2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2004 Poul-Henning Kamp
5 * Copyright (c) 1994,1997 John S. Dyson
6 * Copyright (c) 2013 The FreeBSD Foundation
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * this file contains a new buffer I/O scheme implementing a coherent
36 * VM object and buffer cache scheme. Pains have been taken to make
37 * sure that the performance degradation associated with schemes such
38 * as this is not realized.
40 * Author: John S. Dyson
41 * Significant help during the development and debugging phases
42 * had been provided by David Greenman, also of the FreeBSD core team.
44 * see man buf(9) for more info.
47 #include <sys/cdefs.h>
48 __FBSDID("$FreeBSD$");
50 #include <sys/param.h>
51 #include <sys/systm.h>
54 #include <sys/bitset.h>
56 #include <sys/counter.h>
58 #include <sys/devicestat.h>
59 #include <sys/eventhandler.h>
62 #include <sys/limits.h>
64 #include <sys/malloc.h>
65 #include <sys/mount.h>
66 #include <sys/mutex.h>
67 #include <sys/kernel.h>
68 #include <sys/kthread.h>
70 #include <sys/racct.h>
71 #include <sys/refcount.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
75 #include <sys/sysctl.h>
76 #include <sys/syscallsubr.h>
78 #include <sys/vmmeter.h>
79 #include <sys/vnode.h>
80 #include <sys/watchdog.h>
81 #include <geom/geom.h>
83 #include <vm/vm_param.h>
84 #include <vm/vm_kern.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_map.h>
91 #include <vm/swap_pager.h>
93 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
95 struct bio_ops bioops; /* I/O operation notification */
97 struct buf_ops buf_ops_bio = {
98 .bop_name = "buf_ops_bio",
99 .bop_write = bufwrite,
100 .bop_strategy = bufstrategy,
102 .bop_bdflush = bufbdflush,
106 struct mtx_padalign bq_lock;
107 TAILQ_HEAD(, buf) bq_queue;
109 uint16_t bq_subqueue;
111 } __aligned(CACHE_LINE_SIZE);
113 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock)
114 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq)))
115 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq)))
116 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED)
119 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */
120 struct bufqueue bd_dirtyq;
121 struct bufqueue *bd_cleanq;
122 struct mtx_padalign bd_run_lock;
127 long bd_bufspacethresh;
128 int bd_hifreebuffers;
129 int bd_lofreebuffers;
130 int bd_hidirtybuffers;
131 int bd_lodirtybuffers;
132 int bd_dirtybufthresh;
137 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers;
138 int __aligned(CACHE_LINE_SIZE) bd_running;
139 long __aligned(CACHE_LINE_SIZE) bd_bufspace;
140 int __aligned(CACHE_LINE_SIZE) bd_freebuffers;
141 } __aligned(CACHE_LINE_SIZE);
143 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock)
144 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd)))
145 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd)))
146 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED)
147 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock)
148 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd)))
149 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd)))
150 #define BD_DOMAIN(bd) (bd - bdomain)
152 static char *buf; /* buffer header pool */
156 return ((struct buf *)(buf + (sizeof(struct buf) +
157 sizeof(vm_page_t) * atop(maxbcachebuf)) * i));
160 caddr_t __read_mostly unmapped_buf;
162 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
163 struct proc *bufdaemonproc;
165 static void vm_hold_free_pages(struct buf *bp, int newbsize);
166 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
168 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
169 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
171 static void vfs_clean_pages_dirty_buf(struct buf *bp);
172 static void vfs_setdirty_range(struct buf *bp);
173 static void vfs_vmio_invalidate(struct buf *bp);
174 static void vfs_vmio_truncate(struct buf *bp, int npages);
175 static void vfs_vmio_extend(struct buf *bp, int npages, int size);
176 static int vfs_bio_clcheck(struct vnode *vp, int size,
177 daddr_t lblkno, daddr_t blkno);
178 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int,
179 void (*)(struct buf *));
180 static int buf_flush(struct vnode *vp, struct bufdomain *, int);
181 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int);
182 static void buf_daemon(void);
183 static __inline void bd_wakeup(void);
184 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
185 static void bufkva_reclaim(vmem_t *, int);
186 static void bufkva_free(struct buf *);
187 static int buf_import(void *, void **, int, int, int);
188 static void buf_release(void *, void **, int);
189 static void maxbcachebuf_adjust(void);
190 static inline struct bufdomain *bufdomain(struct buf *);
191 static void bq_remove(struct bufqueue *bq, struct buf *bp);
192 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock);
193 static int buf_recycle(struct bufdomain *, bool kva);
194 static void bq_init(struct bufqueue *bq, int qindex, int cpu,
195 const char *lockname);
196 static void bd_init(struct bufdomain *bd);
197 static int bd_flushall(struct bufdomain *bd);
198 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS);
199 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS);
201 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
202 int vmiodirenable = TRUE;
203 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
204 "Use the VM system for directory writes");
205 long runningbufspace;
206 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
207 "Amount of presently outstanding async buffer io");
208 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
209 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers");
210 static counter_u64_t bufkvaspace;
211 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace,
212 "Kernel virtual memory used for buffers");
213 static long maxbufspace;
214 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace,
215 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace,
216 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L",
217 "Maximum allowed value of bufspace (including metadata)");
218 static long bufmallocspace;
219 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
220 "Amount of malloced memory for buffers");
221 static long maxbufmallocspace;
222 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace,
223 0, "Maximum amount of malloced memory for buffers");
224 static long lobufspace;
225 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace,
226 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace,
227 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L",
228 "Minimum amount of buffers we want to have");
230 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace,
231 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace,
232 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L",
233 "Maximum allowed value of bufspace (excluding metadata)");
235 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh,
236 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh,
237 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L",
238 "Bufspace consumed before waking the daemon to free some");
239 static counter_u64_t buffreekvacnt;
240 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt,
241 "Number of times we have freed the KVA space from some buffer");
242 static counter_u64_t bufdefragcnt;
243 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt,
244 "Number of times we have had to repeat buffer allocation to defragment");
245 static long lorunningspace;
246 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
247 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
248 "Minimum preferred space used for in-progress I/O");
249 static long hirunningspace;
250 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
251 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
252 "Maximum amount of space to use for in-progress I/O");
253 int dirtybufferflushes;
254 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
255 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
257 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
258 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
259 int altbufferflushes;
260 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS,
261 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers");
262 static int recursiveflushes;
263 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS,
264 &recursiveflushes, 0, "Number of flushes skipped due to being recursive");
265 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS);
266 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers,
267 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I",
268 "Number of buffers that are dirty (has unwritten changes) at the moment");
269 static int lodirtybuffers;
270 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers,
271 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers,
272 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I",
273 "How many buffers we want to have free before bufdaemon can sleep");
274 static int hidirtybuffers;
275 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers,
276 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers,
277 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I",
278 "When the number of dirty buffers is considered severe");
280 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh,
281 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh,
282 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I",
283 "Number of bdwrite to bawrite conversions to clear dirty buffers");
284 static int numfreebuffers;
285 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
286 "Number of free buffers");
287 static int lofreebuffers;
288 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers,
289 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers,
290 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I",
291 "Target number of free buffers");
292 static int hifreebuffers;
293 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers,
294 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers,
295 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I",
296 "Threshold for clean buffer recycling");
297 static counter_u64_t getnewbufcalls;
298 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD,
299 &getnewbufcalls, "Number of calls to getnewbuf");
300 static counter_u64_t getnewbufrestarts;
301 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD,
303 "Number of times getnewbuf has had to restart a buffer acquisition");
304 static counter_u64_t mappingrestarts;
305 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD,
307 "Number of times getblk has had to restart a buffer mapping for "
309 static counter_u64_t numbufallocfails;
310 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW,
311 &numbufallocfails, "Number of times buffer allocations failed");
312 static int flushbufqtarget = 100;
313 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
314 "Amount of work to do in flushbufqueues when helping bufdaemon");
315 static counter_u64_t notbufdflushes;
316 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes,
317 "Number of dirty buffer flushes done by the bufdaemon helpers");
318 static long barrierwrites;
319 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS,
320 &barrierwrites, 0, "Number of barrier writes");
321 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
322 &unmapped_buf_allowed, 0,
323 "Permit the use of the unmapped i/o");
324 int maxbcachebuf = MAXBCACHEBUF;
325 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0,
326 "Maximum size of a buffer cache block");
329 * This lock synchronizes access to bd_request.
331 static struct mtx_padalign __exclusive_cache_line bdlock;
334 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
335 * waitrunningbufspace().
337 static struct mtx_padalign __exclusive_cache_line rbreqlock;
340 * Lock that protects bdirtywait.
342 static struct mtx_padalign __exclusive_cache_line bdirtylock;
345 * bufdaemon shutdown request and sleep channel.
347 static bool bd_shutdown;
350 * Wakeup point for bufdaemon, as well as indicator of whether it is already
351 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
354 static int bd_request;
357 * Request for the buf daemon to write more buffers than is indicated by
358 * lodirtybuf. This may be necessary to push out excess dependencies or
359 * defragment the address space where a simple count of the number of dirty
360 * buffers is insufficient to characterize the demand for flushing them.
362 static int bd_speedupreq;
365 * Synchronization (sleep/wakeup) variable for active buffer space requests.
366 * Set when wait starts, cleared prior to wakeup().
367 * Used in runningbufwakeup() and waitrunningbufspace().
369 static int runningbufreq;
372 * Synchronization for bwillwrite() waiters.
374 static int bdirtywait;
377 * Definitions for the buffer free lists.
379 #define QUEUE_NONE 0 /* on no queue */
380 #define QUEUE_EMPTY 1 /* empty buffer headers */
381 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
382 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */
383 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */
385 /* Maximum number of buffer domains. */
386 #define BUF_DOMAINS 8
388 struct bufdomainset bdlodirty; /* Domains > lodirty */
389 struct bufdomainset bdhidirty; /* Domains > hidirty */
391 /* Configured number of clean queues. */
392 static int __read_mostly buf_domains;
394 BITSET_DEFINE(bufdomainset, BUF_DOMAINS);
395 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS];
396 struct bufqueue __exclusive_cache_line bqempty;
399 * per-cpu empty buffer cache.
404 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 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1905 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1906 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1907 for (i = 0; i <= mp_maxid; i++)
1908 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1909 "bufq clean subqueue lock");
1910 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1916 * Removes a buffer from the free list, must be called with the
1917 * correct qlock held.
1920 bq_remove(struct bufqueue *bq, struct buf *bp)
1923 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1924 bp, bp->b_vp, bp->b_flags);
1925 KASSERT(bp->b_qindex != QUEUE_NONE,
1926 ("bq_remove: buffer %p not on a queue.", bp));
1927 KASSERT(bufqueue(bp) == bq,
1928 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1930 BQ_ASSERT_LOCKED(bq);
1931 if (bp->b_qindex != QUEUE_EMPTY) {
1932 BUF_ASSERT_XLOCKED(bp);
1934 KASSERT(bq->bq_len >= 1,
1935 ("queue %d underflow", bp->b_qindex));
1936 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1938 bp->b_qindex = QUEUE_NONE;
1939 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1943 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1947 BQ_ASSERT_LOCKED(bq);
1948 if (bq != bd->bd_cleanq) {
1950 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1951 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1952 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1954 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1956 bd->bd_cleanq->bq_len += bq->bq_len;
1959 if (bd->bd_wanted) {
1961 wakeup(&bd->bd_wanted);
1963 if (bq != bd->bd_cleanq)
1968 bd_flushall(struct bufdomain *bd)
1970 struct bufqueue *bq;
1974 if (bd->bd_lim == 0)
1977 for (i = 0; i <= mp_maxid; i++) {
1978 bq = &bd->bd_subq[i];
1979 if (bq->bq_len == 0)
1991 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1993 struct bufdomain *bd;
1995 if (bp->b_qindex != QUEUE_NONE)
1996 panic("bq_insert: free buffer %p onto another queue?", bp);
1999 if (bp->b_flags & B_AGE) {
2000 /* Place this buf directly on the real queue. */
2001 if (bq->bq_index == QUEUE_CLEAN)
2004 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2007 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2009 bp->b_flags &= ~(B_AGE | B_REUSE);
2011 bp->b_qindex = bq->bq_index;
2012 bp->b_subqueue = bq->bq_subqueue;
2015 * Unlock before we notify so that we don't wakeup a waiter that
2016 * fails a trylock on the buf and sleeps again.
2021 if (bp->b_qindex == QUEUE_CLEAN) {
2023 * Flush the per-cpu queue and notify any waiters.
2025 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2026 bq->bq_len >= bd->bd_lim))
2035 * Free the kva allocation for a buffer.
2039 bufkva_free(struct buf *bp)
2043 if (bp->b_kvasize == 0) {
2044 KASSERT(bp->b_kvabase == unmapped_buf &&
2045 bp->b_data == unmapped_buf,
2046 ("Leaked KVA space on %p", bp));
2047 } else if (buf_mapped(bp))
2048 BUF_CHECK_MAPPED(bp);
2050 BUF_CHECK_UNMAPPED(bp);
2052 if (bp->b_kvasize == 0)
2055 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2056 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2057 counter_u64_add(buffreekvacnt, 1);
2058 bp->b_data = bp->b_kvabase = unmapped_buf;
2065 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2068 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2073 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2074 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2075 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2076 KASSERT(maxsize <= maxbcachebuf,
2077 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2082 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2085 * Buffer map is too fragmented. Request the caller
2086 * to defragment the map.
2090 bp->b_kvabase = (caddr_t)addr;
2091 bp->b_kvasize = maxsize;
2092 counter_u64_add(bufkvaspace, bp->b_kvasize);
2093 if ((gbflags & GB_UNMAPPED) != 0) {
2094 bp->b_data = unmapped_buf;
2095 BUF_CHECK_UNMAPPED(bp);
2097 bp->b_data = bp->b_kvabase;
2098 BUF_CHECK_MAPPED(bp);
2106 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2107 * callback that fires to avoid returning failure.
2110 bufkva_reclaim(vmem_t *vmem, int flags)
2117 for (i = 0; i < 5; i++) {
2118 for (q = 0; q < buf_domains; q++)
2119 if (buf_recycle(&bdomain[q], true) != 0)
2128 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2129 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2130 * the buffer is valid and we do not have to do anything.
2133 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2134 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2142 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2143 if (inmem(vp, *rablkno))
2145 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2146 if ((rabp->b_flags & B_CACHE) != 0) {
2153 racct_add_buf(curproc, rabp, 0);
2154 PROC_UNLOCK(curproc);
2157 td->td_ru.ru_inblock++;
2158 rabp->b_flags |= B_ASYNC;
2159 rabp->b_flags &= ~B_INVAL;
2160 if ((flags & GB_CKHASH) != 0) {
2161 rabp->b_flags |= B_CKHASH;
2162 rabp->b_ckhashcalc = ckhashfunc;
2164 rabp->b_ioflags &= ~BIO_ERROR;
2165 rabp->b_iocmd = BIO_READ;
2166 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2167 rabp->b_rcred = crhold(cred);
2168 vfs_busy_pages(rabp, 0);
2170 rabp->b_iooffset = dbtob(rabp->b_blkno);
2176 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2178 * Get a buffer with the specified data. Look in the cache first. We
2179 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2180 * is set, the buffer is valid and we do not have to do anything, see
2181 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2183 * Always return a NULL buffer pointer (in bpp) when returning an error.
2185 * The blkno parameter is the logical block being requested. Normally
2186 * the mapping of logical block number to disk block address is done
2187 * by calling VOP_BMAP(). However, if the mapping is already known, the
2188 * disk block address can be passed using the dblkno parameter. If the
2189 * disk block address is not known, then the same value should be passed
2190 * for blkno and dblkno.
2193 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2194 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2195 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2199 int error, readwait, rv;
2201 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2204 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2207 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2212 KASSERT(blkno == bp->b_lblkno,
2213 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2214 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2215 flags &= ~GB_NOSPARSE;
2219 * If not found in cache, do some I/O
2222 if ((bp->b_flags & B_CACHE) == 0) {
2225 PROC_LOCK(td->td_proc);
2226 racct_add_buf(td->td_proc, bp, 0);
2227 PROC_UNLOCK(td->td_proc);
2230 td->td_ru.ru_inblock++;
2231 bp->b_iocmd = BIO_READ;
2232 bp->b_flags &= ~B_INVAL;
2233 if ((flags & GB_CKHASH) != 0) {
2234 bp->b_flags |= B_CKHASH;
2235 bp->b_ckhashcalc = ckhashfunc;
2237 if ((flags & GB_CVTENXIO) != 0)
2238 bp->b_xflags |= BX_CVTENXIO;
2239 bp->b_ioflags &= ~BIO_ERROR;
2240 if (bp->b_rcred == NOCRED && cred != NOCRED)
2241 bp->b_rcred = crhold(cred);
2242 vfs_busy_pages(bp, 0);
2243 bp->b_iooffset = dbtob(bp->b_blkno);
2249 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2251 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2265 * Write, release buffer on completion. (Done by iodone
2266 * if async). Do not bother writing anything if the buffer
2269 * Note that we set B_CACHE here, indicating that buffer is
2270 * fully valid and thus cacheable. This is true even of NFS
2271 * now so we set it generally. This could be set either here
2272 * or in biodone() since the I/O is synchronous. We put it
2276 bufwrite(struct buf *bp)
2283 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2284 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2285 bp->b_flags |= B_INVAL | B_RELBUF;
2286 bp->b_flags &= ~B_CACHE;
2290 if (bp->b_flags & B_INVAL) {
2295 if (bp->b_flags & B_BARRIER)
2296 atomic_add_long(&barrierwrites, 1);
2298 oldflags = bp->b_flags;
2300 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2301 ("FFS background buffer should not get here %p", bp));
2305 vp_md = vp->v_vflag & VV_MD;
2310 * Mark the buffer clean. Increment the bufobj write count
2311 * before bundirty() call, to prevent other thread from seeing
2312 * empty dirty list and zero counter for writes in progress,
2313 * falsely indicating that the bufobj is clean.
2315 bufobj_wref(bp->b_bufobj);
2318 bp->b_flags &= ~B_DONE;
2319 bp->b_ioflags &= ~BIO_ERROR;
2320 bp->b_flags |= B_CACHE;
2321 bp->b_iocmd = BIO_WRITE;
2323 vfs_busy_pages(bp, 1);
2326 * Normal bwrites pipeline writes
2328 bp->b_runningbufspace = bp->b_bufsize;
2329 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2334 racct_add_buf(curproc, bp, 1);
2335 PROC_UNLOCK(curproc);
2338 curthread->td_ru.ru_oublock++;
2339 if (oldflags & B_ASYNC)
2341 bp->b_iooffset = dbtob(bp->b_blkno);
2342 buf_track(bp, __func__);
2345 if ((oldflags & B_ASYNC) == 0) {
2346 int rtval = bufwait(bp);
2349 } else if (space > hirunningspace) {
2351 * don't allow the async write to saturate the I/O
2352 * system. We will not deadlock here because
2353 * we are blocking waiting for I/O that is already in-progress
2354 * to complete. We do not block here if it is the update
2355 * or syncer daemon trying to clean up as that can lead
2358 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2359 waitrunningbufspace();
2366 bufbdflush(struct bufobj *bo, struct buf *bp)
2369 struct bufdomain *bd;
2371 bd = &bdomain[bo->bo_domain];
2372 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2373 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2375 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2378 * Try to find a buffer to flush.
2380 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2381 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2383 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2386 panic("bdwrite: found ourselves");
2388 /* Don't countdeps with the bo lock held. */
2389 if (buf_countdeps(nbp, 0)) {
2394 if (nbp->b_flags & B_CLUSTEROK) {
2395 vfs_bio_awrite(nbp);
2400 dirtybufferflushes++;
2409 * Delayed write. (Buffer is marked dirty). Do not bother writing
2410 * anything if the buffer is marked invalid.
2412 * Note that since the buffer must be completely valid, we can safely
2413 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2414 * biodone() in order to prevent getblk from writing the buffer
2415 * out synchronously.
2418 bdwrite(struct buf *bp)
2420 struct thread *td = curthread;
2424 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2425 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2426 KASSERT((bp->b_flags & B_BARRIER) == 0,
2427 ("Barrier request in delayed write %p", bp));
2429 if (bp->b_flags & B_INVAL) {
2435 * If we have too many dirty buffers, don't create any more.
2436 * If we are wildly over our limit, then force a complete
2437 * cleanup. Otherwise, just keep the situation from getting
2438 * out of control. Note that we have to avoid a recursive
2439 * disaster and not try to clean up after our own cleanup!
2443 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2444 td->td_pflags |= TDP_INBDFLUSH;
2446 td->td_pflags &= ~TDP_INBDFLUSH;
2452 * Set B_CACHE, indicating that the buffer is fully valid. This is
2453 * true even of NFS now.
2455 bp->b_flags |= B_CACHE;
2458 * This bmap keeps the system from needing to do the bmap later,
2459 * perhaps when the system is attempting to do a sync. Since it
2460 * is likely that the indirect block -- or whatever other datastructure
2461 * that the filesystem needs is still in memory now, it is a good
2462 * thing to do this. Note also, that if the pageout daemon is
2463 * requesting a sync -- there might not be enough memory to do
2464 * the bmap then... So, this is important to do.
2466 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2467 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2470 buf_track(bp, __func__);
2473 * Set the *dirty* buffer range based upon the VM system dirty
2476 * Mark the buffer pages as clean. We need to do this here to
2477 * satisfy the vnode_pager and the pageout daemon, so that it
2478 * thinks that the pages have been "cleaned". Note that since
2479 * the pages are in a delayed write buffer -- the VFS layer
2480 * "will" see that the pages get written out on the next sync,
2481 * or perhaps the cluster will be completed.
2483 vfs_clean_pages_dirty_buf(bp);
2487 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2488 * due to the softdep code.
2495 * Turn buffer into delayed write request. We must clear BIO_READ and
2496 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2497 * itself to properly update it in the dirty/clean lists. We mark it
2498 * B_DONE to ensure that any asynchronization of the buffer properly
2499 * clears B_DONE ( else a panic will occur later ).
2501 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2502 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2503 * should only be called if the buffer is known-good.
2505 * Since the buffer is not on a queue, we do not update the numfreebuffers
2508 * The buffer must be on QUEUE_NONE.
2511 bdirty(struct buf *bp)
2514 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2515 bp, bp->b_vp, bp->b_flags);
2516 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2517 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2518 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2519 bp->b_flags &= ~(B_RELBUF);
2520 bp->b_iocmd = BIO_WRITE;
2522 if ((bp->b_flags & B_DELWRI) == 0) {
2523 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2532 * Clear B_DELWRI for buffer.
2534 * Since the buffer is not on a queue, we do not update the numfreebuffers
2537 * The buffer must be on QUEUE_NONE.
2541 bundirty(struct buf *bp)
2544 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2545 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2546 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2547 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2549 if (bp->b_flags & B_DELWRI) {
2550 bp->b_flags &= ~B_DELWRI;
2555 * Since it is now being written, we can clear its deferred write flag.
2557 bp->b_flags &= ~B_DEFERRED;
2563 * Asynchronous write. Start output on a buffer, but do not wait for
2564 * it to complete. The buffer is released when the output completes.
2566 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2567 * B_INVAL buffers. Not us.
2570 bawrite(struct buf *bp)
2573 bp->b_flags |= B_ASYNC;
2580 * Asynchronous barrier write. Start output on a buffer, but do not
2581 * wait for it to complete. Place a write barrier after this write so
2582 * that this buffer and all buffers written before it are committed to
2583 * the disk before any buffers written after this write are committed
2584 * to the disk. The buffer is released when the output completes.
2587 babarrierwrite(struct buf *bp)
2590 bp->b_flags |= B_ASYNC | B_BARRIER;
2597 * Synchronous barrier write. Start output on a buffer and wait for
2598 * it to complete. Place a write barrier after this write so that
2599 * this buffer and all buffers written before it are committed to
2600 * the disk before any buffers written after this write are committed
2601 * to the disk. The buffer is released when the output completes.
2604 bbarrierwrite(struct buf *bp)
2607 bp->b_flags |= B_BARRIER;
2608 return (bwrite(bp));
2614 * Called prior to the locking of any vnodes when we are expecting to
2615 * write. We do not want to starve the buffer cache with too many
2616 * dirty buffers so we block here. By blocking prior to the locking
2617 * of any vnodes we attempt to avoid the situation where a locked vnode
2618 * prevents the various system daemons from flushing related buffers.
2624 if (buf_dirty_count_severe()) {
2625 mtx_lock(&bdirtylock);
2626 while (buf_dirty_count_severe()) {
2628 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2631 mtx_unlock(&bdirtylock);
2636 * Return true if we have too many dirty buffers.
2639 buf_dirty_count_severe(void)
2642 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2648 * Release a busy buffer and, if requested, free its resources. The
2649 * buffer will be stashed in the appropriate bufqueue[] allowing it
2650 * to be accessed later as a cache entity or reused for other purposes.
2653 brelse(struct buf *bp)
2655 struct mount *v_mnt;
2659 * Many functions erroneously call brelse with a NULL bp under rare
2660 * error conditions. Simply return when called with a NULL bp.
2664 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2665 bp, bp->b_vp, bp->b_flags);
2666 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2667 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2668 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2669 ("brelse: non-VMIO buffer marked NOREUSE"));
2671 if (BUF_LOCKRECURSED(bp)) {
2673 * Do not process, in particular, do not handle the
2674 * B_INVAL/B_RELBUF and do not release to free list.
2680 if (bp->b_flags & B_MANAGED) {
2685 if (LIST_EMPTY(&bp->b_dep)) {
2686 bp->b_flags &= ~B_IOSTARTED;
2688 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2689 ("brelse: SU io not finished bp %p", bp));
2692 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2693 BO_LOCK(bp->b_bufobj);
2694 bp->b_vflags &= ~BV_BKGRDERR;
2695 BO_UNLOCK(bp->b_bufobj);
2699 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2700 (bp->b_flags & B_INVALONERR)) {
2702 * Forced invalidation of dirty buffer contents, to be used
2703 * after a failed write in the rare case that the loss of the
2704 * contents is acceptable. The buffer is invalidated and
2707 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2708 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2711 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2712 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2713 !(bp->b_flags & B_INVAL)) {
2715 * Failed write, redirty. All errors except ENXIO (which
2716 * means the device is gone) are treated as being
2719 * XXX Treating EIO as transient is not correct; the
2720 * contract with the local storage device drivers is that
2721 * they will only return EIO once the I/O is no longer
2722 * retriable. Network I/O also respects this through the
2723 * guarantees of TCP and/or the internal retries of NFS.
2724 * ENOMEM might be transient, but we also have no way of
2725 * knowing when its ok to retry/reschedule. In general,
2726 * this entire case should be made obsolete through better
2727 * error handling/recovery and resource scheduling.
2729 * Do this also for buffers that failed with ENXIO, but have
2730 * non-empty dependencies - the soft updates code might need
2731 * to access the buffer to untangle them.
2733 * Must clear BIO_ERROR to prevent pages from being scrapped.
2735 bp->b_ioflags &= ~BIO_ERROR;
2737 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2738 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2740 * Either a failed read I/O, or we were asked to free or not
2741 * cache the buffer, or we failed to write to a device that's
2742 * no longer present.
2744 bp->b_flags |= B_INVAL;
2745 if (!LIST_EMPTY(&bp->b_dep))
2747 if (bp->b_flags & B_DELWRI)
2749 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2750 if ((bp->b_flags & B_VMIO) == 0) {
2758 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2759 * is called with B_DELWRI set, the underlying pages may wind up
2760 * getting freed causing a previous write (bdwrite()) to get 'lost'
2761 * because pages associated with a B_DELWRI bp are marked clean.
2763 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2764 * if B_DELWRI is set.
2766 if (bp->b_flags & B_DELWRI)
2767 bp->b_flags &= ~B_RELBUF;
2770 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2771 * constituted, not even NFS buffers now. Two flags effect this. If
2772 * B_INVAL, the struct buf is invalidated but the VM object is kept
2773 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2775 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2776 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2777 * buffer is also B_INVAL because it hits the re-dirtying code above.
2779 * Normally we can do this whether a buffer is B_DELWRI or not. If
2780 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2781 * the commit state and we cannot afford to lose the buffer. If the
2782 * buffer has a background write in progress, we need to keep it
2783 * around to prevent it from being reconstituted and starting a second
2787 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2789 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2790 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2791 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2792 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2793 vfs_vmio_invalidate(bp);
2797 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2798 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2800 bp->b_flags &= ~B_NOREUSE;
2801 if (bp->b_vp != NULL)
2806 * If the buffer has junk contents signal it and eventually
2807 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2810 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2811 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2812 bp->b_flags |= B_INVAL;
2813 if (bp->b_flags & B_INVAL) {
2814 if (bp->b_flags & B_DELWRI)
2820 buf_track(bp, __func__);
2822 /* buffers with no memory */
2823 if (bp->b_bufsize == 0) {
2827 /* buffers with junk contents */
2828 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2829 (bp->b_ioflags & BIO_ERROR)) {
2830 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2831 if (bp->b_vflags & BV_BKGRDINPROG)
2832 panic("losing buffer 2");
2833 qindex = QUEUE_CLEAN;
2834 bp->b_flags |= B_AGE;
2835 /* remaining buffers */
2836 } else if (bp->b_flags & B_DELWRI)
2837 qindex = QUEUE_DIRTY;
2839 qindex = QUEUE_CLEAN;
2841 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2842 panic("brelse: not dirty");
2844 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2845 bp->b_xflags &= ~(BX_CVTENXIO);
2846 /* binsfree unlocks bp. */
2847 binsfree(bp, qindex);
2851 * Release a buffer back to the appropriate queue but do not try to free
2852 * it. The buffer is expected to be used again soon.
2854 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2855 * biodone() to requeue an async I/O on completion. It is also used when
2856 * known good buffers need to be requeued but we think we may need the data
2859 * XXX we should be able to leave the B_RELBUF hint set on completion.
2862 bqrelse(struct buf *bp)
2866 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2867 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2868 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2870 qindex = QUEUE_NONE;
2871 if (BUF_LOCKRECURSED(bp)) {
2872 /* do not release to free list */
2876 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2877 bp->b_xflags &= ~(BX_CVTENXIO);
2879 if (LIST_EMPTY(&bp->b_dep)) {
2880 bp->b_flags &= ~B_IOSTARTED;
2882 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2883 ("bqrelse: SU io not finished bp %p", bp));
2886 if (bp->b_flags & B_MANAGED) {
2887 if (bp->b_flags & B_REMFREE)
2892 /* buffers with stale but valid contents */
2893 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2894 BV_BKGRDERR)) == BV_BKGRDERR) {
2895 BO_LOCK(bp->b_bufobj);
2896 bp->b_vflags &= ~BV_BKGRDERR;
2897 BO_UNLOCK(bp->b_bufobj);
2898 qindex = QUEUE_DIRTY;
2900 if ((bp->b_flags & B_DELWRI) == 0 &&
2901 (bp->b_xflags & BX_VNDIRTY))
2902 panic("bqrelse: not dirty");
2903 if ((bp->b_flags & B_NOREUSE) != 0) {
2907 qindex = QUEUE_CLEAN;
2909 buf_track(bp, __func__);
2910 /* binsfree unlocks bp. */
2911 binsfree(bp, qindex);
2915 buf_track(bp, __func__);
2921 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2922 * restore bogus pages.
2925 vfs_vmio_iodone(struct buf *bp)
2930 struct vnode *vp __unused;
2931 int i, iosize, resid;
2934 obj = bp->b_bufobj->bo_object;
2935 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2936 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2937 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2940 VNPASS(vp->v_holdcnt > 0, vp);
2941 VNPASS(vp->v_object != NULL, vp);
2943 foff = bp->b_offset;
2944 KASSERT(bp->b_offset != NOOFFSET,
2945 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2948 iosize = bp->b_bcount - bp->b_resid;
2949 for (i = 0; i < bp->b_npages; i++) {
2950 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2955 * cleanup bogus pages, restoring the originals
2958 if (m == bogus_page) {
2960 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2962 panic("biodone: page disappeared!");
2964 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2966 * In the write case, the valid and clean bits are
2967 * already changed correctly ( see bdwrite() ), so we
2968 * only need to do this here in the read case.
2970 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2971 resid)) == 0, ("vfs_vmio_iodone: page %p "
2972 "has unexpected dirty bits", m));
2973 vfs_page_set_valid(bp, foff, m);
2975 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2976 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2977 (intmax_t)foff, (uintmax_t)m->pindex));
2980 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2983 vm_object_pip_wakeupn(obj, bp->b_npages);
2984 if (bogus && buf_mapped(bp)) {
2985 BUF_CHECK_MAPPED(bp);
2986 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2987 bp->b_pages, bp->b_npages);
2992 * Perform page invalidation when a buffer is released. The fully invalid
2993 * pages will be reclaimed later in vfs_vmio_truncate().
2996 vfs_vmio_invalidate(struct buf *bp)
3000 int flags, i, resid, poffset, presid;
3002 if (buf_mapped(bp)) {
3003 BUF_CHECK_MAPPED(bp);
3004 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3006 BUF_CHECK_UNMAPPED(bp);
3008 * Get the base offset and length of the buffer. Note that
3009 * in the VMIO case if the buffer block size is not
3010 * page-aligned then b_data pointer may not be page-aligned.
3011 * But our b_pages[] array *IS* page aligned.
3013 * block sizes less then DEV_BSIZE (usually 512) are not
3014 * supported due to the page granularity bits (m->valid,
3015 * m->dirty, etc...).
3017 * See man buf(9) for more information
3019 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3020 obj = bp->b_bufobj->bo_object;
3021 resid = bp->b_bufsize;
3022 poffset = bp->b_offset & PAGE_MASK;
3023 VM_OBJECT_WLOCK(obj);
3024 for (i = 0; i < bp->b_npages; i++) {
3026 if (m == bogus_page)
3027 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3028 bp->b_pages[i] = NULL;
3030 presid = resid > (PAGE_SIZE - poffset) ?
3031 (PAGE_SIZE - poffset) : resid;
3032 KASSERT(presid >= 0, ("brelse: extra page"));
3033 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3034 if (pmap_page_wired_mappings(m) == 0)
3035 vm_page_set_invalid(m, poffset, presid);
3037 vm_page_release_locked(m, flags);
3041 VM_OBJECT_WUNLOCK(obj);
3046 * Page-granular truncation of an existing VMIO buffer.
3049 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3055 if (bp->b_npages == desiredpages)
3058 if (buf_mapped(bp)) {
3059 BUF_CHECK_MAPPED(bp);
3060 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3061 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3063 BUF_CHECK_UNMAPPED(bp);
3066 * The object lock is needed only if we will attempt to free pages.
3068 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3069 if ((bp->b_flags & B_DIRECT) != 0) {
3070 flags |= VPR_TRYFREE;
3071 obj = bp->b_bufobj->bo_object;
3072 VM_OBJECT_WLOCK(obj);
3076 for (i = desiredpages; i < bp->b_npages; i++) {
3078 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3079 bp->b_pages[i] = NULL;
3081 vm_page_release_locked(m, flags);
3083 vm_page_release(m, flags);
3086 VM_OBJECT_WUNLOCK(obj);
3087 bp->b_npages = desiredpages;
3091 * Byte granular extension of VMIO buffers.
3094 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3097 * We are growing the buffer, possibly in a
3098 * byte-granular fashion.
3106 * Step 1, bring in the VM pages from the object, allocating
3107 * them if necessary. We must clear B_CACHE if these pages
3108 * are not valid for the range covered by the buffer.
3110 obj = bp->b_bufobj->bo_object;
3111 if (bp->b_npages < desiredpages) {
3112 KASSERT(desiredpages <= atop(maxbcachebuf),
3113 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3114 bp, desiredpages, maxbcachebuf));
3117 * We must allocate system pages since blocking
3118 * here could interfere with paging I/O, no
3119 * matter which process we are.
3121 * Only exclusive busy can be tested here.
3122 * Blocking on shared busy might lead to
3123 * deadlocks once allocbuf() is called after
3124 * pages are vfs_busy_pages().
3126 (void)vm_page_grab_pages_unlocked(obj,
3127 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3128 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3129 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3130 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3131 bp->b_npages = desiredpages;
3135 * Step 2. We've loaded the pages into the buffer,
3136 * we have to figure out if we can still have B_CACHE
3137 * set. Note that B_CACHE is set according to the
3138 * byte-granular range ( bcount and size ), not the
3139 * aligned range ( newbsize ).
3141 * The VM test is against m->valid, which is DEV_BSIZE
3142 * aligned. Needless to say, the validity of the data
3143 * needs to also be DEV_BSIZE aligned. Note that this
3144 * fails with NFS if the server or some other client
3145 * extends the file's EOF. If our buffer is resized,
3146 * B_CACHE may remain set! XXX
3148 toff = bp->b_bcount;
3149 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3150 while ((bp->b_flags & B_CACHE) && toff < size) {
3153 if (tinc > (size - toff))
3155 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3156 m = bp->b_pages[pi];
3157 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3163 * Step 3, fixup the KVA pmap.
3168 BUF_CHECK_UNMAPPED(bp);
3172 * Check to see if a block at a particular lbn is available for a clustered
3176 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3183 /* If the buf isn't in core skip it */
3184 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3187 /* If the buf is busy we don't want to wait for it */
3188 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3191 /* Only cluster with valid clusterable delayed write buffers */
3192 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3193 (B_DELWRI | B_CLUSTEROK))
3196 if (bpa->b_bufsize != size)
3200 * Check to see if it is in the expected place on disk and that the
3201 * block has been mapped.
3203 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3213 * Implement clustered async writes for clearing out B_DELWRI buffers.
3214 * This is much better then the old way of writing only one buffer at
3215 * a time. Note that we may not be presented with the buffers in the
3216 * correct order, so we search for the cluster in both directions.
3219 vfs_bio_awrite(struct buf *bp)
3224 daddr_t lblkno = bp->b_lblkno;
3225 struct vnode *vp = bp->b_vp;
3233 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3235 * right now we support clustered writing only to regular files. If
3236 * we find a clusterable block we could be in the middle of a cluster
3237 * rather then at the beginning.
3239 if ((vp->v_type == VREG) &&
3240 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3241 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3242 size = vp->v_mount->mnt_stat.f_iosize;
3243 maxcl = maxphys / size;
3246 for (i = 1; i < maxcl; i++)
3247 if (vfs_bio_clcheck(vp, size, lblkno + i,
3248 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3251 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3252 if (vfs_bio_clcheck(vp, size, lblkno - j,
3253 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3259 * this is a possible cluster write
3263 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3269 bp->b_flags |= B_ASYNC;
3271 * default (old) behavior, writing out only one block
3273 * XXX returns b_bufsize instead of b_bcount for nwritten?
3275 nwritten = bp->b_bufsize;
3284 * Allocate KVA for an empty buf header according to gbflags.
3287 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3290 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3292 * In order to keep fragmentation sane we only allocate kva
3293 * in BKVASIZE chunks. XXX with vmem we can do page size.
3295 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3297 if (maxsize != bp->b_kvasize &&
3298 bufkva_alloc(bp, maxsize, gbflags))
3307 * Find and initialize a new buffer header, freeing up existing buffers
3308 * in the bufqueues as necessary. The new buffer is returned locked.
3311 * We have insufficient buffer headers
3312 * We have insufficient buffer space
3313 * buffer_arena is too fragmented ( space reservation fails )
3314 * If we have to flush dirty buffers ( but we try to avoid this )
3316 * The caller is responsible for releasing the reserved bufspace after
3317 * allocbuf() is called.
3320 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3322 struct bufdomain *bd;
3324 bool metadata, reserved;
3327 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3328 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3329 if (!unmapped_buf_allowed)
3330 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3332 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3340 bd = &bdomain[vp->v_bufobj.bo_domain];
3342 counter_u64_add(getnewbufcalls, 1);
3345 if (reserved == false &&
3346 bufspace_reserve(bd, maxsize, metadata) != 0) {
3347 counter_u64_add(getnewbufrestarts, 1);
3351 if ((bp = buf_alloc(bd)) == NULL) {
3352 counter_u64_add(getnewbufrestarts, 1);
3355 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3358 } while (buf_recycle(bd, false) == 0);
3361 bufspace_release(bd, maxsize);
3363 bp->b_flags |= B_INVAL;
3366 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3374 * buffer flushing daemon. Buffers are normally flushed by the
3375 * update daemon but if it cannot keep up this process starts to
3376 * take the load in an attempt to prevent getnewbuf() from blocking.
3378 static struct kproc_desc buf_kp = {
3383 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3386 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3390 flushed = flushbufqueues(vp, bd, target, 0);
3393 * Could not find any buffers without rollback
3394 * dependencies, so just write the first one
3395 * in the hopes of eventually making progress.
3397 if (vp != NULL && target > 2)
3399 flushbufqueues(vp, bd, target, 1);
3405 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3409 if (KERNEL_PANICKED())
3414 wakeup(&bd_request);
3415 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3417 mtx_unlock(&bdlock);
3419 printf("bufdaemon wait error: %d\n", error);
3425 struct bufdomain *bd;
3431 * This process needs to be suspended prior to shutdown sync.
3433 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3434 SHUTDOWN_PRI_LAST + 100);
3437 * Start the buf clean daemons as children threads.
3439 for (i = 0 ; i < buf_domains; i++) {
3442 error = kthread_add((void (*)(void *))bufspace_daemon,
3443 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3445 panic("error %d spawning bufspace daemon", error);
3449 * This process is allowed to take the buffer cache to the limit
3451 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3453 while (!bd_shutdown) {
3455 mtx_unlock(&bdlock);
3458 * Save speedupreq for this pass and reset to capture new
3461 speedupreq = bd_speedupreq;
3465 * Flush each domain sequentially according to its level and
3466 * the speedup request.
3468 for (i = 0; i < buf_domains; i++) {
3471 lodirty = bd->bd_numdirtybuffers / 2;
3473 lodirty = bd->bd_lodirtybuffers;
3474 while (bd->bd_numdirtybuffers > lodirty) {
3475 if (buf_flush(NULL, bd,
3476 bd->bd_numdirtybuffers - lodirty) == 0)
3478 kern_yield(PRI_USER);
3483 * Only clear bd_request if we have reached our low water
3484 * mark. The buf_daemon normally waits 1 second and
3485 * then incrementally flushes any dirty buffers that have
3486 * built up, within reason.
3488 * If we were unable to hit our low water mark and couldn't
3489 * find any flushable buffers, we sleep for a short period
3490 * to avoid endless loops on unlockable buffers.
3495 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3497 * We reached our low water mark, reset the
3498 * request and sleep until we are needed again.
3499 * The sleep is just so the suspend code works.
3503 * Do an extra wakeup in case dirty threshold
3504 * changed via sysctl and the explicit transition
3505 * out of shortfall was missed.
3508 if (runningbufspace <= lorunningspace)
3510 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3513 * We couldn't find any flushable dirty buffers but
3514 * still have too many dirty buffers, we
3515 * have to sleep and try again. (rare)
3517 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3520 wakeup(&bd_shutdown);
3521 mtx_unlock(&bdlock);
3528 * Try to flush a buffer in the dirty queue. We must be careful to
3529 * free up B_INVAL buffers instead of write them, which NFS is
3530 * particularly sensitive to.
3532 static int flushwithdeps = 0;
3533 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3535 "Number of buffers flushed with dependencies that require rollbacks");
3538 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3541 struct bufqueue *bq;
3542 struct buf *sentinel;
3552 bq = &bd->bd_dirtyq;
3554 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3555 sentinel->b_qindex = QUEUE_SENTINEL;
3557 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3559 while (flushed != target) {
3562 bp = TAILQ_NEXT(sentinel, b_freelist);
3564 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3565 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3572 * Skip sentinels inserted by other invocations of the
3573 * flushbufqueues(), taking care to not reorder them.
3575 * Only flush the buffers that belong to the
3576 * vnode locked by the curthread.
3578 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3583 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3589 * BKGRDINPROG can only be set with the buf and bufobj
3590 * locks both held. We tolerate a race to clear it here.
3592 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3593 (bp->b_flags & B_DELWRI) == 0) {
3597 if (bp->b_flags & B_INVAL) {
3604 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3605 if (flushdeps == 0) {
3613 * We must hold the lock on a vnode before writing
3614 * one of its buffers. Otherwise we may confuse, or
3615 * in the case of a snapshot vnode, deadlock the
3618 * The lock order here is the reverse of the normal
3619 * of vnode followed by buf lock. This is ok because
3620 * the NOWAIT will prevent deadlock.
3623 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3629 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3631 ASSERT_VOP_LOCKED(vp, "getbuf");
3633 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3634 vn_lock(vp, LK_TRYUPGRADE);
3637 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3638 bp, bp->b_vp, bp->b_flags);
3639 if (curproc == bufdaemonproc) {
3644 counter_u64_add(notbufdflushes, 1);
3646 vn_finished_write(mp);
3649 flushwithdeps += hasdeps;
3653 * Sleeping on runningbufspace while holding
3654 * vnode lock leads to deadlock.
3656 if (curproc == bufdaemonproc &&
3657 runningbufspace > hirunningspace)
3658 waitrunningbufspace();
3661 vn_finished_write(mp);
3665 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3667 free(sentinel, M_TEMP);
3672 * Check to see if a block is currently memory resident.
3675 incore(struct bufobj *bo, daddr_t blkno)
3677 return (gbincore_unlocked(bo, blkno));
3681 * Returns true if no I/O is needed to access the
3682 * associated VM object. This is like incore except
3683 * it also hunts around in the VM system for the data.
3686 inmem(struct vnode * vp, daddr_t blkno)
3689 vm_offset_t toff, tinc, size;
3694 ASSERT_VOP_LOCKED(vp, "inmem");
3696 if (incore(&vp->v_bufobj, blkno))
3698 if (vp->v_mount == NULL)
3705 if (size > vp->v_mount->mnt_stat.f_iosize)
3706 size = vp->v_mount->mnt_stat.f_iosize;
3707 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3709 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3710 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3716 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3717 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3719 * Consider page validity only if page mapping didn't change
3722 valid = vm_page_is_valid(m,
3723 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3724 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3736 * Set the dirty range for a buffer based on the status of the dirty
3737 * bits in the pages comprising the buffer. The range is limited
3738 * to the size of the buffer.
3740 * Tell the VM system that the pages associated with this buffer
3741 * are clean. This is used for delayed writes where the data is
3742 * going to go to disk eventually without additional VM intevention.
3744 * Note that while we only really need to clean through to b_bcount, we
3745 * just go ahead and clean through to b_bufsize.
3748 vfs_clean_pages_dirty_buf(struct buf *bp)
3750 vm_ooffset_t foff, noff, eoff;
3754 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3757 foff = bp->b_offset;
3758 KASSERT(bp->b_offset != NOOFFSET,
3759 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3761 vfs_busy_pages_acquire(bp);
3762 vfs_setdirty_range(bp);
3763 for (i = 0; i < bp->b_npages; i++) {
3764 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3766 if (eoff > bp->b_offset + bp->b_bufsize)
3767 eoff = bp->b_offset + bp->b_bufsize;
3769 vfs_page_set_validclean(bp, foff, m);
3770 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3773 vfs_busy_pages_release(bp);
3777 vfs_setdirty_range(struct buf *bp)
3779 vm_offset_t boffset;
3780 vm_offset_t eoffset;
3784 * test the pages to see if they have been modified directly
3785 * by users through the VM system.
3787 for (i = 0; i < bp->b_npages; i++)
3788 vm_page_test_dirty(bp->b_pages[i]);
3791 * Calculate the encompassing dirty range, boffset and eoffset,
3792 * (eoffset - boffset) bytes.
3795 for (i = 0; i < bp->b_npages; i++) {
3796 if (bp->b_pages[i]->dirty)
3799 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3801 for (i = bp->b_npages - 1; i >= 0; --i) {
3802 if (bp->b_pages[i]->dirty) {
3806 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3809 * Fit it to the buffer.
3812 if (eoffset > bp->b_bcount)
3813 eoffset = bp->b_bcount;
3816 * If we have a good dirty range, merge with the existing
3820 if (boffset < eoffset) {
3821 if (bp->b_dirtyoff > boffset)
3822 bp->b_dirtyoff = boffset;
3823 if (bp->b_dirtyend < eoffset)
3824 bp->b_dirtyend = eoffset;
3829 * Allocate the KVA mapping for an existing buffer.
3830 * If an unmapped buffer is provided but a mapped buffer is requested, take
3831 * also care to properly setup mappings between pages and KVA.
3834 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3836 int bsize, maxsize, need_mapping, need_kva;
3839 need_mapping = bp->b_data == unmapped_buf &&
3840 (gbflags & GB_UNMAPPED) == 0;
3841 need_kva = bp->b_kvabase == unmapped_buf &&
3842 bp->b_data == unmapped_buf &&
3843 (gbflags & GB_KVAALLOC) != 0;
3844 if (!need_mapping && !need_kva)
3847 BUF_CHECK_UNMAPPED(bp);
3849 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3851 * Buffer is not mapped, but the KVA was already
3852 * reserved at the time of the instantiation. Use the
3859 * Calculate the amount of the address space we would reserve
3860 * if the buffer was mapped.
3862 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3863 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3864 offset = blkno * bsize;
3865 maxsize = size + (offset & PAGE_MASK);
3866 maxsize = imax(maxsize, bsize);
3868 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3869 if ((gbflags & GB_NOWAIT_BD) != 0) {
3871 * XXXKIB: defragmentation cannot
3872 * succeed, not sure what else to do.
3874 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3876 counter_u64_add(mappingrestarts, 1);
3877 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3881 /* b_offset is handled by bpmap_qenter. */
3882 bp->b_data = bp->b_kvabase;
3883 BUF_CHECK_MAPPED(bp);
3889 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3895 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3904 * Get a block given a specified block and offset into a file/device.
3905 * The buffers B_DONE bit will be cleared on return, making it almost
3906 * ready for an I/O initiation. B_INVAL may or may not be set on
3907 * return. The caller should clear B_INVAL prior to initiating a
3910 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3911 * an existing buffer.
3913 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3914 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3915 * and then cleared based on the backing VM. If the previous buffer is
3916 * non-0-sized but invalid, B_CACHE will be cleared.
3918 * If getblk() must create a new buffer, the new buffer is returned with
3919 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3920 * case it is returned with B_INVAL clear and B_CACHE set based on the
3923 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3924 * B_CACHE bit is clear.
3926 * What this means, basically, is that the caller should use B_CACHE to
3927 * determine whether the buffer is fully valid or not and should clear
3928 * B_INVAL prior to issuing a read. If the caller intends to validate
3929 * the buffer by loading its data area with something, the caller needs
3930 * to clear B_INVAL. If the caller does this without issuing an I/O,
3931 * the caller should set B_CACHE ( as an optimization ), else the caller
3932 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3933 * a write attempt or if it was a successful read. If the caller
3934 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3935 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3937 * The blkno parameter is the logical block being requested. Normally
3938 * the mapping of logical block number to disk block address is done
3939 * by calling VOP_BMAP(). However, if the mapping is already known, the
3940 * disk block address can be passed using the dblkno parameter. If the
3941 * disk block address is not known, then the same value should be passed
3942 * for blkno and dblkno.
3945 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3946 int slptimeo, int flags, struct buf **bpp)
3951 int bsize, error, maxsize, vmio;
3954 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3955 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3956 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3957 if (vp->v_type != VCHR)
3958 ASSERT_VOP_LOCKED(vp, "getblk");
3959 if (size > maxbcachebuf)
3960 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3962 if (!unmapped_buf_allowed)
3963 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3968 /* Attempt lockless lookup first. */
3969 bp = gbincore_unlocked(bo, blkno);
3971 goto newbuf_unlocked;
3973 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3978 /* Verify buf identify has not changed since lookup. */
3979 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3980 goto foundbuf_fastpath;
3982 /* It changed, fallback to locked lookup. */
3987 bp = gbincore(bo, blkno);
3992 * Buffer is in-core. If the buffer is not busy nor managed,
3993 * it must be on a queue.
3995 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3996 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
3998 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
4001 error = BUF_TIMELOCK(bp, lockflags,
4002 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
4005 * If we slept and got the lock we have to restart in case
4006 * the buffer changed identities.
4008 if (error == ENOLCK)
4010 /* We timed out or were interrupted. */
4011 else if (error != 0)
4015 /* If recursed, assume caller knows the rules. */
4016 if (BUF_LOCKRECURSED(bp))
4020 * The buffer is locked. B_CACHE is cleared if the buffer is
4021 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4022 * and for a VMIO buffer B_CACHE is adjusted according to the
4025 if (bp->b_flags & B_INVAL)
4026 bp->b_flags &= ~B_CACHE;
4027 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4028 bp->b_flags |= B_CACHE;
4029 if (bp->b_flags & B_MANAGED)
4030 MPASS(bp->b_qindex == QUEUE_NONE);
4035 * check for size inconsistencies for non-VMIO case.
4037 if (bp->b_bcount != size) {
4038 if ((bp->b_flags & B_VMIO) == 0 ||
4039 (size > bp->b_kvasize)) {
4040 if (bp->b_flags & B_DELWRI) {
4041 bp->b_flags |= B_NOCACHE;
4044 if (LIST_EMPTY(&bp->b_dep)) {
4045 bp->b_flags |= B_RELBUF;
4048 bp->b_flags |= B_NOCACHE;
4057 * Handle the case of unmapped buffer which should
4058 * become mapped, or the buffer for which KVA
4059 * reservation is requested.
4061 bp_unmapped_get_kva(bp, blkno, size, flags);
4064 * If the size is inconsistent in the VMIO case, we can resize
4065 * the buffer. This might lead to B_CACHE getting set or
4066 * cleared. If the size has not changed, B_CACHE remains
4067 * unchanged from its previous state.
4071 KASSERT(bp->b_offset != NOOFFSET,
4072 ("getblk: no buffer offset"));
4075 * A buffer with B_DELWRI set and B_CACHE clear must
4076 * be committed before we can return the buffer in
4077 * order to prevent the caller from issuing a read
4078 * ( due to B_CACHE not being set ) and overwriting
4081 * Most callers, including NFS and FFS, need this to
4082 * operate properly either because they assume they
4083 * can issue a read if B_CACHE is not set, or because
4084 * ( for example ) an uncached B_DELWRI might loop due
4085 * to softupdates re-dirtying the buffer. In the latter
4086 * case, B_CACHE is set after the first write completes,
4087 * preventing further loops.
4088 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4089 * above while extending the buffer, we cannot allow the
4090 * buffer to remain with B_CACHE set after the write
4091 * completes or it will represent a corrupt state. To
4092 * deal with this we set B_NOCACHE to scrap the buffer
4095 * We might be able to do something fancy, like setting
4096 * B_CACHE in bwrite() except if B_DELWRI is already set,
4097 * so the below call doesn't set B_CACHE, but that gets real
4098 * confusing. This is much easier.
4101 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4102 bp->b_flags |= B_NOCACHE;
4106 bp->b_flags &= ~B_DONE;
4109 * Buffer is not in-core, create new buffer. The buffer
4110 * returned by getnewbuf() is locked. Note that the returned
4111 * buffer is also considered valid (not marked B_INVAL).
4116 * If the user does not want us to create the buffer, bail out
4119 if (flags & GB_NOCREAT)
4122 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4123 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4124 offset = blkno * bsize;
4125 vmio = vp->v_object != NULL;
4127 maxsize = size + (offset & PAGE_MASK);
4130 /* Do not allow non-VMIO notmapped buffers. */
4131 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4133 maxsize = imax(maxsize, bsize);
4134 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4136 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4137 KASSERT(error != EOPNOTSUPP,
4138 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4143 return (EJUSTRETURN);
4146 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4148 if (slpflag || slptimeo)
4151 * XXX This is here until the sleep path is diagnosed
4152 * enough to work under very low memory conditions.
4154 * There's an issue on low memory, 4BSD+non-preempt
4155 * systems (eg MIPS routers with 32MB RAM) where buffer
4156 * exhaustion occurs without sleeping for buffer
4157 * reclaimation. This just sticks in a loop and
4158 * constantly attempts to allocate a buffer, which
4159 * hits exhaustion and tries to wakeup bufdaemon.
4160 * This never happens because we never yield.
4162 * The real solution is to identify and fix these cases
4163 * so we aren't effectively busy-waiting in a loop
4164 * until the reclaimation path has cycles to run.
4166 kern_yield(PRI_USER);
4171 * This code is used to make sure that a buffer is not
4172 * created while the getnewbuf routine is blocked.
4173 * This can be a problem whether the vnode is locked or not.
4174 * If the buffer is created out from under us, we have to
4175 * throw away the one we just created.
4177 * Note: this must occur before we associate the buffer
4178 * with the vp especially considering limitations in
4179 * the splay tree implementation when dealing with duplicate
4183 if (gbincore(bo, blkno)) {
4185 bp->b_flags |= B_INVAL;
4186 bufspace_release(bufdomain(bp), maxsize);
4192 * Insert the buffer into the hash, so that it can
4193 * be found by incore.
4195 bp->b_lblkno = blkno;
4196 bp->b_blkno = d_blkno;
4197 bp->b_offset = offset;
4202 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4203 * buffer size starts out as 0, B_CACHE will be set by
4204 * allocbuf() for the VMIO case prior to it testing the
4205 * backing store for validity.
4209 bp->b_flags |= B_VMIO;
4210 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4211 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4212 bp, vp->v_object, bp->b_bufobj->bo_object));
4214 bp->b_flags &= ~B_VMIO;
4215 KASSERT(bp->b_bufobj->bo_object == NULL,
4216 ("ARGH! has b_bufobj->bo_object %p %p\n",
4217 bp, bp->b_bufobj->bo_object));
4218 BUF_CHECK_MAPPED(bp);
4222 bufspace_release(bufdomain(bp), maxsize);
4223 bp->b_flags &= ~B_DONE;
4225 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4227 buf_track(bp, __func__);
4228 KASSERT(bp->b_bufobj == bo,
4229 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4235 * Get an empty, disassociated buffer of given size. The buffer is initially
4239 geteblk(int size, int flags)
4244 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4245 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4246 if ((flags & GB_NOWAIT_BD) &&
4247 (curthread->td_pflags & TDP_BUFNEED) != 0)
4251 bufspace_release(bufdomain(bp), maxsize);
4252 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4257 * Truncate the backing store for a non-vmio buffer.
4260 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4263 if (bp->b_flags & B_MALLOC) {
4265 * malloced buffers are not shrunk
4267 if (newbsize == 0) {
4268 bufmallocadjust(bp, 0);
4269 free(bp->b_data, M_BIOBUF);
4270 bp->b_data = bp->b_kvabase;
4271 bp->b_flags &= ~B_MALLOC;
4275 vm_hold_free_pages(bp, newbsize);
4276 bufspace_adjust(bp, newbsize);
4280 * Extend the backing for a non-VMIO buffer.
4283 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4289 * We only use malloced memory on the first allocation.
4290 * and revert to page-allocated memory when the buffer
4293 * There is a potential smp race here that could lead
4294 * to bufmallocspace slightly passing the max. It
4295 * is probably extremely rare and not worth worrying
4298 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4299 bufmallocspace < maxbufmallocspace) {
4300 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4301 bp->b_flags |= B_MALLOC;
4302 bufmallocadjust(bp, newbsize);
4307 * If the buffer is growing on its other-than-first
4308 * allocation then we revert to the page-allocation
4313 if (bp->b_flags & B_MALLOC) {
4314 origbuf = bp->b_data;
4315 origbufsize = bp->b_bufsize;
4316 bp->b_data = bp->b_kvabase;
4317 bufmallocadjust(bp, 0);
4318 bp->b_flags &= ~B_MALLOC;
4319 newbsize = round_page(newbsize);
4321 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4322 (vm_offset_t) bp->b_data + newbsize);
4323 if (origbuf != NULL) {
4324 bcopy(origbuf, bp->b_data, origbufsize);
4325 free(origbuf, M_BIOBUF);
4327 bufspace_adjust(bp, newbsize);
4331 * This code constitutes the buffer memory from either anonymous system
4332 * memory (in the case of non-VMIO operations) or from an associated
4333 * VM object (in the case of VMIO operations). This code is able to
4334 * resize a buffer up or down.
4336 * Note that this code is tricky, and has many complications to resolve
4337 * deadlock or inconsistent data situations. Tread lightly!!!
4338 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4339 * the caller. Calling this code willy nilly can result in the loss of data.
4341 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4342 * B_CACHE for the non-VMIO case.
4345 allocbuf(struct buf *bp, int size)
4349 if (bp->b_bcount == size)
4352 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size,
4353 ("allocbuf: buffer too small %p %#x %#x",
4354 bp, bp->b_kvasize, size));
4356 newbsize = roundup2(size, DEV_BSIZE);
4357 if ((bp->b_flags & B_VMIO) == 0) {
4358 if ((bp->b_flags & B_MALLOC) == 0)
4359 newbsize = round_page(newbsize);
4361 * Just get anonymous memory from the kernel. Don't
4362 * mess with B_CACHE.
4364 if (newbsize < bp->b_bufsize)
4365 vfs_nonvmio_truncate(bp, newbsize);
4366 else if (newbsize > bp->b_bufsize)
4367 vfs_nonvmio_extend(bp, newbsize);
4371 desiredpages = size == 0 ? 0 :
4372 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4374 KASSERT((bp->b_flags & B_MALLOC) == 0,
4375 ("allocbuf: VMIO buffer can't be malloced %p", bp));
4378 * Set B_CACHE initially if buffer is 0 length or will become
4381 if (size == 0 || bp->b_bufsize == 0)
4382 bp->b_flags |= B_CACHE;
4384 if (newbsize < bp->b_bufsize)
4385 vfs_vmio_truncate(bp, desiredpages);
4386 /* XXX This looks as if it should be newbsize > b_bufsize */
4387 else if (size > bp->b_bcount)
4388 vfs_vmio_extend(bp, desiredpages, size);
4389 bufspace_adjust(bp, newbsize);
4391 bp->b_bcount = size; /* requested buffer size. */
4395 extern int inflight_transient_maps;
4397 static struct bio_queue nondump_bios;
4400 biodone(struct bio *bp)
4403 void (*done)(struct bio *);
4404 vm_offset_t start, end;
4406 biotrack(bp, __func__);
4409 * Avoid completing I/O when dumping after a panic since that may
4410 * result in a deadlock in the filesystem or pager code. Note that
4411 * this doesn't affect dumps that were started manually since we aim
4412 * to keep the system usable after it has been resumed.
4414 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4415 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4418 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4419 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4420 bp->bio_flags |= BIO_UNMAPPED;
4421 start = trunc_page((vm_offset_t)bp->bio_data);
4422 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4423 bp->bio_data = unmapped_buf;
4424 pmap_qremove(start, atop(end - start));
4425 vmem_free(transient_arena, start, end - start);
4426 atomic_add_int(&inflight_transient_maps, -1);
4428 done = bp->bio_done;
4430 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4432 bp->bio_flags |= BIO_DONE;
4440 * Wait for a BIO to finish.
4443 biowait(struct bio *bp, const char *wmesg)
4447 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4449 while ((bp->bio_flags & BIO_DONE) == 0)
4450 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4452 if (bp->bio_error != 0)
4453 return (bp->bio_error);
4454 if (!(bp->bio_flags & BIO_ERROR))
4460 biofinish(struct bio *bp, struct devstat *stat, int error)
4464 bp->bio_error = error;
4465 bp->bio_flags |= BIO_ERROR;
4468 devstat_end_transaction_bio(stat, bp);
4472 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4474 biotrack_buf(struct bio *bp, const char *location)
4477 buf_track(bp->bio_track_bp, location);
4484 * Wait for buffer I/O completion, returning error status. The buffer
4485 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4486 * error and cleared.
4489 bufwait(struct buf *bp)
4491 if (bp->b_iocmd == BIO_READ)
4492 bwait(bp, PRIBIO, "biord");
4494 bwait(bp, PRIBIO, "biowr");
4495 if (bp->b_flags & B_EINTR) {
4496 bp->b_flags &= ~B_EINTR;
4499 if (bp->b_ioflags & BIO_ERROR) {
4500 return (bp->b_error ? bp->b_error : EIO);
4509 * Finish I/O on a buffer, optionally calling a completion function.
4510 * This is usually called from an interrupt so process blocking is
4513 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4514 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4515 * assuming B_INVAL is clear.
4517 * For the VMIO case, we set B_CACHE if the op was a read and no
4518 * read error occurred, or if the op was a write. B_CACHE is never
4519 * set if the buffer is invalid or otherwise uncacheable.
4521 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4522 * initiator to leave B_INVAL set to brelse the buffer out of existence
4523 * in the biodone routine.
4526 bufdone(struct buf *bp)
4528 struct bufobj *dropobj;
4529 void (*biodone)(struct buf *);
4531 buf_track(bp, __func__);
4532 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4535 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4537 runningbufwakeup(bp);
4538 if (bp->b_iocmd == BIO_WRITE)
4539 dropobj = bp->b_bufobj;
4540 /* call optional completion function if requested */
4541 if (bp->b_iodone != NULL) {
4542 biodone = bp->b_iodone;
4543 bp->b_iodone = NULL;
4546 bufobj_wdrop(dropobj);
4549 if (bp->b_flags & B_VMIO) {
4551 * Set B_CACHE if the op was a normal read and no error
4552 * occurred. B_CACHE is set for writes in the b*write()
4555 if (bp->b_iocmd == BIO_READ &&
4556 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4557 !(bp->b_ioflags & BIO_ERROR))
4558 bp->b_flags |= B_CACHE;
4559 vfs_vmio_iodone(bp);
4561 if (!LIST_EMPTY(&bp->b_dep))
4563 if ((bp->b_flags & B_CKHASH) != 0) {
4564 KASSERT(bp->b_iocmd == BIO_READ,
4565 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4566 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4567 (*bp->b_ckhashcalc)(bp);
4570 * For asynchronous completions, release the buffer now. The brelse
4571 * will do a wakeup there if necessary - so no need to do a wakeup
4572 * here in the async case. The sync case always needs to do a wakeup.
4574 if (bp->b_flags & B_ASYNC) {
4575 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4576 (bp->b_ioflags & BIO_ERROR))
4583 bufobj_wdrop(dropobj);
4587 * This routine is called in lieu of iodone in the case of
4588 * incomplete I/O. This keeps the busy status for pages
4592 vfs_unbusy_pages(struct buf *bp)
4598 runningbufwakeup(bp);
4599 if (!(bp->b_flags & B_VMIO))
4602 obj = bp->b_bufobj->bo_object;
4603 for (i = 0; i < bp->b_npages; i++) {
4605 if (m == bogus_page) {
4606 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4608 panic("vfs_unbusy_pages: page missing\n");
4610 if (buf_mapped(bp)) {
4611 BUF_CHECK_MAPPED(bp);
4612 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4613 bp->b_pages, bp->b_npages);
4615 BUF_CHECK_UNMAPPED(bp);
4619 vm_object_pip_wakeupn(obj, bp->b_npages);
4623 * vfs_page_set_valid:
4625 * Set the valid bits in a page based on the supplied offset. The
4626 * range is restricted to the buffer's size.
4628 * This routine is typically called after a read completes.
4631 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4636 * Compute the end offset, eoff, such that [off, eoff) does not span a
4637 * page boundary and eoff is not greater than the end of the buffer.
4638 * The end of the buffer, in this case, is our file EOF, not the
4639 * allocation size of the buffer.
4641 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4642 if (eoff > bp->b_offset + bp->b_bcount)
4643 eoff = bp->b_offset + bp->b_bcount;
4646 * Set valid range. This is typically the entire buffer and thus the
4650 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4654 * vfs_page_set_validclean:
4656 * Set the valid bits and clear the dirty bits in a page based on the
4657 * supplied offset. The range is restricted to the buffer's size.
4660 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4662 vm_ooffset_t soff, eoff;
4665 * Start and end offsets in buffer. eoff - soff may not cross a
4666 * page boundary or cross the end of the buffer. The end of the
4667 * buffer, in this case, is our file EOF, not the allocation size
4671 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4672 if (eoff > bp->b_offset + bp->b_bcount)
4673 eoff = bp->b_offset + bp->b_bcount;
4676 * Set valid range. This is typically the entire buffer and thus the
4680 vm_page_set_validclean(
4682 (vm_offset_t) (soff & PAGE_MASK),
4683 (vm_offset_t) (eoff - soff)
4689 * Acquire a shared busy on all pages in the buf.
4692 vfs_busy_pages_acquire(struct buf *bp)
4696 for (i = 0; i < bp->b_npages; i++)
4697 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4701 vfs_busy_pages_release(struct buf *bp)
4705 for (i = 0; i < bp->b_npages; i++)
4706 vm_page_sunbusy(bp->b_pages[i]);
4710 * This routine is called before a device strategy routine.
4711 * It is used to tell the VM system that paging I/O is in
4712 * progress, and treat the pages associated with the buffer
4713 * almost as being exclusive busy. Also the object paging_in_progress
4714 * flag is handled to make sure that the object doesn't become
4717 * Since I/O has not been initiated yet, certain buffer flags
4718 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4719 * and should be ignored.
4722 vfs_busy_pages(struct buf *bp, int clear_modify)
4730 if (!(bp->b_flags & B_VMIO))
4733 obj = bp->b_bufobj->bo_object;
4734 foff = bp->b_offset;
4735 KASSERT(bp->b_offset != NOOFFSET,
4736 ("vfs_busy_pages: no buffer offset"));
4737 if ((bp->b_flags & B_CLUSTER) == 0) {
4738 vm_object_pip_add(obj, bp->b_npages);
4739 vfs_busy_pages_acquire(bp);
4741 if (bp->b_bufsize != 0)
4742 vfs_setdirty_range(bp);
4744 for (i = 0; i < bp->b_npages; i++) {
4746 vm_page_assert_sbusied(m);
4749 * When readying a buffer for a read ( i.e
4750 * clear_modify == 0 ), it is important to do
4751 * bogus_page replacement for valid pages in
4752 * partially instantiated buffers. Partially
4753 * instantiated buffers can, in turn, occur when
4754 * reconstituting a buffer from its VM backing store
4755 * base. We only have to do this if B_CACHE is
4756 * clear ( which causes the I/O to occur in the
4757 * first place ). The replacement prevents the read
4758 * I/O from overwriting potentially dirty VM-backed
4759 * pages. XXX bogus page replacement is, uh, bogus.
4760 * It may not work properly with small-block devices.
4761 * We need to find a better way.
4764 pmap_remove_write(m);
4765 vfs_page_set_validclean(bp, foff, m);
4766 } else if (vm_page_all_valid(m) &&
4767 (bp->b_flags & B_CACHE) == 0) {
4768 bp->b_pages[i] = bogus_page;
4771 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4773 if (bogus && buf_mapped(bp)) {
4774 BUF_CHECK_MAPPED(bp);
4775 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4776 bp->b_pages, bp->b_npages);
4781 * vfs_bio_set_valid:
4783 * Set the range within the buffer to valid. The range is
4784 * relative to the beginning of the buffer, b_offset. Note that
4785 * b_offset itself may be offset from the beginning of the first
4789 vfs_bio_set_valid(struct buf *bp, int base, int size)
4794 if (!(bp->b_flags & B_VMIO))
4798 * Fixup base to be relative to beginning of first page.
4799 * Set initial n to be the maximum number of bytes in the
4800 * first page that can be validated.
4802 base += (bp->b_offset & PAGE_MASK);
4803 n = PAGE_SIZE - (base & PAGE_MASK);
4806 * Busy may not be strictly necessary here because the pages are
4807 * unlikely to be fully valid and the vnode lock will synchronize
4808 * their access via getpages. It is grabbed for consistency with
4809 * other page validation.
4811 vfs_busy_pages_acquire(bp);
4812 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4816 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4821 vfs_busy_pages_release(bp);
4827 * If the specified buffer is a non-VMIO buffer, clear the entire
4828 * buffer. If the specified buffer is a VMIO buffer, clear and
4829 * validate only the previously invalid portions of the buffer.
4830 * This routine essentially fakes an I/O, so we need to clear
4831 * BIO_ERROR and B_INVAL.
4833 * Note that while we only theoretically need to clear through b_bcount,
4834 * we go ahead and clear through b_bufsize.
4837 vfs_bio_clrbuf(struct buf *bp)
4839 int i, j, mask, sa, ea, slide;
4841 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4845 bp->b_flags &= ~B_INVAL;
4846 bp->b_ioflags &= ~BIO_ERROR;
4847 vfs_busy_pages_acquire(bp);
4848 sa = bp->b_offset & PAGE_MASK;
4850 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4851 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4852 ea = slide & PAGE_MASK;
4855 if (bp->b_pages[i] == bogus_page)
4858 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4859 if ((bp->b_pages[i]->valid & mask) == mask)
4861 if ((bp->b_pages[i]->valid & mask) == 0)
4862 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4864 for (; sa < ea; sa += DEV_BSIZE, j++) {
4865 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4866 pmap_zero_page_area(bp->b_pages[i],
4871 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4872 roundup2(ea - sa, DEV_BSIZE));
4874 vfs_busy_pages_release(bp);
4879 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4884 if (buf_mapped(bp)) {
4885 BUF_CHECK_MAPPED(bp);
4886 bzero(bp->b_data + base, size);
4888 BUF_CHECK_UNMAPPED(bp);
4889 n = PAGE_SIZE - (base & PAGE_MASK);
4890 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4894 pmap_zero_page_area(m, base & PAGE_MASK, n);
4903 * Update buffer flags based on I/O request parameters, optionally releasing the
4904 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4905 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4906 * I/O). Otherwise the buffer is released to the cache.
4909 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4912 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4913 ("buf %p non-VMIO noreuse", bp));
4915 if ((ioflag & IO_DIRECT) != 0)
4916 bp->b_flags |= B_DIRECT;
4917 if ((ioflag & IO_EXT) != 0)
4918 bp->b_xflags |= BX_ALTDATA;
4919 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4920 bp->b_flags |= B_RELBUF;
4921 if ((ioflag & IO_NOREUSE) != 0)
4922 bp->b_flags |= B_NOREUSE;
4930 vfs_bio_brelse(struct buf *bp, int ioflag)
4933 b_io_dismiss(bp, ioflag, true);
4937 vfs_bio_set_flags(struct buf *bp, int ioflag)
4940 b_io_dismiss(bp, ioflag, false);
4944 * vm_hold_load_pages and vm_hold_free_pages get pages into
4945 * a buffers address space. The pages are anonymous and are
4946 * not associated with a file object.
4949 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4955 BUF_CHECK_MAPPED(bp);
4957 to = round_page(to);
4958 from = round_page(from);
4959 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4960 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4961 KASSERT(to - from <= maxbcachebuf,
4962 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4963 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4965 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4967 * note: must allocate system pages since blocking here
4968 * could interfere with paging I/O, no matter which
4971 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4972 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4973 pmap_qenter(pg, &p, 1);
4974 bp->b_pages[index] = p;
4976 bp->b_npages = index;
4979 /* Return pages associated with this buf to the vm system */
4981 vm_hold_free_pages(struct buf *bp, int newbsize)
4985 int index, newnpages;
4987 BUF_CHECK_MAPPED(bp);
4989 from = round_page((vm_offset_t)bp->b_data + newbsize);
4990 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4991 if (bp->b_npages > newnpages)
4992 pmap_qremove(from, bp->b_npages - newnpages);
4993 for (index = newnpages; index < bp->b_npages; index++) {
4994 p = bp->b_pages[index];
4995 bp->b_pages[index] = NULL;
4996 vm_page_unwire_noq(p);
4999 bp->b_npages = newnpages;
5003 * Map an IO request into kernel virtual address space.
5005 * All requests are (re)mapped into kernel VA space.
5006 * Notice that we use b_bufsize for the size of the buffer
5007 * to be mapped. b_bcount might be modified by the driver.
5009 * Note that even if the caller determines that the address space should
5010 * be valid, a race or a smaller-file mapped into a larger space may
5011 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5012 * check the return value.
5014 * This function only works with pager buffers.
5017 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5022 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5023 prot = VM_PROT_READ;
5024 if (bp->b_iocmd == BIO_READ)
5025 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5026 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5027 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5030 bp->b_bufsize = len;
5031 bp->b_npages = pidx;
5032 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5033 if (mapbuf || !unmapped_buf_allowed) {
5034 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5035 bp->b_data = bp->b_kvabase + bp->b_offset;
5037 bp->b_data = unmapped_buf;
5042 * Free the io map PTEs associated with this IO operation.
5043 * We also invalidate the TLB entries and restore the original b_addr.
5045 * This function only works with pager buffers.
5048 vunmapbuf(struct buf *bp)
5052 npages = bp->b_npages;
5054 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5055 vm_page_unhold_pages(bp->b_pages, npages);
5057 bp->b_data = unmapped_buf;
5061 bdone(struct buf *bp)
5065 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5067 bp->b_flags |= B_DONE;
5073 bwait(struct buf *bp, u_char pri, const char *wchan)
5077 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5079 while ((bp->b_flags & B_DONE) == 0)
5080 msleep(bp, mtxp, pri, wchan, 0);
5085 bufsync(struct bufobj *bo, int waitfor)
5088 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5092 bufstrategy(struct bufobj *bo, struct buf *bp)
5098 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5099 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5100 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5101 i = VOP_STRATEGY(vp, bp);
5102 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5106 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5109 bufobj_init(struct bufobj *bo, void *private)
5111 static volatile int bufobj_cleanq;
5114 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5115 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5116 bo->bo_private = private;
5117 TAILQ_INIT(&bo->bo_clean.bv_hd);
5118 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5122 bufobj_wrefl(struct bufobj *bo)
5125 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5126 ASSERT_BO_WLOCKED(bo);
5131 bufobj_wref(struct bufobj *bo)
5134 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5141 bufobj_wdrop(struct bufobj *bo)
5144 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5146 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5147 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5148 bo->bo_flag &= ~BO_WWAIT;
5149 wakeup(&bo->bo_numoutput);
5155 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5159 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5160 ASSERT_BO_WLOCKED(bo);
5162 while (bo->bo_numoutput) {
5163 bo->bo_flag |= BO_WWAIT;
5164 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5165 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5173 * Set bio_data or bio_ma for struct bio from the struct buf.
5176 bdata2bio(struct buf *bp, struct bio *bip)
5179 if (!buf_mapped(bp)) {
5180 KASSERT(unmapped_buf_allowed, ("unmapped"));
5181 bip->bio_ma = bp->b_pages;
5182 bip->bio_ma_n = bp->b_npages;
5183 bip->bio_data = unmapped_buf;
5184 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5185 bip->bio_flags |= BIO_UNMAPPED;
5186 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5187 PAGE_SIZE == bp->b_npages,
5188 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5189 (long long)bip->bio_length, bip->bio_ma_n));
5191 bip->bio_data = bp->b_data;
5197 * The MIPS pmap code currently doesn't handle aliased pages.
5198 * The VIPT caches may not handle page aliasing themselves, leading
5199 * to data corruption.
5201 * As such, this code makes a system extremely unhappy if said
5202 * system doesn't support unaliasing the above situation in hardware.
5203 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5204 * this feature at build time, so it has to be handled in software.
5206 * Once the MIPS pmap/cache code grows to support this function on
5207 * earlier chips, it should be flipped back off.
5210 static int buf_pager_relbuf = 1;
5212 static int buf_pager_relbuf = 0;
5214 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5215 &buf_pager_relbuf, 0,
5216 "Make buffer pager release buffers after reading");
5219 * The buffer pager. It uses buffer reads to validate pages.
5221 * In contrast to the generic local pager from vm/vnode_pager.c, this
5222 * pager correctly and easily handles volumes where the underlying
5223 * device block size is greater than the machine page size. The
5224 * buffer cache transparently extends the requested page run to be
5225 * aligned at the block boundary, and does the necessary bogus page
5226 * replacements in the addends to avoid obliterating already valid
5229 * The only non-trivial issue is that the exclusive busy state for
5230 * pages, which is assumed by the vm_pager_getpages() interface, is
5231 * incompatible with the VMIO buffer cache's desire to share-busy the
5232 * pages. This function performs a trivial downgrade of the pages'
5233 * state before reading buffers, and a less trivial upgrade from the
5234 * shared-busy to excl-busy state after the read.
5237 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5238 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5239 vbg_get_blksize_t get_blksize)
5246 vm_ooffset_t la, lb, poff, poffe;
5248 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5251 object = vp->v_object;
5254 la = IDX_TO_OFF(ma[count - 1]->pindex);
5255 if (la >= object->un_pager.vnp.vnp_size)
5256 return (VM_PAGER_BAD);
5259 * Change the meaning of la from where the last requested page starts
5260 * to where it ends, because that's the end of the requested region
5261 * and the start of the potential read-ahead region.
5264 lpart = la > object->un_pager.vnp.vnp_size;
5265 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5268 return (VM_PAGER_ERROR);
5271 * Calculate read-ahead, behind and total pages.
5274 lb = IDX_TO_OFF(ma[0]->pindex);
5275 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5277 if (rbehind != NULL)
5279 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5280 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5281 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5286 VM_CNT_INC(v_vnodein);
5287 VM_CNT_ADD(v_vnodepgsin, pgsin);
5289 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5290 != 0) ? GB_UNMAPPED : 0;
5292 for (i = 0; i < count; i++) {
5293 if (ma[i] != bogus_page)
5294 vm_page_busy_downgrade(ma[i]);
5298 for (i = 0; i < count; i++) {
5300 if (m == bogus_page)
5304 * Pages are shared busy and the object lock is not
5305 * owned, which together allow for the pages'
5306 * invalidation. The racy test for validity avoids
5307 * useless creation of the buffer for the most typical
5308 * case when invalidation is not used in redo or for
5309 * parallel read. The shared->excl upgrade loop at
5310 * the end of the function catches the race in a
5311 * reliable way (protected by the object lock).
5313 if (vm_page_all_valid(m))
5316 poff = IDX_TO_OFF(m->pindex);
5317 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5318 for (; poff < poffe; poff += bsize) {
5319 lbn = get_lblkno(vp, poff);
5324 error = get_blksize(vp, lbn, &bsize);
5326 error = bread_gb(vp, lbn, bsize,
5327 curthread->td_ucred, br_flags, &bp);
5330 if (bp->b_rcred == curthread->td_ucred) {
5331 crfree(bp->b_rcred);
5332 bp->b_rcred = NOCRED;
5334 if (LIST_EMPTY(&bp->b_dep)) {
5336 * Invalidation clears m->valid, but
5337 * may leave B_CACHE flag if the
5338 * buffer existed at the invalidation
5339 * time. In this case, recycle the
5340 * buffer to do real read on next
5341 * bread() after redo.
5343 * Otherwise B_RELBUF is not strictly
5344 * necessary, enable to reduce buf
5347 if (buf_pager_relbuf ||
5348 !vm_page_all_valid(m))
5349 bp->b_flags |= B_RELBUF;
5351 bp->b_flags &= ~B_NOCACHE;
5357 KASSERT(1 /* racy, enable for debugging */ ||
5358 vm_page_all_valid(m) || i == count - 1,
5359 ("buf %d %p invalid", i, m));
5360 if (i == count - 1 && lpart) {
5361 if (!vm_page_none_valid(m) &&
5362 !vm_page_all_valid(m))
5363 vm_page_zero_invalid(m, TRUE);
5370 for (i = 0; i < count; i++) {
5371 if (ma[i] == bogus_page)
5373 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5374 vm_page_sunbusy(ma[i]);
5375 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5380 * Since the pages were only sbusy while neither the
5381 * buffer nor the object lock was held by us, or
5382 * reallocated while vm_page_grab() slept for busy
5383 * relinguish, they could have been invalidated.
5384 * Recheck the valid bits and re-read as needed.
5386 * Note that the last page is made fully valid in the
5387 * read loop, and partial validity for the page at
5388 * index count - 1 could mean that the page was
5389 * invalidated or removed, so we must restart for
5392 if (!vm_page_all_valid(ma[i]))
5395 if (redo && error == 0)
5397 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5400 #include "opt_ddb.h"
5402 #include <ddb/ddb.h>
5404 /* DDB command to show buffer data */
5405 DB_SHOW_COMMAND(buffer, db_show_buffer)
5408 struct buf *bp = (struct buf *)addr;
5409 #ifdef FULL_BUF_TRACKING
5414 db_printf("usage: show buffer <addr>\n");
5418 db_printf("buf at %p\n", bp);
5419 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5420 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5421 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5422 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5423 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5424 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5426 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5427 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5428 "b_vp = %p, b_dep = %p\n",
5429 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5430 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5431 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5432 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5433 bp->b_kvabase, bp->b_kvasize);
5436 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5437 for (i = 0; i < bp->b_npages; i++) {
5441 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5443 (u_long)VM_PAGE_TO_PHYS(m));
5445 db_printf("( ??? )");
5446 if ((i + 1) < bp->b_npages)
5451 BUF_LOCKPRINTINFO(bp);
5452 #if defined(FULL_BUF_TRACKING)
5453 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5455 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5456 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5457 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5459 db_printf(" %2u: %s\n", j,
5460 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5462 #elif defined(BUF_TRACKING)
5463 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5468 DB_SHOW_COMMAND(bufqueues, bufqueues)
5470 struct bufdomain *bd;
5475 db_printf("bqempty: %d\n", bqempty.bq_len);
5477 for (i = 0; i < buf_domains; i++) {
5479 db_printf("Buf domain %d\n", i);
5480 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5481 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5482 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5484 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5485 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5486 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5487 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5488 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5490 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5491 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5492 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5493 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5496 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5497 total += bp->b_bufsize;
5498 db_printf("\tcleanq count\t%d (%ld)\n",
5499 bd->bd_cleanq->bq_len, total);
5501 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5502 total += bp->b_bufsize;
5503 db_printf("\tdirtyq count\t%d (%ld)\n",
5504 bd->bd_dirtyq.bq_len, total);
5505 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5506 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5507 db_printf("\tCPU ");
5508 for (j = 0; j <= mp_maxid; j++)
5509 db_printf("%d, ", bd->bd_subq[j].bq_len);
5513 for (j = 0; j < nbuf; j++) {
5515 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5517 total += bp->b_bufsize;
5520 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5523 for (j = 0; j < nbuf; j++) {
5525 if (bp->b_domain == i) {
5527 total += bp->b_bufsize;
5530 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5534 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5539 for (i = 0; i < nbuf; i++) {
5541 if (BUF_ISLOCKED(bp)) {
5542 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5550 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5556 db_printf("usage: show vnodebufs <addr>\n");
5559 vp = (struct vnode *)addr;
5560 db_printf("Clean buffers:\n");
5561 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5562 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5565 db_printf("Dirty buffers:\n");
5566 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5567 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5572 DB_COMMAND(countfreebufs, db_coundfreebufs)
5575 int i, used = 0, nfree = 0;
5578 db_printf("usage: countfreebufs\n");
5582 for (i = 0; i < nbuf; i++) {
5584 if (bp->b_qindex == QUEUE_EMPTY)
5590 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5592 db_printf("numfreebuffers is %d\n", numfreebuffers);