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;
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 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
409 value = *(long *)arg1;
410 error = sysctl_handle_long(oidp, &value, 0, req);
411 if (error != 0 || req->newptr == NULL)
413 mtx_lock(&rbreqlock);
414 if (arg1 == &hirunningspace) {
415 if (value < lorunningspace)
418 hirunningspace = value;
420 KASSERT(arg1 == &lorunningspace,
421 ("%s: unknown arg1", __func__));
422 if (value > hirunningspace)
425 lorunningspace = value;
427 mtx_unlock(&rbreqlock);
432 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS)
438 value = *(int *)arg1;
439 error = sysctl_handle_int(oidp, &value, 0, req);
440 if (error != 0 || req->newptr == NULL)
442 *(int *)arg1 = value;
443 for (i = 0; i < buf_domains; i++)
444 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
451 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS)
457 value = *(long *)arg1;
458 error = sysctl_handle_long(oidp, &value, 0, req);
459 if (error != 0 || req->newptr == NULL)
461 *(long *)arg1 = value;
462 for (i = 0; i < buf_domains; i++)
463 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) =
469 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
470 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
472 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
479 for (i = 0; i < buf_domains; i++)
480 lvalue += bdomain[i].bd_bufspace;
481 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
482 return (sysctl_handle_long(oidp, &lvalue, 0, req));
483 if (lvalue > INT_MAX)
484 /* On overflow, still write out a long to trigger ENOMEM. */
485 return (sysctl_handle_long(oidp, &lvalue, 0, req));
487 return (sysctl_handle_int(oidp, &ivalue, 0, req));
491 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
497 for (i = 0; i < buf_domains; i++)
498 lvalue += bdomain[i].bd_bufspace;
499 return (sysctl_handle_long(oidp, &lvalue, 0, req));
504 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS)
510 for (i = 0; i < buf_domains; i++)
511 value += bdomain[i].bd_numdirtybuffers;
512 return (sysctl_handle_int(oidp, &value, 0, req));
518 * Wakeup any bwillwrite() waiters.
523 mtx_lock(&bdirtylock);
528 mtx_unlock(&bdirtylock);
534 * Clear a domain from the appropriate bitsets when dirtybuffers
538 bd_clear(struct bufdomain *bd)
541 mtx_lock(&bdirtylock);
542 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers)
543 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
544 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers)
545 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
546 mtx_unlock(&bdirtylock);
552 * Set a domain in the appropriate bitsets when dirtybuffers
556 bd_set(struct bufdomain *bd)
559 mtx_lock(&bdirtylock);
560 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers)
561 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty);
562 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers)
563 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty);
564 mtx_unlock(&bdirtylock);
570 * Decrement the numdirtybuffers count by one and wakeup any
571 * threads blocked in bwillwrite().
574 bdirtysub(struct buf *bp)
576 struct bufdomain *bd;
580 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1);
581 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
583 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
590 * Increment the numdirtybuffers count by one and wakeup the buf
594 bdirtyadd(struct buf *bp)
596 struct bufdomain *bd;
600 * Only do the wakeup once as we cross the boundary. The
601 * buf daemon will keep running until the condition clears.
604 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1);
605 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2)
607 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers)
612 * bufspace_daemon_wakeup:
614 * Wakeup the daemons responsible for freeing clean bufs.
617 bufspace_daemon_wakeup(struct bufdomain *bd)
621 * avoid the lock if the daemon is running.
623 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) {
625 atomic_store_int(&bd->bd_running, 1);
626 wakeup(&bd->bd_running);
634 * Adjust the reported bufspace for a KVA managed buffer, possibly
635 * waking any waiters.
638 bufspace_adjust(struct buf *bp, int bufsize)
640 struct bufdomain *bd;
644 KASSERT((bp->b_flags & B_MALLOC) == 0,
645 ("bufspace_adjust: malloc buf %p", bp));
647 diff = bufsize - bp->b_bufsize;
649 atomic_subtract_long(&bd->bd_bufspace, -diff);
650 } else if (diff > 0) {
651 space = atomic_fetchadd_long(&bd->bd_bufspace, diff);
652 /* Wake up the daemon on the transition. */
653 if (space < bd->bd_bufspacethresh &&
654 space + diff >= bd->bd_bufspacethresh)
655 bufspace_daemon_wakeup(bd);
657 bp->b_bufsize = bufsize;
663 * Reserve bufspace before calling allocbuf(). metadata has a
664 * different space limit than data.
667 bufspace_reserve(struct bufdomain *bd, int size, bool metadata)
673 limit = bd->bd_maxbufspace;
675 limit = bd->bd_hibufspace;
676 space = atomic_fetchadd_long(&bd->bd_bufspace, size);
679 atomic_subtract_long(&bd->bd_bufspace, size);
683 /* Wake up the daemon on the transition. */
684 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh)
685 bufspace_daemon_wakeup(bd);
693 * Release reserved bufspace after bufspace_adjust() has consumed it.
696 bufspace_release(struct bufdomain *bd, int size)
699 atomic_subtract_long(&bd->bd_bufspace, size);
705 * Wait for bufspace, acting as the buf daemon if a locked vnode is
706 * supplied. bd_wanted must be set prior to polling for space. The
707 * operation must be re-tried on return.
710 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags,
711 int slpflag, int slptimeo)
714 int error, fl, norunbuf;
716 if ((gbflags & GB_NOWAIT_BD) != 0)
721 while (bd->bd_wanted) {
722 if (vp != NULL && vp->v_type != VCHR &&
723 (td->td_pflags & TDP_BUFNEED) == 0) {
726 * getblk() is called with a vnode locked, and
727 * some majority of the dirty buffers may as
728 * well belong to the vnode. Flushing the
729 * buffers there would make a progress that
730 * cannot be achieved by the buf_daemon, that
731 * cannot lock the vnode.
733 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
734 (td->td_pflags & TDP_NORUNNINGBUF);
737 * Play bufdaemon. The getnewbuf() function
738 * may be called while the thread owns lock
739 * for another dirty buffer for the same
740 * vnode, which makes it impossible to use
741 * VOP_FSYNC() there, due to the buffer lock
744 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
745 fl = buf_flush(vp, bd, flushbufqtarget);
746 td->td_pflags &= norunbuf;
750 if (bd->bd_wanted == 0)
753 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
754 (PRIBIO + 4) | slpflag, "newbuf", slptimeo);
762 bufspace_daemon_shutdown(void *arg, int howto __unused)
764 struct bufdomain *bd = arg;
767 if (KERNEL_PANICKED())
771 bd->bd_shutdown = true;
772 wakeup(&bd->bd_running);
773 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
774 "bufspace_shutdown", 60 * hz);
777 printf("bufspacedaemon wait error: %d\n", error);
783 * buffer space management daemon. Tries to maintain some marginal
784 * amount of free buffer space so that requesting processes neither
785 * block nor work to reclaim buffers.
788 bufspace_daemon(void *arg)
790 struct bufdomain *bd = arg;
792 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
793 SHUTDOWN_PRI_LAST + 100);
796 while (!bd->bd_shutdown) {
800 * Free buffers from the clean queue until we meet our
803 * Theory of operation: The buffer cache is most efficient
804 * when some free buffer headers and space are always
805 * available to getnewbuf(). This daemon attempts to prevent
806 * the excessive blocking and synchronization associated
807 * with shortfall. It goes through three phases according
810 * 1) The daemon wakes up voluntarily once per-second
811 * during idle periods when the counters are below
812 * the wakeup thresholds (bufspacethresh, lofreebuffers).
814 * 2) The daemon wakes up as we cross the thresholds
815 * ahead of any potential blocking. This may bounce
816 * slightly according to the rate of consumption and
819 * 3) The daemon and consumers are starved for working
820 * clean buffers. This is the 'bufspace' sleep below
821 * which will inefficiently trade bufs with bqrelse
822 * until we return to condition 2.
824 while (bd->bd_bufspace > bd->bd_lobufspace ||
825 bd->bd_freebuffers < bd->bd_hifreebuffers) {
826 if (buf_recycle(bd, false) != 0) {
830 * Speedup dirty if we've run out of clean
831 * buffers. This is possible in particular
832 * because softdep may held many bufs locked
833 * pending writes to other bufs which are
834 * marked for delayed write, exhausting
835 * clean space until they are written.
840 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
841 PRIBIO|PDROP, "bufspace", hz/10);
849 * Re-check our limits and sleep. bd_running must be
850 * cleared prior to checking the limits to avoid missed
851 * wakeups. The waker will adjust one of bufspace or
852 * freebuffers prior to checking bd_running.
857 atomic_store_int(&bd->bd_running, 0);
858 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
859 bd->bd_freebuffers > bd->bd_lofreebuffers) {
860 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
863 /* Avoid spurious wakeups while running. */
864 atomic_store_int(&bd->bd_running, 1);
867 wakeup(&bd->bd_shutdown);
875 * Adjust the reported bufspace for a malloc managed buffer, possibly
876 * waking any waiters.
879 bufmallocadjust(struct buf *bp, int bufsize)
883 KASSERT((bp->b_flags & B_MALLOC) != 0,
884 ("bufmallocadjust: non-malloc buf %p", bp));
885 diff = bufsize - bp->b_bufsize;
887 atomic_subtract_long(&bufmallocspace, -diff);
889 atomic_add_long(&bufmallocspace, diff);
890 bp->b_bufsize = bufsize;
896 * Wake up processes that are waiting on asynchronous writes to fall
897 * below lorunningspace.
903 mtx_lock(&rbreqlock);
906 wakeup(&runningbufreq);
908 mtx_unlock(&rbreqlock);
914 * Decrement the outstanding write count according.
917 runningbufwakeup(struct buf *bp)
921 bspace = bp->b_runningbufspace;
924 space = atomic_fetchadd_long(&runningbufspace, -bspace);
925 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
927 bp->b_runningbufspace = 0;
929 * Only acquire the lock and wakeup on the transition from exceeding
930 * the threshold to falling below it.
932 if (space < lorunningspace)
934 if (space - bspace > lorunningspace)
940 * waitrunningbufspace()
942 * runningbufspace is a measure of the amount of I/O currently
943 * running. This routine is used in async-write situations to
944 * prevent creating huge backups of pending writes to a device.
945 * Only asynchronous writes are governed by this function.
947 * This does NOT turn an async write into a sync write. It waits
948 * for earlier writes to complete and generally returns before the
949 * caller's write has reached the device.
952 waitrunningbufspace(void)
955 mtx_lock(&rbreqlock);
956 while (runningbufspace > hirunningspace) {
958 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
960 mtx_unlock(&rbreqlock);
964 * vfs_buf_test_cache:
966 * Called when a buffer is extended. This function clears the B_CACHE
967 * bit if the newly extended portion of the buffer does not contain
971 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
972 vm_offset_t size, vm_page_t m)
976 * This function and its results are protected by higher level
977 * synchronization requiring vnode and buf locks to page in and
980 if (bp->b_flags & B_CACHE) {
981 int base = (foff + off) & PAGE_MASK;
982 if (vm_page_is_valid(m, base, size) == 0)
983 bp->b_flags &= ~B_CACHE;
987 /* Wake up the buffer daemon if necessary */
993 if (bd_request == 0) {
1001 * Adjust the maxbcachbuf tunable.
1004 maxbcachebuf_adjust(void)
1009 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1012 while (i * 2 <= maxbcachebuf)
1015 if (maxbcachebuf < MAXBSIZE)
1016 maxbcachebuf = MAXBSIZE;
1017 if (maxbcachebuf > maxphys)
1018 maxbcachebuf = maxphys;
1019 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1020 printf("maxbcachebuf=%d\n", maxbcachebuf);
1024 * bd_speedup - speedup the buffer cache flushing code
1033 if (bd_speedupreq == 0 || bd_request == 0)
1038 wakeup(&bd_request);
1039 mtx_unlock(&bdlock);
1043 #define TRANSIENT_DENOM 5
1045 #define TRANSIENT_DENOM 10
1049 * Calculating buffer cache scaling values and reserve space for buffer
1050 * headers. This is called during low level kernel initialization and
1051 * may be called more then once. We CANNOT write to the memory area
1052 * being reserved at this time.
1055 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1058 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1062 * With KASAN enabled, the kernel map is shadowed. Account for this
1063 * when sizing maps based on the amount of physical memory available.
1065 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1066 (KASAN_SHADOW_SCALE + 1);
1070 * physmem_est is in pages. Convert it to kilobytes (assumes
1071 * PAGE_SIZE is >= 1K)
1073 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1075 maxbcachebuf_adjust();
1077 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1078 * For the first 64MB of ram nominally allocate sufficient buffers to
1079 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1080 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1081 * the buffer cache we limit the eventual kva reservation to
1084 * factor represents the 1/4 x ram conversion.
1087 int factor = 4 * BKVASIZE / 1024;
1090 if (physmem_est > 4096)
1091 nbuf += min((physmem_est - 4096) / factor,
1093 if (physmem_est > 65536)
1094 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1095 32 * 1024 * 1024 / (factor * 5));
1097 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1098 nbuf = maxbcache / BKVASIZE;
1103 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1104 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1105 if (nbuf > maxbuf) {
1107 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1113 * Ideal allocation size for the transient bio submap is 10%
1114 * of the maximal space buffer map. This roughly corresponds
1115 * to the amount of the buffer mapped for typical UFS load.
1117 * Clip the buffer map to reserve space for the transient
1118 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1119 * maximum buffer map extent on the platform.
1121 * The fall-back to the maxbuf in case of maxbcache unset,
1122 * allows to not trim the buffer KVA for the architectures
1123 * with ample KVA space.
1125 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1126 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1127 buf_sz = (long)nbuf * BKVASIZE;
1128 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1129 (TRANSIENT_DENOM - 1)) {
1131 * There is more KVA than memory. Do not
1132 * adjust buffer map size, and assign the rest
1133 * of maxbuf to transient map.
1135 biotmap_sz = maxbuf_sz - buf_sz;
1138 * Buffer map spans all KVA we could afford on
1139 * this platform. Give 10% (20% on i386) of
1140 * the buffer map to the transient bio map.
1142 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1143 buf_sz -= biotmap_sz;
1145 if (biotmap_sz / INT_MAX > maxphys)
1146 bio_transient_maxcnt = INT_MAX;
1148 bio_transient_maxcnt = biotmap_sz / maxphys;
1150 * Artificially limit to 1024 simultaneous in-flight I/Os
1151 * using the transient mapping.
1153 if (bio_transient_maxcnt > 1024)
1154 bio_transient_maxcnt = 1024;
1156 nbuf = buf_sz / BKVASIZE;
1160 nswbuf = min(nbuf / 4, 256);
1161 if (nswbuf < NSWBUF_MIN)
1162 nswbuf = NSWBUF_MIN;
1166 * Reserve space for the buffer cache buffers
1169 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1170 atop(maxbcachebuf)) * nbuf;
1176 * Single global constant for BUF_WMESG, to avoid getting multiple
1179 static const char buf_wmesg[] = "bufwait";
1181 /* Initialize the buffer subsystem. Called before use of any buffers. */
1188 KASSERT(maxbcachebuf >= MAXBSIZE,
1189 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1191 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1192 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1193 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1194 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1196 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1198 /* finally, initialize each buffer header and stick on empty q */
1199 for (i = 0; i < nbuf; i++) {
1201 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1202 bp->b_flags = B_INVAL;
1203 bp->b_rcred = NOCRED;
1204 bp->b_wcred = NOCRED;
1205 bp->b_qindex = QUEUE_NONE;
1207 bp->b_subqueue = mp_maxid + 1;
1209 bp->b_data = bp->b_kvabase = unmapped_buf;
1210 LIST_INIT(&bp->b_dep);
1211 BUF_LOCKINIT(bp, buf_wmesg);
1212 bq_insert(&bqempty, bp, false);
1216 * maxbufspace is the absolute maximum amount of buffer space we are
1217 * allowed to reserve in KVM and in real terms. The absolute maximum
1218 * is nominally used by metadata. hibufspace is the nominal maximum
1219 * used by most other requests. The differential is required to
1220 * ensure that metadata deadlocks don't occur.
1222 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1223 * this may result in KVM fragmentation which is not handled optimally
1224 * by the system. XXX This is less true with vmem. We could use
1227 maxbufspace = (long)nbuf * BKVASIZE;
1228 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1229 lobufspace = (hibufspace / 20) * 19; /* 95% */
1230 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1233 * Note: The 16 MiB upper limit for hirunningspace was chosen
1234 * arbitrarily and may need further tuning. It corresponds to
1235 * 128 outstanding write IO requests (if IO size is 128 KiB),
1236 * which fits with many RAID controllers' tagged queuing limits.
1237 * The lower 1 MiB limit is the historical upper limit for
1240 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1241 16 * 1024 * 1024), 1024 * 1024);
1242 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1245 * Limit the amount of malloc memory since it is wired permanently into
1246 * the kernel space. Even though this is accounted for in the buffer
1247 * allocation, we don't want the malloced region to grow uncontrolled.
1248 * The malloc scheme improves memory utilization significantly on
1249 * average (small) directories.
1251 maxbufmallocspace = hibufspace / 20;
1254 * Reduce the chance of a deadlock occurring by limiting the number
1255 * of delayed-write dirty buffers we allow to stack up.
1257 hidirtybuffers = nbuf / 4 + 20;
1258 dirtybufthresh = hidirtybuffers * 9 / 10;
1260 * To support extreme low-memory systems, make sure hidirtybuffers
1261 * cannot eat up all available buffer space. This occurs when our
1262 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1263 * buffer space assuming BKVASIZE'd buffers.
1265 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1266 hidirtybuffers >>= 1;
1268 lodirtybuffers = hidirtybuffers / 2;
1271 * lofreebuffers should be sufficient to avoid stalling waiting on
1272 * buf headers under heavy utilization. The bufs in per-cpu caches
1273 * are counted as free but will be unavailable to threads executing
1276 * hifreebuffers is the free target for the bufspace daemon. This
1277 * should be set appropriately to limit work per-iteration.
1279 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1280 hifreebuffers = (3 * lofreebuffers) / 2;
1281 numfreebuffers = nbuf;
1283 /* Setup the kva and free list allocators. */
1284 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1285 buf_zone = uma_zcache_create("buf free cache",
1286 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1287 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1290 * Size the clean queue according to the amount of buffer space.
1291 * One queue per-256mb up to the max. More queues gives better
1292 * concurrency but less accurate LRU.
1294 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1295 for (i = 0 ; i < buf_domains; i++) {
1296 struct bufdomain *bd;
1300 bd->bd_freebuffers = nbuf / buf_domains;
1301 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1302 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1303 bd->bd_bufspace = 0;
1304 bd->bd_maxbufspace = maxbufspace / buf_domains;
1305 bd->bd_hibufspace = hibufspace / buf_domains;
1306 bd->bd_lobufspace = lobufspace / buf_domains;
1307 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1308 bd->bd_numdirtybuffers = 0;
1309 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1310 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1311 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1312 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1313 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1315 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1316 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1317 mappingrestarts = counter_u64_alloc(M_WAITOK);
1318 numbufallocfails = counter_u64_alloc(M_WAITOK);
1319 notbufdflushes = counter_u64_alloc(M_WAITOK);
1320 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1321 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1322 bufkvaspace = counter_u64_alloc(M_WAITOK);
1327 vfs_buf_check_mapped(struct buf *bp)
1330 KASSERT(bp->b_kvabase != unmapped_buf,
1331 ("mapped buf: b_kvabase was not updated %p", bp));
1332 KASSERT(bp->b_data != unmapped_buf,
1333 ("mapped buf: b_data was not updated %p", bp));
1334 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1335 maxphys, ("b_data + b_offset unmapped %p", bp));
1339 vfs_buf_check_unmapped(struct buf *bp)
1342 KASSERT(bp->b_data == unmapped_buf,
1343 ("unmapped buf: corrupted b_data %p", bp));
1346 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1347 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1349 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1350 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1354 isbufbusy(struct buf *bp)
1356 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1357 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1363 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1366 bufshutdown(int show_busybufs)
1368 static int first_buf_printf = 1;
1370 int i, iter, nbusy, pbusy;
1376 * Sync filesystems for shutdown
1378 wdog_kern_pat(WD_LASTVAL);
1379 kern_sync(curthread);
1382 * With soft updates, some buffers that are
1383 * written will be remarked as dirty until other
1384 * buffers are written.
1386 for (iter = pbusy = 0; iter < 20; iter++) {
1388 for (i = nbuf - 1; i >= 0; i--) {
1394 if (first_buf_printf)
1395 printf("All buffers synced.");
1398 if (first_buf_printf) {
1399 printf("Syncing disks, buffers remaining... ");
1400 first_buf_printf = 0;
1402 printf("%d ", nbusy);
1407 wdog_kern_pat(WD_LASTVAL);
1408 kern_sync(curthread);
1412 * Spin for a while to allow interrupt threads to run.
1414 DELAY(50000 * iter);
1417 * Context switch several times to allow interrupt
1420 for (subiter = 0; subiter < 50 * iter; subiter++) {
1421 thread_lock(curthread);
1429 * Count only busy local buffers to prevent forcing
1430 * a fsck if we're just a client of a wedged NFS server
1433 for (i = nbuf - 1; i >= 0; i--) {
1435 if (isbufbusy(bp)) {
1437 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1438 if (bp->b_dev == NULL) {
1439 TAILQ_REMOVE(&mountlist,
1440 bp->b_vp->v_mount, mnt_list);
1445 if (show_busybufs > 0) {
1447 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1448 nbusy, bp, bp->b_vp, bp->b_flags,
1449 (intmax_t)bp->b_blkno,
1450 (intmax_t)bp->b_lblkno);
1451 BUF_LOCKPRINTINFO(bp);
1452 if (show_busybufs > 1)
1460 * Failed to sync all blocks. Indicate this and don't
1461 * unmount filesystems (thus forcing an fsck on reboot).
1463 printf("Giving up on %d buffers\n", nbusy);
1464 DELAY(5000000); /* 5 seconds */
1467 if (!first_buf_printf)
1468 printf("Final sync complete\n");
1471 * Unmount filesystems and perform swapoff, to quiesce
1472 * the system as much as possible. In particular, no
1473 * I/O should be initiated from top levels since it
1474 * might be abruptly terminated by reset, or otherwise
1475 * erronously handled because other parts of the
1476 * system are disabled.
1478 * Swapoff before unmount, because file-backed swap is
1479 * non-operational after unmount of the underlying
1482 if (!KERNEL_PANICKED()) {
1487 DELAY(100000); /* wait for console output to finish */
1491 bpmap_qenter(struct buf *bp)
1494 BUF_CHECK_MAPPED(bp);
1497 * bp->b_data is relative to bp->b_offset, but
1498 * bp->b_offset may be offset into the first page.
1500 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1501 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1502 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1503 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1506 static inline struct bufdomain *
1507 bufdomain(struct buf *bp)
1510 return (&bdomain[bp->b_domain]);
1513 static struct bufqueue *
1514 bufqueue(struct buf *bp)
1517 switch (bp->b_qindex) {
1520 case QUEUE_SENTINEL:
1525 return (&bufdomain(bp)->bd_dirtyq);
1527 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1531 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1535 * Return the locked bufqueue that bp is a member of.
1537 static struct bufqueue *
1538 bufqueue_acquire(struct buf *bp)
1540 struct bufqueue *bq, *nbq;
1543 * bp can be pushed from a per-cpu queue to the
1544 * cleanq while we're waiting on the lock. Retry
1545 * if the queues don't match.
1563 * Insert the buffer into the appropriate free list. Requires a
1564 * locked buffer on entry and buffer is unlocked before return.
1567 binsfree(struct buf *bp, int qindex)
1569 struct bufdomain *bd;
1570 struct bufqueue *bq;
1572 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1573 ("binsfree: Invalid qindex %d", qindex));
1574 BUF_ASSERT_XLOCKED(bp);
1577 * Handle delayed bremfree() processing.
1579 if (bp->b_flags & B_REMFREE) {
1580 if (bp->b_qindex == qindex) {
1581 bp->b_flags |= B_REUSE;
1582 bp->b_flags &= ~B_REMFREE;
1586 bq = bufqueue_acquire(bp);
1591 if (qindex == QUEUE_CLEAN) {
1592 if (bd->bd_lim != 0)
1593 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1597 bq = &bd->bd_dirtyq;
1598 bq_insert(bq, bp, true);
1604 * Free a buffer to the buf zone once it no longer has valid contents.
1607 buf_free(struct buf *bp)
1610 if (bp->b_flags & B_REMFREE)
1612 if (bp->b_vflags & BV_BKGRDINPROG)
1613 panic("losing buffer 1");
1614 if (bp->b_rcred != NOCRED) {
1615 crfree(bp->b_rcred);
1616 bp->b_rcred = NOCRED;
1618 if (bp->b_wcred != NOCRED) {
1619 crfree(bp->b_wcred);
1620 bp->b_wcred = NOCRED;
1622 if (!LIST_EMPTY(&bp->b_dep))
1625 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1626 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1628 uma_zfree(buf_zone, bp);
1634 * Import bufs into the uma cache from the buf list. The system still
1635 * expects a static array of bufs and much of the synchronization
1636 * around bufs assumes type stable storage. As a result, UMA is used
1637 * only as a per-cpu cache of bufs still maintained on a global list.
1640 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1646 for (i = 0; i < cnt; i++) {
1647 bp = TAILQ_FIRST(&bqempty.bq_queue);
1650 bq_remove(&bqempty, bp);
1653 BQ_UNLOCK(&bqempty);
1661 * Release bufs from the uma cache back to the buffer queues.
1664 buf_release(void *arg, void **store, int cnt)
1666 struct bufqueue *bq;
1672 for (i = 0; i < cnt; i++) {
1674 /* Inline bq_insert() to batch locking. */
1675 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1676 bp->b_flags &= ~(B_AGE | B_REUSE);
1678 bp->b_qindex = bq->bq_index;
1686 * Allocate an empty buffer header.
1689 buf_alloc(struct bufdomain *bd)
1692 int freebufs, error;
1695 * We can only run out of bufs in the buf zone if the average buf
1696 * is less than BKVASIZE. In this case the actual wait/block will
1697 * come from buf_reycle() failing to flush one of these small bufs.
1700 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1702 bp = uma_zalloc(buf_zone, M_NOWAIT);
1704 atomic_add_int(&bd->bd_freebuffers, 1);
1705 bufspace_daemon_wakeup(bd);
1706 counter_u64_add(numbufallocfails, 1);
1710 * Wake-up the bufspace daemon on transition below threshold.
1712 if (freebufs == bd->bd_lofreebuffers)
1713 bufspace_daemon_wakeup(bd);
1715 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL);
1716 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1720 KASSERT(bp->b_vp == NULL,
1721 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1722 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1723 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1724 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1725 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1726 KASSERT(bp->b_npages == 0,
1727 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1728 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1729 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1730 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1732 bp->b_domain = BD_DOMAIN(bd);
1738 bp->b_blkno = bp->b_lblkno = 0;
1739 bp->b_offset = NOOFFSET;
1745 bp->b_dirtyoff = bp->b_dirtyend = 0;
1746 bp->b_bufobj = NULL;
1747 bp->b_data = bp->b_kvabase = unmapped_buf;
1748 bp->b_fsprivate1 = NULL;
1749 bp->b_fsprivate2 = NULL;
1750 bp->b_fsprivate3 = NULL;
1751 LIST_INIT(&bp->b_dep);
1759 * Free a buffer from the given bufqueue. kva controls whether the
1760 * freed buf must own some kva resources. This is used for
1764 buf_recycle(struct bufdomain *bd, bool kva)
1766 struct bufqueue *bq;
1767 struct buf *bp, *nbp;
1770 counter_u64_add(bufdefragcnt, 1);
1774 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1775 ("buf_recycle: Locks don't match"));
1776 nbp = TAILQ_FIRST(&bq->bq_queue);
1779 * Run scan, possibly freeing data and/or kva mappings on the fly
1782 while ((bp = nbp) != NULL) {
1784 * Calculate next bp (we can only use it if we do not
1785 * release the bqlock).
1787 nbp = TAILQ_NEXT(bp, b_freelist);
1790 * If we are defragging then we need a buffer with
1791 * some kva to reclaim.
1793 if (kva && bp->b_kvasize == 0)
1796 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1800 * Implement a second chance algorithm for frequently
1803 if ((bp->b_flags & B_REUSE) != 0) {
1804 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1805 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1806 bp->b_flags &= ~B_REUSE;
1812 * Skip buffers with background writes in progress.
1814 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1819 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1820 ("buf_recycle: inconsistent queue %d bp %p",
1822 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1823 ("getnewbuf: queue domain %d doesn't match request %d",
1824 bp->b_domain, (int)BD_DOMAIN(bd)));
1826 * NOTE: nbp is now entirely invalid. We can only restart
1827 * the scan from this point on.
1833 * Requeue the background write buffer with error and
1836 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1839 nbp = TAILQ_FIRST(&bq->bq_queue);
1842 bp->b_flags |= B_INVAL;
1855 * Mark the buffer for removal from the appropriate free list.
1859 bremfree(struct buf *bp)
1862 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1863 KASSERT((bp->b_flags & B_REMFREE) == 0,
1864 ("bremfree: buffer %p already marked for delayed removal.", bp));
1865 KASSERT(bp->b_qindex != QUEUE_NONE,
1866 ("bremfree: buffer %p not on a queue.", bp));
1867 BUF_ASSERT_XLOCKED(bp);
1869 bp->b_flags |= B_REMFREE;
1875 * Force an immediate removal from a free list. Used only in nfs when
1876 * it abuses the b_freelist pointer.
1879 bremfreef(struct buf *bp)
1881 struct bufqueue *bq;
1883 bq = bufqueue_acquire(bp);
1889 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1892 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1893 TAILQ_INIT(&bq->bq_queue);
1895 bq->bq_index = qindex;
1896 bq->bq_subqueue = subqueue;
1900 bd_init(struct bufdomain *bd)
1904 /* Per-CPU clean buf queues, plus one global queue. */
1905 bd->bd_subq = mallocarray(mp_maxid + 2, sizeof(struct bufqueue),
1906 M_BIOBUF, M_WAITOK | M_ZERO);
1907 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1908 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1909 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1910 for (i = 0; i <= mp_maxid; i++)
1911 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1912 "bufq clean subqueue lock");
1913 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1919 * Removes a buffer from the free list, must be called with the
1920 * correct qlock held.
1923 bq_remove(struct bufqueue *bq, struct buf *bp)
1926 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1927 bp, bp->b_vp, bp->b_flags);
1928 KASSERT(bp->b_qindex != QUEUE_NONE,
1929 ("bq_remove: buffer %p not on a queue.", bp));
1930 KASSERT(bufqueue(bp) == bq,
1931 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1933 BQ_ASSERT_LOCKED(bq);
1934 if (bp->b_qindex != QUEUE_EMPTY) {
1935 BUF_ASSERT_XLOCKED(bp);
1937 KASSERT(bq->bq_len >= 1,
1938 ("queue %d underflow", bp->b_qindex));
1939 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1941 bp->b_qindex = QUEUE_NONE;
1942 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1946 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1950 BQ_ASSERT_LOCKED(bq);
1951 if (bq != bd->bd_cleanq) {
1953 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1954 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1955 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1957 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1959 bd->bd_cleanq->bq_len += bq->bq_len;
1962 if (bd->bd_wanted) {
1964 wakeup(&bd->bd_wanted);
1966 if (bq != bd->bd_cleanq)
1971 bd_flushall(struct bufdomain *bd)
1973 struct bufqueue *bq;
1977 if (bd->bd_lim == 0)
1980 for (i = 0; i <= mp_maxid; i++) {
1981 bq = &bd->bd_subq[i];
1982 if (bq->bq_len == 0)
1994 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1996 struct bufdomain *bd;
1998 if (bp->b_qindex != QUEUE_NONE)
1999 panic("bq_insert: free buffer %p onto another queue?", bp);
2002 if (bp->b_flags & B_AGE) {
2003 /* Place this buf directly on the real queue. */
2004 if (bq->bq_index == QUEUE_CLEAN)
2007 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2010 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2012 bp->b_flags &= ~(B_AGE | B_REUSE);
2014 bp->b_qindex = bq->bq_index;
2015 bp->b_subqueue = bq->bq_subqueue;
2018 * Unlock before we notify so that we don't wakeup a waiter that
2019 * fails a trylock on the buf and sleeps again.
2024 if (bp->b_qindex == QUEUE_CLEAN) {
2026 * Flush the per-cpu queue and notify any waiters.
2028 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2029 bq->bq_len >= bd->bd_lim))
2038 * Free the kva allocation for a buffer.
2042 bufkva_free(struct buf *bp)
2046 if (bp->b_kvasize == 0) {
2047 KASSERT(bp->b_kvabase == unmapped_buf &&
2048 bp->b_data == unmapped_buf,
2049 ("Leaked KVA space on %p", bp));
2050 } else if (buf_mapped(bp))
2051 BUF_CHECK_MAPPED(bp);
2053 BUF_CHECK_UNMAPPED(bp);
2055 if (bp->b_kvasize == 0)
2058 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2059 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2060 counter_u64_add(buffreekvacnt, 1);
2061 bp->b_data = bp->b_kvabase = unmapped_buf;
2068 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2071 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2076 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2077 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2078 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2079 KASSERT(maxsize <= maxbcachebuf,
2080 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2085 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2088 * Buffer map is too fragmented. Request the caller
2089 * to defragment the map.
2093 bp->b_kvabase = (caddr_t)addr;
2094 bp->b_kvasize = maxsize;
2095 counter_u64_add(bufkvaspace, bp->b_kvasize);
2096 if ((gbflags & GB_UNMAPPED) != 0) {
2097 bp->b_data = unmapped_buf;
2098 BUF_CHECK_UNMAPPED(bp);
2100 bp->b_data = bp->b_kvabase;
2101 BUF_CHECK_MAPPED(bp);
2109 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2110 * callback that fires to avoid returning failure.
2113 bufkva_reclaim(vmem_t *vmem, int flags)
2120 for (i = 0; i < 5; i++) {
2121 for (q = 0; q < buf_domains; q++)
2122 if (buf_recycle(&bdomain[q], true) != 0)
2131 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2132 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2133 * the buffer is valid and we do not have to do anything.
2136 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2137 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2145 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2146 if (inmem(vp, *rablkno))
2148 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2149 if ((rabp->b_flags & B_CACHE) != 0) {
2156 racct_add_buf(curproc, rabp, 0);
2157 PROC_UNLOCK(curproc);
2160 td->td_ru.ru_inblock++;
2161 rabp->b_flags |= B_ASYNC;
2162 rabp->b_flags &= ~B_INVAL;
2163 if ((flags & GB_CKHASH) != 0) {
2164 rabp->b_flags |= B_CKHASH;
2165 rabp->b_ckhashcalc = ckhashfunc;
2167 rabp->b_ioflags &= ~BIO_ERROR;
2168 rabp->b_iocmd = BIO_READ;
2169 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2170 rabp->b_rcred = crhold(cred);
2171 vfs_busy_pages(rabp, 0);
2173 rabp->b_iooffset = dbtob(rabp->b_blkno);
2179 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2181 * Get a buffer with the specified data. Look in the cache first. We
2182 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2183 * is set, the buffer is valid and we do not have to do anything, see
2184 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2186 * Always return a NULL buffer pointer (in bpp) when returning an error.
2188 * The blkno parameter is the logical block being requested. Normally
2189 * the mapping of logical block number to disk block address is done
2190 * by calling VOP_BMAP(). However, if the mapping is already known, the
2191 * disk block address can be passed using the dblkno parameter. If the
2192 * disk block address is not known, then the same value should be passed
2193 * for blkno and dblkno.
2196 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2197 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2198 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2202 int error, readwait, rv;
2204 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2207 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2210 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2215 KASSERT(blkno == bp->b_lblkno,
2216 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2217 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2218 flags &= ~GB_NOSPARSE;
2222 * If not found in cache, do some I/O
2225 if ((bp->b_flags & B_CACHE) == 0) {
2228 PROC_LOCK(td->td_proc);
2229 racct_add_buf(td->td_proc, bp, 0);
2230 PROC_UNLOCK(td->td_proc);
2233 td->td_ru.ru_inblock++;
2234 bp->b_iocmd = BIO_READ;
2235 bp->b_flags &= ~B_INVAL;
2236 if ((flags & GB_CKHASH) != 0) {
2237 bp->b_flags |= B_CKHASH;
2238 bp->b_ckhashcalc = ckhashfunc;
2240 if ((flags & GB_CVTENXIO) != 0)
2241 bp->b_xflags |= BX_CVTENXIO;
2242 bp->b_ioflags &= ~BIO_ERROR;
2243 if (bp->b_rcred == NOCRED && cred != NOCRED)
2244 bp->b_rcred = crhold(cred);
2245 vfs_busy_pages(bp, 0);
2246 bp->b_iooffset = dbtob(bp->b_blkno);
2252 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2254 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2268 * Write, release buffer on completion. (Done by iodone
2269 * if async). Do not bother writing anything if the buffer
2272 * Note that we set B_CACHE here, indicating that buffer is
2273 * fully valid and thus cacheable. This is true even of NFS
2274 * now so we set it generally. This could be set either here
2275 * or in biodone() since the I/O is synchronous. We put it
2279 bufwrite(struct buf *bp)
2286 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2287 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2288 bp->b_flags |= B_INVAL | B_RELBUF;
2289 bp->b_flags &= ~B_CACHE;
2293 if (bp->b_flags & B_INVAL) {
2298 if (bp->b_flags & B_BARRIER)
2299 atomic_add_long(&barrierwrites, 1);
2301 oldflags = bp->b_flags;
2303 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2304 ("FFS background buffer should not get here %p", bp));
2308 vp_md = vp->v_vflag & VV_MD;
2313 * Mark the buffer clean. Increment the bufobj write count
2314 * before bundirty() call, to prevent other thread from seeing
2315 * empty dirty list and zero counter for writes in progress,
2316 * falsely indicating that the bufobj is clean.
2318 bufobj_wref(bp->b_bufobj);
2321 bp->b_flags &= ~B_DONE;
2322 bp->b_ioflags &= ~BIO_ERROR;
2323 bp->b_flags |= B_CACHE;
2324 bp->b_iocmd = BIO_WRITE;
2326 vfs_busy_pages(bp, 1);
2329 * Normal bwrites pipeline writes
2331 bp->b_runningbufspace = bp->b_bufsize;
2332 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2337 racct_add_buf(curproc, bp, 1);
2338 PROC_UNLOCK(curproc);
2341 curthread->td_ru.ru_oublock++;
2342 if (oldflags & B_ASYNC)
2344 bp->b_iooffset = dbtob(bp->b_blkno);
2345 buf_track(bp, __func__);
2348 if ((oldflags & B_ASYNC) == 0) {
2349 int rtval = bufwait(bp);
2352 } else if (space > hirunningspace) {
2354 * don't allow the async write to saturate the I/O
2355 * system. We will not deadlock here because
2356 * we are blocking waiting for I/O that is already in-progress
2357 * to complete. We do not block here if it is the update
2358 * or syncer daemon trying to clean up as that can lead
2361 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2362 waitrunningbufspace();
2369 bufbdflush(struct bufobj *bo, struct buf *bp)
2372 struct bufdomain *bd;
2374 bd = &bdomain[bo->bo_domain];
2375 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2376 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2378 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2381 * Try to find a buffer to flush.
2383 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2384 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2386 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2389 panic("bdwrite: found ourselves");
2391 /* Don't countdeps with the bo lock held. */
2392 if (buf_countdeps(nbp, 0)) {
2397 if (nbp->b_flags & B_CLUSTEROK) {
2398 vfs_bio_awrite(nbp);
2403 dirtybufferflushes++;
2412 * Delayed write. (Buffer is marked dirty). Do not bother writing
2413 * anything if the buffer is marked invalid.
2415 * Note that since the buffer must be completely valid, we can safely
2416 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2417 * biodone() in order to prevent getblk from writing the buffer
2418 * out synchronously.
2421 bdwrite(struct buf *bp)
2423 struct thread *td = curthread;
2427 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2428 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2429 KASSERT((bp->b_flags & B_BARRIER) == 0,
2430 ("Barrier request in delayed write %p", bp));
2432 if (bp->b_flags & B_INVAL) {
2438 * If we have too many dirty buffers, don't create any more.
2439 * If we are wildly over our limit, then force a complete
2440 * cleanup. Otherwise, just keep the situation from getting
2441 * out of control. Note that we have to avoid a recursive
2442 * disaster and not try to clean up after our own cleanup!
2446 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2447 td->td_pflags |= TDP_INBDFLUSH;
2449 td->td_pflags &= ~TDP_INBDFLUSH;
2455 * Set B_CACHE, indicating that the buffer is fully valid. This is
2456 * true even of NFS now.
2458 bp->b_flags |= B_CACHE;
2461 * This bmap keeps the system from needing to do the bmap later,
2462 * perhaps when the system is attempting to do a sync. Since it
2463 * is likely that the indirect block -- or whatever other datastructure
2464 * that the filesystem needs is still in memory now, it is a good
2465 * thing to do this. Note also, that if the pageout daemon is
2466 * requesting a sync -- there might not be enough memory to do
2467 * the bmap then... So, this is important to do.
2469 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2470 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2473 buf_track(bp, __func__);
2476 * Set the *dirty* buffer range based upon the VM system dirty
2479 * Mark the buffer pages as clean. We need to do this here to
2480 * satisfy the vnode_pager and the pageout daemon, so that it
2481 * thinks that the pages have been "cleaned". Note that since
2482 * the pages are in a delayed write buffer -- the VFS layer
2483 * "will" see that the pages get written out on the next sync,
2484 * or perhaps the cluster will be completed.
2486 vfs_clean_pages_dirty_buf(bp);
2490 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2491 * due to the softdep code.
2498 * Turn buffer into delayed write request. We must clear BIO_READ and
2499 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2500 * itself to properly update it in the dirty/clean lists. We mark it
2501 * B_DONE to ensure that any asynchronization of the buffer properly
2502 * clears B_DONE ( else a panic will occur later ).
2504 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2505 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2506 * should only be called if the buffer is known-good.
2508 * Since the buffer is not on a queue, we do not update the numfreebuffers
2511 * The buffer must be on QUEUE_NONE.
2514 bdirty(struct buf *bp)
2517 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2518 bp, bp->b_vp, bp->b_flags);
2519 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2520 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2521 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2522 bp->b_flags &= ~(B_RELBUF);
2523 bp->b_iocmd = BIO_WRITE;
2525 if ((bp->b_flags & B_DELWRI) == 0) {
2526 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2535 * Clear B_DELWRI for buffer.
2537 * Since the buffer is not on a queue, we do not update the numfreebuffers
2540 * The buffer must be on QUEUE_NONE.
2544 bundirty(struct buf *bp)
2547 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2548 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2549 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2550 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2552 if (bp->b_flags & B_DELWRI) {
2553 bp->b_flags &= ~B_DELWRI;
2558 * Since it is now being written, we can clear its deferred write flag.
2560 bp->b_flags &= ~B_DEFERRED;
2566 * Asynchronous write. Start output on a buffer, but do not wait for
2567 * it to complete. The buffer is released when the output completes.
2569 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2570 * B_INVAL buffers. Not us.
2573 bawrite(struct buf *bp)
2576 bp->b_flags |= B_ASYNC;
2583 * Asynchronous barrier write. Start output on a buffer, but do not
2584 * wait for it to complete. Place a write barrier after this write so
2585 * that this buffer and all buffers written before it are committed to
2586 * the disk before any buffers written after this write are committed
2587 * to the disk. The buffer is released when the output completes.
2590 babarrierwrite(struct buf *bp)
2593 bp->b_flags |= B_ASYNC | B_BARRIER;
2600 * Synchronous barrier write. Start output on a buffer and wait for
2601 * it to complete. Place a write barrier after this write so that
2602 * this buffer and all buffers written before it are committed to
2603 * the disk before any buffers written after this write are committed
2604 * to the disk. The buffer is released when the output completes.
2607 bbarrierwrite(struct buf *bp)
2610 bp->b_flags |= B_BARRIER;
2611 return (bwrite(bp));
2617 * Called prior to the locking of any vnodes when we are expecting to
2618 * write. We do not want to starve the buffer cache with too many
2619 * dirty buffers so we block here. By blocking prior to the locking
2620 * of any vnodes we attempt to avoid the situation where a locked vnode
2621 * prevents the various system daemons from flushing related buffers.
2627 if (buf_dirty_count_severe()) {
2628 mtx_lock(&bdirtylock);
2629 while (buf_dirty_count_severe()) {
2631 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2634 mtx_unlock(&bdirtylock);
2639 * Return true if we have too many dirty buffers.
2642 buf_dirty_count_severe(void)
2645 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2651 * Release a busy buffer and, if requested, free its resources. The
2652 * buffer will be stashed in the appropriate bufqueue[] allowing it
2653 * to be accessed later as a cache entity or reused for other purposes.
2656 brelse(struct buf *bp)
2658 struct mount *v_mnt;
2662 * Many functions erroneously call brelse with a NULL bp under rare
2663 * error conditions. Simply return when called with a NULL bp.
2667 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2668 bp, bp->b_vp, bp->b_flags);
2669 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2670 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2671 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2672 ("brelse: non-VMIO buffer marked NOREUSE"));
2674 if (BUF_LOCKRECURSED(bp)) {
2676 * Do not process, in particular, do not handle the
2677 * B_INVAL/B_RELBUF and do not release to free list.
2683 if (bp->b_flags & B_MANAGED) {
2688 if (LIST_EMPTY(&bp->b_dep)) {
2689 bp->b_flags &= ~B_IOSTARTED;
2691 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2692 ("brelse: SU io not finished bp %p", bp));
2695 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2696 BO_LOCK(bp->b_bufobj);
2697 bp->b_vflags &= ~BV_BKGRDERR;
2698 BO_UNLOCK(bp->b_bufobj);
2702 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2703 (bp->b_flags & B_INVALONERR)) {
2705 * Forced invalidation of dirty buffer contents, to be used
2706 * after a failed write in the rare case that the loss of the
2707 * contents is acceptable. The buffer is invalidated and
2710 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2711 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2714 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2715 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2716 !(bp->b_flags & B_INVAL)) {
2718 * Failed write, redirty. All errors except ENXIO (which
2719 * means the device is gone) are treated as being
2722 * XXX Treating EIO as transient is not correct; the
2723 * contract with the local storage device drivers is that
2724 * they will only return EIO once the I/O is no longer
2725 * retriable. Network I/O also respects this through the
2726 * guarantees of TCP and/or the internal retries of NFS.
2727 * ENOMEM might be transient, but we also have no way of
2728 * knowing when its ok to retry/reschedule. In general,
2729 * this entire case should be made obsolete through better
2730 * error handling/recovery and resource scheduling.
2732 * Do this also for buffers that failed with ENXIO, but have
2733 * non-empty dependencies - the soft updates code might need
2734 * to access the buffer to untangle them.
2736 * Must clear BIO_ERROR to prevent pages from being scrapped.
2738 bp->b_ioflags &= ~BIO_ERROR;
2740 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2741 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2743 * Either a failed read I/O, or we were asked to free or not
2744 * cache the buffer, or we failed to write to a device that's
2745 * no longer present.
2747 bp->b_flags |= B_INVAL;
2748 if (!LIST_EMPTY(&bp->b_dep))
2750 if (bp->b_flags & B_DELWRI)
2752 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2753 if ((bp->b_flags & B_VMIO) == 0) {
2761 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2762 * is called with B_DELWRI set, the underlying pages may wind up
2763 * getting freed causing a previous write (bdwrite()) to get 'lost'
2764 * because pages associated with a B_DELWRI bp are marked clean.
2766 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2767 * if B_DELWRI is set.
2769 if (bp->b_flags & B_DELWRI)
2770 bp->b_flags &= ~B_RELBUF;
2773 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2774 * constituted, not even NFS buffers now. Two flags effect this. If
2775 * B_INVAL, the struct buf is invalidated but the VM object is kept
2776 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2778 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2779 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2780 * buffer is also B_INVAL because it hits the re-dirtying code above.
2782 * Normally we can do this whether a buffer is B_DELWRI or not. If
2783 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2784 * the commit state and we cannot afford to lose the buffer. If the
2785 * buffer has a background write in progress, we need to keep it
2786 * around to prevent it from being reconstituted and starting a second
2790 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2792 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2793 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2794 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2795 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2796 vfs_vmio_invalidate(bp);
2800 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2801 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2803 bp->b_flags &= ~B_NOREUSE;
2804 if (bp->b_vp != NULL)
2809 * If the buffer has junk contents signal it and eventually
2810 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2813 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2814 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2815 bp->b_flags |= B_INVAL;
2816 if (bp->b_flags & B_INVAL) {
2817 if (bp->b_flags & B_DELWRI)
2823 buf_track(bp, __func__);
2825 /* buffers with no memory */
2826 if (bp->b_bufsize == 0) {
2830 /* buffers with junk contents */
2831 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2832 (bp->b_ioflags & BIO_ERROR)) {
2833 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2834 if (bp->b_vflags & BV_BKGRDINPROG)
2835 panic("losing buffer 2");
2836 qindex = QUEUE_CLEAN;
2837 bp->b_flags |= B_AGE;
2838 /* remaining buffers */
2839 } else if (bp->b_flags & B_DELWRI)
2840 qindex = QUEUE_DIRTY;
2842 qindex = QUEUE_CLEAN;
2844 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2845 panic("brelse: not dirty");
2847 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2848 bp->b_xflags &= ~(BX_CVTENXIO);
2849 /* binsfree unlocks bp. */
2850 binsfree(bp, qindex);
2854 * Release a buffer back to the appropriate queue but do not try to free
2855 * it. The buffer is expected to be used again soon.
2857 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2858 * biodone() to requeue an async I/O on completion. It is also used when
2859 * known good buffers need to be requeued but we think we may need the data
2862 * XXX we should be able to leave the B_RELBUF hint set on completion.
2865 bqrelse(struct buf *bp)
2869 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2870 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2871 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2873 qindex = QUEUE_NONE;
2874 if (BUF_LOCKRECURSED(bp)) {
2875 /* do not release to free list */
2879 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2880 bp->b_xflags &= ~(BX_CVTENXIO);
2882 if (LIST_EMPTY(&bp->b_dep)) {
2883 bp->b_flags &= ~B_IOSTARTED;
2885 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2886 ("bqrelse: SU io not finished bp %p", bp));
2889 if (bp->b_flags & B_MANAGED) {
2890 if (bp->b_flags & B_REMFREE)
2895 /* buffers with stale but valid contents */
2896 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2897 BV_BKGRDERR)) == BV_BKGRDERR) {
2898 BO_LOCK(bp->b_bufobj);
2899 bp->b_vflags &= ~BV_BKGRDERR;
2900 BO_UNLOCK(bp->b_bufobj);
2901 qindex = QUEUE_DIRTY;
2903 if ((bp->b_flags & B_DELWRI) == 0 &&
2904 (bp->b_xflags & BX_VNDIRTY))
2905 panic("bqrelse: not dirty");
2906 if ((bp->b_flags & B_NOREUSE) != 0) {
2910 qindex = QUEUE_CLEAN;
2912 buf_track(bp, __func__);
2913 /* binsfree unlocks bp. */
2914 binsfree(bp, qindex);
2918 buf_track(bp, __func__);
2924 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2925 * restore bogus pages.
2928 vfs_vmio_iodone(struct buf *bp)
2933 struct vnode *vp __unused;
2934 int i, iosize, resid;
2937 obj = bp->b_bufobj->bo_object;
2938 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2939 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2940 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2943 VNPASS(vp->v_holdcnt > 0, vp);
2944 VNPASS(vp->v_object != NULL, vp);
2946 foff = bp->b_offset;
2947 KASSERT(bp->b_offset != NOOFFSET,
2948 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2951 iosize = bp->b_bcount - bp->b_resid;
2952 for (i = 0; i < bp->b_npages; i++) {
2953 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2958 * cleanup bogus pages, restoring the originals
2961 if (m == bogus_page) {
2963 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2965 panic("biodone: page disappeared!");
2967 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2969 * In the write case, the valid and clean bits are
2970 * already changed correctly ( see bdwrite() ), so we
2971 * only need to do this here in the read case.
2973 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2974 resid)) == 0, ("vfs_vmio_iodone: page %p "
2975 "has unexpected dirty bits", m));
2976 vfs_page_set_valid(bp, foff, m);
2978 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2979 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2980 (intmax_t)foff, (uintmax_t)m->pindex));
2983 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2986 vm_object_pip_wakeupn(obj, bp->b_npages);
2987 if (bogus && buf_mapped(bp)) {
2988 BUF_CHECK_MAPPED(bp);
2989 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2990 bp->b_pages, bp->b_npages);
2995 * Perform page invalidation when a buffer is released. The fully invalid
2996 * pages will be reclaimed later in vfs_vmio_truncate().
2999 vfs_vmio_invalidate(struct buf *bp)
3003 int flags, i, resid, poffset, presid;
3005 if (buf_mapped(bp)) {
3006 BUF_CHECK_MAPPED(bp);
3007 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3009 BUF_CHECK_UNMAPPED(bp);
3011 * Get the base offset and length of the buffer. Note that
3012 * in the VMIO case if the buffer block size is not
3013 * page-aligned then b_data pointer may not be page-aligned.
3014 * But our b_pages[] array *IS* page aligned.
3016 * block sizes less then DEV_BSIZE (usually 512) are not
3017 * supported due to the page granularity bits (m->valid,
3018 * m->dirty, etc...).
3020 * See man buf(9) for more information
3022 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3023 obj = bp->b_bufobj->bo_object;
3024 resid = bp->b_bufsize;
3025 poffset = bp->b_offset & PAGE_MASK;
3026 VM_OBJECT_WLOCK(obj);
3027 for (i = 0; i < bp->b_npages; i++) {
3029 if (m == bogus_page)
3030 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3031 bp->b_pages[i] = NULL;
3033 presid = resid > (PAGE_SIZE - poffset) ?
3034 (PAGE_SIZE - poffset) : resid;
3035 KASSERT(presid >= 0, ("brelse: extra page"));
3036 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3037 if (pmap_page_wired_mappings(m) == 0)
3038 vm_page_set_invalid(m, poffset, presid);
3040 vm_page_release_locked(m, flags);
3044 VM_OBJECT_WUNLOCK(obj);
3049 * Page-granular truncation of an existing VMIO buffer.
3052 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3058 if (bp->b_npages == desiredpages)
3061 if (buf_mapped(bp)) {
3062 BUF_CHECK_MAPPED(bp);
3063 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3064 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3066 BUF_CHECK_UNMAPPED(bp);
3069 * The object lock is needed only if we will attempt to free pages.
3071 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3072 if ((bp->b_flags & B_DIRECT) != 0) {
3073 flags |= VPR_TRYFREE;
3074 obj = bp->b_bufobj->bo_object;
3075 VM_OBJECT_WLOCK(obj);
3079 for (i = desiredpages; i < bp->b_npages; i++) {
3081 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3082 bp->b_pages[i] = NULL;
3084 vm_page_release_locked(m, flags);
3086 vm_page_release(m, flags);
3089 VM_OBJECT_WUNLOCK(obj);
3090 bp->b_npages = desiredpages;
3094 * Byte granular extension of VMIO buffers.
3097 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3100 * We are growing the buffer, possibly in a
3101 * byte-granular fashion.
3109 * Step 1, bring in the VM pages from the object, allocating
3110 * them if necessary. We must clear B_CACHE if these pages
3111 * are not valid for the range covered by the buffer.
3113 obj = bp->b_bufobj->bo_object;
3114 if (bp->b_npages < desiredpages) {
3115 KASSERT(desiredpages <= atop(maxbcachebuf),
3116 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3117 bp, desiredpages, maxbcachebuf));
3120 * We must allocate system pages since blocking
3121 * here could interfere with paging I/O, no
3122 * matter which process we are.
3124 * Only exclusive busy can be tested here.
3125 * Blocking on shared busy might lead to
3126 * deadlocks once allocbuf() is called after
3127 * pages are vfs_busy_pages().
3129 (void)vm_page_grab_pages_unlocked(obj,
3130 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3131 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3132 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3133 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3134 bp->b_npages = desiredpages;
3138 * Step 2. We've loaded the pages into the buffer,
3139 * we have to figure out if we can still have B_CACHE
3140 * set. Note that B_CACHE is set according to the
3141 * byte-granular range ( bcount and size ), not the
3142 * aligned range ( newbsize ).
3144 * The VM test is against m->valid, which is DEV_BSIZE
3145 * aligned. Needless to say, the validity of the data
3146 * needs to also be DEV_BSIZE aligned. Note that this
3147 * fails with NFS if the server or some other client
3148 * extends the file's EOF. If our buffer is resized,
3149 * B_CACHE may remain set! XXX
3151 toff = bp->b_bcount;
3152 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3153 while ((bp->b_flags & B_CACHE) && toff < size) {
3156 if (tinc > (size - toff))
3158 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3159 m = bp->b_pages[pi];
3160 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3166 * Step 3, fixup the KVA pmap.
3171 BUF_CHECK_UNMAPPED(bp);
3175 * Check to see if a block at a particular lbn is available for a clustered
3179 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3186 /* If the buf isn't in core skip it */
3187 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3190 /* If the buf is busy we don't want to wait for it */
3191 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3194 /* Only cluster with valid clusterable delayed write buffers */
3195 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3196 (B_DELWRI | B_CLUSTEROK))
3199 if (bpa->b_bufsize != size)
3203 * Check to see if it is in the expected place on disk and that the
3204 * block has been mapped.
3206 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3216 * Implement clustered async writes for clearing out B_DELWRI buffers.
3217 * This is much better then the old way of writing only one buffer at
3218 * a time. Note that we may not be presented with the buffers in the
3219 * correct order, so we search for the cluster in both directions.
3222 vfs_bio_awrite(struct buf *bp)
3227 daddr_t lblkno = bp->b_lblkno;
3228 struct vnode *vp = bp->b_vp;
3236 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3238 * right now we support clustered writing only to regular files. If
3239 * we find a clusterable block we could be in the middle of a cluster
3240 * rather then at the beginning.
3242 if ((vp->v_type == VREG) &&
3243 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3244 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3245 size = vp->v_mount->mnt_stat.f_iosize;
3246 maxcl = maxphys / size;
3249 for (i = 1; i < maxcl; i++)
3250 if (vfs_bio_clcheck(vp, size, lblkno + i,
3251 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3254 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3255 if (vfs_bio_clcheck(vp, size, lblkno - j,
3256 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3262 * this is a possible cluster write
3266 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3272 bp->b_flags |= B_ASYNC;
3274 * default (old) behavior, writing out only one block
3276 * XXX returns b_bufsize instead of b_bcount for nwritten?
3278 nwritten = bp->b_bufsize;
3287 * Allocate KVA for an empty buf header according to gbflags.
3290 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3293 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3295 * In order to keep fragmentation sane we only allocate kva
3296 * in BKVASIZE chunks. XXX with vmem we can do page size.
3298 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3300 if (maxsize != bp->b_kvasize &&
3301 bufkva_alloc(bp, maxsize, gbflags))
3310 * Find and initialize a new buffer header, freeing up existing buffers
3311 * in the bufqueues as necessary. The new buffer is returned locked.
3314 * We have insufficient buffer headers
3315 * We have insufficient buffer space
3316 * buffer_arena is too fragmented ( space reservation fails )
3317 * If we have to flush dirty buffers ( but we try to avoid this )
3319 * The caller is responsible for releasing the reserved bufspace after
3320 * allocbuf() is called.
3323 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3325 struct bufdomain *bd;
3327 bool metadata, reserved;
3330 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3331 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3332 if (!unmapped_buf_allowed)
3333 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3335 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3343 bd = &bdomain[vp->v_bufobj.bo_domain];
3345 counter_u64_add(getnewbufcalls, 1);
3348 if (reserved == false &&
3349 bufspace_reserve(bd, maxsize, metadata) != 0) {
3350 counter_u64_add(getnewbufrestarts, 1);
3354 if ((bp = buf_alloc(bd)) == NULL) {
3355 counter_u64_add(getnewbufrestarts, 1);
3358 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3361 } while (buf_recycle(bd, false) == 0);
3364 bufspace_release(bd, maxsize);
3366 bp->b_flags |= B_INVAL;
3369 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3377 * buffer flushing daemon. Buffers are normally flushed by the
3378 * update daemon but if it cannot keep up this process starts to
3379 * take the load in an attempt to prevent getnewbuf() from blocking.
3381 static struct kproc_desc buf_kp = {
3386 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3389 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3393 flushed = flushbufqueues(vp, bd, target, 0);
3396 * Could not find any buffers without rollback
3397 * dependencies, so just write the first one
3398 * in the hopes of eventually making progress.
3400 if (vp != NULL && target > 2)
3402 flushbufqueues(vp, bd, target, 1);
3408 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3412 if (KERNEL_PANICKED())
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);
3974 goto newbuf_unlocked;
3976 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3981 /* Verify buf identify has not changed since lookup. */
3982 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3983 goto foundbuf_fastpath;
3985 /* It changed, fallback to locked lookup. */
3990 bp = gbincore(bo, blkno);
3995 * Buffer is in-core. If the buffer is not busy nor managed,
3996 * it must be on a queue.
3998 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3999 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
4001 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4004 error = BUF_TIMELOCK(bp, lockflags,
4005 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4008 * If we slept and got the lock we have to restart in case
4009 * the buffer changed identities.
4011 if (error == ENOLCK)
4013 /* We timed out or were interrupted. */
4014 else if (error != 0)
4018 /* If recursed, assume caller knows the rules. */
4019 if (BUF_LOCKRECURSED(bp))
4023 * The buffer is locked. B_CACHE is cleared if the buffer is
4024 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4025 * and for a VMIO buffer B_CACHE is adjusted according to the
4028 if (bp->b_flags & B_INVAL)
4029 bp->b_flags &= ~B_CACHE;
4030 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4031 bp->b_flags |= B_CACHE;
4032 if (bp->b_flags & B_MANAGED)
4033 MPASS(bp->b_qindex == QUEUE_NONE);
4038 * check for size inconsistencies for non-VMIO case.
4040 if (bp->b_bcount != size) {
4041 if ((bp->b_flags & B_VMIO) == 0 ||
4042 (size > bp->b_kvasize)) {
4043 if (bp->b_flags & B_DELWRI) {
4044 bp->b_flags |= B_NOCACHE;
4047 if (LIST_EMPTY(&bp->b_dep)) {
4048 bp->b_flags |= B_RELBUF;
4051 bp->b_flags |= B_NOCACHE;
4060 * Handle the case of unmapped buffer which should
4061 * become mapped, or the buffer for which KVA
4062 * reservation is requested.
4064 bp_unmapped_get_kva(bp, blkno, size, flags);
4067 * If the size is inconsistent in the VMIO case, we can resize
4068 * the buffer. This might lead to B_CACHE getting set or
4069 * cleared. If the size has not changed, B_CACHE remains
4070 * unchanged from its previous state.
4074 KASSERT(bp->b_offset != NOOFFSET,
4075 ("getblk: no buffer offset"));
4078 * A buffer with B_DELWRI set and B_CACHE clear must
4079 * be committed before we can return the buffer in
4080 * order to prevent the caller from issuing a read
4081 * ( due to B_CACHE not being set ) and overwriting
4084 * Most callers, including NFS and FFS, need this to
4085 * operate properly either because they assume they
4086 * can issue a read if B_CACHE is not set, or because
4087 * ( for example ) an uncached B_DELWRI might loop due
4088 * to softupdates re-dirtying the buffer. In the latter
4089 * case, B_CACHE is set after the first write completes,
4090 * preventing further loops.
4091 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4092 * above while extending the buffer, we cannot allow the
4093 * buffer to remain with B_CACHE set after the write
4094 * completes or it will represent a corrupt state. To
4095 * deal with this we set B_NOCACHE to scrap the buffer
4098 * We might be able to do something fancy, like setting
4099 * B_CACHE in bwrite() except if B_DELWRI is already set,
4100 * so the below call doesn't set B_CACHE, but that gets real
4101 * confusing. This is much easier.
4104 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4105 bp->b_flags |= B_NOCACHE;
4109 bp->b_flags &= ~B_DONE;
4112 * Buffer is not in-core, create new buffer. The buffer
4113 * returned by getnewbuf() is locked. Note that the returned
4114 * buffer is also considered valid (not marked B_INVAL).
4119 * If the user does not want us to create the buffer, bail out
4122 if (flags & GB_NOCREAT)
4125 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4126 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4127 offset = blkno * bsize;
4128 vmio = vp->v_object != NULL;
4130 maxsize = size + (offset & PAGE_MASK);
4133 /* Do not allow non-VMIO notmapped buffers. */
4134 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4136 maxsize = imax(maxsize, bsize);
4137 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4139 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4140 KASSERT(error != EOPNOTSUPP,
4141 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4146 return (EJUSTRETURN);
4149 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4151 if (slpflag || slptimeo)
4154 * XXX This is here until the sleep path is diagnosed
4155 * enough to work under very low memory conditions.
4157 * There's an issue on low memory, 4BSD+non-preempt
4158 * systems (eg MIPS routers with 32MB RAM) where buffer
4159 * exhaustion occurs without sleeping for buffer
4160 * reclaimation. This just sticks in a loop and
4161 * constantly attempts to allocate a buffer, which
4162 * hits exhaustion and tries to wakeup bufdaemon.
4163 * This never happens because we never yield.
4165 * The real solution is to identify and fix these cases
4166 * so we aren't effectively busy-waiting in a loop
4167 * until the reclaimation path has cycles to run.
4169 kern_yield(PRI_USER);
4174 * This code is used to make sure that a buffer is not
4175 * created while the getnewbuf routine is blocked.
4176 * This can be a problem whether the vnode is locked or not.
4177 * If the buffer is created out from under us, we have to
4178 * throw away the one we just created.
4180 * Note: this must occur before we associate the buffer
4181 * with the vp especially considering limitations in
4182 * the splay tree implementation when dealing with duplicate
4186 if (gbincore(bo, blkno)) {
4188 bp->b_flags |= B_INVAL;
4189 bufspace_release(bufdomain(bp), maxsize);
4195 * Insert the buffer into the hash, so that it can
4196 * be found by incore.
4198 bp->b_lblkno = blkno;
4199 bp->b_blkno = d_blkno;
4200 bp->b_offset = offset;
4205 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4206 * buffer size starts out as 0, B_CACHE will be set by
4207 * allocbuf() for the VMIO case prior to it testing the
4208 * backing store for validity.
4212 bp->b_flags |= B_VMIO;
4213 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4214 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4215 bp, vp->v_object, bp->b_bufobj->bo_object));
4217 bp->b_flags &= ~B_VMIO;
4218 KASSERT(bp->b_bufobj->bo_object == NULL,
4219 ("ARGH! has b_bufobj->bo_object %p %p\n",
4220 bp, bp->b_bufobj->bo_object));
4221 BUF_CHECK_MAPPED(bp);
4225 bufspace_release(bufdomain(bp), maxsize);
4226 bp->b_flags &= ~B_DONE;
4228 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4230 buf_track(bp, __func__);
4231 KASSERT(bp->b_bufobj == bo,
4232 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4238 * Get an empty, disassociated buffer of given size. The buffer is initially
4242 geteblk(int size, int flags)
4247 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4248 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4249 if ((flags & GB_NOWAIT_BD) &&
4250 (curthread->td_pflags & TDP_BUFNEED) != 0)
4254 bufspace_release(bufdomain(bp), maxsize);
4255 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4260 * Truncate the backing store for a non-vmio buffer.
4263 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4266 if (bp->b_flags & B_MALLOC) {
4268 * malloced buffers are not shrunk
4270 if (newbsize == 0) {
4271 bufmallocadjust(bp, 0);
4272 free(bp->b_data, M_BIOBUF);
4273 bp->b_data = bp->b_kvabase;
4274 bp->b_flags &= ~B_MALLOC;
4278 vm_hold_free_pages(bp, newbsize);
4279 bufspace_adjust(bp, newbsize);
4283 * Extend the backing for a non-VMIO buffer.
4286 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4292 * We only use malloced memory on the first allocation.
4293 * and revert to page-allocated memory when the buffer
4296 * There is a potential smp race here that could lead
4297 * to bufmallocspace slightly passing the max. It
4298 * is probably extremely rare and not worth worrying
4301 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4302 bufmallocspace < maxbufmallocspace) {
4303 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4304 bp->b_flags |= B_MALLOC;
4305 bufmallocadjust(bp, newbsize);
4310 * If the buffer is growing on its other-than-first
4311 * allocation then we revert to the page-allocation
4316 if (bp->b_flags & B_MALLOC) {
4317 origbuf = bp->b_data;
4318 origbufsize = bp->b_bufsize;
4319 bp->b_data = bp->b_kvabase;
4320 bufmallocadjust(bp, 0);
4321 bp->b_flags &= ~B_MALLOC;
4322 newbsize = round_page(newbsize);
4324 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4325 (vm_offset_t) bp->b_data + newbsize);
4326 if (origbuf != NULL) {
4327 bcopy(origbuf, bp->b_data, origbufsize);
4328 free(origbuf, M_BIOBUF);
4330 bufspace_adjust(bp, newbsize);
4334 * This code constitutes the buffer memory from either anonymous system
4335 * memory (in the case of non-VMIO operations) or from an associated
4336 * VM object (in the case of VMIO operations). This code is able to
4337 * resize a buffer up or down.
4339 * Note that this code is tricky, and has many complications to resolve
4340 * deadlock or inconsistent data situations. Tread lightly!!!
4341 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4342 * the caller. Calling this code willy nilly can result in the loss of data.
4344 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4345 * B_CACHE for the non-VMIO case.
4348 allocbuf(struct buf *bp, int size)
4352 if (bp->b_bcount == size)
4355 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size,
4356 ("allocbuf: buffer too small %p %#x %#x",
4357 bp, bp->b_kvasize, size));
4359 newbsize = roundup2(size, DEV_BSIZE);
4360 if ((bp->b_flags & B_VMIO) == 0) {
4361 if ((bp->b_flags & B_MALLOC) == 0)
4362 newbsize = round_page(newbsize);
4364 * Just get anonymous memory from the kernel. Don't
4365 * mess with B_CACHE.
4367 if (newbsize < bp->b_bufsize)
4368 vfs_nonvmio_truncate(bp, newbsize);
4369 else if (newbsize > bp->b_bufsize)
4370 vfs_nonvmio_extend(bp, newbsize);
4374 desiredpages = size == 0 ? 0 :
4375 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4377 KASSERT((bp->b_flags & B_MALLOC) == 0,
4378 ("allocbuf: VMIO buffer can't be malloced %p", bp));
4381 * Set B_CACHE initially if buffer is 0 length or will become
4384 if (size == 0 || bp->b_bufsize == 0)
4385 bp->b_flags |= B_CACHE;
4387 if (newbsize < bp->b_bufsize)
4388 vfs_vmio_truncate(bp, desiredpages);
4389 /* XXX This looks as if it should be newbsize > b_bufsize */
4390 else if (size > bp->b_bcount)
4391 vfs_vmio_extend(bp, desiredpages, size);
4392 bufspace_adjust(bp, newbsize);
4394 bp->b_bcount = size; /* requested buffer size. */
4398 extern int inflight_transient_maps;
4400 static struct bio_queue nondump_bios;
4403 biodone(struct bio *bp)
4406 void (*done)(struct bio *);
4407 vm_offset_t start, end;
4409 biotrack(bp, __func__);
4412 * Avoid completing I/O when dumping after a panic since that may
4413 * result in a deadlock in the filesystem or pager code. Note that
4414 * this doesn't affect dumps that were started manually since we aim
4415 * to keep the system usable after it has been resumed.
4417 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4418 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4421 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4422 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4423 bp->bio_flags |= BIO_UNMAPPED;
4424 start = trunc_page((vm_offset_t)bp->bio_data);
4425 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4426 bp->bio_data = unmapped_buf;
4427 pmap_qremove(start, atop(end - start));
4428 vmem_free(transient_arena, start, end - start);
4429 atomic_add_int(&inflight_transient_maps, -1);
4431 done = bp->bio_done;
4433 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4435 bp->bio_flags |= BIO_DONE;
4443 * Wait for a BIO to finish.
4446 biowait(struct bio *bp, const char *wmesg)
4450 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4452 while ((bp->bio_flags & BIO_DONE) == 0)
4453 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4455 if (bp->bio_error != 0)
4456 return (bp->bio_error);
4457 if (!(bp->bio_flags & BIO_ERROR))
4463 biofinish(struct bio *bp, struct devstat *stat, int error)
4467 bp->bio_error = error;
4468 bp->bio_flags |= BIO_ERROR;
4471 devstat_end_transaction_bio(stat, bp);
4475 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4477 biotrack_buf(struct bio *bp, const char *location)
4480 buf_track(bp->bio_track_bp, location);
4487 * Wait for buffer I/O completion, returning error status. The buffer
4488 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4489 * error and cleared.
4492 bufwait(struct buf *bp)
4494 if (bp->b_iocmd == BIO_READ)
4495 bwait(bp, PRIBIO, "biord");
4497 bwait(bp, PRIBIO, "biowr");
4498 if (bp->b_flags & B_EINTR) {
4499 bp->b_flags &= ~B_EINTR;
4502 if (bp->b_ioflags & BIO_ERROR) {
4503 return (bp->b_error ? bp->b_error : EIO);
4512 * Finish I/O on a buffer, optionally calling a completion function.
4513 * This is usually called from an interrupt so process blocking is
4516 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4517 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4518 * assuming B_INVAL is clear.
4520 * For the VMIO case, we set B_CACHE if the op was a read and no
4521 * read error occurred, or if the op was a write. B_CACHE is never
4522 * set if the buffer is invalid or otherwise uncacheable.
4524 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4525 * initiator to leave B_INVAL set to brelse the buffer out of existence
4526 * in the biodone routine.
4529 bufdone(struct buf *bp)
4531 struct bufobj *dropobj;
4532 void (*biodone)(struct buf *);
4534 buf_track(bp, __func__);
4535 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4538 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4540 runningbufwakeup(bp);
4541 if (bp->b_iocmd == BIO_WRITE)
4542 dropobj = bp->b_bufobj;
4543 /* call optional completion function if requested */
4544 if (bp->b_iodone != NULL) {
4545 biodone = bp->b_iodone;
4546 bp->b_iodone = NULL;
4549 bufobj_wdrop(dropobj);
4552 if (bp->b_flags & B_VMIO) {
4554 * Set B_CACHE if the op was a normal read and no error
4555 * occurred. B_CACHE is set for writes in the b*write()
4558 if (bp->b_iocmd == BIO_READ &&
4559 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4560 !(bp->b_ioflags & BIO_ERROR))
4561 bp->b_flags |= B_CACHE;
4562 vfs_vmio_iodone(bp);
4564 if (!LIST_EMPTY(&bp->b_dep))
4566 if ((bp->b_flags & B_CKHASH) != 0) {
4567 KASSERT(bp->b_iocmd == BIO_READ,
4568 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4569 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4570 (*bp->b_ckhashcalc)(bp);
4573 * For asynchronous completions, release the buffer now. The brelse
4574 * will do a wakeup there if necessary - so no need to do a wakeup
4575 * here in the async case. The sync case always needs to do a wakeup.
4577 if (bp->b_flags & B_ASYNC) {
4578 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4579 (bp->b_ioflags & BIO_ERROR))
4586 bufobj_wdrop(dropobj);
4590 * This routine is called in lieu of iodone in the case of
4591 * incomplete I/O. This keeps the busy status for pages
4595 vfs_unbusy_pages(struct buf *bp)
4601 runningbufwakeup(bp);
4602 if (!(bp->b_flags & B_VMIO))
4605 obj = bp->b_bufobj->bo_object;
4606 for (i = 0; i < bp->b_npages; i++) {
4608 if (m == bogus_page) {
4609 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4611 panic("vfs_unbusy_pages: page missing\n");
4613 if (buf_mapped(bp)) {
4614 BUF_CHECK_MAPPED(bp);
4615 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4616 bp->b_pages, bp->b_npages);
4618 BUF_CHECK_UNMAPPED(bp);
4622 vm_object_pip_wakeupn(obj, bp->b_npages);
4626 * vfs_page_set_valid:
4628 * Set the valid bits in a page based on the supplied offset. The
4629 * range is restricted to the buffer's size.
4631 * This routine is typically called after a read completes.
4634 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4639 * Compute the end offset, eoff, such that [off, eoff) does not span a
4640 * page boundary and eoff is not greater than the end of the buffer.
4641 * The end of the buffer, in this case, is our file EOF, not the
4642 * allocation size of the buffer.
4644 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4645 if (eoff > bp->b_offset + bp->b_bcount)
4646 eoff = bp->b_offset + bp->b_bcount;
4649 * Set valid range. This is typically the entire buffer and thus the
4653 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4657 * vfs_page_set_validclean:
4659 * Set the valid bits and clear the dirty bits in a page based on the
4660 * supplied offset. The range is restricted to the buffer's size.
4663 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4665 vm_ooffset_t soff, eoff;
4668 * Start and end offsets in buffer. eoff - soff may not cross a
4669 * page boundary or cross the end of the buffer. The end of the
4670 * buffer, in this case, is our file EOF, not the allocation size
4674 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4675 if (eoff > bp->b_offset + bp->b_bcount)
4676 eoff = bp->b_offset + bp->b_bcount;
4679 * Set valid range. This is typically the entire buffer and thus the
4683 vm_page_set_validclean(
4685 (vm_offset_t) (soff & PAGE_MASK),
4686 (vm_offset_t) (eoff - soff)
4692 * Acquire a shared busy on all pages in the buf.
4695 vfs_busy_pages_acquire(struct buf *bp)
4699 for (i = 0; i < bp->b_npages; i++)
4700 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4704 vfs_busy_pages_release(struct buf *bp)
4708 for (i = 0; i < bp->b_npages; i++)
4709 vm_page_sunbusy(bp->b_pages[i]);
4713 * This routine is called before a device strategy routine.
4714 * It is used to tell the VM system that paging I/O is in
4715 * progress, and treat the pages associated with the buffer
4716 * almost as being exclusive busy. Also the object paging_in_progress
4717 * flag is handled to make sure that the object doesn't become
4720 * Since I/O has not been initiated yet, certain buffer flags
4721 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4722 * and should be ignored.
4725 vfs_busy_pages(struct buf *bp, int clear_modify)
4733 if (!(bp->b_flags & B_VMIO))
4736 obj = bp->b_bufobj->bo_object;
4737 foff = bp->b_offset;
4738 KASSERT(bp->b_offset != NOOFFSET,
4739 ("vfs_busy_pages: no buffer offset"));
4740 if ((bp->b_flags & B_CLUSTER) == 0) {
4741 vm_object_pip_add(obj, bp->b_npages);
4742 vfs_busy_pages_acquire(bp);
4744 if (bp->b_bufsize != 0)
4745 vfs_setdirty_range(bp);
4747 for (i = 0; i < bp->b_npages; i++) {
4749 vm_page_assert_sbusied(m);
4752 * When readying a buffer for a read ( i.e
4753 * clear_modify == 0 ), it is important to do
4754 * bogus_page replacement for valid pages in
4755 * partially instantiated buffers. Partially
4756 * instantiated buffers can, in turn, occur when
4757 * reconstituting a buffer from its VM backing store
4758 * base. We only have to do this if B_CACHE is
4759 * clear ( which causes the I/O to occur in the
4760 * first place ). The replacement prevents the read
4761 * I/O from overwriting potentially dirty VM-backed
4762 * pages. XXX bogus page replacement is, uh, bogus.
4763 * It may not work properly with small-block devices.
4764 * We need to find a better way.
4767 pmap_remove_write(m);
4768 vfs_page_set_validclean(bp, foff, m);
4769 } else if (vm_page_all_valid(m) &&
4770 (bp->b_flags & B_CACHE) == 0) {
4771 bp->b_pages[i] = bogus_page;
4774 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4776 if (bogus && buf_mapped(bp)) {
4777 BUF_CHECK_MAPPED(bp);
4778 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4779 bp->b_pages, bp->b_npages);
4784 * vfs_bio_set_valid:
4786 * Set the range within the buffer to valid. The range is
4787 * relative to the beginning of the buffer, b_offset. Note that
4788 * b_offset itself may be offset from the beginning of the first
4792 vfs_bio_set_valid(struct buf *bp, int base, int size)
4797 if (!(bp->b_flags & B_VMIO))
4801 * Fixup base to be relative to beginning of first page.
4802 * Set initial n to be the maximum number of bytes in the
4803 * first page that can be validated.
4805 base += (bp->b_offset & PAGE_MASK);
4806 n = PAGE_SIZE - (base & PAGE_MASK);
4809 * Busy may not be strictly necessary here because the pages are
4810 * unlikely to be fully valid and the vnode lock will synchronize
4811 * their access via getpages. It is grabbed for consistency with
4812 * other page validation.
4814 vfs_busy_pages_acquire(bp);
4815 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4819 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4824 vfs_busy_pages_release(bp);
4830 * If the specified buffer is a non-VMIO buffer, clear the entire
4831 * buffer. If the specified buffer is a VMIO buffer, clear and
4832 * validate only the previously invalid portions of the buffer.
4833 * This routine essentially fakes an I/O, so we need to clear
4834 * BIO_ERROR and B_INVAL.
4836 * Note that while we only theoretically need to clear through b_bcount,
4837 * we go ahead and clear through b_bufsize.
4840 vfs_bio_clrbuf(struct buf *bp)
4842 int i, j, mask, sa, ea, slide;
4844 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4848 bp->b_flags &= ~B_INVAL;
4849 bp->b_ioflags &= ~BIO_ERROR;
4850 vfs_busy_pages_acquire(bp);
4851 sa = bp->b_offset & PAGE_MASK;
4853 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4854 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4855 ea = slide & PAGE_MASK;
4858 if (bp->b_pages[i] == bogus_page)
4861 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4862 if ((bp->b_pages[i]->valid & mask) == mask)
4864 if ((bp->b_pages[i]->valid & mask) == 0)
4865 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4867 for (; sa < ea; sa += DEV_BSIZE, j++) {
4868 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4869 pmap_zero_page_area(bp->b_pages[i],
4874 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4875 roundup2(ea - sa, DEV_BSIZE));
4877 vfs_busy_pages_release(bp);
4882 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4887 if (buf_mapped(bp)) {
4888 BUF_CHECK_MAPPED(bp);
4889 bzero(bp->b_data + base, size);
4891 BUF_CHECK_UNMAPPED(bp);
4892 n = PAGE_SIZE - (base & PAGE_MASK);
4893 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4897 pmap_zero_page_area(m, base & PAGE_MASK, n);
4906 * Update buffer flags based on I/O request parameters, optionally releasing the
4907 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4908 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4909 * I/O). Otherwise the buffer is released to the cache.
4912 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4915 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4916 ("buf %p non-VMIO noreuse", bp));
4918 if ((ioflag & IO_DIRECT) != 0)
4919 bp->b_flags |= B_DIRECT;
4920 if ((ioflag & IO_EXT) != 0)
4921 bp->b_xflags |= BX_ALTDATA;
4922 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4923 bp->b_flags |= B_RELBUF;
4924 if ((ioflag & IO_NOREUSE) != 0)
4925 bp->b_flags |= B_NOREUSE;
4933 vfs_bio_brelse(struct buf *bp, int ioflag)
4936 b_io_dismiss(bp, ioflag, true);
4940 vfs_bio_set_flags(struct buf *bp, int ioflag)
4943 b_io_dismiss(bp, ioflag, false);
4947 * vm_hold_load_pages and vm_hold_free_pages get pages into
4948 * a buffers address space. The pages are anonymous and are
4949 * not associated with a file object.
4952 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4958 BUF_CHECK_MAPPED(bp);
4960 to = round_page(to);
4961 from = round_page(from);
4962 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4963 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4964 KASSERT(to - from <= maxbcachebuf,
4965 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4966 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4968 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4970 * note: must allocate system pages since blocking here
4971 * could interfere with paging I/O, no matter which
4974 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4975 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4976 pmap_qenter(pg, &p, 1);
4977 bp->b_pages[index] = p;
4979 bp->b_npages = index;
4982 /* Return pages associated with this buf to the vm system */
4984 vm_hold_free_pages(struct buf *bp, int newbsize)
4988 int index, newnpages;
4990 BUF_CHECK_MAPPED(bp);
4992 from = round_page((vm_offset_t)bp->b_data + newbsize);
4993 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4994 if (bp->b_npages > newnpages)
4995 pmap_qremove(from, bp->b_npages - newnpages);
4996 for (index = newnpages; index < bp->b_npages; index++) {
4997 p = bp->b_pages[index];
4998 bp->b_pages[index] = NULL;
4999 vm_page_unwire_noq(p);
5002 bp->b_npages = newnpages;
5006 * Map an IO request into kernel virtual address space.
5008 * All requests are (re)mapped into kernel VA space.
5009 * Notice that we use b_bufsize for the size of the buffer
5010 * to be mapped. b_bcount might be modified by the driver.
5012 * Note that even if the caller determines that the address space should
5013 * be valid, a race or a smaller-file mapped into a larger space may
5014 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5015 * check the return value.
5017 * This function only works with pager buffers.
5020 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5025 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5026 prot = VM_PROT_READ;
5027 if (bp->b_iocmd == BIO_READ)
5028 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5029 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5030 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5033 bp->b_bufsize = len;
5034 bp->b_npages = pidx;
5035 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5036 if (mapbuf || !unmapped_buf_allowed) {
5037 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5038 bp->b_data = bp->b_kvabase + bp->b_offset;
5040 bp->b_data = unmapped_buf;
5045 * Free the io map PTEs associated with this IO operation.
5046 * We also invalidate the TLB entries and restore the original b_addr.
5048 * This function only works with pager buffers.
5051 vunmapbuf(struct buf *bp)
5055 npages = bp->b_npages;
5057 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5058 vm_page_unhold_pages(bp->b_pages, npages);
5060 bp->b_data = unmapped_buf;
5064 bdone(struct buf *bp)
5068 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5070 bp->b_flags |= B_DONE;
5076 bwait(struct buf *bp, u_char pri, const char *wchan)
5080 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5082 while ((bp->b_flags & B_DONE) == 0)
5083 msleep(bp, mtxp, pri, wchan, 0);
5088 bufsync(struct bufobj *bo, int waitfor)
5091 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5095 bufstrategy(struct bufobj *bo, struct buf *bp)
5101 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5102 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5103 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5104 i = VOP_STRATEGY(vp, bp);
5105 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5109 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5112 bufobj_init(struct bufobj *bo, void *private)
5114 static volatile int bufobj_cleanq;
5117 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5118 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5119 bo->bo_private = private;
5120 TAILQ_INIT(&bo->bo_clean.bv_hd);
5121 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5125 bufobj_wrefl(struct bufobj *bo)
5128 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5129 ASSERT_BO_WLOCKED(bo);
5134 bufobj_wref(struct bufobj *bo)
5137 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5144 bufobj_wdrop(struct bufobj *bo)
5147 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5149 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5150 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5151 bo->bo_flag &= ~BO_WWAIT;
5152 wakeup(&bo->bo_numoutput);
5158 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5162 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5163 ASSERT_BO_WLOCKED(bo);
5165 while (bo->bo_numoutput) {
5166 bo->bo_flag |= BO_WWAIT;
5167 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5168 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5176 * Set bio_data or bio_ma for struct bio from the struct buf.
5179 bdata2bio(struct buf *bp, struct bio *bip)
5182 if (!buf_mapped(bp)) {
5183 KASSERT(unmapped_buf_allowed, ("unmapped"));
5184 bip->bio_ma = bp->b_pages;
5185 bip->bio_ma_n = bp->b_npages;
5186 bip->bio_data = unmapped_buf;
5187 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5188 bip->bio_flags |= BIO_UNMAPPED;
5189 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5190 PAGE_SIZE == bp->b_npages,
5191 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5192 (long long)bip->bio_length, bip->bio_ma_n));
5194 bip->bio_data = bp->b_data;
5200 * The MIPS pmap code currently doesn't handle aliased pages.
5201 * The VIPT caches may not handle page aliasing themselves, leading
5202 * to data corruption.
5204 * As such, this code makes a system extremely unhappy if said
5205 * system doesn't support unaliasing the above situation in hardware.
5206 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5207 * this feature at build time, so it has to be handled in software.
5209 * Once the MIPS pmap/cache code grows to support this function on
5210 * earlier chips, it should be flipped back off.
5213 static int buf_pager_relbuf = 1;
5215 static int buf_pager_relbuf = 0;
5217 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5218 &buf_pager_relbuf, 0,
5219 "Make buffer pager release buffers after reading");
5222 * The buffer pager. It uses buffer reads to validate pages.
5224 * In contrast to the generic local pager from vm/vnode_pager.c, this
5225 * pager correctly and easily handles volumes where the underlying
5226 * device block size is greater than the machine page size. The
5227 * buffer cache transparently extends the requested page run to be
5228 * aligned at the block boundary, and does the necessary bogus page
5229 * replacements in the addends to avoid obliterating already valid
5232 * The only non-trivial issue is that the exclusive busy state for
5233 * pages, which is assumed by the vm_pager_getpages() interface, is
5234 * incompatible with the VMIO buffer cache's desire to share-busy the
5235 * pages. This function performs a trivial downgrade of the pages'
5236 * state before reading buffers, and a less trivial upgrade from the
5237 * shared-busy to excl-busy state after the read.
5240 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5241 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5242 vbg_get_blksize_t get_blksize)
5249 vm_ooffset_t la, lb, poff, poffe;
5251 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5254 object = vp->v_object;
5257 la = IDX_TO_OFF(ma[count - 1]->pindex);
5258 if (la >= object->un_pager.vnp.vnp_size)
5259 return (VM_PAGER_BAD);
5262 * Change the meaning of la from where the last requested page starts
5263 * to where it ends, because that's the end of the requested region
5264 * and the start of the potential read-ahead region.
5267 lpart = la > object->un_pager.vnp.vnp_size;
5268 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5271 return (VM_PAGER_ERROR);
5274 * Calculate read-ahead, behind and total pages.
5277 lb = IDX_TO_OFF(ma[0]->pindex);
5278 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5280 if (rbehind != NULL)
5282 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5283 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5284 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5289 VM_CNT_INC(v_vnodein);
5290 VM_CNT_ADD(v_vnodepgsin, pgsin);
5292 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5293 != 0) ? GB_UNMAPPED : 0;
5295 for (i = 0; i < count; i++) {
5296 if (ma[i] != bogus_page)
5297 vm_page_busy_downgrade(ma[i]);
5301 for (i = 0; i < count; i++) {
5303 if (m == bogus_page)
5307 * Pages are shared busy and the object lock is not
5308 * owned, which together allow for the pages'
5309 * invalidation. The racy test for validity avoids
5310 * useless creation of the buffer for the most typical
5311 * case when invalidation is not used in redo or for
5312 * parallel read. The shared->excl upgrade loop at
5313 * the end of the function catches the race in a
5314 * reliable way (protected by the object lock).
5316 if (vm_page_all_valid(m))
5319 poff = IDX_TO_OFF(m->pindex);
5320 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5321 for (; poff < poffe; poff += bsize) {
5322 lbn = get_lblkno(vp, poff);
5327 error = get_blksize(vp, lbn, &bsize);
5329 error = bread_gb(vp, lbn, bsize,
5330 curthread->td_ucred, br_flags, &bp);
5333 if (bp->b_rcred == curthread->td_ucred) {
5334 crfree(bp->b_rcred);
5335 bp->b_rcred = NOCRED;
5337 if (LIST_EMPTY(&bp->b_dep)) {
5339 * Invalidation clears m->valid, but
5340 * may leave B_CACHE flag if the
5341 * buffer existed at the invalidation
5342 * time. In this case, recycle the
5343 * buffer to do real read on next
5344 * bread() after redo.
5346 * Otherwise B_RELBUF is not strictly
5347 * necessary, enable to reduce buf
5350 if (buf_pager_relbuf ||
5351 !vm_page_all_valid(m))
5352 bp->b_flags |= B_RELBUF;
5354 bp->b_flags &= ~B_NOCACHE;
5360 KASSERT(1 /* racy, enable for debugging */ ||
5361 vm_page_all_valid(m) || i == count - 1,
5362 ("buf %d %p invalid", i, m));
5363 if (i == count - 1 && lpart) {
5364 if (!vm_page_none_valid(m) &&
5365 !vm_page_all_valid(m))
5366 vm_page_zero_invalid(m, TRUE);
5373 for (i = 0; i < count; i++) {
5374 if (ma[i] == bogus_page)
5376 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5377 vm_page_sunbusy(ma[i]);
5378 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5383 * Since the pages were only sbusy while neither the
5384 * buffer nor the object lock was held by us, or
5385 * reallocated while vm_page_grab() slept for busy
5386 * relinguish, they could have been invalidated.
5387 * Recheck the valid bits and re-read as needed.
5389 * Note that the last page is made fully valid in the
5390 * read loop, and partial validity for the page at
5391 * index count - 1 could mean that the page was
5392 * invalidated or removed, so we must restart for
5395 if (!vm_page_all_valid(ma[i]))
5398 if (redo && error == 0)
5400 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5403 #include "opt_ddb.h"
5405 #include <ddb/ddb.h>
5407 /* DDB command to show buffer data */
5408 DB_SHOW_COMMAND(buffer, db_show_buffer)
5411 struct buf *bp = (struct buf *)addr;
5412 #ifdef FULL_BUF_TRACKING
5417 db_printf("usage: show buffer <addr>\n");
5421 db_printf("buf at %p\n", bp);
5422 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5423 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5424 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5425 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5426 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5427 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5429 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5430 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5431 "b_vp = %p, b_dep = %p\n",
5432 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5433 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5434 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5435 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5436 bp->b_kvabase, bp->b_kvasize);
5439 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5440 for (i = 0; i < bp->b_npages; i++) {
5444 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5446 (u_long)VM_PAGE_TO_PHYS(m));
5448 db_printf("( ??? )");
5449 if ((i + 1) < bp->b_npages)
5454 BUF_LOCKPRINTINFO(bp);
5455 #if defined(FULL_BUF_TRACKING)
5456 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5458 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5459 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5460 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5462 db_printf(" %2u: %s\n", j,
5463 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5465 #elif defined(BUF_TRACKING)
5466 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5471 DB_SHOW_COMMAND(bufqueues, bufqueues)
5473 struct bufdomain *bd;
5478 db_printf("bqempty: %d\n", bqempty.bq_len);
5480 for (i = 0; i < buf_domains; i++) {
5482 db_printf("Buf domain %d\n", i);
5483 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5484 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5485 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5487 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5488 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5489 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5490 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5491 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5493 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5494 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5495 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5496 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5499 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5500 total += bp->b_bufsize;
5501 db_printf("\tcleanq count\t%d (%ld)\n",
5502 bd->bd_cleanq->bq_len, total);
5504 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5505 total += bp->b_bufsize;
5506 db_printf("\tdirtyq count\t%d (%ld)\n",
5507 bd->bd_dirtyq.bq_len, total);
5508 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5509 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5510 db_printf("\tCPU ");
5511 for (j = 0; j <= mp_maxid; j++)
5512 db_printf("%d, ", bd->bd_subq[j].bq_len);
5516 for (j = 0; j < nbuf; j++) {
5518 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5520 total += bp->b_bufsize;
5523 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5526 for (j = 0; j < nbuf; j++) {
5528 if (bp->b_domain == i) {
5530 total += bp->b_bufsize;
5533 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5537 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5542 for (i = 0; i < nbuf; i++) {
5544 if (BUF_ISLOCKED(bp)) {
5545 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5553 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5559 db_printf("usage: show vnodebufs <addr>\n");
5562 vp = (struct vnode *)addr;
5563 db_printf("Clean buffers:\n");
5564 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5565 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5568 db_printf("Dirty buffers:\n");
5569 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5570 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5575 DB_COMMAND(countfreebufs, db_coundfreebufs)
5578 int i, used = 0, nfree = 0;
5581 db_printf("usage: countfreebufs\n");
5585 for (i = 0; i < nbuf; i++) {
5587 if (bp->b_qindex == QUEUE_EMPTY)
5593 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5595 db_printf("numfreebuffers is %d\n", numfreebuffers);