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;
137 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
138 int __aligned(CACHE_LINE_SIZE) bd_running;
139 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
140 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
141 } __aligned(CACHE_LINE_SIZE);
143 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
144 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
145 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
146 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
147 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
148 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
149 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
150 #define BD_DOMAIN(bd) (bd - bdomain)
152 static char *buf; /* buffer header pool */
156 return ((struct buf *)(buf + (sizeof(struct buf) +
157 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
160 caddr_t __read_mostly unmapped_buf;
162 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
163 struct proc *bufdaemonproc;
165 static void vm_hold_free_pages(struct buf *bp, int newbsize);
166 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
168 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
169 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
171 static void vfs_clean_pages_dirty_buf(struct buf *bp);
172 static void vfs_setdirty_range(struct buf *bp);
173 static void vfs_vmio_invalidate(struct buf *bp);
174 static void vfs_vmio_truncate(struct buf *bp, int npages);
175 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
176 static int vfs_bio_clcheck(struct vnode *vp, int size,
177 daddr_t lblkno, daddr_t blkno);
178 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
179 void (*)(struct buf *));
180 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
181 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
182 static void buf_daemon(void);
183 static __inline void bd_wakeup(void);
184 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
185 static void bufkva_reclaim(vmem_t *, int);
186 static void bufkva_free(struct buf *);
187 static int buf_import(void *, void **, int, int, int);
188 static void buf_release(void *, void **, int);
189 static void maxbcachebuf_adjust(void);
190 static inline struct bufdomain *bufdomain(struct buf *);
191 static void bq_remove(struct bufqueue *bq, struct buf *bp);
192 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
193 static int buf_recycle(struct bufdomain *, bool kva);
194 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
195 const char *lockname);
196 static void bd_init(struct bufdomain *bd);
197 static int bd_flushall(struct bufdomain *bd);
198 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
199 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
201 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
202 int vmiodirenable = TRUE;
203 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
204 "Use the VM system for directory writes");
205 long runningbufspace;
206 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
207 "Amount of presently outstanding async buffer io");
208 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
209 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
210 static counter_u64_t bufkvaspace;
211 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
212 "Kernel virtual memory used for buffers");
213 static long maxbufspace;
214 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
215 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
216 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
217 "Maximum allowed value of bufspace (including metadata)");
218 static long bufmallocspace;
219 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
220 "Amount of malloced memory for buffers");
221 static long maxbufmallocspace;
222 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
223 0, "Maximum amount of malloced memory for buffers");
224 static long lobufspace;
225 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
226 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
227 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
228 "Minimum amount of buffers we want to have");
230 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
231 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
232 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
233 "Maximum allowed value of bufspace (excluding metadata)");
235 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
236 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
237 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
238 "Bufspace consumed before waking the daemon to free some");
239 static counter_u64_t buffreekvacnt;
240 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
241 "Number of times we have freed the KVA space from some buffer");
242 static counter_u64_t bufdefragcnt;
243 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
244 "Number of times we have had to repeat buffer allocation to defragment");
245 static long lorunningspace;
246 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
247 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
248 "Minimum preferred space used for in-progress I/O");
249 static long hirunningspace;
250 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
251 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
252 "Maximum amount of space to use for in-progress I/O");
253 int dirtybufferflushes;
254 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
255 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
257 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
258 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
259 int altbufferflushes;
260 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
261 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
262 static int recursiveflushes;
263 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
264 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
265 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
266 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
267 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
268 "Number of buffers that are dirty (has unwritten changes) at the moment");
269 static int lodirtybuffers;
270 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
271 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
272 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
273 "How many buffers we want to have free before bufdaemon can sleep");
274 static int hidirtybuffers;
275 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
276 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
277 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
278 "When the number of dirty buffers is considered severe");
280 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
281 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
282 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
283 "Number of bdwrite to bawrite conversions to clear dirty buffers");
284 static int numfreebuffers;
285 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
286 "Number of free buffers");
287 static int lofreebuffers;
288 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
289 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
290 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
291 "Target number of free buffers");
292 static int hifreebuffers;
293 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
294 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
295 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
296 "Threshold for clean buffer recycling");
297 static counter_u64_t getnewbufcalls;
298 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
299 &getnewbufcalls, "Number of calls to getnewbuf");
300 static counter_u64_t getnewbufrestarts;
301 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
303 "Number of times getnewbuf has had to restart a buffer acquisition");
304 static counter_u64_t mappingrestarts;
305 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
307 "Number of times getblk has had to restart a buffer mapping for "
309 static counter_u64_t numbufallocfails;
310 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
311 &numbufallocfails, "Number of times buffer allocations failed");
312 static int flushbufqtarget = 100;
313 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
314 "Amount of work to do in flushbufqueues when helping bufdaemon");
315 static counter_u64_t notbufdflushes;
316 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
317 "Number of dirty buffer flushes done by the bufdaemon helpers");
318 static long barrierwrites;
319 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
320 &barrierwrites, 0, "Number of barrier writes");
321 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
322 &unmapped_buf_allowed, 0,
323 "Permit the use of the unmapped i/o");
324 int maxbcachebuf = MAXBCACHEBUF;
325 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
326 "Maximum size of a buffer cache block");
329 * This lock synchronizes access to bd_request.
331 static struct mtx_padalign __exclusive_cache_line bdlock;
334 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
335 * waitrunningbufspace().
337 static struct mtx_padalign __exclusive_cache_line rbreqlock;
340 * Lock that protects bdirtywait.
342 static struct mtx_padalign __exclusive_cache_line bdirtylock;
345 * bufdaemon shutdown request and sleep channel.
347 static bool bd_shutdown;
350 * Wakeup point for bufdaemon, as well as indicator of whether it is already
351 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
354 static int bd_request;
357 * Request for the buf daemon to write more buffers than is indicated by
358 * lodirtybuf. This may be necessary to push out excess dependencies or
359 * defragment the address space where a simple count of the number of dirty
360 * buffers is insufficient to characterize the demand for flushing them.
362 static int bd_speedupreq;
365 * Synchronization (sleep/wakeup) variable for active buffer space requests.
366 * Set when wait starts, cleared prior to wakeup().
367 * Used in runningbufwakeup() and waitrunningbufspace().
369 static int runningbufreq;
372 * Synchronization for bwillwrite() waiters.
374 static int bdirtywait;
377 * Definitions for the buffer free lists.
379 #define QUEUE_NONE 0 /* on no queue */
380 #define QUEUE_EMPTY 1 /* empty buffer headers */
381 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
382 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
383 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
385 /* Maximum number of buffer domains. */
386 #define BUF_DOMAINS 8
388 struct bufdomainset bdlodirty; /* Domains > lodirty */
389 struct bufdomainset bdhidirty; /* Domains > hidirty */
391 /* Configured number of clean queues. */
392 static int __read_mostly buf_domains;
394 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
395 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
396 struct bufqueue __exclusive_cache_line bqempty;
399 * per-cpu empty buffer cache.
404 * Single global constant for BUF_WMESG, to avoid getting multiple references.
405 * buf_wmesg is referred from macros.
407 const char *buf_wmesg = BUF_WMESG;
410 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
415 value = *(long *)arg1;
416 error = sysctl_handle_long(oidp, &value, 0, req);
417 if (error != 0 || req->newptr == NULL)
419 mtx_lock(&rbreqlock);
420 if (arg1 == &hirunningspace) {
421 if (value < lorunningspace)
424 hirunningspace = value;
426 KASSERT(arg1 == &lorunningspace,
427 ("%s: unknown arg1", __func__));
428 if (value > hirunningspace)
431 lorunningspace = value;
433 mtx_unlock(&rbreqlock);
438 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
444 value = *(int *)arg1;
445 error = sysctl_handle_int(oidp, &value, 0, req);
446 if (error != 0 || req->newptr == NULL)
448 *(int *)arg1 = value;
449 for (i = 0; i < buf_domains; i++)
450 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
457 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
463 value = *(long *)arg1;
464 error = sysctl_handle_long(oidp, &value, 0, req);
465 if (error != 0 || req->newptr == NULL)
467 *(long *)arg1 = value;
468 for (i = 0; i < buf_domains; i++)
469 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
475 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
476 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
478 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
485 for (i = 0; i < buf_domains; i++)
486 lvalue += bdomain[i].bd_bufspace;
487 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
488 return (sysctl_handle_long(oidp, &lvalue, 0, req));
489 if (lvalue > INT_MAX)
490 /* On overflow, still write out a long to trigger ENOMEM. */
491 return (sysctl_handle_long(oidp, &lvalue, 0, req));
493 return (sysctl_handle_int(oidp, &ivalue, 0, req));
497 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
503 for (i = 0; i < buf_domains; i++)
504 lvalue += bdomain[i].bd_bufspace;
505 return (sysctl_handle_long(oidp, &lvalue, 0, req));
510 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
516 for (i = 0; i < buf_domains; i++)
517 value += bdomain[i].bd_numdirtybuffers;
518 return (sysctl_handle_int(oidp, &value, 0, req));
524 * Wakeup any bwillwrite() waiters.
529 mtx_lock(&bdirtylock);
534 mtx_unlock(&bdirtylock);
540 * Clear a domain from the appropriate bitsets when dirtybuffers
544 bd_clear(struct bufdomain *bd)
547 mtx_lock(&bdirtylock);
548 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
549 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
550 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
551 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
552 mtx_unlock(&bdirtylock);
558 * Set a domain in the appropriate bitsets when dirtybuffers
562 bd_set(struct bufdomain *bd)
565 mtx_lock(&bdirtylock);
566 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
567 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
568 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
569 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
570 mtx_unlock(&bdirtylock);
576 * Decrement the numdirtybuffers count by one and wakeup any
577 * threads blocked in bwillwrite().
580 bdirtysub(struct buf *bp)
582 struct bufdomain *bd;
586 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
587 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
589 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
596 * Increment the numdirtybuffers count by one and wakeup the buf
600 bdirtyadd(struct buf *bp)
602 struct bufdomain *bd;
606 * Only do the wakeup once as we cross the boundary. The
607 * buf daemon will keep running until the condition clears.
610 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
611 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
613 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
618 * bufspace_daemon_wakeup:
620 * Wakeup the daemons responsible for freeing clean bufs.
623 bufspace_daemon_wakeup(struct bufdomain *bd)
627 * avoid the lock if the daemon is running.
629 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
631 atomic_store_int(&bd->bd_running, 1);
632 wakeup(&bd->bd_running);
640 * Adjust the reported bufspace for a KVA managed buffer, possibly
641 * waking any waiters.
644 bufspace_adjust(struct buf *bp, int bufsize)
646 struct bufdomain *bd;
650 KASSERT((bp->b_flags & B_MALLOC) == 0,
651 ("bufspace_adjust: malloc buf %p", bp));
653 diff = bufsize - bp->b_bufsize;
655 atomic_subtract_long(&bd->bd_bufspace, -diff);
656 } else if (diff > 0) {
657 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
658 /* Wake up the daemon on the transition. */
659 if (space < bd->bd_bufspacethresh &&
660 space + diff >= bd->bd_bufspacethresh)
661 bufspace_daemon_wakeup(bd);
663 bp->b_bufsize = bufsize;
669 * Reserve bufspace before calling allocbuf(). metadata has a
670 * different space limit than data.
673 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
679 limit = bd->bd_maxbufspace;
681 limit = bd->bd_hibufspace;
682 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
685 atomic_subtract_long(&bd->bd_bufspace, size);
689 /* Wake up the daemon on the transition. */
690 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
691 bufspace_daemon_wakeup(bd);
699 * Release reserved bufspace after bufspace_adjust() has consumed it.
702 bufspace_release(struct bufdomain *bd, int size)
705 atomic_subtract_long(&bd->bd_bufspace, size);
711 * Wait for bufspace, acting as the buf daemon if a locked vnode is
712 * supplied. bd_wanted must be set prior to polling for space. The
713 * operation must be re-tried on return.
716 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
717 int slpflag, int slptimeo)
720 int error, fl, norunbuf;
722 if ((gbflags & GB_NOWAIT_BD) != 0)
727 while (bd->bd_wanted) {
728 if (vp != NULL && vp->v_type != VCHR &&
729 (td->td_pflags & TDP_BUFNEED) == 0) {
732 * getblk() is called with a vnode locked, and
733 * some majority of the dirty buffers may as
734 * well belong to the vnode. Flushing the
735 * buffers there would make a progress that
736 * cannot be achieved by the buf_daemon, that
737 * cannot lock the vnode.
739 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
740 (td->td_pflags & TDP_NORUNNINGBUF);
743 * Play bufdaemon. The getnewbuf() function
744 * may be called while the thread owns lock
745 * for another dirty buffer for the same
746 * vnode, which makes it impossible to use
747 * VOP_FSYNC() there, due to the buffer lock
750 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
751 fl = buf_flush(vp, bd, flushbufqtarget);
752 td->td_pflags &= norunbuf;
756 if (bd->bd_wanted == 0)
759 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
760 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
768 bufspace_daemon_shutdown(void *arg, int howto __unused)
770 struct bufdomain *bd = arg;
774 bd->bd_shutdown = true;
775 wakeup(&bd->bd_running);
776 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
777 "bufspace_shutdown", 60 * hz);
780 printf("bufspacedaemon wait error: %d\n", error);
786 * buffer space management daemon. Tries to maintain some marginal
787 * amount of free buffer space so that requesting processes neither
788 * block nor work to reclaim buffers.
791 bufspace_daemon(void *arg)
793 struct bufdomain *bd = arg;
795 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
796 SHUTDOWN_PRI_LAST + 100);
799 while (!bd->bd_shutdown) {
803 * Free buffers from the clean queue until we meet our
806 * Theory of operation: The buffer cache is most efficient
807 * when some free buffer headers and space are always
808 * available to getnewbuf(). This daemon attempts to prevent
809 * the excessive blocking and synchronization associated
810 * with shortfall. It goes through three phases according
813 * 1) The daemon wakes up voluntarily once per-second
814 * during idle periods when the counters are below
815 * the wakeup thresholds (bufspacethresh, lofreebuffers).
817 * 2) The daemon wakes up as we cross the thresholds
818 * ahead of any potential blocking. This may bounce
819 * slightly according to the rate of consumption and
822 * 3) The daemon and consumers are starved for working
823 * clean buffers. This is the 'bufspace' sleep below
824 * which will inefficiently trade bufs with bqrelse
825 * until we return to condition 2.
827 while (bd->bd_bufspace > bd->bd_lobufspace ||
828 bd->bd_freebuffers < bd->bd_hifreebuffers) {
829 if (buf_recycle(bd, false) != 0) {
833 * Speedup dirty if we've run out of clean
834 * buffers. This is possible in particular
835 * because softdep may held many bufs locked
836 * pending writes to other bufs which are
837 * marked for delayed write, exhausting
838 * clean space until they are written.
843 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
844 PRIBIO|PDROP, "bufspace", hz/10);
852 * Re-check our limits and sleep. bd_running must be
853 * cleared prior to checking the limits to avoid missed
854 * wakeups. The waker will adjust one of bufspace or
855 * freebuffers prior to checking bd_running.
860 atomic_store_int(&bd->bd_running, 0);
861 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
862 bd->bd_freebuffers > bd->bd_lofreebuffers) {
863 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
866 /* Avoid spurious wakeups while running. */
867 atomic_store_int(&bd->bd_running, 1);
870 wakeup(&bd->bd_shutdown);
878 * Adjust the reported bufspace for a malloc managed buffer, possibly
879 * waking any waiters.
882 bufmallocadjust(struct buf *bp, int bufsize)
886 KASSERT((bp->b_flags & B_MALLOC) != 0,
887 ("bufmallocadjust: non-malloc buf %p", bp));
888 diff = bufsize - bp->b_bufsize;
890 atomic_subtract_long(&bufmallocspace, -diff);
892 atomic_add_long(&bufmallocspace, diff);
893 bp->b_bufsize = bufsize;
899 * Wake up processes that are waiting on asynchronous writes to fall
900 * below lorunningspace.
906 mtx_lock(&rbreqlock);
909 wakeup(&runningbufreq);
911 mtx_unlock(&rbreqlock);
917 * Decrement the outstanding write count according.
920 runningbufwakeup(struct buf *bp)
924 bspace = bp->b_runningbufspace;
927 space = atomic_fetchadd_long(&runningbufspace, -bspace);
928 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
930 bp->b_runningbufspace = 0;
932 * Only acquire the lock and wakeup on the transition from exceeding
933 * the threshold to falling below it.
935 if (space < lorunningspace)
937 if (space - bspace > lorunningspace)
943 * waitrunningbufspace()
945 * runningbufspace is a measure of the amount of I/O currently
946 * running. This routine is used in async-write situations to
947 * prevent creating huge backups of pending writes to a device.
948 * Only asynchronous writes are governed by this function.
950 * This does NOT turn an async write into a sync write. It waits
951 * for earlier writes to complete and generally returns before the
952 * caller's write has reached the device.
955 waitrunningbufspace(void)
958 mtx_lock(&rbreqlock);
959 while (runningbufspace > hirunningspace) {
961 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
963 mtx_unlock(&rbreqlock);
967 * vfs_buf_test_cache:
969 * Called when a buffer is extended. This function clears the B_CACHE
970 * bit if the newly extended portion of the buffer does not contain
974 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
975 vm_offset_t size, vm_page_t m)
979 * This function and its results are protected by higher level
980 * synchronization requiring vnode and buf locks to page in and
983 if (bp->b_flags & B_CACHE) {
984 int base = (foff + off) & PAGE_MASK;
985 if (vm_page_is_valid(m, base, size) == 0)
986 bp->b_flags &= ~B_CACHE;
990 /* Wake up the buffer daemon if necessary */
996 if (bd_request == 0) {
1000 mtx_unlock(&bdlock);
1004 * Adjust the maxbcachbuf tunable.
1007 maxbcachebuf_adjust(void)
1012 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1015 while (i * 2 <= maxbcachebuf)
1018 if (maxbcachebuf < MAXBSIZE)
1019 maxbcachebuf = MAXBSIZE;
1020 if (maxbcachebuf > maxphys)
1021 maxbcachebuf = maxphys;
1022 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1023 printf("maxbcachebuf=%d\n", maxbcachebuf);
1027 * bd_speedup - speedup the buffer cache flushing code
1036 if (bd_speedupreq == 0 || bd_request == 0)
1041 wakeup(&bd_request);
1042 mtx_unlock(&bdlock);
1046 #define TRANSIENT_DENOM 5
1048 #define TRANSIENT_DENOM 10
1052 * Calculating buffer cache scaling values and reserve space for buffer
1053 * headers. This is called during low level kernel initialization and
1054 * may be called more then once. We CANNOT write to the memory area
1055 * being reserved at this time.
1058 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1061 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1064 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for
1065 * this when sizing maps based on the amount of physical memory
1069 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1070 (KASAN_SHADOW_SCALE + 1);
1071 #elif defined(KMSAN)
1075 * KMSAN cannot reliably determine whether buffer data is initialized
1076 * unless it is updated through a KVA mapping.
1078 unmapped_buf_allowed = 0;
1082 * physmem_est is in pages. Convert it to kilobytes (assumes
1083 * PAGE_SIZE is >= 1K)
1085 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1087 maxbcachebuf_adjust();
1089 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1090 * For the first 64MB of ram nominally allocate sufficient buffers to
1091 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1092 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1093 * the buffer cache we limit the eventual kva reservation to
1096 * factor represents the 1/4 x ram conversion.
1099 int factor = 4 * BKVASIZE / 1024;
1102 if (physmem_est > 4096)
1103 nbuf += min((physmem_est - 4096) / factor,
1105 if (physmem_est > 65536)
1106 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1107 32 * 1024 * 1024 / (factor * 5));
1109 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1110 nbuf = maxbcache / BKVASIZE;
1115 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1116 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1117 if (nbuf > maxbuf) {
1119 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1125 * Ideal allocation size for the transient bio submap is 10%
1126 * of the maximal space buffer map. This roughly corresponds
1127 * to the amount of the buffer mapped for typical UFS load.
1129 * Clip the buffer map to reserve space for the transient
1130 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1131 * maximum buffer map extent on the platform.
1133 * The fall-back to the maxbuf in case of maxbcache unset,
1134 * allows to not trim the buffer KVA for the architectures
1135 * with ample KVA space.
1137 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1138 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1139 buf_sz = (long)nbuf * BKVASIZE;
1140 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1141 (TRANSIENT_DENOM - 1)) {
1143 * There is more KVA than memory. Do not
1144 * adjust buffer map size, and assign the rest
1145 * of maxbuf to transient map.
1147 biotmap_sz = maxbuf_sz - buf_sz;
1150 * Buffer map spans all KVA we could afford on
1151 * this platform. Give 10% (20% on i386) of
1152 * the buffer map to the transient bio map.
1154 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1155 buf_sz -= biotmap_sz;
1157 if (biotmap_sz / INT_MAX > maxphys)
1158 bio_transient_maxcnt = INT_MAX;
1160 bio_transient_maxcnt = biotmap_sz / maxphys;
1162 * Artificially limit to 1024 simultaneous in-flight I/Os
1163 * using the transient mapping.
1165 if (bio_transient_maxcnt > 1024)
1166 bio_transient_maxcnt = 1024;
1168 nbuf = buf_sz / BKVASIZE;
1172 nswbuf = min(nbuf / 4, 256);
1173 if (nswbuf < NSWBUF_MIN)
1174 nswbuf = NSWBUF_MIN;
1178 * Reserve space for the buffer cache buffers
1181 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1182 atop(maxbcachebuf)) * nbuf;
1187 /* Initialize the buffer subsystem. Called before use of any buffers. */
1194 KASSERT(maxbcachebuf >= MAXBSIZE,
1195 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1197 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1198 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1199 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1200 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1202 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1204 /* finally, initialize each buffer header and stick on empty q */
1205 for (i = 0; i < nbuf; i++) {
1207 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1208 bp->b_flags = B_INVAL;
1209 bp->b_rcred = NOCRED;
1210 bp->b_wcred = NOCRED;
1211 bp->b_qindex = QUEUE_NONE;
1213 bp->b_subqueue = mp_maxid + 1;
1215 bp->b_data = bp->b_kvabase = unmapped_buf;
1216 LIST_INIT(&bp->b_dep);
1218 bq_insert(&bqempty, bp, false);
1222 * maxbufspace is the absolute maximum amount of buffer space we are
1223 * allowed to reserve in KVM and in real terms. The absolute maximum
1224 * is nominally used by metadata. hibufspace is the nominal maximum
1225 * used by most other requests. The differential is required to
1226 * ensure that metadata deadlocks don't occur.
1228 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1229 * this may result in KVM fragmentation which is not handled optimally
1230 * by the system. XXX This is less true with vmem. We could use
1233 maxbufspace = (long)nbuf * BKVASIZE;
1234 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1235 lobufspace = (hibufspace / 20) * 19; /* 95% */
1236 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1239 * Note: The 16 MiB upper limit for hirunningspace was chosen
1240 * arbitrarily and may need further tuning. It corresponds to
1241 * 128 outstanding write IO requests (if IO size is 128 KiB),
1242 * which fits with many RAID controllers' tagged queuing limits.
1243 * The lower 1 MiB limit is the historical upper limit for
1246 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1247 16 * 1024 * 1024), 1024 * 1024);
1248 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1251 * Limit the amount of malloc memory since it is wired permanently into
1252 * the kernel space. Even though this is accounted for in the buffer
1253 * allocation, we don't want the malloced region to grow uncontrolled.
1254 * The malloc scheme improves memory utilization significantly on
1255 * average (small) directories.
1257 maxbufmallocspace = hibufspace / 20;
1260 * Reduce the chance of a deadlock occurring by limiting the number
1261 * of delayed-write dirty buffers we allow to stack up.
1263 hidirtybuffers = nbuf / 4 + 20;
1264 dirtybufthresh = hidirtybuffers * 9 / 10;
1266 * To support extreme low-memory systems, make sure hidirtybuffers
1267 * cannot eat up all available buffer space. This occurs when our
1268 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1269 * buffer space assuming BKVASIZE'd buffers.
1271 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1272 hidirtybuffers >>= 1;
1274 lodirtybuffers = hidirtybuffers / 2;
1277 * lofreebuffers should be sufficient to avoid stalling waiting on
1278 * buf headers under heavy utilization. The bufs in per-cpu caches
1279 * are counted as free but will be unavailable to threads executing
1282 * hifreebuffers is the free target for the bufspace daemon. This
1283 * should be set appropriately to limit work per-iteration.
1285 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1286 hifreebuffers = (3 * lofreebuffers) / 2;
1287 numfreebuffers = nbuf;
1289 /* Setup the kva and free list allocators. */
1290 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1291 buf_zone = uma_zcache_create("buf free cache",
1292 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1293 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1296 * Size the clean queue according to the amount of buffer space.
1297 * One queue per-256mb up to the max. More queues gives better
1298 * concurrency but less accurate LRU.
1300 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1301 for (i = 0 ; i < buf_domains; i++) {
1302 struct bufdomain *bd;
1306 bd->bd_freebuffers = nbuf / buf_domains;
1307 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1308 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1309 bd->bd_bufspace = 0;
1310 bd->bd_maxbufspace = maxbufspace / buf_domains;
1311 bd->bd_hibufspace = hibufspace / buf_domains;
1312 bd->bd_lobufspace = lobufspace / buf_domains;
1313 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1314 bd->bd_numdirtybuffers = 0;
1315 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1316 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1317 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1318 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1319 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1321 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1322 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1323 mappingrestarts = counter_u64_alloc(M_WAITOK);
1324 numbufallocfails = counter_u64_alloc(M_WAITOK);
1325 notbufdflushes = counter_u64_alloc(M_WAITOK);
1326 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1327 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1328 bufkvaspace = counter_u64_alloc(M_WAITOK);
1333 vfs_buf_check_mapped(struct buf *bp)
1336 KASSERT(bp->b_kvabase != unmapped_buf,
1337 ("mapped buf: b_kvabase was not updated %p", bp));
1338 KASSERT(bp->b_data != unmapped_buf,
1339 ("mapped buf: b_data was not updated %p", bp));
1340 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1341 maxphys, ("b_data + b_offset unmapped %p", bp));
1345 vfs_buf_check_unmapped(struct buf *bp)
1348 KASSERT(bp->b_data == unmapped_buf,
1349 ("unmapped buf: corrupted b_data %p", bp));
1352 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1353 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1355 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1356 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1360 isbufbusy(struct buf *bp)
1362 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1363 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1369 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1372 bufshutdown(int show_busybufs)
1374 static int first_buf_printf = 1;
1376 int i, iter, nbusy, pbusy;
1382 * Sync filesystems for shutdown
1384 wdog_kern_pat(WD_LASTVAL);
1385 kern_sync(curthread);
1388 * With soft updates, some buffers that are
1389 * written will be remarked as dirty until other
1390 * buffers are written.
1392 for (iter = pbusy = 0; iter < 20; iter++) {
1394 for (i = nbuf - 1; i >= 0; i--) {
1400 if (first_buf_printf)
1401 printf("All buffers synced.");
1404 if (first_buf_printf) {
1405 printf("Syncing disks, buffers remaining... ");
1406 first_buf_printf = 0;
1408 printf("%d ", nbusy);
1413 wdog_kern_pat(WD_LASTVAL);
1414 kern_sync(curthread);
1418 * Spin for a while to allow interrupt threads to run.
1420 DELAY(50000 * iter);
1423 * Context switch several times to allow interrupt
1426 for (subiter = 0; subiter < 50 * iter; subiter++) {
1427 thread_lock(curthread);
1435 * Count only busy local buffers to prevent forcing
1436 * a fsck if we're just a client of a wedged NFS server
1439 for (i = nbuf - 1; i >= 0; i--) {
1441 if (isbufbusy(bp)) {
1443 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1444 if (bp->b_dev == NULL) {
1445 TAILQ_REMOVE(&mountlist,
1446 bp->b_vp->v_mount, mnt_list);
1451 if (show_busybufs > 0) {
1453 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1454 nbusy, bp, bp->b_vp, bp->b_flags,
1455 (intmax_t)bp->b_blkno,
1456 (intmax_t)bp->b_lblkno);
1457 BUF_LOCKPRINTINFO(bp);
1458 if (show_busybufs > 1)
1466 * Failed to sync all blocks. Indicate this and don't
1467 * unmount filesystems (thus forcing an fsck on reboot).
1469 printf("Giving up on %d buffers\n", nbusy);
1470 DELAY(5000000); /* 5 seconds */
1473 if (!first_buf_printf)
1474 printf("Final sync complete\n");
1477 * Unmount filesystems and perform swapoff, to quiesce
1478 * the system as much as possible. In particular, no
1479 * I/O should be initiated from top levels since it
1480 * might be abruptly terminated by reset, or otherwise
1481 * erronously handled because other parts of the
1482 * system are disabled.
1484 * Swapoff before unmount, because file-backed swap is
1485 * non-operational after unmount of the underlying
1488 if (!KERNEL_PANICKED()) {
1493 DELAY(100000); /* wait for console output to finish */
1497 bpmap_qenter(struct buf *bp)
1500 BUF_CHECK_MAPPED(bp);
1503 * bp->b_data is relative to bp->b_offset, but
1504 * bp->b_offset may be offset into the first page.
1506 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1507 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1508 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1509 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1512 static inline struct bufdomain *
1513 bufdomain(struct buf *bp)
1516 return (&bdomain[bp->b_domain]);
1519 static struct bufqueue *
1520 bufqueue(struct buf *bp)
1523 switch (bp->b_qindex) {
1526 case QUEUE_SENTINEL:
1531 return (&bufdomain(bp)->bd_dirtyq);
1533 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1537 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1541 * Return the locked bufqueue that bp is a member of.
1543 static struct bufqueue *
1544 bufqueue_acquire(struct buf *bp)
1546 struct bufqueue *bq, *nbq;
1549 * bp can be pushed from a per-cpu queue to the
1550 * cleanq while we're waiting on the lock. Retry
1551 * if the queues don't match.
1569 * Insert the buffer into the appropriate free list. Requires a
1570 * locked buffer on entry and buffer is unlocked before return.
1573 binsfree(struct buf *bp, int qindex)
1575 struct bufdomain *bd;
1576 struct bufqueue *bq;
1578 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1579 ("binsfree: Invalid qindex %d", qindex));
1580 BUF_ASSERT_XLOCKED(bp);
1583 * Handle delayed bremfree() processing.
1585 if (bp->b_flags & B_REMFREE) {
1586 if (bp->b_qindex == qindex) {
1587 bp->b_flags |= B_REUSE;
1588 bp->b_flags &= ~B_REMFREE;
1592 bq = bufqueue_acquire(bp);
1597 if (qindex == QUEUE_CLEAN) {
1598 if (bd->bd_lim != 0)
1599 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1603 bq = &bd->bd_dirtyq;
1604 bq_insert(bq, bp, true);
1610 * Free a buffer to the buf zone once it no longer has valid contents.
1613 buf_free(struct buf *bp)
1616 if (bp->b_flags & B_REMFREE)
1618 if (bp->b_vflags & BV_BKGRDINPROG)
1619 panic("losing buffer 1");
1620 if (bp->b_rcred != NOCRED) {
1621 crfree(bp->b_rcred);
1622 bp->b_rcred = NOCRED;
1624 if (bp->b_wcred != NOCRED) {
1625 crfree(bp->b_wcred);
1626 bp->b_wcred = NOCRED;
1628 if (!LIST_EMPTY(&bp->b_dep))
1631 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1632 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1634 uma_zfree(buf_zone, bp);
1640 * Import bufs into the uma cache from the buf list. The system still
1641 * expects a static array of bufs and much of the synchronization
1642 * around bufs assumes type stable storage. As a result, UMA is used
1643 * only as a per-cpu cache of bufs still maintained on a global list.
1646 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1652 for (i = 0; i < cnt; i++) {
1653 bp = TAILQ_FIRST(&bqempty.bq_queue);
1656 bq_remove(&bqempty, bp);
1659 BQ_UNLOCK(&bqempty);
1667 * Release bufs from the uma cache back to the buffer queues.
1670 buf_release(void *arg, void **store, int cnt)
1672 struct bufqueue *bq;
1678 for (i = 0; i < cnt; i++) {
1680 /* Inline bq_insert() to batch locking. */
1681 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1682 bp->b_flags &= ~(B_AGE | B_REUSE);
1684 bp->b_qindex = bq->bq_index;
1692 * Allocate an empty buffer header.
1695 buf_alloc(struct bufdomain *bd)
1698 int freebufs, error;
1701 * We can only run out of bufs in the buf zone if the average buf
1702 * is less than BKVASIZE. In this case the actual wait/block will
1703 * come from buf_reycle() failing to flush one of these small bufs.
1706 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1708 bp = uma_zalloc(buf_zone, M_NOWAIT);
1710 atomic_add_int(&bd->bd_freebuffers, 1);
1711 bufspace_daemon_wakeup(bd);
1712 counter_u64_add(numbufallocfails, 1);
1716 * Wake-up the bufspace daemon on transition below threshold.
1718 if (freebufs == bd->bd_lofreebuffers)
1719 bufspace_daemon_wakeup(bd);
1721 error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL);
1722 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1726 KASSERT(bp->b_vp == NULL,
1727 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1728 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1729 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1730 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1731 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1732 KASSERT(bp->b_npages == 0,
1733 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1734 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1735 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1736 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1738 bp->b_domain = BD_DOMAIN(bd);
1744 bp->b_blkno = bp->b_lblkno = 0;
1745 bp->b_offset = NOOFFSET;
1751 bp->b_dirtyoff = bp->b_dirtyend = 0;
1752 bp->b_bufobj = NULL;
1753 bp->b_data = bp->b_kvabase = unmapped_buf;
1754 bp->b_fsprivate1 = NULL;
1755 bp->b_fsprivate2 = NULL;
1756 bp->b_fsprivate3 = NULL;
1757 LIST_INIT(&bp->b_dep);
1765 * Free a buffer from the given bufqueue. kva controls whether the
1766 * freed buf must own some kva resources. This is used for
1770 buf_recycle(struct bufdomain *bd, bool kva)
1772 struct bufqueue *bq;
1773 struct buf *bp, *nbp;
1776 counter_u64_add(bufdefragcnt, 1);
1780 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1781 ("buf_recycle: Locks don't match"));
1782 nbp = TAILQ_FIRST(&bq->bq_queue);
1785 * Run scan, possibly freeing data and/or kva mappings on the fly
1788 while ((bp = nbp) != NULL) {
1790 * Calculate next bp (we can only use it if we do not
1791 * release the bqlock).
1793 nbp = TAILQ_NEXT(bp, b_freelist);
1796 * If we are defragging then we need a buffer with
1797 * some kva to reclaim.
1799 if (kva && bp->b_kvasize == 0)
1802 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1806 * Implement a second chance algorithm for frequently
1809 if ((bp->b_flags & B_REUSE) != 0) {
1810 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1811 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1812 bp->b_flags &= ~B_REUSE;
1818 * Skip buffers with background writes in progress.
1820 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1825 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1826 ("buf_recycle: inconsistent queue %d bp %p",
1828 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1829 ("getnewbuf: queue domain %d doesn't match request %d",
1830 bp->b_domain, (int)BD_DOMAIN(bd)));
1832 * NOTE: nbp is now entirely invalid. We can only restart
1833 * the scan from this point on.
1839 * Requeue the background write buffer with error and
1842 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1845 nbp = TAILQ_FIRST(&bq->bq_queue);
1848 bp->b_flags |= B_INVAL;
1861 * Mark the buffer for removal from the appropriate free list.
1865 bremfree(struct buf *bp)
1868 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1869 KASSERT((bp->b_flags & B_REMFREE) == 0,
1870 ("bremfree: buffer %p already marked for delayed removal.", bp));
1871 KASSERT(bp->b_qindex != QUEUE_NONE,
1872 ("bremfree: buffer %p not on a queue.", bp));
1873 BUF_ASSERT_XLOCKED(bp);
1875 bp->b_flags |= B_REMFREE;
1881 * Force an immediate removal from a free list. Used only in nfs when
1882 * it abuses the b_freelist pointer.
1885 bremfreef(struct buf *bp)
1887 struct bufqueue *bq;
1889 bq = bufqueue_acquire(bp);
1895 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1898 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1899 TAILQ_INIT(&bq->bq_queue);
1901 bq->bq_index = qindex;
1902 bq->bq_subqueue = subqueue;
1906 bd_init(struct bufdomain *bd)
1910 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1911 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1912 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1913 for (i = 0; i <= mp_maxid; i++)
1914 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1915 "bufq clean subqueue lock");
1916 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1922 * Removes a buffer from the free list, must be called with the
1923 * correct qlock held.
1926 bq_remove(struct bufqueue *bq, struct buf *bp)
1929 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1930 bp, bp->b_vp, bp->b_flags);
1931 KASSERT(bp->b_qindex != QUEUE_NONE,
1932 ("bq_remove: buffer %p not on a queue.", bp));
1933 KASSERT(bufqueue(bp) == bq,
1934 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1936 BQ_ASSERT_LOCKED(bq);
1937 if (bp->b_qindex != QUEUE_EMPTY) {
1938 BUF_ASSERT_XLOCKED(bp);
1940 KASSERT(bq->bq_len >= 1,
1941 ("queue %d underflow", bp->b_qindex));
1942 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1944 bp->b_qindex = QUEUE_NONE;
1945 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1949 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1953 BQ_ASSERT_LOCKED(bq);
1954 if (bq != bd->bd_cleanq) {
1956 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1957 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1958 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1960 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1962 bd->bd_cleanq->bq_len += bq->bq_len;
1965 if (bd->bd_wanted) {
1967 wakeup(&bd->bd_wanted);
1969 if (bq != bd->bd_cleanq)
1974 bd_flushall(struct bufdomain *bd)
1976 struct bufqueue *bq;
1980 if (bd->bd_lim == 0)
1983 for (i = 0; i <= mp_maxid; i++) {
1984 bq = &bd->bd_subq[i];
1985 if (bq->bq_len == 0)
1997 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1999 struct bufdomain *bd;
2001 if (bp->b_qindex != QUEUE_NONE)
2002 panic("bq_insert: free buffer %p onto another queue?", bp);
2005 if (bp->b_flags & B_AGE) {
2006 /* Place this buf directly on the real queue. */
2007 if (bq->bq_index == QUEUE_CLEAN)
2010 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2013 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2015 bp->b_flags &= ~(B_AGE | B_REUSE);
2017 bp->b_qindex = bq->bq_index;
2018 bp->b_subqueue = bq->bq_subqueue;
2021 * Unlock before we notify so that we don't wakeup a waiter that
2022 * fails a trylock on the buf and sleeps again.
2027 if (bp->b_qindex == QUEUE_CLEAN) {
2029 * Flush the per-cpu queue and notify any waiters.
2031 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2032 bq->bq_len >= bd->bd_lim))
2041 * Free the kva allocation for a buffer.
2045 bufkva_free(struct buf *bp)
2049 if (bp->b_kvasize == 0) {
2050 KASSERT(bp->b_kvabase == unmapped_buf &&
2051 bp->b_data == unmapped_buf,
2052 ("Leaked KVA space on %p", bp));
2053 } else if (buf_mapped(bp))
2054 BUF_CHECK_MAPPED(bp);
2056 BUF_CHECK_UNMAPPED(bp);
2058 if (bp->b_kvasize == 0)
2061 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2062 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2063 counter_u64_add(buffreekvacnt, 1);
2064 bp->b_data = bp->b_kvabase = unmapped_buf;
2071 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2074 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2079 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2080 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2081 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2082 KASSERT(maxsize <= maxbcachebuf,
2083 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2088 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2091 * Buffer map is too fragmented. Request the caller
2092 * to defragment the map.
2096 bp->b_kvabase = (caddr_t)addr;
2097 bp->b_kvasize = maxsize;
2098 counter_u64_add(bufkvaspace, bp->b_kvasize);
2099 if ((gbflags & GB_UNMAPPED) != 0) {
2100 bp->b_data = unmapped_buf;
2101 BUF_CHECK_UNMAPPED(bp);
2103 bp->b_data = bp->b_kvabase;
2104 BUF_CHECK_MAPPED(bp);
2112 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2113 * callback that fires to avoid returning failure.
2116 bufkva_reclaim(vmem_t *vmem, int flags)
2123 for (i = 0; i < 5; i++) {
2124 for (q = 0; q < buf_domains; q++)
2125 if (buf_recycle(&bdomain[q], true) != 0)
2134 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2135 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2136 * the buffer is valid and we do not have to do anything.
2139 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2140 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2148 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2149 if (inmem(vp, *rablkno))
2151 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2152 if ((rabp->b_flags & B_CACHE) != 0) {
2159 racct_add_buf(curproc, rabp, 0);
2160 PROC_UNLOCK(curproc);
2163 td->td_ru.ru_inblock++;
2164 rabp->b_flags |= B_ASYNC;
2165 rabp->b_flags &= ~B_INVAL;
2166 if ((flags & GB_CKHASH) != 0) {
2167 rabp->b_flags |= B_CKHASH;
2168 rabp->b_ckhashcalc = ckhashfunc;
2170 rabp->b_ioflags &= ~BIO_ERROR;
2171 rabp->b_iocmd = BIO_READ;
2172 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2173 rabp->b_rcred = crhold(cred);
2174 vfs_busy_pages(rabp, 0);
2176 rabp->b_iooffset = dbtob(rabp->b_blkno);
2182 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2184 * Get a buffer with the specified data. Look in the cache first. We
2185 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2186 * is set, the buffer is valid and we do not have to do anything, see
2187 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2189 * Always return a NULL buffer pointer (in bpp) when returning an error.
2191 * The blkno parameter is the logical block being requested. Normally
2192 * the mapping of logical block number to disk block address is done
2193 * by calling VOP_BMAP(). However, if the mapping is already known, the
2194 * disk block address can be passed using the dblkno parameter. If the
2195 * disk block address is not known, then the same value should be passed
2196 * for blkno and dblkno.
2199 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2200 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2201 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2205 int error, readwait, rv;
2207 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2210 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2213 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2218 KASSERT(blkno == bp->b_lblkno,
2219 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2220 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2221 flags &= ~GB_NOSPARSE;
2225 * If not found in cache, do some I/O
2228 if ((bp->b_flags & B_CACHE) == 0) {
2231 PROC_LOCK(td->td_proc);
2232 racct_add_buf(td->td_proc, bp, 0);
2233 PROC_UNLOCK(td->td_proc);
2236 td->td_ru.ru_inblock++;
2237 bp->b_iocmd = BIO_READ;
2238 bp->b_flags &= ~B_INVAL;
2239 if ((flags & GB_CKHASH) != 0) {
2240 bp->b_flags |= B_CKHASH;
2241 bp->b_ckhashcalc = ckhashfunc;
2243 if ((flags & GB_CVTENXIO) != 0)
2244 bp->b_xflags |= BX_CVTENXIO;
2245 bp->b_ioflags &= ~BIO_ERROR;
2246 if (bp->b_rcred == NOCRED && cred != NOCRED)
2247 bp->b_rcred = crhold(cred);
2248 vfs_busy_pages(bp, 0);
2249 bp->b_iooffset = dbtob(bp->b_blkno);
2255 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2257 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2271 * Write, release buffer on completion. (Done by iodone
2272 * if async). Do not bother writing anything if the buffer
2275 * Note that we set B_CACHE here, indicating that buffer is
2276 * fully valid and thus cacheable. This is true even of NFS
2277 * now so we set it generally. This could be set either here
2278 * or in biodone() since the I/O is synchronous. We put it
2282 bufwrite(struct buf *bp)
2289 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2290 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2291 bp->b_flags |= B_INVAL | B_RELBUF;
2292 bp->b_flags &= ~B_CACHE;
2296 if (bp->b_flags & B_INVAL) {
2301 if (bp->b_flags & B_BARRIER)
2302 atomic_add_long(&barrierwrites, 1);
2304 oldflags = bp->b_flags;
2306 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2307 ("FFS background buffer should not get here %p", bp));
2311 vp_md = vp->v_vflag & VV_MD;
2316 * Mark the buffer clean. Increment the bufobj write count
2317 * before bundirty() call, to prevent other thread from seeing
2318 * empty dirty list and zero counter for writes in progress,
2319 * falsely indicating that the bufobj is clean.
2321 bufobj_wref(bp->b_bufobj);
2324 bp->b_flags &= ~B_DONE;
2325 bp->b_ioflags &= ~BIO_ERROR;
2326 bp->b_flags |= B_CACHE;
2327 bp->b_iocmd = BIO_WRITE;
2329 vfs_busy_pages(bp, 1);
2332 * Normal bwrites pipeline writes
2334 bp->b_runningbufspace = bp->b_bufsize;
2335 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2340 racct_add_buf(curproc, bp, 1);
2341 PROC_UNLOCK(curproc);
2344 curthread->td_ru.ru_oublock++;
2345 if (oldflags & B_ASYNC)
2347 bp->b_iooffset = dbtob(bp->b_blkno);
2348 buf_track(bp, __func__);
2351 if ((oldflags & B_ASYNC) == 0) {
2352 int rtval = bufwait(bp);
2355 } else if (space > hirunningspace) {
2357 * don't allow the async write to saturate the I/O
2358 * system. We will not deadlock here because
2359 * we are blocking waiting for I/O that is already in-progress
2360 * to complete. We do not block here if it is the update
2361 * or syncer daemon trying to clean up as that can lead
2364 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2365 waitrunningbufspace();
2372 bufbdflush(struct bufobj *bo, struct buf *bp)
2375 struct bufdomain *bd;
2377 bd = &bdomain[bo->bo_domain];
2378 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2379 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2381 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2384 * Try to find a buffer to flush.
2386 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2387 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2389 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2392 panic("bdwrite: found ourselves");
2394 /* Don't countdeps with the bo lock held. */
2395 if (buf_countdeps(nbp, 0)) {
2400 if (nbp->b_flags & B_CLUSTEROK) {
2401 vfs_bio_awrite(nbp);
2406 dirtybufferflushes++;
2415 * Delayed write. (Buffer is marked dirty). Do not bother writing
2416 * anything if the buffer is marked invalid.
2418 * Note that since the buffer must be completely valid, we can safely
2419 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2420 * biodone() in order to prevent getblk from writing the buffer
2421 * out synchronously.
2424 bdwrite(struct buf *bp)
2426 struct thread *td = curthread;
2430 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2431 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2432 KASSERT((bp->b_flags & B_BARRIER) == 0,
2433 ("Barrier request in delayed write %p", bp));
2435 if (bp->b_flags & B_INVAL) {
2441 * If we have too many dirty buffers, don't create any more.
2442 * If we are wildly over our limit, then force a complete
2443 * cleanup. Otherwise, just keep the situation from getting
2444 * out of control. Note that we have to avoid a recursive
2445 * disaster and not try to clean up after our own cleanup!
2449 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2450 td->td_pflags |= TDP_INBDFLUSH;
2452 td->td_pflags &= ~TDP_INBDFLUSH;
2458 * Set B_CACHE, indicating that the buffer is fully valid. This is
2459 * true even of NFS now.
2461 bp->b_flags |= B_CACHE;
2464 * This bmap keeps the system from needing to do the bmap later,
2465 * perhaps when the system is attempting to do a sync. Since it
2466 * is likely that the indirect block -- or whatever other datastructure
2467 * that the filesystem needs is still in memory now, it is a good
2468 * thing to do this. Note also, that if the pageout daemon is
2469 * requesting a sync -- there might not be enough memory to do
2470 * the bmap then... So, this is important to do.
2472 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2473 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2476 buf_track(bp, __func__);
2479 * Set the *dirty* buffer range based upon the VM system dirty
2482 * Mark the buffer pages as clean. We need to do this here to
2483 * satisfy the vnode_pager and the pageout daemon, so that it
2484 * thinks that the pages have been "cleaned". Note that since
2485 * the pages are in a delayed write buffer -- the VFS layer
2486 * "will" see that the pages get written out on the next sync,
2487 * or perhaps the cluster will be completed.
2489 vfs_clean_pages_dirty_buf(bp);
2493 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2494 * due to the softdep code.
2501 * Turn buffer into delayed write request. We must clear BIO_READ and
2502 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2503 * itself to properly update it in the dirty/clean lists. We mark it
2504 * B_DONE to ensure that any asynchronization of the buffer properly
2505 * clears B_DONE ( else a panic will occur later ).
2507 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2508 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2509 * should only be called if the buffer is known-good.
2511 * Since the buffer is not on a queue, we do not update the numfreebuffers
2514 * The buffer must be on QUEUE_NONE.
2517 bdirty(struct buf *bp)
2520 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2521 bp, bp->b_vp, bp->b_flags);
2522 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2523 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2524 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2525 bp->b_flags &= ~(B_RELBUF);
2526 bp->b_iocmd = BIO_WRITE;
2528 if ((bp->b_flags & B_DELWRI) == 0) {
2529 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2538 * Clear B_DELWRI for buffer.
2540 * Since the buffer is not on a queue, we do not update the numfreebuffers
2543 * The buffer must be on QUEUE_NONE.
2547 bundirty(struct buf *bp)
2550 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2551 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2552 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2553 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2555 if (bp->b_flags & B_DELWRI) {
2556 bp->b_flags &= ~B_DELWRI;
2561 * Since it is now being written, we can clear its deferred write flag.
2563 bp->b_flags &= ~B_DEFERRED;
2569 * Asynchronous write. Start output on a buffer, but do not wait for
2570 * it to complete. The buffer is released when the output completes.
2572 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2573 * B_INVAL buffers. Not us.
2576 bawrite(struct buf *bp)
2579 bp->b_flags |= B_ASYNC;
2586 * Asynchronous barrier write. Start output on a buffer, but do not
2587 * wait for it to complete. Place a write barrier after this write so
2588 * that this buffer and all buffers written before it are committed to
2589 * the disk before any buffers written after this write are committed
2590 * to the disk. The buffer is released when the output completes.
2593 babarrierwrite(struct buf *bp)
2596 bp->b_flags |= B_ASYNC | B_BARRIER;
2603 * Synchronous barrier write. Start output on a buffer and wait for
2604 * it to complete. Place a write barrier after this write so that
2605 * this buffer and all buffers written before it are committed to
2606 * the disk before any buffers written after this write are committed
2607 * to the disk. The buffer is released when the output completes.
2610 bbarrierwrite(struct buf *bp)
2613 bp->b_flags |= B_BARRIER;
2614 return (bwrite(bp));
2620 * Called prior to the locking of any vnodes when we are expecting to
2621 * write. We do not want to starve the buffer cache with too many
2622 * dirty buffers so we block here. By blocking prior to the locking
2623 * of any vnodes we attempt to avoid the situation where a locked vnode
2624 * prevents the various system daemons from flushing related buffers.
2630 if (buf_dirty_count_severe()) {
2631 mtx_lock(&bdirtylock);
2632 while (buf_dirty_count_severe()) {
2634 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2637 mtx_unlock(&bdirtylock);
2642 * Return true if we have too many dirty buffers.
2645 buf_dirty_count_severe(void)
2648 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2654 * Release a busy buffer and, if requested, free its resources. The
2655 * buffer will be stashed in the appropriate bufqueue[] allowing it
2656 * to be accessed later as a cache entity or reused for other purposes.
2659 brelse(struct buf *bp)
2661 struct mount *v_mnt;
2665 * Many functions erroneously call brelse with a NULL bp under rare
2666 * error conditions. Simply return when called with a NULL bp.
2670 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2671 bp, bp->b_vp, bp->b_flags);
2672 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2673 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2674 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2675 ("brelse: non-VMIO buffer marked NOREUSE"));
2677 if (BUF_LOCKRECURSED(bp)) {
2679 * Do not process, in particular, do not handle the
2680 * B_INVAL/B_RELBUF and do not release to free list.
2686 if (bp->b_flags & B_MANAGED) {
2691 if (LIST_EMPTY(&bp->b_dep)) {
2692 bp->b_flags &= ~B_IOSTARTED;
2694 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2695 ("brelse: SU io not finished bp %p", bp));
2698 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2699 BO_LOCK(bp->b_bufobj);
2700 bp->b_vflags &= ~BV_BKGRDERR;
2701 BO_UNLOCK(bp->b_bufobj);
2705 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2706 (bp->b_flags & B_INVALONERR)) {
2708 * Forced invalidation of dirty buffer contents, to be used
2709 * after a failed write in the rare case that the loss of the
2710 * contents is acceptable. The buffer is invalidated and
2713 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2714 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2717 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2718 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2719 !(bp->b_flags & B_INVAL)) {
2721 * Failed write, redirty. All errors except ENXIO (which
2722 * means the device is gone) are treated as being
2725 * XXX Treating EIO as transient is not correct; the
2726 * contract with the local storage device drivers is that
2727 * they will only return EIO once the I/O is no longer
2728 * retriable. Network I/O also respects this through the
2729 * guarantees of TCP and/or the internal retries of NFS.
2730 * ENOMEM might be transient, but we also have no way of
2731 * knowing when its ok to retry/reschedule. In general,
2732 * this entire case should be made obsolete through better
2733 * error handling/recovery and resource scheduling.
2735 * Do this also for buffers that failed with ENXIO, but have
2736 * non-empty dependencies - the soft updates code might need
2737 * to access the buffer to untangle them.
2739 * Must clear BIO_ERROR to prevent pages from being scrapped.
2741 bp->b_ioflags &= ~BIO_ERROR;
2743 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2744 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2746 * Either a failed read I/O, or we were asked to free or not
2747 * cache the buffer, or we failed to write to a device that's
2748 * no longer present.
2750 bp->b_flags |= B_INVAL;
2751 if (!LIST_EMPTY(&bp->b_dep))
2753 if (bp->b_flags & B_DELWRI)
2755 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2756 if ((bp->b_flags & B_VMIO) == 0) {
2764 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2765 * is called with B_DELWRI set, the underlying pages may wind up
2766 * getting freed causing a previous write (bdwrite()) to get 'lost'
2767 * because pages associated with a B_DELWRI bp are marked clean.
2769 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2770 * if B_DELWRI is set.
2772 if (bp->b_flags & B_DELWRI)
2773 bp->b_flags &= ~B_RELBUF;
2776 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2777 * constituted, not even NFS buffers now. Two flags effect this. If
2778 * B_INVAL, the struct buf is invalidated but the VM object is kept
2779 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2781 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2782 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2783 * buffer is also B_INVAL because it hits the re-dirtying code above.
2785 * Normally we can do this whether a buffer is B_DELWRI or not. If
2786 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2787 * the commit state and we cannot afford to lose the buffer. If the
2788 * buffer has a background write in progress, we need to keep it
2789 * around to prevent it from being reconstituted and starting a second
2793 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2795 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2796 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2797 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2798 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2799 vfs_vmio_invalidate(bp);
2803 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2804 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2806 bp->b_flags &= ~B_NOREUSE;
2807 if (bp->b_vp != NULL)
2812 * If the buffer has junk contents signal it and eventually
2813 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2816 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2817 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2818 bp->b_flags |= B_INVAL;
2819 if (bp->b_flags & B_INVAL) {
2820 if (bp->b_flags & B_DELWRI)
2826 buf_track(bp, __func__);
2828 /* buffers with no memory */
2829 if (bp->b_bufsize == 0) {
2833 /* buffers with junk contents */
2834 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2835 (bp->b_ioflags & BIO_ERROR)) {
2836 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2837 if (bp->b_vflags & BV_BKGRDINPROG)
2838 panic("losing buffer 2");
2839 qindex = QUEUE_CLEAN;
2840 bp->b_flags |= B_AGE;
2841 /* remaining buffers */
2842 } else if (bp->b_flags & B_DELWRI)
2843 qindex = QUEUE_DIRTY;
2845 qindex = QUEUE_CLEAN;
2847 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2848 panic("brelse: not dirty");
2850 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2851 bp->b_xflags &= ~(BX_CVTENXIO);
2852 /* binsfree unlocks bp. */
2853 binsfree(bp, qindex);
2857 * Release a buffer back to the appropriate queue but do not try to free
2858 * it. The buffer is expected to be used again soon.
2860 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2861 * biodone() to requeue an async I/O on completion. It is also used when
2862 * known good buffers need to be requeued but we think we may need the data
2865 * XXX we should be able to leave the B_RELBUF hint set on completion.
2868 bqrelse(struct buf *bp)
2872 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2873 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2874 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2876 qindex = QUEUE_NONE;
2877 if (BUF_LOCKRECURSED(bp)) {
2878 /* do not release to free list */
2882 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2883 bp->b_xflags &= ~(BX_CVTENXIO);
2885 if (LIST_EMPTY(&bp->b_dep)) {
2886 bp->b_flags &= ~B_IOSTARTED;
2888 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2889 ("bqrelse: SU io not finished bp %p", bp));
2892 if (bp->b_flags & B_MANAGED) {
2893 if (bp->b_flags & B_REMFREE)
2898 /* buffers with stale but valid contents */
2899 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2900 BV_BKGRDERR)) == BV_BKGRDERR) {
2901 BO_LOCK(bp->b_bufobj);
2902 bp->b_vflags &= ~BV_BKGRDERR;
2903 BO_UNLOCK(bp->b_bufobj);
2904 qindex = QUEUE_DIRTY;
2906 if ((bp->b_flags & B_DELWRI) == 0 &&
2907 (bp->b_xflags & BX_VNDIRTY))
2908 panic("bqrelse: not dirty");
2909 if ((bp->b_flags & B_NOREUSE) != 0) {
2913 qindex = QUEUE_CLEAN;
2915 buf_track(bp, __func__);
2916 /* binsfree unlocks bp. */
2917 binsfree(bp, qindex);
2921 buf_track(bp, __func__);
2927 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2928 * restore bogus pages.
2931 vfs_vmio_iodone(struct buf *bp)
2936 struct vnode *vp __unused;
2937 int i, iosize, resid;
2940 obj = bp->b_bufobj->bo_object;
2941 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2942 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2943 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2946 VNPASS(vp->v_holdcnt > 0, vp);
2947 VNPASS(vp->v_object != NULL, vp);
2949 foff = bp->b_offset;
2950 KASSERT(bp->b_offset != NOOFFSET,
2951 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2954 iosize = bp->b_bcount - bp->b_resid;
2955 for (i = 0; i < bp->b_npages; i++) {
2956 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2961 * cleanup bogus pages, restoring the originals
2964 if (m == bogus_page) {
2966 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2968 panic("biodone: page disappeared!");
2970 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2972 * In the write case, the valid and clean bits are
2973 * already changed correctly ( see bdwrite() ), so we
2974 * only need to do this here in the read case.
2976 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2977 resid)) == 0, ("vfs_vmio_iodone: page %p "
2978 "has unexpected dirty bits", m));
2979 vfs_page_set_valid(bp, foff, m);
2981 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2982 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2983 (intmax_t)foff, (uintmax_t)m->pindex));
2986 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2989 vm_object_pip_wakeupn(obj, bp->b_npages);
2990 if (bogus && buf_mapped(bp)) {
2991 BUF_CHECK_MAPPED(bp);
2992 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2993 bp->b_pages, bp->b_npages);
2998 * Perform page invalidation when a buffer is released. The fully invalid
2999 * pages will be reclaimed later in vfs_vmio_truncate().
3002 vfs_vmio_invalidate(struct buf *bp)
3006 int flags, i, resid, poffset, presid;
3008 if (buf_mapped(bp)) {
3009 BUF_CHECK_MAPPED(bp);
3010 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3012 BUF_CHECK_UNMAPPED(bp);
3014 * Get the base offset and length of the buffer. Note that
3015 * in the VMIO case if the buffer block size is not
3016 * page-aligned then b_data pointer may not be page-aligned.
3017 * But our b_pages[] array *IS* page aligned.
3019 * block sizes less then DEV_BSIZE (usually 512) are not
3020 * supported due to the page granularity bits (m->valid,
3021 * m->dirty, etc...).
3023 * See man buf(9) for more information
3025 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3026 obj = bp->b_bufobj->bo_object;
3027 resid = bp->b_bufsize;
3028 poffset = bp->b_offset & PAGE_MASK;
3029 VM_OBJECT_WLOCK(obj);
3030 for (i = 0; i < bp->b_npages; i++) {
3032 if (m == bogus_page)
3033 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3034 bp->b_pages[i] = NULL;
3036 presid = resid > (PAGE_SIZE - poffset) ?
3037 (PAGE_SIZE - poffset) : resid;
3038 KASSERT(presid >= 0, ("brelse: extra page"));
3039 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3040 if (pmap_page_wired_mappings(m) == 0)
3041 vm_page_set_invalid(m, poffset, presid);
3043 vm_page_release_locked(m, flags);
3047 VM_OBJECT_WUNLOCK(obj);
3052 * Page-granular truncation of an existing VMIO buffer.
3055 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3061 if (bp->b_npages == desiredpages)
3064 if (buf_mapped(bp)) {
3065 BUF_CHECK_MAPPED(bp);
3066 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3067 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3069 BUF_CHECK_UNMAPPED(bp);
3072 * The object lock is needed only if we will attempt to free pages.
3074 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3075 if ((bp->b_flags & B_DIRECT) != 0) {
3076 flags |= VPR_TRYFREE;
3077 obj = bp->b_bufobj->bo_object;
3078 VM_OBJECT_WLOCK(obj);
3082 for (i = desiredpages; i < bp->b_npages; i++) {
3084 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3085 bp->b_pages[i] = NULL;
3087 vm_page_release_locked(m, flags);
3089 vm_page_release(m, flags);
3092 VM_OBJECT_WUNLOCK(obj);
3093 bp->b_npages = desiredpages;
3097 * Byte granular extension of VMIO buffers.
3100 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3103 * We are growing the buffer, possibly in a
3104 * byte-granular fashion.
3112 * Step 1, bring in the VM pages from the object, allocating
3113 * them if necessary. We must clear B_CACHE if these pages
3114 * are not valid for the range covered by the buffer.
3116 obj = bp->b_bufobj->bo_object;
3117 if (bp->b_npages < desiredpages) {
3118 KASSERT(desiredpages <= atop(maxbcachebuf),
3119 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3120 bp, desiredpages, maxbcachebuf));
3123 * We must allocate system pages since blocking
3124 * here could interfere with paging I/O, no
3125 * matter which process we are.
3127 * Only exclusive busy can be tested here.
3128 * Blocking on shared busy might lead to
3129 * deadlocks once allocbuf() is called after
3130 * pages are vfs_busy_pages().
3132 (void)vm_page_grab_pages_unlocked(obj,
3133 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3134 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3135 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3136 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3137 bp->b_npages = desiredpages;
3141 * Step 2. We've loaded the pages into the buffer,
3142 * we have to figure out if we can still have B_CACHE
3143 * set. Note that B_CACHE is set according to the
3144 * byte-granular range ( bcount and size ), not the
3145 * aligned range ( newbsize ).
3147 * The VM test is against m->valid, which is DEV_BSIZE
3148 * aligned. Needless to say, the validity of the data
3149 * needs to also be DEV_BSIZE aligned. Note that this
3150 * fails with NFS if the server or some other client
3151 * extends the file's EOF. If our buffer is resized,
3152 * B_CACHE may remain set! XXX
3154 toff = bp->b_bcount;
3155 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3156 while ((bp->b_flags & B_CACHE) && toff < size) {
3159 if (tinc > (size - toff))
3161 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3162 m = bp->b_pages[pi];
3163 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3169 * Step 3, fixup the KVA pmap.
3174 BUF_CHECK_UNMAPPED(bp);
3178 * Check to see if a block at a particular lbn is available for a clustered
3182 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3189 /* If the buf isn't in core skip it */
3190 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3193 /* If the buf is busy we don't want to wait for it */
3194 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3197 /* Only cluster with valid clusterable delayed write buffers */
3198 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3199 (B_DELWRI | B_CLUSTEROK))
3202 if (bpa->b_bufsize != size)
3206 * Check to see if it is in the expected place on disk and that the
3207 * block has been mapped.
3209 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3219 * Implement clustered async writes for clearing out B_DELWRI buffers.
3220 * This is much better then the old way of writing only one buffer at
3221 * a time. Note that we may not be presented with the buffers in the
3222 * correct order, so we search for the cluster in both directions.
3225 vfs_bio_awrite(struct buf *bp)
3230 daddr_t lblkno = bp->b_lblkno;
3231 struct vnode *vp = bp->b_vp;
3239 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3241 * right now we support clustered writing only to regular files. If
3242 * we find a clusterable block we could be in the middle of a cluster
3243 * rather then at the beginning.
3245 if ((vp->v_type == VREG) &&
3246 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3247 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3248 size = vp->v_mount->mnt_stat.f_iosize;
3249 maxcl = maxphys / size;
3252 for (i = 1; i < maxcl; i++)
3253 if (vfs_bio_clcheck(vp, size, lblkno + i,
3254 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3257 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3258 if (vfs_bio_clcheck(vp, size, lblkno - j,
3259 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3265 * this is a possible cluster write
3269 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3275 bp->b_flags |= B_ASYNC;
3277 * default (old) behavior, writing out only one block
3279 * XXX returns b_bufsize instead of b_bcount for nwritten?
3281 nwritten = bp->b_bufsize;
3290 * Allocate KVA for an empty buf header according to gbflags.
3293 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3296 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3298 * In order to keep fragmentation sane we only allocate kva
3299 * in BKVASIZE chunks. XXX with vmem we can do page size.
3301 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3303 if (maxsize != bp->b_kvasize &&
3304 bufkva_alloc(bp, maxsize, gbflags))
3313 * Find and initialize a new buffer header, freeing up existing buffers
3314 * in the bufqueues as necessary. The new buffer is returned locked.
3317 * We have insufficient buffer headers
3318 * We have insufficient buffer space
3319 * buffer_arena is too fragmented ( space reservation fails )
3320 * If we have to flush dirty buffers ( but we try to avoid this )
3322 * The caller is responsible for releasing the reserved bufspace after
3323 * allocbuf() is called.
3326 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3328 struct bufdomain *bd;
3330 bool metadata, reserved;
3333 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3334 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3335 if (!unmapped_buf_allowed)
3336 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3338 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3346 bd = &bdomain[vp->v_bufobj.bo_domain];
3348 counter_u64_add(getnewbufcalls, 1);
3351 if (reserved == false &&
3352 bufspace_reserve(bd, maxsize, metadata) != 0) {
3353 counter_u64_add(getnewbufrestarts, 1);
3357 if ((bp = buf_alloc(bd)) == NULL) {
3358 counter_u64_add(getnewbufrestarts, 1);
3361 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3364 } while (buf_recycle(bd, false) == 0);
3367 bufspace_release(bd, maxsize);
3369 bp->b_flags |= B_INVAL;
3372 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3380 * buffer flushing daemon. Buffers are normally flushed by the
3381 * update daemon but if it cannot keep up this process starts to
3382 * take the load in an attempt to prevent getnewbuf() from blocking.
3384 static struct kproc_desc buf_kp = {
3389 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3392 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3396 flushed = flushbufqueues(vp, bd, target, 0);
3399 * Could not find any buffers without rollback
3400 * dependencies, so just write the first one
3401 * in the hopes of eventually making progress.
3403 if (vp != NULL && target > 2)
3405 flushbufqueues(vp, bd, target, 1);
3411 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3417 wakeup(&bd_request);
3418 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3420 mtx_unlock(&bdlock);
3422 printf("bufdaemon wait error: %d\n", error);
3428 struct bufdomain *bd;
3434 * This process needs to be suspended prior to shutdown sync.
3436 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3437 SHUTDOWN_PRI_LAST + 100);
3440 * Start the buf clean daemons as children threads.
3442 for (i = 0 ; i < buf_domains; i++) {
3445 error = kthread_add((void (*)(void *))bufspace_daemon,
3446 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3448 panic("error %d spawning bufspace daemon", error);
3452 * This process is allowed to take the buffer cache to the limit
3454 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3456 while (!bd_shutdown) {
3458 mtx_unlock(&bdlock);
3461 * Save speedupreq for this pass and reset to capture new
3464 speedupreq = bd_speedupreq;
3468 * Flush each domain sequentially according to its level and
3469 * the speedup request.
3471 for (i = 0; i < buf_domains; i++) {
3474 lodirty = bd->bd_numdirtybuffers / 2;
3476 lodirty = bd->bd_lodirtybuffers;
3477 while (bd->bd_numdirtybuffers > lodirty) {
3478 if (buf_flush(NULL, bd,
3479 bd->bd_numdirtybuffers - lodirty) == 0)
3481 kern_yield(PRI_USER);
3486 * Only clear bd_request if we have reached our low water
3487 * mark. The buf_daemon normally waits 1 second and
3488 * then incrementally flushes any dirty buffers that have
3489 * built up, within reason.
3491 * If we were unable to hit our low water mark and couldn't
3492 * find any flushable buffers, we sleep for a short period
3493 * to avoid endless loops on unlockable buffers.
3498 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3500 * We reached our low water mark, reset the
3501 * request and sleep until we are needed again.
3502 * The sleep is just so the suspend code works.
3506 * Do an extra wakeup in case dirty threshold
3507 * changed via sysctl and the explicit transition
3508 * out of shortfall was missed.
3511 if (runningbufspace <= lorunningspace)
3513 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3516 * We couldn't find any flushable dirty buffers but
3517 * still have too many dirty buffers, we
3518 * have to sleep and try again. (rare)
3520 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3523 wakeup(&bd_shutdown);
3524 mtx_unlock(&bdlock);
3531 * Try to flush a buffer in the dirty queue. We must be careful to
3532 * free up B_INVAL buffers instead of write them, which NFS is
3533 * particularly sensitive to.
3535 static int flushwithdeps = 0;
3536 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3538 "Number of buffers flushed with dependencies that require rollbacks");
3541 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3544 struct bufqueue *bq;
3545 struct buf *sentinel;
3555 bq = &bd->bd_dirtyq;
3557 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3558 sentinel->b_qindex = QUEUE_SENTINEL;
3560 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3562 while (flushed != target) {
3565 bp = TAILQ_NEXT(sentinel, b_freelist);
3567 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3568 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3575 * Skip sentinels inserted by other invocations of the
3576 * flushbufqueues(), taking care to not reorder them.
3578 * Only flush the buffers that belong to the
3579 * vnode locked by the curthread.
3581 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3586 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3592 * BKGRDINPROG can only be set with the buf and bufobj
3593 * locks both held. We tolerate a race to clear it here.
3595 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3596 (bp->b_flags & B_DELWRI) == 0) {
3600 if (bp->b_flags & B_INVAL) {
3607 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3608 if (flushdeps == 0) {
3616 * We must hold the lock on a vnode before writing
3617 * one of its buffers. Otherwise we may confuse, or
3618 * in the case of a snapshot vnode, deadlock the
3621 * The lock order here is the reverse of the normal
3622 * of vnode followed by buf lock. This is ok because
3623 * the NOWAIT will prevent deadlock.
3626 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3632 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3634 ASSERT_VOP_LOCKED(vp, "getbuf");
3636 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3637 vn_lock(vp, LK_TRYUPGRADE);
3640 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3641 bp, bp->b_vp, bp->b_flags);
3642 if (curproc == bufdaemonproc) {
3647 counter_u64_add(notbufdflushes, 1);
3649 vn_finished_write(mp);
3652 flushwithdeps += hasdeps;
3656 * Sleeping on runningbufspace while holding
3657 * vnode lock leads to deadlock.
3659 if (curproc == bufdaemonproc &&
3660 runningbufspace > hirunningspace)
3661 waitrunningbufspace();
3664 vn_finished_write(mp);
3668 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3670 free(sentinel, M_TEMP);
3675 * Check to see if a block is currently memory resident.
3678 incore(struct bufobj *bo, daddr_t blkno)
3680 return (gbincore_unlocked(bo, blkno));
3684 * Returns true if no I/O is needed to access the
3685 * associated VM object. This is like incore except
3686 * it also hunts around in the VM system for the data.
3689 inmem(struct vnode * vp, daddr_t blkno)
3692 vm_offset_t toff, tinc, size;
3697 ASSERT_VOP_LOCKED(vp, "inmem");
3699 if (incore(&vp->v_bufobj, blkno))
3701 if (vp->v_mount == NULL)
3708 if (size > vp->v_mount->mnt_stat.f_iosize)
3709 size = vp->v_mount->mnt_stat.f_iosize;
3710 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3712 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3713 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3719 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3720 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3722 * Consider page validity only if page mapping didn't change
3725 valid = vm_page_is_valid(m,
3726 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3727 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3739 * Set the dirty range for a buffer based on the status of the dirty
3740 * bits in the pages comprising the buffer. The range is limited
3741 * to the size of the buffer.
3743 * Tell the VM system that the pages associated with this buffer
3744 * are clean. This is used for delayed writes where the data is
3745 * going to go to disk eventually without additional VM intevention.
3747 * Note that while we only really need to clean through to b_bcount, we
3748 * just go ahead and clean through to b_bufsize.
3751 vfs_clean_pages_dirty_buf(struct buf *bp)
3753 vm_ooffset_t foff, noff, eoff;
3757 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3760 foff = bp->b_offset;
3761 KASSERT(bp->b_offset != NOOFFSET,
3762 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3764 vfs_busy_pages_acquire(bp);
3765 vfs_setdirty_range(bp);
3766 for (i = 0; i < bp->b_npages; i++) {
3767 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3769 if (eoff > bp->b_offset + bp->b_bufsize)
3770 eoff = bp->b_offset + bp->b_bufsize;
3772 vfs_page_set_validclean(bp, foff, m);
3773 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3776 vfs_busy_pages_release(bp);
3780 vfs_setdirty_range(struct buf *bp)
3782 vm_offset_t boffset;
3783 vm_offset_t eoffset;
3787 * test the pages to see if they have been modified directly
3788 * by users through the VM system.
3790 for (i = 0; i < bp->b_npages; i++)
3791 vm_page_test_dirty(bp->b_pages[i]);
3794 * Calculate the encompassing dirty range, boffset and eoffset,
3795 * (eoffset - boffset) bytes.
3798 for (i = 0; i < bp->b_npages; i++) {
3799 if (bp->b_pages[i]->dirty)
3802 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3804 for (i = bp->b_npages - 1; i >= 0; --i) {
3805 if (bp->b_pages[i]->dirty) {
3809 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3812 * Fit it to the buffer.
3815 if (eoffset > bp->b_bcount)
3816 eoffset = bp->b_bcount;
3819 * If we have a good dirty range, merge with the existing
3823 if (boffset < eoffset) {
3824 if (bp->b_dirtyoff > boffset)
3825 bp->b_dirtyoff = boffset;
3826 if (bp->b_dirtyend < eoffset)
3827 bp->b_dirtyend = eoffset;
3832 * Allocate the KVA mapping for an existing buffer.
3833 * If an unmapped buffer is provided but a mapped buffer is requested, take
3834 * also care to properly setup mappings between pages and KVA.
3837 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3839 int bsize, maxsize, need_mapping, need_kva;
3842 need_mapping = bp->b_data == unmapped_buf &&
3843 (gbflags & GB_UNMAPPED) == 0;
3844 need_kva = bp->b_kvabase == unmapped_buf &&
3845 bp->b_data == unmapped_buf &&
3846 (gbflags & GB_KVAALLOC) != 0;
3847 if (!need_mapping && !need_kva)
3850 BUF_CHECK_UNMAPPED(bp);
3852 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3854 * Buffer is not mapped, but the KVA was already
3855 * reserved at the time of the instantiation. Use the
3862 * Calculate the amount of the address space we would reserve
3863 * if the buffer was mapped.
3865 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3866 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3867 offset = blkno * bsize;
3868 maxsize = size + (offset & PAGE_MASK);
3869 maxsize = imax(maxsize, bsize);
3871 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3872 if ((gbflags & GB_NOWAIT_BD) != 0) {
3874 * XXXKIB: defragmentation cannot
3875 * succeed, not sure what else to do.
3877 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3879 counter_u64_add(mappingrestarts, 1);
3880 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3884 /* b_offset is handled by bpmap_qenter. */
3885 bp->b_data = bp->b_kvabase;
3886 BUF_CHECK_MAPPED(bp);
3892 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3898 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3907 * Get a block given a specified block and offset into a file/device.
3908 * The buffers B_DONE bit will be cleared on return, making it almost
3909 * ready for an I/O initiation. B_INVAL may or may not be set on
3910 * return. The caller should clear B_INVAL prior to initiating a
3913 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3914 * an existing buffer.
3916 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3917 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3918 * and then cleared based on the backing VM. If the previous buffer is
3919 * non-0-sized but invalid, B_CACHE will be cleared.
3921 * If getblk() must create a new buffer, the new buffer is returned with
3922 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3923 * case it is returned with B_INVAL clear and B_CACHE set based on the
3926 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3927 * B_CACHE bit is clear.
3929 * What this means, basically, is that the caller should use B_CACHE to
3930 * determine whether the buffer is fully valid or not and should clear
3931 * B_INVAL prior to issuing a read. If the caller intends to validate
3932 * the buffer by loading its data area with something, the caller needs
3933 * to clear B_INVAL. If the caller does this without issuing an I/O,
3934 * the caller should set B_CACHE ( as an optimization ), else the caller
3935 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3936 * a write attempt or if it was a successful read. If the caller
3937 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3938 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3940 * The blkno parameter is the logical block being requested. Normally
3941 * the mapping of logical block number to disk block address is done
3942 * by calling VOP_BMAP(). However, if the mapping is already known, the
3943 * disk block address can be passed using the dblkno parameter. If the
3944 * disk block address is not known, then the same value should be passed
3945 * for blkno and dblkno.
3948 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3949 int slptimeo, int flags, struct buf **bpp)
3954 int bsize, error, maxsize, vmio;
3957 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3958 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3959 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3960 if (vp->v_type != VCHR)
3961 ASSERT_VOP_LOCKED(vp, "getblk");
3962 if (size > maxbcachebuf)
3963 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3965 if (!unmapped_buf_allowed)
3966 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3971 /* Attempt lockless lookup first. */
3972 bp = gbincore_unlocked(bo, blkno);
3975 * With GB_NOCREAT we must be sure about not finding the buffer
3976 * as it may have been reassigned during unlocked lookup.
3978 if ((flags & GB_NOCREAT) != 0)
3980 goto newbuf_unlocked;
3983 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3988 /* Verify buf identify has not changed since lookup. */
3989 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3990 goto foundbuf_fastpath;
3992 /* It changed, fallback to locked lookup. */
3997 bp = gbincore(bo, blkno);
4002 * Buffer is in-core. If the buffer is not busy nor managed,
4003 * it must be on a queue.
4005 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
4006 ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL);
4008 error = BUF_TIMELOCK(bp, lockflags,
4009 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4012 * If we slept and got the lock we have to restart in case
4013 * the buffer changed identities.
4015 if (error == ENOLCK)
4017 /* We timed out or were interrupted. */
4018 else if (error != 0)
4022 /* If recursed, assume caller knows the rules. */
4023 if (BUF_LOCKRECURSED(bp))
4027 * The buffer is locked. B_CACHE is cleared if the buffer is
4028 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4029 * and for a VMIO buffer B_CACHE is adjusted according to the
4032 if (bp->b_flags & B_INVAL)
4033 bp->b_flags &= ~B_CACHE;
4034 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4035 bp->b_flags |= B_CACHE;
4036 if (bp->b_flags & B_MANAGED)
4037 MPASS(bp->b_qindex == QUEUE_NONE);
4042 * check for size inconsistencies for non-VMIO case.
4044 if (bp->b_bcount != size) {
4045 if ((bp->b_flags & B_VMIO) == 0 ||
4046 (size > bp->b_kvasize)) {
4047 if (bp->b_flags & B_DELWRI) {
4048 bp->b_flags |= B_NOCACHE;
4051 if (LIST_EMPTY(&bp->b_dep)) {
4052 bp->b_flags |= B_RELBUF;
4055 bp->b_flags |= B_NOCACHE;
4064 * Handle the case of unmapped buffer which should
4065 * become mapped, or the buffer for which KVA
4066 * reservation is requested.
4068 bp_unmapped_get_kva(bp, blkno, size, flags);
4071 * If the size is inconsistent in the VMIO case, we can resize
4072 * the buffer. This might lead to B_CACHE getting set or
4073 * cleared. If the size has not changed, B_CACHE remains
4074 * unchanged from its previous state.
4078 KASSERT(bp->b_offset != NOOFFSET,
4079 ("getblk: no buffer offset"));
4082 * A buffer with B_DELWRI set and B_CACHE clear must
4083 * be committed before we can return the buffer in
4084 * order to prevent the caller from issuing a read
4085 * ( due to B_CACHE not being set ) and overwriting
4088 * Most callers, including NFS and FFS, need this to
4089 * operate properly either because they assume they
4090 * can issue a read if B_CACHE is not set, or because
4091 * ( for example ) an uncached B_DELWRI might loop due
4092 * to softupdates re-dirtying the buffer. In the latter
4093 * case, B_CACHE is set after the first write completes,
4094 * preventing further loops.
4095 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4096 * above while extending the buffer, we cannot allow the
4097 * buffer to remain with B_CACHE set after the write
4098 * completes or it will represent a corrupt state. To
4099 * deal with this we set B_NOCACHE to scrap the buffer
4102 * We might be able to do something fancy, like setting
4103 * B_CACHE in bwrite() except if B_DELWRI is already set,
4104 * so the below call doesn't set B_CACHE, but that gets real
4105 * confusing. This is much easier.
4108 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4109 bp->b_flags |= B_NOCACHE;
4113 bp->b_flags &= ~B_DONE;
4116 * Buffer is not in-core, create new buffer. The buffer
4117 * returned by getnewbuf() is locked. Note that the returned
4118 * buffer is also considered valid (not marked B_INVAL).
4123 * If the user does not want us to create the buffer, bail out
4126 if (flags & GB_NOCREAT)
4129 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4130 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4131 offset = blkno * bsize;
4132 vmio = vp->v_object != NULL;
4134 maxsize = size + (offset & PAGE_MASK);
4137 /* Do not allow non-VMIO notmapped buffers. */
4138 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4140 maxsize = imax(maxsize, bsize);
4141 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4143 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4144 KASSERT(error != EOPNOTSUPP,
4145 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4150 return (EJUSTRETURN);
4153 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4155 if (slpflag || slptimeo)
4158 * XXX This is here until the sleep path is diagnosed
4159 * enough to work under very low memory conditions.
4161 * There's an issue on low memory, 4BSD+non-preempt
4162 * systems (eg MIPS routers with 32MB RAM) where buffer
4163 * exhaustion occurs without sleeping for buffer
4164 * reclaimation. This just sticks in a loop and
4165 * constantly attempts to allocate a buffer, which
4166 * hits exhaustion and tries to wakeup bufdaemon.
4167 * This never happens because we never yield.
4169 * The real solution is to identify and fix these cases
4170 * so we aren't effectively busy-waiting in a loop
4171 * until the reclaimation path has cycles to run.
4173 kern_yield(PRI_USER);
4178 * This code is used to make sure that a buffer is not
4179 * created while the getnewbuf routine is blocked.
4180 * This can be a problem whether the vnode is locked or not.
4181 * If the buffer is created out from under us, we have to
4182 * throw away the one we just created.
4184 * Note: this must occur before we associate the buffer
4185 * with the vp especially considering limitations in
4186 * the splay tree implementation when dealing with duplicate
4190 if (gbincore(bo, blkno)) {
4192 bp->b_flags |= B_INVAL;
4193 bufspace_release(bufdomain(bp), maxsize);
4199 * Insert the buffer into the hash, so that it can
4200 * be found by incore.
4202 bp->b_lblkno = blkno;
4203 bp->b_blkno = d_blkno;
4204 bp->b_offset = offset;
4209 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4210 * buffer size starts out as 0, B_CACHE will be set by
4211 * allocbuf() for the VMIO case prior to it testing the
4212 * backing store for validity.
4216 bp->b_flags |= B_VMIO;
4217 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4218 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4219 bp, vp->v_object, bp->b_bufobj->bo_object));
4221 bp->b_flags &= ~B_VMIO;
4222 KASSERT(bp->b_bufobj->bo_object == NULL,
4223 ("ARGH! has b_bufobj->bo_object %p %p\n",
4224 bp, bp->b_bufobj->bo_object));
4225 BUF_CHECK_MAPPED(bp);
4229 bufspace_release(bufdomain(bp), maxsize);
4230 bp->b_flags &= ~B_DONE;
4232 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4234 buf_track(bp, __func__);
4235 KASSERT(bp->b_bufobj == bo,
4236 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4242 * Get an empty, disassociated buffer of given size. The buffer is initially
4246 geteblk(int size, int flags)
4251 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4252 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4253 if ((flags & GB_NOWAIT_BD) &&
4254 (curthread->td_pflags & TDP_BUFNEED) != 0)
4258 bufspace_release(bufdomain(bp), maxsize);
4259 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4264 * Truncate the backing store for a non-vmio buffer.
4267 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4270 if (bp->b_flags & B_MALLOC) {
4272 * malloced buffers are not shrunk
4274 if (newbsize == 0) {
4275 bufmallocadjust(bp, 0);
4276 free(bp->b_data, M_BIOBUF);
4277 bp->b_data = bp->b_kvabase;
4278 bp->b_flags &= ~B_MALLOC;
4282 vm_hold_free_pages(bp, newbsize);
4283 bufspace_adjust(bp, newbsize);
4287 * Extend the backing for a non-VMIO buffer.
4290 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4296 * We only use malloced memory on the first allocation.
4297 * and revert to page-allocated memory when the buffer
4300 * There is a potential smp race here that could lead
4301 * to bufmallocspace slightly passing the max. It
4302 * is probably extremely rare and not worth worrying
4305 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4306 bufmallocspace < maxbufmallocspace) {
4307 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4308 bp->b_flags |= B_MALLOC;
4309 bufmallocadjust(bp, newbsize);
4314 * If the buffer is growing on its other-than-first
4315 * allocation then we revert to the page-allocation
4320 if (bp->b_flags & B_MALLOC) {
4321 origbuf = bp->b_data;
4322 origbufsize = bp->b_bufsize;
4323 bp->b_data = bp->b_kvabase;
4324 bufmallocadjust(bp, 0);
4325 bp->b_flags &= ~B_MALLOC;
4326 newbsize = round_page(newbsize);
4328 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4329 (vm_offset_t) bp->b_data + newbsize);
4330 if (origbuf != NULL) {
4331 bcopy(origbuf, bp->b_data, origbufsize);
4332 free(origbuf, M_BIOBUF);
4334 bufspace_adjust(bp, newbsize);
4338 * This code constitutes the buffer memory from either anonymous system
4339 * memory (in the case of non-VMIO operations) or from an associated
4340 * VM object (in the case of VMIO operations). This code is able to
4341 * resize a buffer up or down.
4343 * Note that this code is tricky, and has many complications to resolve
4344 * deadlock or inconsistent data situations. Tread lightly!!!
4345 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4346 * the caller. Calling this code willy nilly can result in the loss of data.
4348 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4349 * B_CACHE for the non-VMIO case.
4352 allocbuf(struct buf *bp, int size)
4356 if (bp->b_bcount == size)
4359 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4360 panic("allocbuf: buffer too small");
4362 newbsize = roundup2(size, DEV_BSIZE);
4363 if ((bp->b_flags & B_VMIO) == 0) {
4364 if ((bp->b_flags & B_MALLOC) == 0)
4365 newbsize = round_page(newbsize);
4367 * Just get anonymous memory from the kernel. Don't
4368 * mess with B_CACHE.
4370 if (newbsize < bp->b_bufsize)
4371 vfs_nonvmio_truncate(bp, newbsize);
4372 else if (newbsize > bp->b_bufsize)
4373 vfs_nonvmio_extend(bp, newbsize);
4377 desiredpages = (size == 0) ? 0 :
4378 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4380 if (bp->b_flags & B_MALLOC)
4381 panic("allocbuf: VMIO buffer can't be malloced");
4383 * Set B_CACHE initially if buffer is 0 length or will become
4386 if (size == 0 || bp->b_bufsize == 0)
4387 bp->b_flags |= B_CACHE;
4389 if (newbsize < bp->b_bufsize)
4390 vfs_vmio_truncate(bp, desiredpages);
4391 /* XXX This looks as if it should be newbsize > b_bufsize */
4392 else if (size > bp->b_bcount)
4393 vfs_vmio_extend(bp, desiredpages, size);
4394 bufspace_adjust(bp, newbsize);
4396 bp->b_bcount = size; /* requested buffer size. */
4400 extern int inflight_transient_maps;
4402 static struct bio_queue nondump_bios;
4405 biodone(struct bio *bp)
4408 void (*done)(struct bio *);
4409 vm_offset_t start, end;
4411 biotrack(bp, __func__);
4414 * Avoid completing I/O when dumping after a panic since that may
4415 * result in a deadlock in the filesystem or pager code. Note that
4416 * this doesn't affect dumps that were started manually since we aim
4417 * to keep the system usable after it has been resumed.
4419 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4420 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4423 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4424 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4425 bp->bio_flags |= BIO_UNMAPPED;
4426 start = trunc_page((vm_offset_t)bp->bio_data);
4427 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4428 bp->bio_data = unmapped_buf;
4429 pmap_qremove(start, atop(end - start));
4430 vmem_free(transient_arena, start, end - start);
4431 atomic_add_int(&inflight_transient_maps, -1);
4433 done = bp->bio_done;
4435 * The check for done == biodone is to allow biodone to be
4436 * used as a bio_done routine.
4438 if (done == NULL || done == biodone) {
4439 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4441 bp->bio_flags |= BIO_DONE;
4449 * Wait for a BIO to finish.
4452 biowait(struct bio *bp, const char *wmesg)
4456 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4458 while ((bp->bio_flags & BIO_DONE) == 0)
4459 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4461 if (bp->bio_error != 0)
4462 return (bp->bio_error);
4463 if (!(bp->bio_flags & BIO_ERROR))
4469 biofinish(struct bio *bp, struct devstat *stat, int error)
4473 bp->bio_error = error;
4474 bp->bio_flags |= BIO_ERROR;
4477 devstat_end_transaction_bio(stat, bp);
4481 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4483 biotrack_buf(struct bio *bp, const char *location)
4486 buf_track(bp->bio_track_bp, location);
4493 * Wait for buffer I/O completion, returning error status. The buffer
4494 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4495 * error and cleared.
4498 bufwait(struct buf *bp)
4500 if (bp->b_iocmd == BIO_READ)
4501 bwait(bp, PRIBIO, "biord");
4503 bwait(bp, PRIBIO, "biowr");
4504 if (bp->b_flags & B_EINTR) {
4505 bp->b_flags &= ~B_EINTR;
4508 if (bp->b_ioflags & BIO_ERROR) {
4509 return (bp->b_error ? bp->b_error : EIO);
4518 * Finish I/O on a buffer, optionally calling a completion function.
4519 * This is usually called from an interrupt so process blocking is
4522 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4523 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4524 * assuming B_INVAL is clear.
4526 * For the VMIO case, we set B_CACHE if the op was a read and no
4527 * read error occurred, or if the op was a write. B_CACHE is never
4528 * set if the buffer is invalid or otherwise uncacheable.
4530 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4531 * initiator to leave B_INVAL set to brelse the buffer out of existence
4532 * in the biodone routine.
4535 bufdone(struct buf *bp)
4537 struct bufobj *dropobj;
4538 void (*biodone)(struct buf *);
4540 buf_track(bp, __func__);
4541 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4544 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4546 runningbufwakeup(bp);
4547 if (bp->b_iocmd == BIO_WRITE)
4548 dropobj = bp->b_bufobj;
4549 /* call optional completion function if requested */
4550 if (bp->b_iodone != NULL) {
4551 biodone = bp->b_iodone;
4552 bp->b_iodone = NULL;
4555 bufobj_wdrop(dropobj);
4558 if (bp->b_flags & B_VMIO) {
4560 * Set B_CACHE if the op was a normal read and no error
4561 * occurred. B_CACHE is set for writes in the b*write()
4564 if (bp->b_iocmd == BIO_READ &&
4565 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4566 !(bp->b_ioflags & BIO_ERROR))
4567 bp->b_flags |= B_CACHE;
4568 vfs_vmio_iodone(bp);
4570 if (!LIST_EMPTY(&bp->b_dep))
4572 if ((bp->b_flags & B_CKHASH) != 0) {
4573 KASSERT(bp->b_iocmd == BIO_READ,
4574 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4575 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4576 (*bp->b_ckhashcalc)(bp);
4579 * For asynchronous completions, release the buffer now. The brelse
4580 * will do a wakeup there if necessary - so no need to do a wakeup
4581 * here in the async case. The sync case always needs to do a wakeup.
4583 if (bp->b_flags & B_ASYNC) {
4584 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4585 (bp->b_ioflags & BIO_ERROR))
4592 bufobj_wdrop(dropobj);
4596 * This routine is called in lieu of iodone in the case of
4597 * incomplete I/O. This keeps the busy status for pages
4601 vfs_unbusy_pages(struct buf *bp)
4607 runningbufwakeup(bp);
4608 if (!(bp->b_flags & B_VMIO))
4611 obj = bp->b_bufobj->bo_object;
4612 for (i = 0; i < bp->b_npages; i++) {
4614 if (m == bogus_page) {
4615 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4617 panic("vfs_unbusy_pages: page missing\n");
4619 if (buf_mapped(bp)) {
4620 BUF_CHECK_MAPPED(bp);
4621 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4622 bp->b_pages, bp->b_npages);
4624 BUF_CHECK_UNMAPPED(bp);
4628 vm_object_pip_wakeupn(obj, bp->b_npages);
4632 * vfs_page_set_valid:
4634 * Set the valid bits in a page based on the supplied offset. The
4635 * range is restricted to the buffer's size.
4637 * This routine is typically called after a read completes.
4640 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4645 * Compute the end offset, eoff, such that [off, eoff) does not span a
4646 * page boundary and eoff is not greater than the end of the buffer.
4647 * The end of the buffer, in this case, is our file EOF, not the
4648 * allocation size of the buffer.
4650 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4651 if (eoff > bp->b_offset + bp->b_bcount)
4652 eoff = bp->b_offset + bp->b_bcount;
4655 * Set valid range. This is typically the entire buffer and thus the
4659 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4663 * vfs_page_set_validclean:
4665 * Set the valid bits and clear the dirty bits in a page based on the
4666 * supplied offset. The range is restricted to the buffer's size.
4669 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4671 vm_ooffset_t soff, eoff;
4674 * Start and end offsets in buffer. eoff - soff may not cross a
4675 * page boundary or cross the end of the buffer. The end of the
4676 * buffer, in this case, is our file EOF, not the allocation size
4680 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4681 if (eoff > bp->b_offset + bp->b_bcount)
4682 eoff = bp->b_offset + bp->b_bcount;
4685 * Set valid range. This is typically the entire buffer and thus the
4689 vm_page_set_validclean(
4691 (vm_offset_t) (soff & PAGE_MASK),
4692 (vm_offset_t) (eoff - soff)
4698 * Acquire a shared busy on all pages in the buf.
4701 vfs_busy_pages_acquire(struct buf *bp)
4705 for (i = 0; i < bp->b_npages; i++)
4706 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4710 vfs_busy_pages_release(struct buf *bp)
4714 for (i = 0; i < bp->b_npages; i++)
4715 vm_page_sunbusy(bp->b_pages[i]);
4719 * This routine is called before a device strategy routine.
4720 * It is used to tell the VM system that paging I/O is in
4721 * progress, and treat the pages associated with the buffer
4722 * almost as being exclusive busy. Also the object paging_in_progress
4723 * flag is handled to make sure that the object doesn't become
4726 * Since I/O has not been initiated yet, certain buffer flags
4727 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4728 * and should be ignored.
4731 vfs_busy_pages(struct buf *bp, int clear_modify)
4739 if (!(bp->b_flags & B_VMIO))
4742 obj = bp->b_bufobj->bo_object;
4743 foff = bp->b_offset;
4744 KASSERT(bp->b_offset != NOOFFSET,
4745 ("vfs_busy_pages: no buffer offset"));
4746 if ((bp->b_flags & B_CLUSTER) == 0) {
4747 vm_object_pip_add(obj, bp->b_npages);
4748 vfs_busy_pages_acquire(bp);
4750 if (bp->b_bufsize != 0)
4751 vfs_setdirty_range(bp);
4753 for (i = 0; i < bp->b_npages; i++) {
4755 vm_page_assert_sbusied(m);
4758 * When readying a buffer for a read ( i.e
4759 * clear_modify == 0 ), it is important to do
4760 * bogus_page replacement for valid pages in
4761 * partially instantiated buffers. Partially
4762 * instantiated buffers can, in turn, occur when
4763 * reconstituting a buffer from its VM backing store
4764 * base. We only have to do this if B_CACHE is
4765 * clear ( which causes the I/O to occur in the
4766 * first place ). The replacement prevents the read
4767 * I/O from overwriting potentially dirty VM-backed
4768 * pages. XXX bogus page replacement is, uh, bogus.
4769 * It may not work properly with small-block devices.
4770 * We need to find a better way.
4773 pmap_remove_write(m);
4774 vfs_page_set_validclean(bp, foff, m);
4775 } else if (vm_page_all_valid(m) &&
4776 (bp->b_flags & B_CACHE) == 0) {
4777 bp->b_pages[i] = bogus_page;
4780 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4782 if (bogus && buf_mapped(bp)) {
4783 BUF_CHECK_MAPPED(bp);
4784 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4785 bp->b_pages, bp->b_npages);
4790 * vfs_bio_set_valid:
4792 * Set the range within the buffer to valid. The range is
4793 * relative to the beginning of the buffer, b_offset. Note that
4794 * b_offset itself may be offset from the beginning of the first
4798 vfs_bio_set_valid(struct buf *bp, int base, int size)
4803 if (!(bp->b_flags & B_VMIO))
4807 * Fixup base to be relative to beginning of first page.
4808 * Set initial n to be the maximum number of bytes in the
4809 * first page that can be validated.
4811 base += (bp->b_offset & PAGE_MASK);
4812 n = PAGE_SIZE - (base & PAGE_MASK);
4815 * Busy may not be strictly necessary here because the pages are
4816 * unlikely to be fully valid and the vnode lock will synchronize
4817 * their access via getpages. It is grabbed for consistency with
4818 * other page validation.
4820 vfs_busy_pages_acquire(bp);
4821 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4825 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4830 vfs_busy_pages_release(bp);
4836 * If the specified buffer is a non-VMIO buffer, clear the entire
4837 * buffer. If the specified buffer is a VMIO buffer, clear and
4838 * validate only the previously invalid portions of the buffer.
4839 * This routine essentially fakes an I/O, so we need to clear
4840 * BIO_ERROR and B_INVAL.
4842 * Note that while we only theoretically need to clear through b_bcount,
4843 * we go ahead and clear through b_bufsize.
4846 vfs_bio_clrbuf(struct buf *bp)
4848 int i, j, mask, sa, ea, slide;
4850 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4854 bp->b_flags &= ~B_INVAL;
4855 bp->b_ioflags &= ~BIO_ERROR;
4856 vfs_busy_pages_acquire(bp);
4857 sa = bp->b_offset & PAGE_MASK;
4859 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4860 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4861 ea = slide & PAGE_MASK;
4864 if (bp->b_pages[i] == bogus_page)
4867 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4868 if ((bp->b_pages[i]->valid & mask) == mask)
4870 if ((bp->b_pages[i]->valid & mask) == 0)
4871 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4873 for (; sa < ea; sa += DEV_BSIZE, j++) {
4874 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4875 pmap_zero_page_area(bp->b_pages[i],
4880 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4881 roundup2(ea - sa, DEV_BSIZE));
4883 vfs_busy_pages_release(bp);
4888 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4893 if (buf_mapped(bp)) {
4894 BUF_CHECK_MAPPED(bp);
4895 bzero(bp->b_data + base, size);
4897 BUF_CHECK_UNMAPPED(bp);
4898 n = PAGE_SIZE - (base & PAGE_MASK);
4899 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4903 pmap_zero_page_area(m, base & PAGE_MASK, n);
4912 * Update buffer flags based on I/O request parameters, optionally releasing the
4913 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4914 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4915 * I/O). Otherwise the buffer is released to the cache.
4918 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4921 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4922 ("buf %p non-VMIO noreuse", bp));
4924 if ((ioflag & IO_DIRECT) != 0)
4925 bp->b_flags |= B_DIRECT;
4926 if ((ioflag & IO_EXT) != 0)
4927 bp->b_xflags |= BX_ALTDATA;
4928 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4929 bp->b_flags |= B_RELBUF;
4930 if ((ioflag & IO_NOREUSE) != 0)
4931 bp->b_flags |= B_NOREUSE;
4939 vfs_bio_brelse(struct buf *bp, int ioflag)
4942 b_io_dismiss(bp, ioflag, true);
4946 vfs_bio_set_flags(struct buf *bp, int ioflag)
4949 b_io_dismiss(bp, ioflag, false);
4953 * vm_hold_load_pages and vm_hold_free_pages get pages into
4954 * a buffers address space. The pages are anonymous and are
4955 * not associated with a file object.
4958 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4964 BUF_CHECK_MAPPED(bp);
4966 to = round_page(to);
4967 from = round_page(from);
4968 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4969 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4970 KASSERT(to - from <= maxbcachebuf,
4971 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4972 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4974 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4976 * note: must allocate system pages since blocking here
4977 * could interfere with paging I/O, no matter which
4980 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4981 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4982 pmap_qenter(pg, &p, 1);
4983 bp->b_pages[index] = p;
4985 bp->b_npages = index;
4988 /* Return pages associated with this buf to the vm system */
4990 vm_hold_free_pages(struct buf *bp, int newbsize)
4994 int index, newnpages;
4996 BUF_CHECK_MAPPED(bp);
4998 from = round_page((vm_offset_t)bp->b_data + newbsize);
4999 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
5000 if (bp->b_npages > newnpages)
5001 pmap_qremove(from, bp->b_npages - newnpages);
5002 for (index = newnpages; index < bp->b_npages; index++) {
5003 p = bp->b_pages[index];
5004 bp->b_pages[index] = NULL;
5005 vm_page_unwire_noq(p);
5008 bp->b_npages = newnpages;
5012 * Map an IO request into kernel virtual address space.
5014 * All requests are (re)mapped into kernel VA space.
5015 * Notice that we use b_bufsize for the size of the buffer
5016 * to be mapped. b_bcount might be modified by the driver.
5018 * Note that even if the caller determines that the address space should
5019 * be valid, a race or a smaller-file mapped into a larger space may
5020 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5021 * check the return value.
5023 * This function only works with pager buffers.
5026 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5031 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5032 prot = VM_PROT_READ;
5033 if (bp->b_iocmd == BIO_READ)
5034 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5035 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5036 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5039 bp->b_bufsize = len;
5040 bp->b_npages = pidx;
5041 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5042 if (mapbuf || !unmapped_buf_allowed) {
5043 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5044 bp->b_data = bp->b_kvabase + bp->b_offset;
5046 bp->b_data = unmapped_buf;
5051 * Free the io map PTEs associated with this IO operation.
5052 * We also invalidate the TLB entries and restore the original b_addr.
5054 * This function only works with pager buffers.
5057 vunmapbuf(struct buf *bp)
5061 npages = bp->b_npages;
5063 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5064 vm_page_unhold_pages(bp->b_pages, npages);
5066 bp->b_data = unmapped_buf;
5070 bdone(struct buf *bp)
5074 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5076 bp->b_flags |= B_DONE;
5082 bwait(struct buf *bp, u_char pri, const char *wchan)
5086 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5088 while ((bp->b_flags & B_DONE) == 0)
5089 msleep(bp, mtxp, pri, wchan, 0);
5094 bufsync(struct bufobj *bo, int waitfor)
5097 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5101 bufstrategy(struct bufobj *bo, struct buf *bp)
5107 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5108 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5109 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5110 i = VOP_STRATEGY(vp, bp);
5111 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5115 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5118 bufobj_init(struct bufobj *bo, void *private)
5120 static volatile int bufobj_cleanq;
5123 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5124 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5125 bo->bo_private = private;
5126 TAILQ_INIT(&bo->bo_clean.bv_hd);
5127 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5131 bufobj_wrefl(struct bufobj *bo)
5134 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5135 ASSERT_BO_WLOCKED(bo);
5140 bufobj_wref(struct bufobj *bo)
5143 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5150 bufobj_wdrop(struct bufobj *bo)
5153 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5155 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5156 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5157 bo->bo_flag &= ~BO_WWAIT;
5158 wakeup(&bo->bo_numoutput);
5164 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5168 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5169 ASSERT_BO_WLOCKED(bo);
5171 while (bo->bo_numoutput) {
5172 bo->bo_flag |= BO_WWAIT;
5173 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5174 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5182 * Set bio_data or bio_ma for struct bio from the struct buf.
5185 bdata2bio(struct buf *bp, struct bio *bip)
5188 if (!buf_mapped(bp)) {
5189 KASSERT(unmapped_buf_allowed, ("unmapped"));
5190 bip->bio_ma = bp->b_pages;
5191 bip->bio_ma_n = bp->b_npages;
5192 bip->bio_data = unmapped_buf;
5193 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5194 bip->bio_flags |= BIO_UNMAPPED;
5195 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5196 PAGE_SIZE == bp->b_npages,
5197 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5198 (long long)bip->bio_length, bip->bio_ma_n));
5200 bip->bio_data = bp->b_data;
5206 * The MIPS pmap code currently doesn't handle aliased pages.
5207 * The VIPT caches may not handle page aliasing themselves, leading
5208 * to data corruption.
5210 * As such, this code makes a system extremely unhappy if said
5211 * system doesn't support unaliasing the above situation in hardware.
5212 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5213 * this feature at build time, so it has to be handled in software.
5215 * Once the MIPS pmap/cache code grows to support this function on
5216 * earlier chips, it should be flipped back off.
5219 static int buf_pager_relbuf = 1;
5221 static int buf_pager_relbuf = 0;
5223 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5224 &buf_pager_relbuf, 0,
5225 "Make buffer pager release buffers after reading");
5228 * The buffer pager. It uses buffer reads to validate pages.
5230 * In contrast to the generic local pager from vm/vnode_pager.c, this
5231 * pager correctly and easily handles volumes where the underlying
5232 * device block size is greater than the machine page size. The
5233 * buffer cache transparently extends the requested page run to be
5234 * aligned at the block boundary, and does the necessary bogus page
5235 * replacements in the addends to avoid obliterating already valid
5238 * The only non-trivial issue is that the exclusive busy state for
5239 * pages, which is assumed by the vm_pager_getpages() interface, is
5240 * incompatible with the VMIO buffer cache's desire to share-busy the
5241 * pages. This function performs a trivial downgrade of the pages'
5242 * state before reading buffers, and a less trivial upgrade from the
5243 * shared-busy to excl-busy state after the read.
5246 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5247 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5248 vbg_get_blksize_t get_blksize)
5255 vm_ooffset_t la, lb, poff, poffe;
5257 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5260 object = vp->v_object;
5263 la = IDX_TO_OFF(ma[count - 1]->pindex);
5264 if (la >= object->un_pager.vnp.vnp_size)
5265 return (VM_PAGER_BAD);
5268 * Change the meaning of la from where the last requested page starts
5269 * to where it ends, because that's the end of the requested region
5270 * and the start of the potential read-ahead region.
5273 lpart = la > object->un_pager.vnp.vnp_size;
5274 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5277 return (VM_PAGER_ERROR);
5280 * Calculate read-ahead, behind and total pages.
5283 lb = IDX_TO_OFF(ma[0]->pindex);
5284 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5286 if (rbehind != NULL)
5288 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5289 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5290 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5295 VM_CNT_INC(v_vnodein);
5296 VM_CNT_ADD(v_vnodepgsin, pgsin);
5298 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5299 != 0) ? GB_UNMAPPED : 0;
5301 for (i = 0; i < count; i++) {
5302 if (ma[i] != bogus_page)
5303 vm_page_busy_downgrade(ma[i]);
5307 for (i = 0; i < count; i++) {
5309 if (m == bogus_page)
5313 * Pages are shared busy and the object lock is not
5314 * owned, which together allow for the pages'
5315 * invalidation. The racy test for validity avoids
5316 * useless creation of the buffer for the most typical
5317 * case when invalidation is not used in redo or for
5318 * parallel read. The shared->excl upgrade loop at
5319 * the end of the function catches the race in a
5320 * reliable way (protected by the object lock).
5322 if (vm_page_all_valid(m))
5325 poff = IDX_TO_OFF(m->pindex);
5326 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5327 for (; poff < poffe; poff += bsize) {
5328 lbn = get_lblkno(vp, poff);
5333 error = get_blksize(vp, lbn, &bsize);
5335 error = bread_gb(vp, lbn, bsize,
5336 curthread->td_ucred, br_flags, &bp);
5339 if (bp->b_rcred == curthread->td_ucred) {
5340 crfree(bp->b_rcred);
5341 bp->b_rcred = NOCRED;
5343 if (LIST_EMPTY(&bp->b_dep)) {
5345 * Invalidation clears m->valid, but
5346 * may leave B_CACHE flag if the
5347 * buffer existed at the invalidation
5348 * time. In this case, recycle the
5349 * buffer to do real read on next
5350 * bread() after redo.
5352 * Otherwise B_RELBUF is not strictly
5353 * necessary, enable to reduce buf
5356 if (buf_pager_relbuf ||
5357 !vm_page_all_valid(m))
5358 bp->b_flags |= B_RELBUF;
5360 bp->b_flags &= ~B_NOCACHE;
5366 KASSERT(1 /* racy, enable for debugging */ ||
5367 vm_page_all_valid(m) || i == count - 1,
5368 ("buf %d %p invalid", i, m));
5369 if (i == count - 1 && lpart) {
5370 if (!vm_page_none_valid(m) &&
5371 !vm_page_all_valid(m))
5372 vm_page_zero_invalid(m, TRUE);
5379 for (i = 0; i < count; i++) {
5380 if (ma[i] == bogus_page)
5382 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5383 vm_page_sunbusy(ma[i]);
5384 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5389 * Since the pages were only sbusy while neither the
5390 * buffer nor the object lock was held by us, or
5391 * reallocated while vm_page_grab() slept for busy
5392 * relinguish, they could have been invalidated.
5393 * Recheck the valid bits and re-read as needed.
5395 * Note that the last page is made fully valid in the
5396 * read loop, and partial validity for the page at
5397 * index count - 1 could mean that the page was
5398 * invalidated or removed, so we must restart for
5401 if (!vm_page_all_valid(ma[i]))
5404 if (redo && error == 0)
5406 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5409 #include "opt_ddb.h"
5411 #include <ddb/ddb.h>
5413 /* DDB command to show buffer data */
5414 DB_SHOW_COMMAND(buffer, db_show_buffer)
5417 struct buf *bp = (struct buf *)addr;
5418 #ifdef FULL_BUF_TRACKING
5423 db_printf("usage: show buffer <addr>\n");
5427 db_printf("buf at %p\n", bp);
5428 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5429 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5430 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5431 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5432 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5433 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5435 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5436 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5437 "b_vp = %p, b_dep = %p\n",
5438 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5439 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5440 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5441 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5442 bp->b_kvabase, bp->b_kvasize);
5445 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5446 for (i = 0; i < bp->b_npages; i++) {
5450 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5452 (u_long)VM_PAGE_TO_PHYS(m));
5454 db_printf("( ??? )");
5455 if ((i + 1) < bp->b_npages)
5460 BUF_LOCKPRINTINFO(bp);
5461 #if defined(FULL_BUF_TRACKING)
5462 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5464 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5465 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5466 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5468 db_printf(" %2u: %s\n", j,
5469 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5471 #elif defined(BUF_TRACKING)
5472 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5477 DB_SHOW_COMMAND(bufqueues, bufqueues)
5479 struct bufdomain *bd;
5484 db_printf("bqempty: %d\n", bqempty.bq_len);
5486 for (i = 0; i < buf_domains; i++) {
5488 db_printf("Buf domain %d\n", i);
5489 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5490 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5491 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5493 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5494 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5495 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5496 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5497 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5499 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5500 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5501 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5502 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5505 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5506 total += bp->b_bufsize;
5507 db_printf("\tcleanq count\t%d (%ld)\n",
5508 bd->bd_cleanq->bq_len, total);
5510 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5511 total += bp->b_bufsize;
5512 db_printf("\tdirtyq count\t%d (%ld)\n",
5513 bd->bd_dirtyq.bq_len, total);
5514 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5515 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5516 db_printf("\tCPU ");
5517 for (j = 0; j <= mp_maxid; j++)
5518 db_printf("%d, ", bd->bd_subq[j].bq_len);
5522 for (j = 0; j < nbuf; j++) {
5524 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5526 total += bp->b_bufsize;
5529 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5532 for (j = 0; j < nbuf; j++) {
5534 if (bp->b_domain == i) {
5536 total += bp->b_bufsize;
5539 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5543 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5548 for (i = 0; i < nbuf; i++) {
5550 if (BUF_ISLOCKED(bp)) {
5551 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5559 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5565 db_printf("usage: show vnodebufs <addr>\n");
5568 vp = (struct vnode *)addr;
5569 db_printf("Clean buffers:\n");
5570 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5571 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5574 db_printf("Dirty buffers:\n");
5575 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5576 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5581 DB_COMMAND(countfreebufs, db_coundfreebufs)
5584 int i, used = 0, nfree = 0;
5587 db_printf("usage: countfreebufs\n");
5591 for (i = 0; i < nbuf; i++) {
5593 if (bp->b_qindex == QUEUE_EMPTY)
5599 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5601 db_printf("numfreebuffers is %d\n", numfreebuffers);