);
1402 * Spin for a while to allow interrupt threads to run.
1404 DELAY(50000 * iter);
1407 * Context switch several times to allow interrupt
1410 for (subiter = 0; subiter < 50 * iter; subiter++) {
1411 thread_lock(curthread);
1419 * Count only busy local buffers to prevent forcing
1420 * a fsck if we're just a client of a wedged NFS server
1423 for (i = nbuf - 1; i >= 0; i--) {
1425 if (isbufbusy(bp)) {
1427 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1428 if (bp->b_dev == NULL) {
1429 TAILQ_REMOVE(&mountlist,
1430 bp->b_vp->v_mount, mnt_list);
1435 if (show_busybufs > 0) {
1437 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1438 nbusy, bp, bp->b_vp, bp->b_flags,
1439 (intmax_t)bp->b_blkno,
1440 (intmax_t)bp->b_lblkno);
1441 BUF_LOCKPRINTINFO(bp);
1442 if (show_busybufs > 1)
1450 * Failed to sync all blocks. Indicate this and don't
1451 * unmount filesystems (thus forcing an fsck on reboot).
1453 printf("Giving up on %d buffers\n", nbusy);
1454 DELAY(5000000); /* 5 seconds */
1456 if (!first_buf_printf)
1457 printf("Final sync complete\n");
1459 * Unmount filesystems
1461 if (!KERNEL_PANICKED())
1465 DELAY(100000); /* wait for console output to finish */
1469 bpmap_qenter(struct buf *bp)
1472 BUF_CHECK_MAPPED(bp);
1475 * bp->b_data is relative to bp->b_offset, but
1476 * bp->b_offset may be offset into the first page.
1478 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1479 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1480 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1481 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1484 static inline struct bufdomain *
1485 bufdomain(struct buf *bp)
1488 return (&bdomain[bp->b_domain]);
1491 static struct bufqueue *
1492 bufqueue(struct buf *bp)
1495 switch (bp->b_qindex) {
1498 case QUEUE_SENTINEL:
1503 return (&bufdomain(bp)->bd_dirtyq);
1505 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1509 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1513 * Return the locked bufqueue that bp is a member of.
1515 static struct bufqueue *
1516 bufqueue_acquire(struct buf *bp)
1518 struct bufqueue *bq, *nbq;
1521 * bp can be pushed from a per-cpu queue to the
1522 * cleanq while we're waiting on the lock. Retry
1523 * if the queues don't match.
1541 * Insert the buffer into the appropriate free list. Requires a
1542 * locked buffer on entry and buffer is unlocked before return.
1545 binsfree(struct buf *bp, int qindex)
1547 struct bufdomain *bd;
1548 struct bufqueue *bq;
1550 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1551 ("binsfree: Invalid qindex %d", qindex));
1552 BUF_ASSERT_XLOCKED(bp);
1555 * Handle delayed bremfree() processing.
1557 if (bp->b_flags & B_REMFREE) {
1558 if (bp->b_qindex == qindex) {
1559 bp->b_flags |= B_REUSE;
1560 bp->b_flags &= ~B_REMFREE;
1564 bq = bufqueue_acquire(bp);
1569 if (qindex == QUEUE_CLEAN) {
1570 if (bd->bd_lim != 0)
1571 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1575 bq = &bd->bd_dirtyq;
1576 bq_insert(bq, bp, true);
1582 * Free a buffer to the buf zone once it no longer has valid contents.
1585 buf_free(struct buf *bp)
1588 if (bp->b_flags & B_REMFREE)
1590 if (bp->b_vflags & BV_BKGRDINPROG)
1591 panic("losing buffer 1");
1592 if (bp->b_rcred != NOCRED) {
1593 crfree(bp->b_rcred);
1594 bp->b_rcred = NOCRED;
1596 if (bp->b_wcred != NOCRED) {
1597 crfree(bp->b_wcred);
1598 bp->b_wcred = NOCRED;
1600 if (!LIST_EMPTY(&bp->b_dep))
1603 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1604 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1606 uma_zfree(buf_zone, bp);
1612 * Import bufs into the uma cache from the buf list. The system still
1613 * expects a static array of bufs and much of the synchronization
1614 * around bufs assumes type stable storage. As a result, UMA is used
1615 * only as a per-cpu cache of bufs still maintained on a global list.
1618 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1624 for (i = 0; i < cnt; i++) {
1625 bp = TAILQ_FIRST(&bqempty.bq_queue);
1628 bq_remove(&bqempty, bp);
1631 BQ_UNLOCK(&bqempty);
1639 * Release bufs from the uma cache back to the buffer queues.
1642 buf_release(void *arg, void **store, int cnt)
1644 struct bufqueue *bq;
1650 for (i = 0; i < cnt; i++) {
1652 /* Inline bq_insert() to batch locking. */
1653 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1654 bp->b_flags &= ~(B_AGE | B_REUSE);
1656 bp->b_qindex = bq->bq_index;
1664 * Allocate an empty buffer header.
1667 buf_alloc(struct bufdomain *bd)
1670 int freebufs, error;
1673 * We can only run out of bufs in the buf zone if the average buf
1674 * is less than BKVASIZE. In this case the actual wait/block will
1675 * come from buf_reycle() failing to flush one of these small bufs.
1678 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1680 bp = uma_zalloc(buf_zone, M_NOWAIT);
1682 atomic_add_int(&bd->bd_freebuffers, 1);
1683 bufspace_daemon_wakeup(bd);
1684 counter_u64_add(numbufallocfails, 1);
1688 * Wake-up the bufspace daemon on transition below threshold.
1690 if (freebufs == bd->bd_lofreebuffers)
1691 bufspace_daemon_wakeup(bd);
1693 error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1694 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1698 KASSERT(bp->b_vp == NULL,
1699 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1700 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1701 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1702 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1703 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1704 KASSERT(bp->b_npages == 0,
1705 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1706 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1707 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1708 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1710 bp->b_domain = BD_DOMAIN(bd);
1716 bp->b_blkno = bp->b_lblkno = 0;
1717 bp->b_offset = NOOFFSET;
1723 bp->b_dirtyoff = bp->b_dirtyend = 0;
1724 bp->b_bufobj = NULL;
1725 bp->b_data = bp->b_kvabase = unmapped_buf;
1726 bp->b_fsprivate1 = NULL;
1727 bp->b_fsprivate2 = NULL;
1728 bp->b_fsprivate3 = NULL;
1729 LIST_INIT(&bp->b_dep);
1737 * Free a buffer from the given bufqueue. kva controls whether the
1738 * freed buf must own some kva resources. This is used for
1742 buf_recycle(struct bufdomain *bd, bool kva)
1744 struct bufqueue *bq;
1745 struct buf *bp, *nbp;
1748 counter_u64_add(bufdefragcnt, 1);
1752 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1753 ("buf_recycle: Locks don't match"));
1754 nbp = TAILQ_FIRST(&bq->bq_queue);
1757 * Run scan, possibly freeing data and/or kva mappings on the fly
1760 while ((bp = nbp) != NULL) {
1762 * Calculate next bp (we can only use it if we do not
1763 * release the bqlock).
1765 nbp = TAILQ_NEXT(bp, b_freelist);
1768 * If we are defragging then we need a buffer with
1769 * some kva to reclaim.
1771 if (kva && bp->b_kvasize == 0)
1774 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1778 * Implement a second chance algorithm for frequently
1781 if ((bp->b_flags & B_REUSE) != 0) {
1782 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1783 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1784 bp->b_flags &= ~B_REUSE;
1790 * Skip buffers with background writes in progress.
1792 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1797 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1798 ("buf_recycle: inconsistent queue %d bp %p",
1800 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1801 ("getnewbuf: queue domain %d doesn't match request %d",
1802 bp->b_domain, (int)BD_DOMAIN(bd)));
1804 * NOTE: nbp is now entirely invalid. We can only restart
1805 * the scan from this point on.
1811 * Requeue the background write buffer with error and
1814 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1817 nbp = TAILQ_FIRST(&bq->bq_queue);
1820 bp->b_flags |= B_INVAL;
1833 * Mark the buffer for removal from the appropriate free list.
1837 bremfree(struct buf *bp)
1840 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1841 KASSERT((bp->b_flags & B_REMFREE) == 0,
1842 ("bremfree: buffer %p already marked for delayed removal.", bp));
1843 KASSERT(bp->b_qindex != QUEUE_NONE,
1844 ("bremfree: buffer %p not on a queue.", bp));
1845 BUF_ASSERT_XLOCKED(bp);
1847 bp->b_flags |= B_REMFREE;
1853 * Force an immediate removal from a free list. Used only in nfs when
1854 * it abuses the b_freelist pointer.
1857 bremfreef(struct buf *bp)
1859 struct bufqueue *bq;
1861 bq = bufqueue_acquire(bp);
1867 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1870 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1871 TAILQ_INIT(&bq->bq_queue);
1873 bq->bq_index = qindex;
1874 bq->bq_subqueue = subqueue;
1878 bd_init(struct bufdomain *bd)
1882 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1883 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1884 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1885 for (i = 0; i <= mp_maxid; i++)
1886 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1887 "bufq clean subqueue lock");
1888 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1894 * Removes a buffer from the free list, must be called with the
1895 * correct qlock held.
1898 bq_remove(struct bufqueue *bq, struct buf *bp)
1901 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1902 bp, bp->b_vp, bp->b_flags);
1903 KASSERT(bp->b_qindex != QUEUE_NONE,
1904 ("bq_remove: buffer %p not on a queue.", bp));
1905 KASSERT(bufqueue(bp) == bq,
1906 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1908 BQ_ASSERT_LOCKED(bq);
1909 if (bp->b_qindex != QUEUE_EMPTY) {
1910 BUF_ASSERT_XLOCKED(bp);
1912 KASSERT(bq->bq_len >= 1,
1913 ("queue %d underflow", bp->b_qindex));
1914 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1916 bp->b_qindex = QUEUE_NONE;
1917 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1921 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1925 BQ_ASSERT_LOCKED(bq);
1926 if (bq != bd->bd_cleanq) {
1928 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1929 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1930 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1932 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1934 bd->bd_cleanq->bq_len += bq->bq_len;
1937 if (bd->bd_wanted) {
1939 wakeup(&bd->bd_wanted);
1941 if (bq != bd->bd_cleanq)
1946 bd_flushall(struct bufdomain *bd)
1948 struct bufqueue *bq;
1952 if (bd->bd_lim == 0)
1955 for (i = 0; i <= mp_maxid; i++) {
1956 bq = &bd->bd_subq[i];
1957 if (bq->bq_len == 0)
1969 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1971 struct bufdomain *bd;
1973 if (bp->b_qindex != QUEUE_NONE)
1974 panic("bq_insert: free buffer %p onto another queue?", bp);
1977 if (bp->b_flags & B_AGE) {
1978 /* Place this buf directly on the real queue. */
1979 if (bq->bq_index == QUEUE_CLEAN)
1982 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
1985 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1987 bp->b_flags &= ~(B_AGE | B_REUSE);
1989 bp->b_qindex = bq->bq_index;
1990 bp->b_subqueue = bq->bq_subqueue;
1993 * Unlock before we notify so that we don't wakeup a waiter that
1994 * fails a trylock on the buf and sleeps again.
1999 if (bp->b_qindex == QUEUE_CLEAN) {
2001 * Flush the per-cpu queue and notify any waiters.
2003 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2004 bq->bq_len >= bd->bd_lim))
2013 * Free the kva allocation for a buffer.
2017 bufkva_free(struct buf *bp)
2021 if (bp->b_kvasize == 0) {
2022 KASSERT(bp->b_kvabase == unmapped_buf &&
2023 bp->b_data == unmapped_buf,
2024 ("Leaked KVA space on %p", bp));
2025 } else if (buf_mapped(bp))
2026 BUF_CHECK_MAPPED(bp);
2028 BUF_CHECK_UNMAPPED(bp);
2030 if (bp->b_kvasize == 0)
2033 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2034 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2035 counter_u64_add(buffreekvacnt, 1);
2036 bp->b_data = bp->b_kvabase = unmapped_buf;
2043 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2046 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2051 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2052 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2053 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2054 KASSERT(maxsize <= maxbcachebuf,
2055 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2060 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2063 * Buffer map is too fragmented. Request the caller
2064 * to defragment the map.
2068 bp->b_kvabase = (caddr_t)addr;
2069 bp->b_kvasize = maxsize;
2070 counter_u64_add(bufkvaspace, bp->b_kvasize);
2071 if ((gbflags & GB_UNMAPPED) != 0) {
2072 bp->b_data = unmapped_buf;
2073 BUF_CHECK_UNMAPPED(bp);
2075 bp->b_data = bp->b_kvabase;
2076 BUF_CHECK_MAPPED(bp);
2084 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2085 * callback that fires to avoid returning failure.
2088 bufkva_reclaim(vmem_t *vmem, int flags)
2095 for (i = 0; i < 5; i++) {
2096 for (q = 0; q < buf_domains; q++)
2097 if (buf_recycle(&bdomain[q], true) != 0)
2106 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2107 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2108 * the buffer is valid and we do not have to do anything.
2111 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2112 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2120 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2121 if (inmem(vp, *rablkno))
2123 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2124 if ((rabp->b_flags & B_CACHE) != 0) {
2131 racct_add_buf(curproc, rabp, 0);
2132 PROC_UNLOCK(curproc);
2135 td->td_ru.ru_inblock++;
2136 rabp->b_flags |= B_ASYNC;
2137 rabp->b_flags &= ~B_INVAL;
2138 if ((flags & GB_CKHASH) != 0) {
2139 rabp->b_flags |= B_CKHASH;
2140 rabp->b_ckhashcalc = ckhashfunc;
2142 rabp->b_ioflags &= ~BIO_ERROR;
2143 rabp->b_iocmd = BIO_READ;
2144 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2145 rabp->b_rcred = crhold(cred);
2146 vfs_busy_pages(rabp, 0);
2148 rabp->b_iooffset = dbtob(rabp->b_blkno);
2154 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2156 * Get a buffer with the specified data. Look in the cache first. We
2157 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2158 * is set, the buffer is valid and we do not have to do anything, see
2159 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2161 * Always return a NULL buffer pointer (in bpp) when returning an error.
2163 * The blkno parameter is the logical block being requested. Normally
2164 * the mapping of logical block number to disk block address is done
2165 * by calling VOP_BMAP(). However, if the mapping is already known, the
2166 * disk block address can be passed using the dblkno parameter. If the
2167 * disk block address is not known, then the same value should be passed
2168 * for blkno and dblkno.
2171 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2172 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2173 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2177 int error, readwait, rv;
2179 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2182 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2185 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2190 KASSERT(blkno == bp->b_lblkno,
2191 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2192 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2193 flags &= ~GB_NOSPARSE;
2197 * If not found in cache, do some I/O
2200 if ((bp->b_flags & B_CACHE) == 0) {
2203 PROC_LOCK(td->td_proc);
2204 racct_add_buf(td->td_proc, bp, 0);
2205 PROC_UNLOCK(td->td_proc);
2208 td->td_ru.ru_inblock++;
2209 bp->b_iocmd = BIO_READ;
2210 bp->b_flags &= ~B_INVAL;
2211 if ((flags & GB_CKHASH) != 0) {
2212 bp->b_flags |= B_CKHASH;
2213 bp->b_ckhashcalc = ckhashfunc;
2215 if ((flags & GB_CVTENXIO) != 0)
2216 bp->b_xflags |= BX_CVTENXIO;
2217 bp->b_ioflags &= ~BIO_ERROR;
2218 if (bp->b_rcred == NOCRED && cred != NOCRED)
2219 bp->b_rcred = crhold(cred);
2220 vfs_busy_pages(bp, 0);
2221 bp->b_iooffset = dbtob(bp->b_blkno);
2227 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2229 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2243 * Write, release buffer on completion. (Done by iodone
2244 * if async). Do not bother writing anything if the buffer
2247 * Note that we set B_CACHE here, indicating that buffer is
2248 * fully valid and thus cacheable. This is true even of NFS
2249 * now so we set it generally. This could be set either here
2250 * or in biodone() since the I/O is synchronous. We put it
2254 bufwrite(struct buf *bp)
2261 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2262 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2263 bp->b_flags |= B_INVAL | B_RELBUF;
2264 bp->b_flags &= ~B_CACHE;
2268 if (bp->b_flags & B_INVAL) {
2273 if (bp->b_flags & B_BARRIER)
2274 atomic_add_long(&barrierwrites, 1);
2276 oldflags = bp->b_flags;
2278 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2279 ("FFS background buffer should not get here %p", bp));
2283 vp_md = vp->v_vflag & VV_MD;
2288 * Mark the buffer clean. Increment the bufobj write count
2289 * before bundirty() call, to prevent other thread from seeing
2290 * empty dirty list and zero counter for writes in progress,
2291 * falsely indicating that the bufobj is clean.
2293 bufobj_wref(bp->b_bufobj);
2296 bp->b_flags &= ~B_DONE;
2297 bp->b_ioflags &= ~BIO_ERROR;
2298 bp->b_flags |= B_CACHE;
2299 bp->b_iocmd = BIO_WRITE;
2301 vfs_busy_pages(bp, 1);
2304 * Normal bwrites pipeline writes
2306 bp->b_runningbufspace = bp->b_bufsize;
2307 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2312 racct_add_buf(curproc, bp, 1);
2313 PROC_UNLOCK(curproc);
2316 curthread->td_ru.ru_oublock++;
2317 if (oldflags & B_ASYNC)
2319 bp->b_iooffset = dbtob(bp->b_blkno);
2320 buf_track(bp, __func__);
2323 if ((oldflags & B_ASYNC) == 0) {
2324 int rtval = bufwait(bp);
2327 } else if (space > hirunningspace) {
2329 * don't allow the async write to saturate the I/O
2330 * system. We will not deadlock here because
2331 * we are blocking waiting for I/O that is already in-progress
2332 * to complete. We do not block here if it is the update
2333 * or syncer daemon trying to clean up as that can lead
2336 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2337 waitrunningbufspace();
2344 bufbdflush(struct bufobj *bo, struct buf *bp)
2347 struct bufdomain *bd;
2349 bd = &bdomain[bo->bo_domain];
2350 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2351 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2353 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2356 * Try to find a buffer to flush.
2358 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2359 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2361 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2364 panic("bdwrite: found ourselves");
2366 /* Don't countdeps with the bo lock held. */
2367 if (buf_countdeps(nbp, 0)) {
2372 if (nbp->b_flags & B_CLUSTEROK) {
2373 vfs_bio_awrite(nbp);
2378 dirtybufferflushes++;
2387 * Delayed write. (Buffer is marked dirty). Do not bother writing
2388 * anything if the buffer is marked invalid.
2390 * Note that since the buffer must be completely valid, we can safely
2391 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2392 * biodone() in order to prevent getblk from writing the buffer
2393 * out synchronously.
2396 bdwrite(struct buf *bp)
2398 struct thread *td = curthread;
2402 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2403 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2404 KASSERT((bp->b_flags & B_BARRIER) == 0,
2405 ("Barrier request in delayed write %p", bp));
2407 if (bp->b_flags & B_INVAL) {
2413 * If we have too many dirty buffers, don't create any more.
2414 * If we are wildly over our limit, then force a complete
2415 * cleanup. Otherwise, just keep the situation from getting
2416 * out of control. Note that we have to avoid a recursive
2417 * disaster and not try to clean up after our own cleanup!
2421 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2422 td->td_pflags |= TDP_INBDFLUSH;
2424 td->td_pflags &= ~TDP_INBDFLUSH;
2430 * Set B_CACHE, indicating that the buffer is fully valid. This is
2431 * true even of NFS now.
2433 bp->b_flags |= B_CACHE;
2436 * This bmap keeps the system from needing to do the bmap later,
2437 * perhaps when the system is attempting to do a sync. Since it
2438 * is likely that the indirect block -- or whatever other datastructure
2439 * that the filesystem needs is still in memory now, it is a good
2440 * thing to do this. Note also, that if the pageout daemon is
2441 * requesting a sync -- there might not be enough memory to do
2442 * the bmap then... So, this is important to do.
2444 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2445 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2448 buf_track(bp, __func__);
2451 * Set the *dirty* buffer range based upon the VM system dirty
2454 * Mark the buffer pages as clean. We need to do this here to
2455 * satisfy the vnode_pager and the pageout daemon, so that it
2456 * thinks that the pages have been "cleaned". Note that since
2457 * the pages are in a delayed write buffer -- the VFS layer
2458 * "will" see that the pages get written out on the next sync,
2459 * or perhaps the cluster will be completed.
2461 vfs_clean_pages_dirty_buf(bp);
2465 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2466 * due to the softdep code.
2473 * Turn buffer into delayed write request. We must clear BIO_READ and
2474 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2475 * itself to properly update it in the dirty/clean lists. We mark it
2476 * B_DONE to ensure that any asynchronization of the buffer properly
2477 * clears B_DONE ( else a panic will occur later ).
2479 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2480 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2481 * should only be called if the buffer is known-good.
2483 * Since the buffer is not on a queue, we do not update the numfreebuffers
2486 * The buffer must be on QUEUE_NONE.
2489 bdirty(struct buf *bp)
2492 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2493 bp, bp->b_vp, bp->b_flags);
2494 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2495 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2496 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2497 bp->b_flags &= ~(B_RELBUF);
2498 bp->b_iocmd = BIO_WRITE;
2500 if ((bp->b_flags & B_DELWRI) == 0) {
2501 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2510 * Clear B_DELWRI for buffer.
2512 * Since the buffer is not on a queue, we do not update the numfreebuffers
2515 * The buffer must be on QUEUE_NONE.
2519 bundirty(struct buf *bp)
2522 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2523 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2524 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2525 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2527 if (bp->b_flags & B_DELWRI) {
2528 bp->b_flags &= ~B_DELWRI;
2533 * Since it is now being written, we can clear its deferred write flag.
2535 bp->b_flags &= ~B_DEFERRED;
2541 * Asynchronous write. Start output on a buffer, but do not wait for
2542 * it to complete. The buffer is released when the output completes.
2544 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2545 * B_INVAL buffers. Not us.
2548 bawrite(struct buf *bp)
2551 bp->b_flags |= B_ASYNC;
2558 * Asynchronous barrier write. Start output on a buffer, but do not
2559 * wait for it to complete. Place a write barrier after this write so
2560 * that this buffer and all buffers written before it are committed to
2561 * the disk before any buffers written after this write are committed
2562 * to the disk. The buffer is released when the output completes.
2565 babarrierwrite(struct buf *bp)
2568 bp->b_flags |= B_ASYNC | B_BARRIER;
2575 * Synchronous barrier write. Start output on a buffer and wait for
2576 * it to complete. Place a write barrier after this write so that
2577 * this buffer and all buffers written before it are committed to
2578 * the disk before any buffers written after this write are committed
2579 * to the disk. The buffer is released when the output completes.
2582 bbarrierwrite(struct buf *bp)
2585 bp->b_flags |= B_BARRIER;
2586 return (bwrite(bp));
2592 * Called prior to the locking of any vnodes when we are expecting to
2593 * write. We do not want to starve the buffer cache with too many
2594 * dirty buffers so we block here. By blocking prior to the locking
2595 * of any vnodes we attempt to avoid the situation where a locked vnode
2596 * prevents the various system daemons from flushing related buffers.
2602 if (buf_dirty_count_severe()) {
2603 mtx_lock(&bdirtylock);
2604 while (buf_dirty_count_severe()) {
2606 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2609 mtx_unlock(&bdirtylock);
2614 * Return true if we have too many dirty buffers.
2617 buf_dirty_count_severe(void)
2620 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2626 * Release a busy buffer and, if requested, free its resources. The
2627 * buffer will be stashed in the appropriate bufqueue[] allowing it
2628 * to be accessed later as a cache entity or reused for other purposes.
2631 brelse(struct buf *bp)
2633 struct mount *v_mnt;
2637 * Many functions erroneously call brelse with a NULL bp under rare
2638 * error conditions. Simply return when called with a NULL bp.
2642 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2643 bp, bp->b_vp, bp->b_flags);
2644 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2645 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2646 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2647 ("brelse: non-VMIO buffer marked NOREUSE"));
2649 if (BUF_LOCKRECURSED(bp)) {
2651 * Do not process, in particular, do not handle the
2652 * B_INVAL/B_RELBUF and do not release to free list.
2658 if (bp->b_flags & B_MANAGED) {
2663 if (LIST_EMPTY(&bp->b_dep)) {
2664 bp->b_flags &= ~B_IOSTARTED;
2666 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2667 ("brelse: SU io not finished bp %p", bp));
2670 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2671 BO_LOCK(bp->b_bufobj);
2672 bp->b_vflags &= ~BV_BKGRDERR;
2673 BO_UNLOCK(bp->b_bufobj);
2677 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2678 (bp->b_flags & B_INVALONERR)) {
2680 * Forced invalidation of dirty buffer contents, to be used
2681 * after a failed write in the rare case that the loss of the
2682 * contents is acceptable. The buffer is invalidated and
2685 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2686 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2689 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2690 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2691 !(bp->b_flags & B_INVAL)) {
2693 * Failed write, redirty. All errors except ENXIO (which
2694 * means the device is gone) are treated as being
2697 * XXX Treating EIO as transient is not correct; the
2698 * contract with the local storage device drivers is that
2699 * they will only return EIO once the I/O is no longer
2700 * retriable. Network I/O also respects this through the
2701 * guarantees of TCP and/or the internal retries of NFS.
2702 * ENOMEM might be transient, but we also have no way of
2703 * knowing when its ok to retry/reschedule. In general,
2704 * this entire case should be made obsolete through better
2705 * error handling/recovery and resource scheduling.
2707 * Do this also for buffers that failed with ENXIO, but have
2708 * non-empty dependencies - the soft updates code might need
2709 * to access the buffer to untangle them.
2711 * Must clear BIO_ERROR to prevent pages from being scrapped.
2713 bp->b_ioflags &= ~BIO_ERROR;
2715 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2716 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2718 * Either a failed read I/O, or we were asked to free or not
2719 * cache the buffer, or we failed to write to a device that's
2720 * no longer present.
2722 bp->b_flags |= B_INVAL;
2723 if (!LIST_EMPTY(&bp->b_dep))
2725 if (bp->b_flags & B_DELWRI)
2727 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2728 if ((bp->b_flags & B_VMIO) == 0) {
2736 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2737 * is called with B_DELWRI set, the underlying pages may wind up
2738 * getting freed causing a previous write (bdwrite()) to get 'lost'
2739 * because pages associated with a B_DELWRI bp are marked clean.
2741 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2742 * if B_DELWRI is set.
2744 if (bp->b_flags & B_DELWRI)
2745 bp->b_flags &= ~B_RELBUF;
2748 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2749 * constituted, not even NFS buffers now. Two flags effect this. If
2750 * B_INVAL, the struct buf is invalidated but the VM object is kept
2751 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2753 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2754 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2755 * buffer is also B_INVAL because it hits the re-dirtying code above.
2757 * Normally we can do this whether a buffer is B_DELWRI or not. If
2758 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2759 * the commit state and we cannot afford to lose the buffer. If the
2760 * buffer has a background write in progress, we need to keep it
2761 * around to prevent it from being reconstituted and starting a second
2765 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2767 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2768 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2769 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2770 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2771 vfs_vmio_invalidate(bp);
2775 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2776 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2778 bp->b_flags &= ~B_NOREUSE;
2779 if (bp->b_vp != NULL)
2784 * If the buffer has junk contents signal it and eventually
2785 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2788 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2789 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2790 bp->b_flags |= B_INVAL;
2791 if (bp->b_flags & B_INVAL) {
2792 if (bp->b_flags & B_DELWRI)
2798 buf_track(bp, __func__);
2800 /* buffers with no memory */
2801 if (bp->b_bufsize == 0) {
2805 /* buffers with junk contents */
2806 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2807 (bp->b_ioflags & BIO_ERROR)) {
2808 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2809 if (bp->b_vflags & BV_BKGRDINPROG)
2810 panic("losing buffer 2");
2811 qindex = QUEUE_CLEAN;
2812 bp->b_flags |= B_AGE;
2813 /* remaining buffers */
2814 } else if (bp->b_flags & B_DELWRI)
2815 qindex = QUEUE_DIRTY;
2817 qindex = QUEUE_CLEAN;
2819 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2820 panic("brelse: not dirty");
2822 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2823 bp->b_xflags &= ~(BX_CVTENXIO);
2824 /* binsfree unlocks bp. */
2825 binsfree(bp, qindex);
2829 * Release a buffer back to the appropriate queue but do not try to free
2830 * it. The buffer is expected to be used again soon.
2832 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2833 * biodone() to requeue an async I/O on completion. It is also used when
2834 * known good buffers need to be requeued but we think we may need the data
2837 * XXX we should be able to leave the B_RELBUF hint set on completion.
2840 bqrelse(struct buf *bp)
2844 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2845 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2846 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2848 qindex = QUEUE_NONE;
2849 if (BUF_LOCKRECURSED(bp)) {
2850 /* do not release to free list */
2854 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2855 bp->b_xflags &= ~(BX_CVTENXIO);
2857 if (LIST_EMPTY(&bp->b_dep)) {
2858 bp->b_flags &= ~B_IOSTARTED;
2860 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2861 ("bqrelse: SU io not finished bp %p", bp));
2864 if (bp->b_flags & B_MANAGED) {
2865 if (bp->b_flags & B_REMFREE)
2870 /* buffers with stale but valid contents */
2871 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2872 BV_BKGRDERR)) == BV_BKGRDERR) {
2873 BO_LOCK(bp->b_bufobj);
2874 bp->b_vflags &= ~BV_BKGRDERR;
2875 BO_UNLOCK(bp->b_bufobj);
2876 qindex = QUEUE_DIRTY;
2878 if ((bp->b_flags & B_DELWRI) == 0 &&
2879 (bp->b_xflags & BX_VNDIRTY))
2880 panic("bqrelse: not dirty");
2881 if ((bp->b_flags & B_NOREUSE) != 0) {
2885 qindex = QUEUE_CLEAN;
2887 buf_track(bp, __func__);
2888 /* binsfree unlocks bp. */
2889 binsfree(bp, qindex);
2893 buf_track(bp, __func__);
2899 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2900 * restore bogus pages.
2903 vfs_vmio_iodone(struct buf *bp)
2908 struct vnode *vp __unused;
2909 int i, iosize, resid;
2912 obj = bp->b_bufobj->bo_object;
2913 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2914 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2915 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2918 VNPASS(vp->v_holdcnt > 0, vp);
2919 VNPASS(vp->v_object != NULL, vp);
2921 foff = bp->b_offset;
2922 KASSERT(bp->b_offset != NOOFFSET,
2923 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2926 iosize = bp->b_bcount - bp->b_resid;
2927 for (i = 0; i < bp->b_npages; i++) {
2928 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2933 * cleanup bogus pages, restoring the originals
2936 if (m == bogus_page) {
2938 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2940 panic("biodone: page disappeared!");
2942 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2944 * In the write case, the valid and clean bits are
2945 * already changed correctly ( see bdwrite() ), so we
2946 * only need to do this here in the read case.
2948 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2949 resid)) == 0, ("vfs_vmio_iodone: page %p "
2950 "has unexpected dirty bits", m));
2951 vfs_page_set_valid(bp, foff, m);
2953 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2954 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2955 (intmax_t)foff, (uintmax_t)m->pindex));
2958 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2961 vm_object_pip_wakeupn(obj, bp->b_npages);
2962 if (bogus && buf_mapped(bp)) {
2963 BUF_CHECK_MAPPED(bp);
2964 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2965 bp->b_pages, bp->b_npages);
2970 * Perform page invalidation when a buffer is released. The fully invalid
2971 * pages will be reclaimed later in vfs_vmio_truncate().
2974 vfs_vmio_invalidate(struct buf *bp)
2978 int flags, i, resid, poffset, presid;
2980 if (buf_mapped(bp)) {
2981 BUF_CHECK_MAPPED(bp);
2982 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
2984 BUF_CHECK_UNMAPPED(bp);
2986 * Get the base offset and length of the buffer. Note that
2987 * in the VMIO case if the buffer block size is not
2988 * page-aligned then b_data pointer may not be page-aligned.
2989 * But our b_pages[] array *IS* page aligned.
2991 * block sizes less then DEV_BSIZE (usually 512) are not
2992 * supported due to the page granularity bits (m->valid,
2993 * m->dirty, etc...).
2995 * See man buf(9) for more information
2997 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
2998 obj = bp->b_bufobj->bo_object;
2999 resid = bp->b_bufsize;
3000 poffset = bp->b_offset & PAGE_MASK;
3001 VM_OBJECT_WLOCK(obj);
3002 for (i = 0; i < bp->b_npages; i++) {
3004 if (m == bogus_page)
3005 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3006 bp->b_pages[i] = NULL;
3008 presid = resid > (PAGE_SIZE - poffset) ?
3009 (PAGE_SIZE - poffset) : resid;
3010 KASSERT(presid >= 0, ("brelse: extra page"));
3011 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3012 if (pmap_page_wired_mappings(m) == 0)
3013 vm_page_set_invalid(m, poffset, presid);
3015 vm_page_release_locked(m, flags);
3019 VM_OBJECT_WUNLOCK(obj);
3024 * Page-granular truncation of an existing VMIO buffer.
3027 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3033 if (bp->b_npages == desiredpages)
3036 if (buf_mapped(bp)) {
3037 BUF_CHECK_MAPPED(bp);
3038 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3039 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3041 BUF_CHECK_UNMAPPED(bp);
3044 * The object lock is needed only if we will attempt to free pages.
3046 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3047 if ((bp->b_flags & B_DIRECT) != 0) {
3048 flags |= VPR_TRYFREE;
3049 obj = bp->b_bufobj->bo_object;
3050 VM_OBJECT_WLOCK(obj);
3054 for (i = desiredpages; i < bp->b_npages; i++) {
3056 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3057 bp->b_pages[i] = NULL;
3059 vm_page_release_locked(m, flags);
3061 vm_page_release(m, flags);
3064 VM_OBJECT_WUNLOCK(obj);
3065 bp->b_npages = desiredpages;
3069 * Byte granular extension of VMIO buffers.
3072 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3075 * We are growing the buffer, possibly in a
3076 * byte-granular fashion.
3084 * Step 1, bring in the VM pages from the object, allocating
3085 * them if necessary. We must clear B_CACHE if these pages
3086 * are not valid for the range covered by the buffer.
3088 obj = bp->b_bufobj->bo_object;
3089 if (bp->b_npages < desiredpages) {
3090 KASSERT(desiredpages <= atop(maxbcachebuf),
3091 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3092 bp, desiredpages, maxbcachebuf));
3095 * We must allocate system pages since blocking
3096 * here could interfere with paging I/O, no
3097 * matter which process we are.
3099 * Only exclusive busy can be tested here.
3100 * Blocking on shared busy might lead to
3101 * deadlocks once allocbuf() is called after
3102 * pages are vfs_busy_pages().
3104 (void)vm_page_grab_pages_unlocked(obj,
3105 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3106 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3107 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3108 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3109 bp->b_npages = desiredpages;
3113 * Step 2. We've loaded the pages into the buffer,
3114 * we have to figure out if we can still have B_CACHE
3115 * set. Note that B_CACHE is set according to the
3116 * byte-granular range ( bcount and size ), not the
3117 * aligned range ( newbsize ).
3119 * The VM test is against m->valid, which is DEV_BSIZE
3120 * aligned. Needless to say, the validity of the data
3121 * needs to also be DEV_BSIZE aligned. Note that this
3122 * fails with NFS if the server or some other client
3123 * extends the file's EOF. If our buffer is resized,
3124 * B_CACHE may remain set! XXX
3126 toff = bp->b_bcount;
3127 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3128 while ((bp->b_flags & B_CACHE) && toff < size) {
3131 if (tinc > (size - toff))
3133 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3134 m = bp->b_pages[pi];
3135 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3141 * Step 3, fixup the KVA pmap.
3146 BUF_CHECK_UNMAPPED(bp);
3150 * Check to see if a block at a particular lbn is available for a clustered
3154 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3161 /* If the buf isn't in core skip it */
3162 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3165 /* If the buf is busy we don't want to wait for it */
3166 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3169 /* Only cluster with valid clusterable delayed write buffers */
3170 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3171 (B_DELWRI | B_CLUSTEROK))
3174 if (bpa->b_bufsize != size)
3178 * Check to see if it is in the expected place on disk and that the
3179 * block has been mapped.
3181 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3191 * Implement clustered async writes for clearing out B_DELWRI buffers.
3192 * This is much better then the old way of writing only one buffer at
3193 * a time. Note that we may not be presented with the buffers in the
3194 * correct order, so we search for the cluster in both directions.
3197 vfs_bio_awrite(struct buf *bp)
3202 daddr_t lblkno = bp->b_lblkno;
3203 struct vnode *vp = bp->b_vp;
3211 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3213 * right now we support clustered writing only to regular files. If
3214 * we find a clusterable block we could be in the middle of a cluster
3215 * rather then at the beginning.
3217 if ((vp->v_type == VREG) &&
3218 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3219 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3220 size = vp->v_mount->mnt_stat.f_iosize;
3221 maxcl = maxphys / size;
3224 for (i = 1; i < maxcl; i++)
3225 if (vfs_bio_clcheck(vp, size, lblkno + i,
3226 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3229 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3230 if (vfs_bio_clcheck(vp, size, lblkno - j,
3231 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3237 * this is a possible cluster write
3241 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3247 bp->b_flags |= B_ASYNC;
3249 * default (old) behavior, writing out only one block
3251 * XXX returns b_bufsize instead of b_bcount for nwritten?
3253 nwritten = bp->b_bufsize;
3262 * Allocate KVA for an empty buf header according to gbflags.
3265 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3268 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3270 * In order to keep fragmentation sane we only allocate kva
3271 * in BKVASIZE chunks. XXX with vmem we can do page size.
3273 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3275 if (maxsize != bp->b_kvasize &&
3276 bufkva_alloc(bp, maxsize, gbflags))
3285 * Find and initialize a new buffer header, freeing up existing buffers
3286 * in the bufqueues as necessary. The new buffer is returned locked.
3289 * We have insufficient buffer headers
3290 * We have insufficient buffer space
3291 * buffer_arena is too fragmented ( space reservation fails )
3292 * If we have to flush dirty buffers ( but we try to avoid this )
3294 * The caller is responsible for releasing the reserved bufspace after
3295 * allocbuf() is called.
3298 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3300 struct bufdomain *bd;
3302 bool metadata, reserved;
3305 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3306 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3307 if (!unmapped_buf_allowed)
3308 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3310 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3318 bd = &bdomain[vp->v_bufobj.bo_domain];
3320 counter_u64_add(getnewbufcalls, 1);
3323 if (reserved == false &&
3324 bufspace_reserve(bd, maxsize, metadata) != 0) {
3325 counter_u64_add(getnewbufrestarts, 1);
3329 if ((bp = buf_alloc(bd)) == NULL) {
3330 counter_u64_add(getnewbufrestarts, 1);
3333 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3336 } while (buf_recycle(bd, false) == 0);
3339 bufspace_release(bd, maxsize);
3341 bp->b_flags |= B_INVAL;
3344 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3352 * buffer flushing daemon. Buffers are normally flushed by the
3353 * update daemon but if it cannot keep up this process starts to
3354 * take the load in an attempt to prevent getnewbuf() from blocking.
3356 static struct kproc_desc buf_kp = {
3361 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3364 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3368 flushed = flushbufqueues(vp, bd, target, 0);
3371 * Could not find any buffers without rollback
3372 * dependencies, so just write the first one
3373 * in the hopes of eventually making progress.
3375 if (vp != NULL && target > 2)
3377 flushbufqueues(vp, bd, target, 1);
3385 struct bufdomain *bd;
3391 * This process needs to be suspended prior to shutdown sync.
3393 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread,
3394 SHUTDOWN_PRI_LAST + 100);
3397 * Start the buf clean daemons as children threads.
3399 for (i = 0 ; i < buf_domains; i++) {
3402 error = kthread_add((void (*)(void *))bufspace_daemon,
3403 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3405 panic("error %d spawning bufspace daemon", error);
3409 * This process is allowed to take the buffer cache to the limit
3411 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3415 mtx_unlock(&bdlock);
3417 kthread_suspend_check();
3420 * Save speedupreq for this pass and reset to capture new
3423 speedupreq = bd_speedupreq;
3427 * Flush each domain sequentially according to its level and
3428 * the speedup request.
3430 for (i = 0; i < buf_domains; i++) {
3433 lodirty = bd->bd_numdirtybuffers / 2;
3435 lodirty = bd->bd_lodirtybuffers;
3436 while (bd->bd_numdirtybuffers > lodirty) {
3437 if (buf_flush(NULL, bd,
3438 bd->bd_numdirtybuffers - lodirty) == 0)
3440 kern_yield(PRI_USER);
3445 * Only clear bd_request if we have reached our low water
3446 * mark. The buf_daemon normally waits 1 second and
3447 * then incrementally flushes any dirty buffers that have
3448 * built up, within reason.
3450 * If we were unable to hit our low water mark and couldn't
3451 * find any flushable buffers, we sleep for a short period
3452 * to avoid endless loops on unlockable buffers.
3455 if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3457 * We reached our low water mark, reset the
3458 * request and sleep until we are needed again.
3459 * The sleep is just so the suspend code works.
3463 * Do an extra wakeup in case dirty threshold
3464 * changed via sysctl and the explicit transition
3465 * out of shortfall was missed.
3468 if (runningbufspace <= lorunningspace)
3470 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3473 * We couldn't find any flushable dirty buffers but
3474 * still have too many dirty buffers, we
3475 * have to sleep and try again. (rare)
3477 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3485 * Try to flush a buffer in the dirty queue. We must be careful to
3486 * free up B_INVAL buffers instead of write them, which NFS is
3487 * particularly sensitive to.
3489 static int flushwithdeps = 0;
3490 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3492 "Number of buffers flushed with dependecies that require rollbacks");
3495 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3498 struct bufqueue *bq;
3499 struct buf *sentinel;
3509 bq = &bd->bd_dirtyq;
3511 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3512 sentinel->b_qindex = QUEUE_SENTINEL;
3514 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3516 while (flushed != target) {
3519 bp = TAILQ_NEXT(sentinel, b_freelist);
3521 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3522 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3529 * Skip sentinels inserted by other invocations of the
3530 * flushbufqueues(), taking care to not reorder them.
3532 * Only flush the buffers that belong to the
3533 * vnode locked by the curthread.
3535 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3540 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3546 * BKGRDINPROG can only be set with the buf and bufobj
3547 * locks both held. We tolerate a race to clear it here.
3549 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3550 (bp->b_flags & B_DELWRI) == 0) {
3554 if (bp->b_flags & B_INVAL) {
3561 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3562 if (flushdeps == 0) {
3570 * We must hold the lock on a vnode before writing
3571 * one of its buffers. Otherwise we may confuse, or
3572 * in the case of a snapshot vnode, deadlock the
3575 * The lock order here is the reverse of the normal
3576 * of vnode followed by buf lock. This is ok because
3577 * the NOWAIT will prevent deadlock.
3580 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3586 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3588 ASSERT_VOP_LOCKED(vp, "getbuf");
3590 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3591 vn_lock(vp, LK_TRYUPGRADE);
3594 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3595 bp, bp->b_vp, bp->b_flags);
3596 if (curproc == bufdaemonproc) {
3601 counter_u64_add(notbufdflushes, 1);
3603 vn_finished_write(mp);
3606 flushwithdeps += hasdeps;
3610 * Sleeping on runningbufspace while holding
3611 * vnode lock leads to deadlock.
3613 if (curproc == bufdaemonproc &&
3614 runningbufspace > hirunningspace)
3615 waitrunningbufspace();
3618 vn_finished_write(mp);
3622 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3624 free(sentinel, M_TEMP);
3629 * Check to see if a block is currently memory resident.
3632 incore(struct bufobj *bo, daddr_t blkno)
3634 return (gbincore_unlocked(bo, blkno));
3638 * Returns true if no I/O is needed to access the
3639 * associated VM object. This is like incore except
3640 * it also hunts around in the VM system for the data.
3643 inmem(struct vnode * vp, daddr_t blkno)
3646 vm_offset_t toff, tinc, size;
3651 ASSERT_VOP_LOCKED(vp, "inmem");
3653 if (incore(&vp->v_bufobj, blkno))
3655 if (vp->v_mount == NULL)
3662 if (size > vp->v_mount->mnt_stat.f_iosize)
3663 size = vp->v_mount->mnt_stat.f_iosize;
3664 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3666 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3667 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3673 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3674 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3676 * Consider page validity only if page mapping didn't change
3679 valid = vm_page_is_valid(m,
3680 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3681 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3693 * Set the dirty range for a buffer based on the status of the dirty
3694 * bits in the pages comprising the buffer. The range is limited
3695 * to the size of the buffer.
3697 * Tell the VM system that the pages associated with this buffer
3698 * are clean. This is used for delayed writes where the data is
3699 * going to go to disk eventually without additional VM intevention.
3701 * Note that while we only really need to clean through to b_bcount, we
3702 * just go ahead and clean through to b_bufsize.
3705 vfs_clean_pages_dirty_buf(struct buf *bp)
3707 vm_ooffset_t foff, noff, eoff;
3711 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3714 foff = bp->b_offset;
3715 KASSERT(bp->b_offset != NOOFFSET,
3716 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3718 vfs_busy_pages_acquire(bp);
3719 vfs_setdirty_range(bp);
3720 for (i = 0; i < bp->b_npages; i++) {
3721 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3723 if (eoff > bp->b_offset + bp->b_bufsize)
3724 eoff = bp->b_offset + bp->b_bufsize;
3726 vfs_page_set_validclean(bp, foff, m);
3727 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3730 vfs_busy_pages_release(bp);
3734 vfs_setdirty_range(struct buf *bp)
3736 vm_offset_t boffset;
3737 vm_offset_t eoffset;
3741 * test the pages to see if they have been modified directly
3742 * by users through the VM system.
3744 for (i = 0; i < bp->b_npages; i++)
3745 vm_page_test_dirty(bp->b_pages[i]);
3748 * Calculate the encompassing dirty range, boffset and eoffset,
3749 * (eoffset - boffset) bytes.
3752 for (i = 0; i < bp->b_npages; i++) {
3753 if (bp->b_pages[i]->dirty)
3756 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3758 for (i = bp->b_npages - 1; i >= 0; --i) {
3759 if (bp->b_pages[i]->dirty) {
3763 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3766 * Fit it to the buffer.
3769 if (eoffset > bp->b_bcount)
3770 eoffset = bp->b_bcount;
3773 * If we have a good dirty range, merge with the existing
3777 if (boffset < eoffset) {
3778 if (bp->b_dirtyoff > boffset)
3779 bp->b_dirtyoff = boffset;
3780 if (bp->b_dirtyend < eoffset)
3781 bp->b_dirtyend = eoffset;
3786 * Allocate the KVA mapping for an existing buffer.
3787 * If an unmapped buffer is provided but a mapped buffer is requested, take
3788 * also care to properly setup mappings between pages and KVA.
3791 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3793 int bsize, maxsize, need_mapping, need_kva;
3796 need_mapping = bp->b_data == unmapped_buf &&
3797 (gbflags & GB_UNMAPPED) == 0;
3798 need_kva = bp->b_kvabase == unmapped_buf &&
3799 bp->b_data == unmapped_buf &&
3800 (gbflags & GB_KVAALLOC) != 0;
3801 if (!need_mapping && !need_kva)
3804 BUF_CHECK_UNMAPPED(bp);
3806 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3808 * Buffer is not mapped, but the KVA was already
3809 * reserved at the time of the instantiation. Use the
3816 * Calculate the amount of the address space we would reserve
3817 * if the buffer was mapped.
3819 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3820 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3821 offset = blkno * bsize;
3822 maxsize = size + (offset & PAGE_MASK);
3823 maxsize = imax(maxsize, bsize);
3825 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3826 if ((gbflags & GB_NOWAIT_BD) != 0) {
3828 * XXXKIB: defragmentation cannot
3829 * succeed, not sure what else to do.
3831 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3833 counter_u64_add(mappingrestarts, 1);
3834 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3838 /* b_offset is handled by bpmap_qenter. */
3839 bp->b_data = bp->b_kvabase;
3840 BUF_CHECK_MAPPED(bp);
3846 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3852 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3861 * Get a block given a specified block and offset into a file/device.
3862 * The buffers B_DONE bit will be cleared on return, making it almost
3863 * ready for an I/O initiation. B_INVAL may or may not be set on
3864 * return. The caller should clear B_INVAL prior to initiating a
3867 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3868 * an existing buffer.
3870 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3871 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3872 * and then cleared based on the backing VM. If the previous buffer is
3873 * non-0-sized but invalid, B_CACHE will be cleared.
3875 * If getblk() must create a new buffer, the new buffer is returned with
3876 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3877 * case it is returned with B_INVAL clear and B_CACHE set based on the
3880 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3881 * B_CACHE bit is clear.
3883 * What this means, basically, is that the caller should use B_CACHE to
3884 * determine whether the buffer is fully valid or not and should clear
3885 * B_INVAL prior to issuing a read. If the caller intends to validate
3886 * the buffer by loading its data area with something, the caller needs
3887 * to clear B_INVAL. If the caller does this without issuing an I/O,
3888 * the caller should set B_CACHE ( as an optimization ), else the caller
3889 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3890 * a write attempt or if it was a successful read. If the caller
3891 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3892 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3894 * The blkno parameter is the logical block being requested. Normally
3895 * the mapping of logical block number to disk block address is done
3896 * by calling VOP_BMAP(). However, if the mapping is already known, the
3897 * disk block address can be passed using the dblkno parameter. If the
3898 * disk block address is not known, then the same value should be passed
3899 * for blkno and dblkno.
3902 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3903 int slptimeo, int flags, struct buf **bpp)
3908 int bsize, error, maxsize, vmio;
3911 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3912 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3913 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3914 ASSERT_VOP_LOCKED(vp, "getblk");
3915 if (size > maxbcachebuf)
3916 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3918 if (!unmapped_buf_allowed)
3919 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3924 /* Attempt lockless lookup first. */
3925 bp = gbincore_unlocked(bo, blkno);
3928 * With GB_NOCREAT we must be sure about not finding the buffer
3929 * as it may have been reassigned during unlocked lookup.
3931 if ((flags & GB_NOCREAT) != 0)
3933 goto newbuf_unlocked;
3936 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3941 /* Verify buf identify has not changed since lookup. */
3942 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3943 goto foundbuf_fastpath;
3945 /* It changed, fallback to locked lookup. */
3950 bp = gbincore(bo, blkno);
3955 * Buffer is in-core. If the buffer is not busy nor managed,
3956 * it must be on a queue.
3958 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3959 ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
3961 error = BUF_TIMELOCK(bp, lockflags,
3962 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3965 * If we slept and got the lock we have to restart in case
3966 * the buffer changed identities.
3968 if (error == ENOLCK)
3970 /* We timed out or were interrupted. */
3971 else if (error != 0)
3975 /* If recursed, assume caller knows the rules. */
3976 if (BUF_LOCKRECURSED(bp))
3980 * The buffer is locked. B_CACHE is cleared if the buffer is
3981 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3982 * and for a VMIO buffer B_CACHE is adjusted according to the
3985 if (bp->b_flags & B_INVAL)
3986 bp->b_flags &= ~B_CACHE;
3987 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3988 bp->b_flags |= B_CACHE;
3989 if (bp->b_flags & B_MANAGED)
3990 MPASS(bp->b_qindex == QUEUE_NONE);
3995 * check for size inconsistencies for non-VMIO case.
3997 if (bp->b_bcount != size) {
3998 if ((bp->b_flags & B_VMIO) == 0 ||
3999 (size > bp->b_kvasize)) {
4000 if (bp->b_flags & B_DELWRI) {
4001 bp->b_flags |= B_NOCACHE;
4004 if (LIST_EMPTY(&bp->b_dep)) {
4005 bp->b_flags |= B_RELBUF;
4008 bp->b_flags |= B_NOCACHE;
4017 * Handle the case of unmapped buffer which should
4018 * become mapped, or the buffer for which KVA
4019 * reservation is requested.
4021 bp_unmapped_get_kva(bp, blkno, size, flags);
4024 * If the size is inconsistent in the VMIO case, we can resize
4025 * the buffer. This might lead to B_CACHE getting set or
4026 * cleared. If the size has not changed, B_CACHE remains
4027 * unchanged from its previous state.
4031 KASSERT(bp->b_offset != NOOFFSET,
4032 ("getblk: no buffer offset"));
4035 * A buffer with B_DELWRI set and B_CACHE clear must
4036 * be committed before we can return the buffer in
4037 * order to prevent the caller from issuing a read
4038 * ( due to B_CACHE not being set ) and overwriting
4041 * Most callers, including NFS and FFS, need this to
4042 * operate properly either because they assume they
4043 * can issue a read if B_CACHE is not set, or because
4044 * ( for example ) an uncached B_DELWRI might loop due
4045 * to softupdates re-dirtying the buffer. In the latter
4046 * case, B_CACHE is set after the first write completes,
4047 * preventing further loops.
4048 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4049 * above while extending the buffer, we cannot allow the
4050 * buffer to remain with B_CACHE set after the write
4051 * completes or it will represent a corrupt state. To
4052 * deal with this we set B_NOCACHE to scrap the buffer
4055 * We might be able to do something fancy, like setting
4056 * B_CACHE in bwrite() except if B_DELWRI is already set,
4057 * so the below call doesn't set B_CACHE, but that gets real
4058 * confusing. This is much easier.
4061 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4062 bp->b_flags |= B_NOCACHE;
4066 bp->b_flags &= ~B_DONE;
4069 * Buffer is not in-core, create new buffer. The buffer
4070 * returned by getnewbuf() is locked. Note that the returned
4071 * buffer is also considered valid (not marked B_INVAL).
4076 * If the user does not want us to create the buffer, bail out
4079 if (flags & GB_NOCREAT)
4082 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4083 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4084 offset = blkno * bsize;
4085 vmio = vp->v_object != NULL;
4087 maxsize = size + (offset & PAGE_MASK);
4090 /* Do not allow non-VMIO notmapped buffers. */
4091 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4093 maxsize = imax(maxsize, bsize);
4094 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4096 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4097 KASSERT(error != EOPNOTSUPP,
4098 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4103 return (EJUSTRETURN);
4106 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4108 if (slpflag || slptimeo)
4111 * XXX This is here until the sleep path is diagnosed
4112 * enough to work under very low memory conditions.
4114 * There's an issue on low memory, 4BSD+non-preempt
4115 * systems (eg MIPS routers with 32MB RAM) where buffer
4116 * exhaustion occurs without sleeping for buffer
4117 * reclaimation. This just sticks in a loop and
4118 * constantly attempts to allocate a buffer, which
4119 * hits exhaustion and tries to wakeup bufdaemon.
4120 * This never happens because we never yield.
4122 * The real solution is to identify and fix these cases
4123 * so we aren't effectively busy-waiting in a loop
4124 * until the reclaimation path has cycles to run.
4126 kern_yield(PRI_USER);
4131 * This code is used to make sure that a buffer is not
4132 * created while the getnewbuf routine is blocked.
4133 * This can be a problem whether the vnode is locked or not.
4134 * If the buffer is created out from under us, we have to
4135 * throw away the one we just created.
4137 * Note: this must occur before we associate the buffer
4138 * with the vp especially considering limitations in
4139 * the splay tree implementation when dealing with duplicate
4143 if (gbincore(bo, blkno)) {
4145 bp->b_flags |= B_INVAL;
4146 bufspace_release(bufdomain(bp), maxsize);
4152 * Insert the buffer into the hash, so that it can
4153 * be found by incore.
4155 bp->b_lblkno = blkno;
4156 bp->b_blkno = d_blkno;
4157 bp->b_offset = offset;
4162 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4163 * buffer size starts out as 0, B_CACHE will be set by
4164 * allocbuf() for the VMIO case prior to it testing the
4165 * backing store for validity.
4169 bp->b_flags |= B_VMIO;
4170 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4171 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4172 bp, vp->v_object, bp->b_bufobj->bo_object));
4174 bp->b_flags &= ~B_VMIO;
4175 KASSERT(bp->b_bufobj->bo_object == NULL,
4176 ("ARGH! has b_bufobj->bo_object %p %p\n",
4177 bp, bp->b_bufobj->bo_object));
4178 BUF_CHECK_MAPPED(bp);
4182 bufspace_release(bufdomain(bp), maxsize);
4183 bp->b_flags &= ~B_DONE;
4185 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4187 buf_track(bp, __func__);
4188 KASSERT(bp->b_bufobj == bo,
4189 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4195 * Get an empty, disassociated buffer of given size. The buffer is initially
4199 geteblk(int size, int flags)
4204 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4205 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4206 if ((flags & GB_NOWAIT_BD) &&
4207 (curthread->td_pflags & TDP_BUFNEED) != 0)
4211 bufspace_release(bufdomain(bp), maxsize);
4212 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4217 * Truncate the backing store for a non-vmio buffer.
4220 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4223 if (bp->b_flags & B_MALLOC) {
4225 * malloced buffers are not shrunk
4227 if (newbsize == 0) {
4228 bufmallocadjust(bp, 0);
4229 free(bp->b_data, M_BIOBUF);
4230 bp->b_data = bp->b_kvabase;
4231 bp->b_flags &= ~B_MALLOC;
4235 vm_hold_free_pages(bp, newbsize);
4236 bufspace_adjust(bp, newbsize);
4240 * Extend the backing for a non-VMIO buffer.
4243 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4249 * We only use malloced memory on the first allocation.
4250 * and revert to page-allocated memory when the buffer
4253 * There is a potential smp race here that could lead
4254 * to bufmallocspace slightly passing the max. It
4255 * is probably extremely rare and not worth worrying
4258 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4259 bufmallocspace < maxbufmallocspace) {
4260 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4261 bp->b_flags |= B_MALLOC;
4262 bufmallocadjust(bp, newbsize);
4267 * If the buffer is growing on its other-than-first
4268 * allocation then we revert to the page-allocation
4273 if (bp->b_flags & B_MALLOC) {
4274 origbuf = bp->b_data;
4275 origbufsize = bp->b_bufsize;
4276 bp->b_data = bp->b_kvabase;
4277 bufmallocadjust(bp, 0);
4278 bp->b_flags &= ~B_MALLOC;
4279 newbsize = round_page(newbsize);
4281 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4282 (vm_offset_t) bp->b_data + newbsize);
4283 if (origbuf != NULL) {
4284 bcopy(origbuf, bp->b_data, origbufsize);
4285 free(origbuf, M_BIOBUF);
4287 bufspace_adjust(bp, newbsize);
4291 * This code constitutes the buffer memory from either anonymous system
4292 * memory (in the case of non-VMIO operations) or from an associated
4293 * VM object (in the case of VMIO operations). This code is able to
4294 * resize a buffer up or down.
4296 * Note that this code is tricky, and has many complications to resolve
4297 * deadlock or inconsistent data situations. Tread lightly!!!
4298 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4299 * the caller. Calling this code willy nilly can result in the loss of data.
4301 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4302 * B_CACHE for the non-VMIO case.
4305 allocbuf(struct buf *bp, int size)
4309 if (bp->b_bcount == size)
4312 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4313 panic("allocbuf: buffer too small");
4315 newbsize = roundup2(size, DEV_BSIZE);
4316 if ((bp->b_flags & B_VMIO) == 0) {
4317 if ((bp->b_flags & B_MALLOC) == 0)
4318 newbsize = round_page(newbsize);
4320 * Just get anonymous memory from the kernel. Don't
4321 * mess with B_CACHE.
4323 if (newbsize < bp->b_bufsize)
4324 vfs_nonvmio_truncate(bp, newbsize);
4325 else if (newbsize > bp->b_bufsize)
4326 vfs_nonvmio_extend(bp, newbsize);
4330 desiredpages = (size == 0) ? 0 :
4331 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4333 if (bp->b_flags & B_MALLOC)
4334 panic("allocbuf: VMIO buffer can't be malloced");
4336 * Set B_CACHE initially if buffer is 0 length or will become
4339 if (size == 0 || bp->b_bufsize == 0)
4340 bp->b_flags |= B_CACHE;
4342 if (newbsize < bp->b_bufsize)
4343 vfs_vmio_truncate(bp, desiredpages);
4344 /* XXX This looks as if it should be newbsize > b_bufsize */
4345 else if (size > bp->b_bcount)
4346 vfs_vmio_extend(bp, desiredpages, size);
4347 bufspace_adjust(bp, newbsize);
4349 bp->b_bcount = size; /* requested buffer size. */
4353 extern int inflight_transient_maps;
4355 static struct bio_queue nondump_bios;
4358 biodone(struct bio *bp)
4361 void (*done)(struct bio *);
4362 vm_offset_t start, end;
4364 biotrack(bp, __func__);
4367 * Avoid completing I/O when dumping after a panic since that may
4368 * result in a deadlock in the filesystem or pager code. Note that
4369 * this doesn't affect dumps that were started manually since we aim
4370 * to keep the system usable after it has been resumed.
4372 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4373 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4376 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4377 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4378 bp->bio_flags |= BIO_UNMAPPED;
4379 start = trunc_page((vm_offset_t)bp->bio_data);
4380 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4381 bp->bio_data = unmapped_buf;
4382 pmap_qremove(start, atop(end - start));
4383 vmem_free(transient_arena, start, end - start);
4384 atomic_add_int(&inflight_transient_maps, -1);
4386 done = bp->bio_done;
4388 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4390 bp->bio_flags |= BIO_DONE;
4398 * Wait for a BIO to finish.
4401 biowait(struct bio *bp, const char *wchan)
4405 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4407 while ((bp->bio_flags & BIO_DONE) == 0)
4408 msleep(bp, mtxp, PRIBIO, wchan, 0);
4410 if (bp->bio_error != 0)
4411 return (bp->bio_error);
4412 if (!(bp->bio_flags & BIO_ERROR))
4418 biofinish(struct bio *bp, struct devstat *stat, int error)
4422 bp->bio_error = error;
4423 bp->bio_flags |= BIO_ERROR;
4426 devstat_end_transaction_bio(stat, bp);
4430 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4432 biotrack_buf(struct bio *bp, const char *location)
4435 buf_track(bp->bio_track_bp, location);
4442 * Wait for buffer I/O completion, returning error status. The buffer
4443 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4444 * error and cleared.
4447 bufwait(struct buf *bp)
4449 if (bp->b_iocmd == BIO_READ)
4450 bwait(bp, PRIBIO, "biord");
4452 bwait(bp, PRIBIO, "biowr");
4453 if (bp->b_flags & B_EINTR) {
4454 bp->b_flags &= ~B_EINTR;
4457 if (bp->b_ioflags & BIO_ERROR) {
4458 return (bp->b_error ? bp->b_error : EIO);
4467 * Finish I/O on a buffer, optionally calling a completion function.
4468 * This is usually called from an interrupt so process blocking is
4471 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4472 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4473 * assuming B_INVAL is clear.
4475 * For the VMIO case, we set B_CACHE if the op was a read and no
4476 * read error occurred, or if the op was a write. B_CACHE is never
4477 * set if the buffer is invalid or otherwise uncacheable.
4479 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4480 * initiator to leave B_INVAL set to brelse the buffer out of existence
4481 * in the biodone routine.
4484 bufdone(struct buf *bp)
4486 struct bufobj *dropobj;
4487 void (*biodone)(struct buf *);
4489 buf_track(bp, __func__);
4490 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4493 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4495 runningbufwakeup(bp);
4496 if (bp->b_iocmd == BIO_WRITE)
4497 dropobj = bp->b_bufobj;
4498 /* call optional completion function if requested */
4499 if (bp->b_iodone != NULL) {
4500 biodone = bp->b_iodone;
4501 bp->b_iodone = NULL;
4504 bufobj_wdrop(dropobj);
4507 if (bp->b_flags & B_VMIO) {
4509 * Set B_CACHE if the op was a normal read and no error
4510 * occurred. B_CACHE is set for writes in the b*write()
4513 if (bp->b_iocmd == BIO_READ &&
4514 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4515 !(bp->b_ioflags & BIO_ERROR))
4516 bp->b_flags |= B_CACHE;
4517 vfs_vmio_iodone(bp);
4519 if (!LIST_EMPTY(&bp->b_dep))
4521 if ((bp->b_flags & B_CKHASH) != 0) {
4522 KASSERT(bp->b_iocmd == BIO_READ,
4523 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4524 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4525 (*bp->b_ckhashcalc)(bp);
4528 * For asynchronous completions, release the buffer now. The brelse
4529 * will do a wakeup there if necessary - so no need to do a wakeup
4530 * here in the async case. The sync case always needs to do a wakeup.
4532 if (bp->b_flags & B_ASYNC) {
4533 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4534 (bp->b_ioflags & BIO_ERROR))
4541 bufobj_wdrop(dropobj);
4545 * This routine is called in lieu of iodone in the case of
4546 * incomplete I/O. This keeps the busy status for pages
4550 vfs_unbusy_pages(struct buf *bp)
4556 runningbufwakeup(bp);
4557 if (!(bp->b_flags & B_VMIO))
4560 obj = bp->b_bufobj->bo_object;
4561 for (i = 0; i < bp->b_npages; i++) {
4563 if (m == bogus_page) {
4564 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4566 panic("vfs_unbusy_pages: page missing\n");
4568 if (buf_mapped(bp)) {
4569 BUF_CHECK_MAPPED(bp);
4570 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4571 bp->b_pages, bp->b_npages);
4573 BUF_CHECK_UNMAPPED(bp);
4577 vm_object_pip_wakeupn(obj, bp->b_npages);
4581 * vfs_page_set_valid:
4583 * Set the valid bits in a page based on the supplied offset. The
4584 * range is restricted to the buffer's size.
4586 * This routine is typically called after a read completes.
4589 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4594 * Compute the end offset, eoff, such that [off, eoff) does not span a
4595 * page boundary and eoff is not greater than the end of the buffer.
4596 * The end of the buffer, in this case, is our file EOF, not the
4597 * allocation size of the buffer.
4599 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4600 if (eoff > bp->b_offset + bp->b_bcount)
4601 eoff = bp->b_offset + bp->b_bcount;
4604 * Set valid range. This is typically the entire buffer and thus the
4608 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4612 * vfs_page_set_validclean:
4614 * Set the valid bits and clear the dirty bits in a page based on the
4615 * supplied offset. The range is restricted to the buffer's size.
4618 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4620 vm_ooffset_t soff, eoff;
4623 * Start and end offsets in buffer. eoff - soff may not cross a
4624 * page boundary or cross the end of the buffer. The end of the
4625 * buffer, in this case, is our file EOF, not the allocation size
4629 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4630 if (eoff > bp->b_offset + bp->b_bcount)
4631 eoff = bp->b_offset + bp->b_bcount;
4634 * Set valid range. This is typically the entire buffer and thus the
4638 vm_page_set_validclean(
4640 (vm_offset_t) (soff & PAGE_MASK),
4641 (vm_offset_t) (eoff - soff)
4647 * Acquire a shared busy on all pages in the buf.
4650 vfs_busy_pages_acquire(struct buf *bp)
4654 for (i = 0; i < bp->b_npages; i++)
4655 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4659 vfs_busy_pages_release(struct buf *bp)
4663 for (i = 0; i < bp->b_npages; i++)
4664 vm_page_sunbusy(bp->b_pages[i]);
4668 * This routine is called before a device strategy routine.
4669 * It is used to tell the VM system that paging I/O is in
4670 * progress, and treat the pages associated with the buffer
4671 * almost as being exclusive busy. Also the object paging_in_progress
4672 * flag is handled to make sure that the object doesn't become
4675 * Since I/O has not been initiated yet, certain buffer flags
4676 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4677 * and should be ignored.
4680 vfs_busy_pages(struct buf *bp, int clear_modify)
4688 if (!(bp->b_flags & B_VMIO))
4691 obj = bp->b_bufobj->bo_object;
4692 foff = bp->b_offset;
4693 KASSERT(bp->b_offset != NOOFFSET,
4694 ("vfs_busy_pages: no buffer offset"));
4695 if ((bp->b_flags & B_CLUSTER) == 0) {
4696 vm_object_pip_add(obj, bp->b_npages);
4697 vfs_busy_pages_acquire(bp);
4699 if (bp->b_bufsize != 0)
4700 vfs_setdirty_range(bp);
4702 for (i = 0; i < bp->b_npages; i++) {
4704 vm_page_assert_sbusied(m);
4707 * When readying a buffer for a read ( i.e
4708 * clear_modify == 0 ), it is important to do
4709 * bogus_page replacement for valid pages in
4710 * partially instantiated buffers. Partially
4711 * instantiated buffers can, in turn, occur when
4712 * reconstituting a buffer from its VM backing store
4713 * base. We only have to do this if B_CACHE is
4714 * clear ( which causes the I/O to occur in the
4715 * first place ). The replacement prevents the read
4716 * I/O from overwriting potentially dirty VM-backed
4717 * pages. XXX bogus page replacement is, uh, bogus.
4718 * It may not work properly with small-block devices.
4719 * We need to find a better way.
4722 pmap_remove_write(m);
4723 vfs_page_set_validclean(bp, foff, m);
4724 } else if (vm_page_all_valid(m) &&
4725 (bp->b_flags & B_CACHE) == 0) {
4726 bp->b_pages[i] = bogus_page;
4729 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4731 if (bogus && buf_mapped(bp)) {
4732 BUF_CHECK_MAPPED(bp);
4733 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4734 bp->b_pages, bp->b_npages);
4739 * vfs_bio_set_valid:
4741 * Set the range within the buffer to valid. The range is
4742 * relative to the beginning of the buffer, b_offset. Note that
4743 * b_offset itself may be offset from the beginning of the first
4747 vfs_bio_set_valid(struct buf *bp, int base, int size)
4752 if (!(bp->b_flags & B_VMIO))
4756 * Fixup base to be relative to beginning of first page.
4757 * Set initial n to be the maximum number of bytes in the
4758 * first page that can be validated.
4760 base += (bp->b_offset & PAGE_MASK);
4761 n = PAGE_SIZE - (base & PAGE_MASK);
4764 * Busy may not be strictly necessary here because the pages are
4765 * unlikely to be fully valid and the vnode lock will synchronize
4766 * their access via getpages. It is grabbed for consistency with
4767 * other page validation.
4769 vfs_busy_pages_acquire(bp);
4770 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4774 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4779 vfs_busy_pages_release(bp);
4785 * If the specified buffer is a non-VMIO buffer, clear the entire
4786 * buffer. If the specified buffer is a VMIO buffer, clear and
4787 * validate only the previously invalid portions of the buffer.
4788 * This routine essentially fakes an I/O, so we need to clear
4789 * BIO_ERROR and B_INVAL.
4791 * Note that while we only theoretically need to clear through b_bcount,
4792 * we go ahead and clear through b_bufsize.
4795 vfs_bio_clrbuf(struct buf *bp)
4797 int i, j, mask, sa, ea, slide;
4799 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4803 bp->b_flags &= ~B_INVAL;
4804 bp->b_ioflags &= ~BIO_ERROR;
4805 vfs_busy_pages_acquire(bp);
4806 sa = bp->b_offset & PAGE_MASK;
4808 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4809 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4810 ea = slide & PAGE_MASK;
4813 if (bp->b_pages[i] == bogus_page)
4816 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4817 if ((bp->b_pages[i]->valid & mask) == mask)
4819 if ((bp->b_pages[i]->valid & mask) == 0)
4820 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4822 for (; sa < ea; sa += DEV_BSIZE, j++) {
4823 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4824 pmap_zero_page_area(bp->b_pages[i],
4829 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4830 roundup2(ea - sa, DEV_BSIZE));
4832 vfs_busy_pages_release(bp);
4837 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4842 if (buf_mapped(bp)) {
4843 BUF_CHECK_MAPPED(bp);
4844 bzero(bp->b_data + base, size);
4846 BUF_CHECK_UNMAPPED(bp);
4847 n = PAGE_SIZE - (base & PAGE_MASK);
4848 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4852 pmap_zero_page_area(m, base & PAGE_MASK, n);
4861 * Update buffer flags based on I/O request parameters, optionally releasing the
4862 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4863 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4864 * I/O). Otherwise the buffer is released to the cache.
4867 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4870 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4871 ("buf %p non-VMIO noreuse", bp));
4873 if ((ioflag & IO_DIRECT) != 0)
4874 bp->b_flags |= B_DIRECT;
4875 if ((ioflag & IO_EXT) != 0)
4876 bp->b_xflags |= BX_ALTDATA;
4877 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4878 bp->b_flags |= B_RELBUF;
4879 if ((ioflag & IO_NOREUSE) != 0)
4880 bp->b_flags |= B_NOREUSE;
4888 vfs_bio_brelse(struct buf *bp, int ioflag)
4891 b_io_dismiss(bp, ioflag, true);
4895 vfs_bio_set_flags(struct buf *bp, int ioflag)
4898 b_io_dismiss(bp, ioflag, false);
4902 * vm_hold_load_pages and vm_hold_free_pages get pages into
4903 * a buffers address space. The pages are anonymous and are
4904 * not associated with a file object.
4907 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4913 BUF_CHECK_MAPPED(bp);
4915 to = round_page(to);
4916 from = round_page(from);
4917 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4918 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4919 KASSERT(to - from <= maxbcachebuf,
4920 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4921 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4923 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4925 * note: must allocate system pages since blocking here
4926 * could interfere with paging I/O, no matter which
4929 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4930 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) |
4932 pmap_qenter(pg, &p, 1);
4933 bp->b_pages[index] = p;
4935 bp->b_npages = index;
4938 /* Return pages associated with this buf to the vm system */
4940 vm_hold_free_pages(struct buf *bp, int newbsize)
4944 int index, newnpages;
4946 BUF_CHECK_MAPPED(bp);
4948 from = round_page((vm_offset_t)bp->b_data + newbsize);
4949 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4950 if (bp->b_npages > newnpages)
4951 pmap_qremove(from, bp->b_npages - newnpages);
4952 for (index = newnpages; index < bp->b_npages; index++) {
4953 p = bp->b_pages[index];
4954 bp->b_pages[index] = NULL;
4955 vm_page_unwire_noq(p);
4958 bp->b_npages = newnpages;
4962 * Map an IO request into kernel virtual address space.
4964 * All requests are (re)mapped into kernel VA space.
4965 * Notice that we use b_bufsize for the size of the buffer
4966 * to be mapped. b_bcount might be modified by the driver.
4968 * Note that even if the caller determines that the address space should
4969 * be valid, a race or a smaller-file mapped into a larger space may
4970 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4971 * check the return value.
4973 * This function only works with pager buffers.
4976 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
4981 MPASS((bp->b_flags & B_MAXPHYS) != 0);
4982 prot = VM_PROT_READ;
4983 if (bp->b_iocmd == BIO_READ)
4984 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4985 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4986 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
4989 bp->b_bufsize = len;
4990 bp->b_npages = pidx;
4991 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
4992 if (mapbuf || !unmapped_buf_allowed) {
4993 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
4994 bp->b_data = bp->b_kvabase + bp->b_offset;
4996 bp->b_data = unmapped_buf;
5001 * Free the io map PTEs associated with this IO operation.
5002 * We also invalidate the TLB entries and restore the original b_addr.
5004 * This function only works with pager buffers.
5007 vunmapbuf(struct buf *bp)
5011 npages = bp->b_npages;
5013 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5014 vm_page_unhold_pages(bp->b_pages, npages);
5016 bp->b_data = unmapped_buf;
5020 bdone(struct buf *bp)
5024 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5026 bp->b_flags |= B_DONE;
5032 bwait(struct buf *bp, u_char pri, const char *wchan)
5036 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5038 while ((bp->b_flags & B_DONE) == 0)
5039 msleep(bp, mtxp, pri, wchan, 0);
5044 bufsync(struct bufobj *bo, int waitfor)
5047 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5051 bufstrategy(struct bufobj *bo, struct buf *bp)
5057 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5058 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5059 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5060 i = VOP_STRATEGY(vp, bp);
5061 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5065 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5068 bufobj_init(struct bufobj *bo, void *private)
5070 static volatile int bufobj_cleanq;
5073 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5074 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5075 bo->bo_private = private;
5076 TAILQ_INIT(&bo->bo_clean.bv_hd);
5077 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5081 bufobj_wrefl(struct bufobj *bo)
5084 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5085 ASSERT_BO_WLOCKED(bo);
5090 bufobj_wref(struct bufobj *bo)
5093 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5100 bufobj_wdrop(struct bufobj *bo)
5103 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5105 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5106 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5107 bo->bo_flag &= ~BO_WWAIT;
5108 wakeup(&bo->bo_numoutput);
5114 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5118 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5119 ASSERT_BO_WLOCKED(bo);
5121 while (bo->bo_numoutput) {
5122 bo->bo_flag |= BO_WWAIT;
5123 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5124 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5132 * Set bio_data or bio_ma for struct bio from the struct buf.
5135 bdata2bio(struct buf *bp, struct bio *bip)
5138 if (!buf_mapped(bp)) {
5139 KASSERT(unmapped_buf_allowed, ("unmapped"));
5140 bip->bio_ma = bp->b_pages;
5141 bip->bio_ma_n = bp->b_npages;
5142 bip->bio_data = unmapped_buf;
5143 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5144 bip->bio_flags |= BIO_UNMAPPED;
5145 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5146 PAGE_SIZE == bp->b_npages,
5147 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5148 (long long)bip->bio_length, bip->bio_ma_n));
5150 bip->bio_data = bp->b_data;
5156 * The MIPS pmap code currently doesn't handle aliased pages.
5157 * The VIPT caches may not handle page aliasing themselves, leading
5158 * to data corruption.
5160 * As such, this code makes a system extremely unhappy if said
5161 * system doesn't support unaliasing the above situation in hardware.
5162 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5163 * this feature at build time, so it has to be handled in software.
5165 * Once the MIPS pmap/cache code grows to support this function on
5166 * earlier chips, it should be flipped back off.
5169 static int buf_pager_relbuf = 1;
5171 static int buf_pager_relbuf = 0;
5173 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5174 &buf_pager_relbuf, 0,
5175 "Make buffer pager release buffers after reading");
5178 * The buffer pager. It uses buffer reads to validate pages.
5180 * In contrast to the generic local pager from vm/vnode_pager.c, this
5181 * pager correctly and easily handles volumes where the underlying
5182 * device block size is greater than the machine page size. The
5183 * buffer cache transparently extends the requested page run to be
5184 * aligned at the block boundary, and does the necessary bogus page
5185 * replacements in the addends to avoid obliterating already valid
5188 * The only non-trivial issue is that the exclusive busy state for
5189 * pages, which is assumed by the vm_pager_getpages() interface, is
5190 * incompatible with the VMIO buffer cache's desire to share-busy the
5191 * pages. This function performs a trivial downgrade of the pages'
5192 * state before reading buffers, and a less trivial upgrade from the
5193 * shared-busy to excl-busy state after the read.
5196 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5197 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5198 vbg_get_blksize_t get_blksize)
5205 vm_ooffset_t la, lb, poff, poffe;
5207 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5210 object = vp->v_object;
5213 la = IDX_TO_OFF(ma[count - 1]->pindex);
5214 if (la >= object->un_pager.vnp.vnp_size)
5215 return (VM_PAGER_BAD);
5218 * Change the meaning of la from where the last requested page starts
5219 * to where it ends, because that's the end of the requested region
5220 * and the start of the potential read-ahead region.
5223 lpart = la > object->un_pager.vnp.vnp_size;
5224 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5227 return (VM_PAGER_ERROR);
5230 * Calculate read-ahead, behind and total pages.
5233 lb = IDX_TO_OFF(ma[0]->pindex);
5234 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5236 if (rbehind != NULL)
5238 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5239 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5240 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5245 VM_CNT_INC(v_vnodein);
5246 VM_CNT_ADD(v_vnodepgsin, pgsin);
5248 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5249 != 0) ? GB_UNMAPPED : 0;
5251 for (i = 0; i < count; i++) {
5252 if (ma[i] != bogus_page)
5253 vm_page_busy_downgrade(ma[i]);
5257 for (i = 0; i < count; i++) {
5259 if (m == bogus_page)
5263 * Pages are shared busy and the object lock is not
5264 * owned, which together allow for the pages'
5265 * invalidation. The racy test for validity avoids
5266 * useless creation of the buffer for the most typical
5267 * case when invalidation is not used in redo or for
5268 * parallel read. The shared->excl upgrade loop at
5269 * the end of the function catches the race in a
5270 * reliable way (protected by the object lock).
5272 if (vm_page_all_valid(m))
5275 poff = IDX_TO_OFF(m->pindex);
5276 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5277 for (; poff < poffe; poff += bsize) {
5278 lbn = get_lblkno(vp, poff);
5283 error = get_blksize(vp, lbn, &bsize);
5285 error = bread_gb(vp, lbn, bsize,
5286 curthread->td_ucred, br_flags, &bp);
5289 if (bp->b_rcred == curthread->td_ucred) {
5290 crfree(bp->b_rcred);
5291 bp->b_rcred = NOCRED;
5293 if (LIST_EMPTY(&bp->b_dep)) {
5295 * Invalidation clears m->valid, but
5296 * may leave B_CACHE flag if the
5297 * buffer existed at the invalidation
5298 * time. In this case, recycle the
5299 * buffer to do real read on next
5300 * bread() after redo.
5302 * Otherwise B_RELBUF is not strictly
5303 * necessary, enable to reduce buf
5306 if (buf_pager_relbuf ||
5307 !vm_page_all_valid(m))
5308 bp->b_flags |= B_RELBUF;
5310 bp->b_flags &= ~B_NOCACHE;
5316 KASSERT(1 /* racy, enable for debugging */ ||
5317 vm_page_all_valid(m) || i == count - 1,
5318 ("buf %d %p invalid", i, m));
5319 if (i == count - 1 && lpart) {
5320 if (!vm_page_none_valid(m) &&
5321 !vm_page_all_valid(m))
5322 vm_page_zero_invalid(m, TRUE);
5329 for (i = 0; i < count; i++) {
5330 if (ma[i] == bogus_page)
5332 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5333 vm_page_sunbusy(ma[i]);
5334 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5339 * Since the pages were only sbusy while neither the
5340 * buffer nor the object lock was held by us, or
5341 * reallocated while vm_page_grab() slept for busy
5342 * relinguish, they could have been invalidated.
5343 * Recheck the valid bits and re-read as needed.
5345 * Note that the last page is made fully valid in the
5346 * read loop, and partial validity for the page at
5347 * index count - 1 could mean that the page was
5348 * invalidated or removed, so we must restart for
5351 if (!vm_page_all_valid(ma[i]))
5354 if (redo && error == 0)
5356 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5359 #include "opt_ddb.h"
5361 #include <ddb/ddb.h>
5363 /* DDB command to show buffer data */
5364 DB_SHOW_COMMAND(buffer, db_show_buffer)
5367 struct buf *bp = (struct buf *)addr;
5368 #ifdef FULL_BUF_TRACKING
5373 db_printf("usage: show buffer <addr>\n");
5377 db_printf("buf at %p\n", bp);
5378 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5379 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5380 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5381 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5382 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5383 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5385 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5386 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5387 "b_vp = %p, b_dep = %p\n",
5388 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5389 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5390 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5391 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5392 bp->b_kvabase, bp->b_kvasize);
5395 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5396 for (i = 0; i < bp->b_npages; i++) {
5400 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5402 (u_long)VM_PAGE_TO_PHYS(m));
5404 db_printf("( ??? )");
5405 if ((i + 1) < bp->b_npages)
5410 BUF_LOCKPRINTINFO(bp);
5411 #if defined(FULL_BUF_TRACKING)
5412 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5414 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5415 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5416 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5418 db_printf(" %2u: %s\n", j,
5419 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5421 #elif defined(BUF_TRACKING)
5422 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5427 DB_SHOW_COMMAND(bufqueues, bufqueues)
5429 struct bufdomain *bd;
5434 db_printf("bqempty: %d\n", bqempty.bq_len);
5436 for (i = 0; i < buf_domains; i++) {
5438 db_printf("Buf domain %d\n", i);
5439 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5440 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5441 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5443 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5444 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5445 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5446 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5447 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5449 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5450 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5451 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5452 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5455 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5456 total += bp->b_bufsize;
5457 db_printf("\tcleanq count\t%d (%ld)\n",
5458 bd->bd_cleanq->bq_len, total);
5460 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5461 total += bp->b_bufsize;
5462 db_printf("\tdirtyq count\t%d (%ld)\n",
5463 bd->bd_dirtyq.bq_len, total);
5464 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5465 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5466 db_printf("\tCPU ");
5467 for (j = 0; j <= mp_maxid; j++)
5468 db_printf("%d, ", bd->bd_subq[j].bq_len);
5472 for (j = 0; j < nbuf; j++) {
5474 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5476 total += bp->b_bufsize;
5479 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5482 for (j = 0; j < nbuf; j++) {
5484 if (bp->b_domain == i) {
5486 total += bp->b_bufsize;
5489 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5493 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5498 for (i = 0; i < nbuf; i++) {
5500 if (BUF_ISLOCKED(bp)) {
5501 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5509 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5515 db_printf("usage: show vnodebufs <addr>\n");
5518 vp = (struct vnode *)addr;
5519 db_printf("Clean buffers:\n");
5520 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5521 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5524 db_printf("Dirty buffers:\n");
5525 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5526 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5531 DB_COMMAND(countfreebufs, db_coundfreebufs)
5534 int i, used = 0, nfree = 0;
5537 db_printf("usage: countfreebufs\n");
5541 for (i = 0; i < nbuf; i++) {
5543 if (bp->b_qindex == QUEUE_EMPTY)
5549 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5551 db_printf("numfreebuffers is %d\n", numfreebuffers);