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;
768 bd->bd_shutdown = true;
769 wakeup(&bd->bd_running);
770 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0,
771 "bufspace_shutdown", 60 * hz);
774 printf("bufspacedaemon wait error: %d\n", error);
780 * buffer space management daemon. Tries to maintain some marginal
781 * amount of free buffer space so that requesting processes neither
782 * block nor work to reclaim buffers.
785 bufspace_daemon(void *arg)
787 struct bufdomain *bd = arg;
789 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd,
790 SHUTDOWN_PRI_LAST + 100);
793 while (!bd->bd_shutdown) {
797 * Free buffers from the clean queue until we meet our
800 * Theory of operation: The buffer cache is most efficient
801 * when some free buffer headers and space are always
802 * available to getnewbuf(). This daemon attempts to prevent
803 * the excessive blocking and synchronization associated
804 * with shortfall. It goes through three phases according
807 * 1) The daemon wakes up voluntarily once per-second
808 * during idle periods when the counters are below
809 * the wakeup thresholds (bufspacethresh, lofreebuffers).
811 * 2) The daemon wakes up as we cross the thresholds
812 * ahead of any potential blocking. This may bounce
813 * slightly according to the rate of consumption and
816 * 3) The daemon and consumers are starved for working
817 * clean buffers. This is the 'bufspace' sleep below
818 * which will inefficiently trade bufs with bqrelse
819 * until we return to condition 2.
821 while (bd->bd_bufspace > bd->bd_lobufspace ||
822 bd->bd_freebuffers < bd->bd_hifreebuffers) {
823 if (buf_recycle(bd, false) != 0) {
827 * Speedup dirty if we've run out of clean
828 * buffers. This is possible in particular
829 * because softdep may held many bufs locked
830 * pending writes to other bufs which are
831 * marked for delayed write, exhausting
832 * clean space until they are written.
837 msleep(&bd->bd_wanted, BD_LOCKPTR(bd),
838 PRIBIO|PDROP, "bufspace", hz/10);
846 * Re-check our limits and sleep. bd_running must be
847 * cleared prior to checking the limits to avoid missed
848 * wakeups. The waker will adjust one of bufspace or
849 * freebuffers prior to checking bd_running.
854 atomic_store_int(&bd->bd_running, 0);
855 if (bd->bd_bufspace < bd->bd_bufspacethresh &&
856 bd->bd_freebuffers > bd->bd_lofreebuffers) {
857 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd),
860 /* Avoid spurious wakeups while running. */
861 atomic_store_int(&bd->bd_running, 1);
864 wakeup(&bd->bd_shutdown);
872 * Adjust the reported bufspace for a malloc managed buffer, possibly
873 * waking any waiters.
876 bufmallocadjust(struct buf *bp, int bufsize)
880 KASSERT((bp->b_flags & B_MALLOC) != 0,
881 ("bufmallocadjust: non-malloc buf %p", bp));
882 diff = bufsize - bp->b_bufsize;
884 atomic_subtract_long(&bufmallocspace, -diff);
886 atomic_add_long(&bufmallocspace, diff);
887 bp->b_bufsize = bufsize;
893 * Wake up processes that are waiting on asynchronous writes to fall
894 * below lorunningspace.
900 mtx_lock(&rbreqlock);
903 wakeup(&runningbufreq);
905 mtx_unlock(&rbreqlock);
911 * Decrement the outstanding write count according.
914 runningbufwakeup(struct buf *bp)
918 bspace = bp->b_runningbufspace;
921 space = atomic_fetchadd_long(&runningbufspace, -bspace);
922 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
924 bp->b_runningbufspace = 0;
926 * Only acquire the lock and wakeup on the transition from exceeding
927 * the threshold to falling below it.
929 if (space < lorunningspace)
931 if (space - bspace > lorunningspace)
937 * waitrunningbufspace()
939 * runningbufspace is a measure of the amount of I/O currently
940 * running. This routine is used in async-write situations to
941 * prevent creating huge backups of pending writes to a device.
942 * Only asynchronous writes are governed by this function.
944 * This does NOT turn an async write into a sync write. It waits
945 * for earlier writes to complete and generally returns before the
946 * caller's write has reached the device.
949 waitrunningbufspace(void)
952 mtx_lock(&rbreqlock);
953 while (runningbufspace > hirunningspace) {
955 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
957 mtx_unlock(&rbreqlock);
961 * vfs_buf_test_cache:
963 * Called when a buffer is extended. This function clears the B_CACHE
964 * bit if the newly extended portion of the buffer does not contain
968 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off,
969 vm_offset_t size, vm_page_t m)
973 * This function and its results are protected by higher level
974 * synchronization requiring vnode and buf locks to page in and
977 if (bp->b_flags & B_CACHE) {
978 int base = (foff + off) & PAGE_MASK;
979 if (vm_page_is_valid(m, base, size) == 0)
980 bp->b_flags &= ~B_CACHE;
984 /* Wake up the buffer daemon if necessary */
990 if (bd_request == 0) {
998 * Adjust the maxbcachbuf tunable.
1001 maxbcachebuf_adjust(void)
1006 * maxbcachebuf must be a power of 2 >= MAXBSIZE.
1009 while (i * 2 <= maxbcachebuf)
1012 if (maxbcachebuf < MAXBSIZE)
1013 maxbcachebuf = MAXBSIZE;
1014 if (maxbcachebuf > maxphys)
1015 maxbcachebuf = maxphys;
1016 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF)
1017 printf("maxbcachebuf=%d\n", maxbcachebuf);
1021 * bd_speedup - speedup the buffer cache flushing code
1030 if (bd_speedupreq == 0 || bd_request == 0)
1035 wakeup(&bd_request);
1036 mtx_unlock(&bdlock);
1040 #define TRANSIENT_DENOM 5
1042 #define TRANSIENT_DENOM 10
1046 * Calculating buffer cache scaling values and reserve space for buffer
1047 * headers. This is called during low level kernel initialization and
1048 * may be called more then once. We CANNOT write to the memory area
1049 * being reserved at this time.
1052 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
1055 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
1059 * With KASAN enabled, the kernel map is shadowed. Account for this
1060 * when sizing maps based on the amount of physical memory available.
1062 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) /
1063 (KASAN_SHADOW_SCALE + 1);
1067 * physmem_est is in pages. Convert it to kilobytes (assumes
1068 * PAGE_SIZE is >= 1K)
1070 physmem_est = physmem_est * (PAGE_SIZE / 1024);
1072 maxbcachebuf_adjust();
1074 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
1075 * For the first 64MB of ram nominally allocate sufficient buffers to
1076 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
1077 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
1078 * the buffer cache we limit the eventual kva reservation to
1081 * factor represents the 1/4 x ram conversion.
1084 int factor = 4 * BKVASIZE / 1024;
1087 if (physmem_est > 4096)
1088 nbuf += min((physmem_est - 4096) / factor,
1090 if (physmem_est > 65536)
1091 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
1092 32 * 1024 * 1024 / (factor * 5));
1094 if (maxbcache && nbuf > maxbcache / BKVASIZE)
1095 nbuf = maxbcache / BKVASIZE;
1100 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
1101 maxbuf = (LONG_MAX / 3) / BKVASIZE;
1102 if (nbuf > maxbuf) {
1104 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
1110 * Ideal allocation size for the transient bio submap is 10%
1111 * of the maximal space buffer map. This roughly corresponds
1112 * to the amount of the buffer mapped for typical UFS load.
1114 * Clip the buffer map to reserve space for the transient
1115 * BIOs, if its extent is bigger than 90% (80% on i386) of the
1116 * maximum buffer map extent on the platform.
1118 * The fall-back to the maxbuf in case of maxbcache unset,
1119 * allows to not trim the buffer KVA for the architectures
1120 * with ample KVA space.
1122 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
1123 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
1124 buf_sz = (long)nbuf * BKVASIZE;
1125 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
1126 (TRANSIENT_DENOM - 1)) {
1128 * There is more KVA than memory. Do not
1129 * adjust buffer map size, and assign the rest
1130 * of maxbuf to transient map.
1132 biotmap_sz = maxbuf_sz - buf_sz;
1135 * Buffer map spans all KVA we could afford on
1136 * this platform. Give 10% (20% on i386) of
1137 * the buffer map to the transient bio map.
1139 biotmap_sz = buf_sz / TRANSIENT_DENOM;
1140 buf_sz -= biotmap_sz;
1142 if (biotmap_sz / INT_MAX > maxphys)
1143 bio_transient_maxcnt = INT_MAX;
1145 bio_transient_maxcnt = biotmap_sz / maxphys;
1147 * Artificially limit to 1024 simultaneous in-flight I/Os
1148 * using the transient mapping.
1150 if (bio_transient_maxcnt > 1024)
1151 bio_transient_maxcnt = 1024;
1153 nbuf = buf_sz / BKVASIZE;
1157 nswbuf = min(nbuf / 4, 256);
1158 if (nswbuf < NSWBUF_MIN)
1159 nswbuf = NSWBUF_MIN;
1163 * Reserve space for the buffer cache buffers
1166 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) *
1167 atop(maxbcachebuf)) * nbuf;
1173 * Single global constant for BUF_WMESG, to avoid getting multiple
1176 static const char buf_wmesg[] = "bufwait";
1178 /* Initialize the buffer subsystem. Called before use of any buffers. */
1185 KASSERT(maxbcachebuf >= MAXBSIZE,
1186 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf,
1188 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock");
1189 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
1190 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
1191 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
1193 unmapped_buf = (caddr_t)kva_alloc(maxphys);
1195 /* finally, initialize each buffer header and stick on empty q */
1196 for (i = 0; i < nbuf; i++) {
1198 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf));
1199 bp->b_flags = B_INVAL;
1200 bp->b_rcred = NOCRED;
1201 bp->b_wcred = NOCRED;
1202 bp->b_qindex = QUEUE_NONE;
1204 bp->b_subqueue = mp_maxid + 1;
1206 bp->b_data = bp->b_kvabase = unmapped_buf;
1207 LIST_INIT(&bp->b_dep);
1208 BUF_LOCKINIT(bp, buf_wmesg);
1209 bq_insert(&bqempty, bp, false);
1213 * maxbufspace is the absolute maximum amount of buffer space we are
1214 * allowed to reserve in KVM and in real terms. The absolute maximum
1215 * is nominally used by metadata. hibufspace is the nominal maximum
1216 * used by most other requests. The differential is required to
1217 * ensure that metadata deadlocks don't occur.
1219 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
1220 * this may result in KVM fragmentation which is not handled optimally
1221 * by the system. XXX This is less true with vmem. We could use
1224 maxbufspace = (long)nbuf * BKVASIZE;
1225 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10);
1226 lobufspace = (hibufspace / 20) * 19; /* 95% */
1227 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2;
1230 * Note: The 16 MiB upper limit for hirunningspace was chosen
1231 * arbitrarily and may need further tuning. It corresponds to
1232 * 128 outstanding write IO requests (if IO size is 128 KiB),
1233 * which fits with many RAID controllers' tagged queuing limits.
1234 * The lower 1 MiB limit is the historical upper limit for
1237 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf),
1238 16 * 1024 * 1024), 1024 * 1024);
1239 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf);
1242 * Limit the amount of malloc memory since it is wired permanently into
1243 * the kernel space. Even though this is accounted for in the buffer
1244 * allocation, we don't want the malloced region to grow uncontrolled.
1245 * The malloc scheme improves memory utilization significantly on
1246 * average (small) directories.
1248 maxbufmallocspace = hibufspace / 20;
1251 * Reduce the chance of a deadlock occurring by limiting the number
1252 * of delayed-write dirty buffers we allow to stack up.
1254 hidirtybuffers = nbuf / 4 + 20;
1255 dirtybufthresh = hidirtybuffers * 9 / 10;
1257 * To support extreme low-memory systems, make sure hidirtybuffers
1258 * cannot eat up all available buffer space. This occurs when our
1259 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our
1260 * buffer space assuming BKVASIZE'd buffers.
1262 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
1263 hidirtybuffers >>= 1;
1265 lodirtybuffers = hidirtybuffers / 2;
1268 * lofreebuffers should be sufficient to avoid stalling waiting on
1269 * buf headers under heavy utilization. The bufs in per-cpu caches
1270 * are counted as free but will be unavailable to threads executing
1273 * hifreebuffers is the free target for the bufspace daemon. This
1274 * should be set appropriately to limit work per-iteration.
1276 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus);
1277 hifreebuffers = (3 * lofreebuffers) / 2;
1278 numfreebuffers = nbuf;
1280 /* Setup the kva and free list allocators. */
1281 vmem_set_reclaim(buffer_arena, bufkva_reclaim);
1282 buf_zone = uma_zcache_create("buf free cache",
1283 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf),
1284 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0);
1287 * Size the clean queue according to the amount of buffer space.
1288 * One queue per-256mb up to the max. More queues gives better
1289 * concurrency but less accurate LRU.
1291 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS);
1292 for (i = 0 ; i < buf_domains; i++) {
1293 struct bufdomain *bd;
1297 bd->bd_freebuffers = nbuf / buf_domains;
1298 bd->bd_hifreebuffers = hifreebuffers / buf_domains;
1299 bd->bd_lofreebuffers = lofreebuffers / buf_domains;
1300 bd->bd_bufspace = 0;
1301 bd->bd_maxbufspace = maxbufspace / buf_domains;
1302 bd->bd_hibufspace = hibufspace / buf_domains;
1303 bd->bd_lobufspace = lobufspace / buf_domains;
1304 bd->bd_bufspacethresh = bufspacethresh / buf_domains;
1305 bd->bd_numdirtybuffers = 0;
1306 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains;
1307 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains;
1308 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains;
1309 /* Don't allow more than 2% of bufs in the per-cpu caches. */
1310 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus;
1312 getnewbufcalls = counter_u64_alloc(M_WAITOK);
1313 getnewbufrestarts = counter_u64_alloc(M_WAITOK);
1314 mappingrestarts = counter_u64_alloc(M_WAITOK);
1315 numbufallocfails = counter_u64_alloc(M_WAITOK);
1316 notbufdflushes = counter_u64_alloc(M_WAITOK);
1317 buffreekvacnt = counter_u64_alloc(M_WAITOK);
1318 bufdefragcnt = counter_u64_alloc(M_WAITOK);
1319 bufkvaspace = counter_u64_alloc(M_WAITOK);
1324 vfs_buf_check_mapped(struct buf *bp)
1327 KASSERT(bp->b_kvabase != unmapped_buf,
1328 ("mapped buf: b_kvabase was not updated %p", bp));
1329 KASSERT(bp->b_data != unmapped_buf,
1330 ("mapped buf: b_data was not updated %p", bp));
1331 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf +
1332 maxphys, ("b_data + b_offset unmapped %p", bp));
1336 vfs_buf_check_unmapped(struct buf *bp)
1339 KASSERT(bp->b_data == unmapped_buf,
1340 ("unmapped buf: corrupted b_data %p", bp));
1343 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
1344 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
1346 #define BUF_CHECK_MAPPED(bp) do {} while (0)
1347 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
1351 isbufbusy(struct buf *bp)
1353 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) ||
1354 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI))
1360 * Shutdown the system cleanly to prepare for reboot, halt, or power off.
1363 bufshutdown(int show_busybufs)
1365 static int first_buf_printf = 1;
1367 int i, iter, nbusy, pbusy;
1373 * Sync filesystems for shutdown
1375 wdog_kern_pat(WD_LASTVAL);
1376 kern_sync(curthread);
1379 * With soft updates, some buffers that are
1380 * written will be remarked as dirty until other
1381 * buffers are written.
1383 for (iter = pbusy = 0; iter < 20; iter++) {
1385 for (i = nbuf - 1; i >= 0; i--) {
1391 if (first_buf_printf)
1392 printf("All buffers synced.");
1395 if (first_buf_printf) {
1396 printf("Syncing disks, buffers remaining... ");
1397 first_buf_printf = 0;
1399 printf("%d ", nbusy);
1404 wdog_kern_pat(WD_LASTVAL);
1405 kern_sync(curthread);
1409 * Spin for a while to allow interrupt threads to run.
1411 DELAY(50000 * iter);
1414 * Context switch several times to allow interrupt
1417 for (subiter = 0; subiter < 50 * iter; subiter++) {
1418 thread_lock(curthread);
1426 * Count only busy local buffers to prevent forcing
1427 * a fsck if we're just a client of a wedged NFS server
1430 for (i = nbuf - 1; i >= 0; i--) {
1432 if (isbufbusy(bp)) {
1434 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */
1435 if (bp->b_dev == NULL) {
1436 TAILQ_REMOVE(&mountlist,
1437 bp->b_vp->v_mount, mnt_list);
1442 if (show_busybufs > 0) {
1444 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:",
1445 nbusy, bp, bp->b_vp, bp->b_flags,
1446 (intmax_t)bp->b_blkno,
1447 (intmax_t)bp->b_lblkno);
1448 BUF_LOCKPRINTINFO(bp);
1449 if (show_busybufs > 1)
1457 * Failed to sync all blocks. Indicate this and don't
1458 * unmount filesystems (thus forcing an fsck on reboot).
1460 printf("Giving up on %d buffers\n", nbusy);
1461 DELAY(5000000); /* 5 seconds */
1464 if (!first_buf_printf)
1465 printf("Final sync complete\n");
1468 * Unmount filesystems and perform swapoff, to quiesce
1469 * the system as much as possible. In particular, no
1470 * I/O should be initiated from top levels since it
1471 * might be abruptly terminated by reset, or otherwise
1472 * erronously handled because other parts of the
1473 * system are disabled.
1475 * Swapoff before unmount, because file-backed swap is
1476 * non-operational after unmount of the underlying
1479 if (!KERNEL_PANICKED()) {
1484 DELAY(100000); /* wait for console output to finish */
1488 bpmap_qenter(struct buf *bp)
1491 BUF_CHECK_MAPPED(bp);
1494 * bp->b_data is relative to bp->b_offset, but
1495 * bp->b_offset may be offset into the first page.
1497 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
1498 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
1499 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
1500 (vm_offset_t)(bp->b_offset & PAGE_MASK));
1503 static inline struct bufdomain *
1504 bufdomain(struct buf *bp)
1507 return (&bdomain[bp->b_domain]);
1510 static struct bufqueue *
1511 bufqueue(struct buf *bp)
1514 switch (bp->b_qindex) {
1517 case QUEUE_SENTINEL:
1522 return (&bufdomain(bp)->bd_dirtyq);
1524 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]);
1528 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex);
1532 * Return the locked bufqueue that bp is a member of.
1534 static struct bufqueue *
1535 bufqueue_acquire(struct buf *bp)
1537 struct bufqueue *bq, *nbq;
1540 * bp can be pushed from a per-cpu queue to the
1541 * cleanq while we're waiting on the lock. Retry
1542 * if the queues don't match.
1560 * Insert the buffer into the appropriate free list. Requires a
1561 * locked buffer on entry and buffer is unlocked before return.
1564 binsfree(struct buf *bp, int qindex)
1566 struct bufdomain *bd;
1567 struct bufqueue *bq;
1569 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY,
1570 ("binsfree: Invalid qindex %d", qindex));
1571 BUF_ASSERT_XLOCKED(bp);
1574 * Handle delayed bremfree() processing.
1576 if (bp->b_flags & B_REMFREE) {
1577 if (bp->b_qindex == qindex) {
1578 bp->b_flags |= B_REUSE;
1579 bp->b_flags &= ~B_REMFREE;
1583 bq = bufqueue_acquire(bp);
1588 if (qindex == QUEUE_CLEAN) {
1589 if (bd->bd_lim != 0)
1590 bq = &bd->bd_subq[PCPU_GET(cpuid)];
1594 bq = &bd->bd_dirtyq;
1595 bq_insert(bq, bp, true);
1601 * Free a buffer to the buf zone once it no longer has valid contents.
1604 buf_free(struct buf *bp)
1607 if (bp->b_flags & B_REMFREE)
1609 if (bp->b_vflags & BV_BKGRDINPROG)
1610 panic("losing buffer 1");
1611 if (bp->b_rcred != NOCRED) {
1612 crfree(bp->b_rcred);
1613 bp->b_rcred = NOCRED;
1615 if (bp->b_wcred != NOCRED) {
1616 crfree(bp->b_wcred);
1617 bp->b_wcred = NOCRED;
1619 if (!LIST_EMPTY(&bp->b_dep))
1622 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1);
1623 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1625 uma_zfree(buf_zone, bp);
1631 * Import bufs into the uma cache from the buf list. The system still
1632 * expects a static array of bufs and much of the synchronization
1633 * around bufs assumes type stable storage. As a result, UMA is used
1634 * only as a per-cpu cache of bufs still maintained on a global list.
1637 buf_import(void *arg, void **store, int cnt, int domain, int flags)
1643 for (i = 0; i < cnt; i++) {
1644 bp = TAILQ_FIRST(&bqempty.bq_queue);
1647 bq_remove(&bqempty, bp);
1650 BQ_UNLOCK(&bqempty);
1658 * Release bufs from the uma cache back to the buffer queues.
1661 buf_release(void *arg, void **store, int cnt)
1663 struct bufqueue *bq;
1669 for (i = 0; i < cnt; i++) {
1671 /* Inline bq_insert() to batch locking. */
1672 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1673 bp->b_flags &= ~(B_AGE | B_REUSE);
1675 bp->b_qindex = bq->bq_index;
1683 * Allocate an empty buffer header.
1686 buf_alloc(struct bufdomain *bd)
1689 int freebufs, error;
1692 * We can only run out of bufs in the buf zone if the average buf
1693 * is less than BKVASIZE. In this case the actual wait/block will
1694 * come from buf_reycle() failing to flush one of these small bufs.
1697 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1);
1699 bp = uma_zalloc(buf_zone, M_NOWAIT);
1701 atomic_add_int(&bd->bd_freebuffers, 1);
1702 bufspace_daemon_wakeup(bd);
1703 counter_u64_add(numbufallocfails, 1);
1707 * Wake-up the bufspace daemon on transition below threshold.
1709 if (freebufs == bd->bd_lofreebuffers)
1710 bufspace_daemon_wakeup(bd);
1712 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL);
1713 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp,
1717 KASSERT(bp->b_vp == NULL,
1718 ("bp: %p still has vnode %p.", bp, bp->b_vp));
1719 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0,
1720 ("invalid buffer %p flags %#x", bp, bp->b_flags));
1721 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1722 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
1723 KASSERT(bp->b_npages == 0,
1724 ("bp: %p still has %d vm pages\n", bp, bp->b_npages));
1725 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp));
1726 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp));
1727 MPASS((bp->b_flags & B_MAXPHYS) == 0);
1729 bp->b_domain = BD_DOMAIN(bd);
1735 bp->b_blkno = bp->b_lblkno = 0;
1736 bp->b_offset = NOOFFSET;
1742 bp->b_dirtyoff = bp->b_dirtyend = 0;
1743 bp->b_bufobj = NULL;
1744 bp->b_data = bp->b_kvabase = unmapped_buf;
1745 bp->b_fsprivate1 = NULL;
1746 bp->b_fsprivate2 = NULL;
1747 bp->b_fsprivate3 = NULL;
1748 LIST_INIT(&bp->b_dep);
1756 * Free a buffer from the given bufqueue. kva controls whether the
1757 * freed buf must own some kva resources. This is used for
1761 buf_recycle(struct bufdomain *bd, bool kva)
1763 struct bufqueue *bq;
1764 struct buf *bp, *nbp;
1767 counter_u64_add(bufdefragcnt, 1);
1771 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd),
1772 ("buf_recycle: Locks don't match"));
1773 nbp = TAILQ_FIRST(&bq->bq_queue);
1776 * Run scan, possibly freeing data and/or kva mappings on the fly
1779 while ((bp = nbp) != NULL) {
1781 * Calculate next bp (we can only use it if we do not
1782 * release the bqlock).
1784 nbp = TAILQ_NEXT(bp, b_freelist);
1787 * If we are defragging then we need a buffer with
1788 * some kva to reclaim.
1790 if (kva && bp->b_kvasize == 0)
1793 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1797 * Implement a second chance algorithm for frequently
1800 if ((bp->b_flags & B_REUSE) != 0) {
1801 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1802 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
1803 bp->b_flags &= ~B_REUSE;
1809 * Skip buffers with background writes in progress.
1811 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
1816 KASSERT(bp->b_qindex == QUEUE_CLEAN,
1817 ("buf_recycle: inconsistent queue %d bp %p",
1819 KASSERT(bp->b_domain == BD_DOMAIN(bd),
1820 ("getnewbuf: queue domain %d doesn't match request %d",
1821 bp->b_domain, (int)BD_DOMAIN(bd)));
1823 * NOTE: nbp is now entirely invalid. We can only restart
1824 * the scan from this point on.
1830 * Requeue the background write buffer with error and
1833 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
1836 nbp = TAILQ_FIRST(&bq->bq_queue);
1839 bp->b_flags |= B_INVAL;
1852 * Mark the buffer for removal from the appropriate free list.
1856 bremfree(struct buf *bp)
1859 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1860 KASSERT((bp->b_flags & B_REMFREE) == 0,
1861 ("bremfree: buffer %p already marked for delayed removal.", bp));
1862 KASSERT(bp->b_qindex != QUEUE_NONE,
1863 ("bremfree: buffer %p not on a queue.", bp));
1864 BUF_ASSERT_XLOCKED(bp);
1866 bp->b_flags |= B_REMFREE;
1872 * Force an immediate removal from a free list. Used only in nfs when
1873 * it abuses the b_freelist pointer.
1876 bremfreef(struct buf *bp)
1878 struct bufqueue *bq;
1880 bq = bufqueue_acquire(bp);
1886 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname)
1889 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF);
1890 TAILQ_INIT(&bq->bq_queue);
1892 bq->bq_index = qindex;
1893 bq->bq_subqueue = subqueue;
1897 bd_init(struct bufdomain *bd)
1901 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1];
1902 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock");
1903 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock");
1904 for (i = 0; i <= mp_maxid; i++)
1905 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i,
1906 "bufq clean subqueue lock");
1907 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF);
1913 * Removes a buffer from the free list, must be called with the
1914 * correct qlock held.
1917 bq_remove(struct bufqueue *bq, struct buf *bp)
1920 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X",
1921 bp, bp->b_vp, bp->b_flags);
1922 KASSERT(bp->b_qindex != QUEUE_NONE,
1923 ("bq_remove: buffer %p not on a queue.", bp));
1924 KASSERT(bufqueue(bp) == bq,
1925 ("bq_remove: Remove buffer %p from wrong queue.", bp));
1927 BQ_ASSERT_LOCKED(bq);
1928 if (bp->b_qindex != QUEUE_EMPTY) {
1929 BUF_ASSERT_XLOCKED(bp);
1931 KASSERT(bq->bq_len >= 1,
1932 ("queue %d underflow", bp->b_qindex));
1933 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1935 bp->b_qindex = QUEUE_NONE;
1936 bp->b_flags &= ~(B_REMFREE | B_REUSE);
1940 bd_flush(struct bufdomain *bd, struct bufqueue *bq)
1944 BQ_ASSERT_LOCKED(bq);
1945 if (bq != bd->bd_cleanq) {
1947 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) {
1948 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist);
1949 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp,
1951 bp->b_subqueue = bd->bd_cleanq->bq_subqueue;
1953 bd->bd_cleanq->bq_len += bq->bq_len;
1956 if (bd->bd_wanted) {
1958 wakeup(&bd->bd_wanted);
1960 if (bq != bd->bd_cleanq)
1965 bd_flushall(struct bufdomain *bd)
1967 struct bufqueue *bq;
1971 if (bd->bd_lim == 0)
1974 for (i = 0; i <= mp_maxid; i++) {
1975 bq = &bd->bd_subq[i];
1976 if (bq->bq_len == 0)
1988 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock)
1990 struct bufdomain *bd;
1992 if (bp->b_qindex != QUEUE_NONE)
1993 panic("bq_insert: free buffer %p onto another queue?", bp);
1996 if (bp->b_flags & B_AGE) {
1997 /* Place this buf directly on the real queue. */
1998 if (bq->bq_index == QUEUE_CLEAN)
2001 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist);
2004 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist);
2006 bp->b_flags &= ~(B_AGE | B_REUSE);
2008 bp->b_qindex = bq->bq_index;
2009 bp->b_subqueue = bq->bq_subqueue;
2012 * Unlock before we notify so that we don't wakeup a waiter that
2013 * fails a trylock on the buf and sleeps again.
2018 if (bp->b_qindex == QUEUE_CLEAN) {
2020 * Flush the per-cpu queue and notify any waiters.
2022 if (bd->bd_wanted || (bq != bd->bd_cleanq &&
2023 bq->bq_len >= bd->bd_lim))
2032 * Free the kva allocation for a buffer.
2036 bufkva_free(struct buf *bp)
2040 if (bp->b_kvasize == 0) {
2041 KASSERT(bp->b_kvabase == unmapped_buf &&
2042 bp->b_data == unmapped_buf,
2043 ("Leaked KVA space on %p", bp));
2044 } else if (buf_mapped(bp))
2045 BUF_CHECK_MAPPED(bp);
2047 BUF_CHECK_UNMAPPED(bp);
2049 if (bp->b_kvasize == 0)
2052 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize);
2053 counter_u64_add(bufkvaspace, -bp->b_kvasize);
2054 counter_u64_add(buffreekvacnt, 1);
2055 bp->b_data = bp->b_kvabase = unmapped_buf;
2062 * Allocate the buffer KVA and set b_kvasize and b_kvabase.
2065 bufkva_alloc(struct buf *bp, int maxsize, int gbflags)
2070 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0,
2071 ("Invalid gbflags 0x%x in %s", gbflags, __func__));
2072 MPASS((bp->b_flags & B_MAXPHYS) == 0);
2073 KASSERT(maxsize <= maxbcachebuf,
2074 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf));
2079 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr);
2082 * Buffer map is too fragmented. Request the caller
2083 * to defragment the map.
2087 bp->b_kvabase = (caddr_t)addr;
2088 bp->b_kvasize = maxsize;
2089 counter_u64_add(bufkvaspace, bp->b_kvasize);
2090 if ((gbflags & GB_UNMAPPED) != 0) {
2091 bp->b_data = unmapped_buf;
2092 BUF_CHECK_UNMAPPED(bp);
2094 bp->b_data = bp->b_kvabase;
2095 BUF_CHECK_MAPPED(bp);
2103 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem
2104 * callback that fires to avoid returning failure.
2107 bufkva_reclaim(vmem_t *vmem, int flags)
2114 for (i = 0; i < 5; i++) {
2115 for (q = 0; q < buf_domains; q++)
2116 if (buf_recycle(&bdomain[q], true) != 0)
2125 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
2126 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
2127 * the buffer is valid and we do not have to do anything.
2130 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt,
2131 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *))
2139 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
2140 if (inmem(vp, *rablkno))
2142 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
2143 if ((rabp->b_flags & B_CACHE) != 0) {
2150 racct_add_buf(curproc, rabp, 0);
2151 PROC_UNLOCK(curproc);
2154 td->td_ru.ru_inblock++;
2155 rabp->b_flags |= B_ASYNC;
2156 rabp->b_flags &= ~B_INVAL;
2157 if ((flags & GB_CKHASH) != 0) {
2158 rabp->b_flags |= B_CKHASH;
2159 rabp->b_ckhashcalc = ckhashfunc;
2161 rabp->b_ioflags &= ~BIO_ERROR;
2162 rabp->b_iocmd = BIO_READ;
2163 if (rabp->b_rcred == NOCRED && cred != NOCRED)
2164 rabp->b_rcred = crhold(cred);
2165 vfs_busy_pages(rabp, 0);
2167 rabp->b_iooffset = dbtob(rabp->b_blkno);
2173 * Entry point for bread() and breadn() via #defines in sys/buf.h.
2175 * Get a buffer with the specified data. Look in the cache first. We
2176 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
2177 * is set, the buffer is valid and we do not have to do anything, see
2178 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
2180 * Always return a NULL buffer pointer (in bpp) when returning an error.
2182 * The blkno parameter is the logical block being requested. Normally
2183 * the mapping of logical block number to disk block address is done
2184 * by calling VOP_BMAP(). However, if the mapping is already known, the
2185 * disk block address can be passed using the dblkno parameter. If the
2186 * disk block address is not known, then the same value should be passed
2187 * for blkno and dblkno.
2190 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size,
2191 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags,
2192 void (*ckhashfunc)(struct buf *), struct buf **bpp)
2196 int error, readwait, rv;
2198 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
2201 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags
2204 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp);
2209 KASSERT(blkno == bp->b_lblkno,
2210 ("getblkx returned buffer for blkno %jd instead of blkno %jd",
2211 (intmax_t)bp->b_lblkno, (intmax_t)blkno));
2212 flags &= ~GB_NOSPARSE;
2216 * If not found in cache, do some I/O
2219 if ((bp->b_flags & B_CACHE) == 0) {
2222 PROC_LOCK(td->td_proc);
2223 racct_add_buf(td->td_proc, bp, 0);
2224 PROC_UNLOCK(td->td_proc);
2227 td->td_ru.ru_inblock++;
2228 bp->b_iocmd = BIO_READ;
2229 bp->b_flags &= ~B_INVAL;
2230 if ((flags & GB_CKHASH) != 0) {
2231 bp->b_flags |= B_CKHASH;
2232 bp->b_ckhashcalc = ckhashfunc;
2234 if ((flags & GB_CVTENXIO) != 0)
2235 bp->b_xflags |= BX_CVTENXIO;
2236 bp->b_ioflags &= ~BIO_ERROR;
2237 if (bp->b_rcred == NOCRED && cred != NOCRED)
2238 bp->b_rcred = crhold(cred);
2239 vfs_busy_pages(bp, 0);
2240 bp->b_iooffset = dbtob(bp->b_blkno);
2246 * Attempt to initiate asynchronous I/O on read-ahead blocks.
2248 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc);
2262 * Write, release buffer on completion. (Done by iodone
2263 * if async). Do not bother writing anything if the buffer
2266 * Note that we set B_CACHE here, indicating that buffer is
2267 * fully valid and thus cacheable. This is true even of NFS
2268 * now so we set it generally. This could be set either here
2269 * or in biodone() since the I/O is synchronous. We put it
2273 bufwrite(struct buf *bp)
2280 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2281 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) {
2282 bp->b_flags |= B_INVAL | B_RELBUF;
2283 bp->b_flags &= ~B_CACHE;
2287 if (bp->b_flags & B_INVAL) {
2292 if (bp->b_flags & B_BARRIER)
2293 atomic_add_long(&barrierwrites, 1);
2295 oldflags = bp->b_flags;
2297 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
2298 ("FFS background buffer should not get here %p", bp));
2302 vp_md = vp->v_vflag & VV_MD;
2307 * Mark the buffer clean. Increment the bufobj write count
2308 * before bundirty() call, to prevent other thread from seeing
2309 * empty dirty list and zero counter for writes in progress,
2310 * falsely indicating that the bufobj is clean.
2312 bufobj_wref(bp->b_bufobj);
2315 bp->b_flags &= ~B_DONE;
2316 bp->b_ioflags &= ~BIO_ERROR;
2317 bp->b_flags |= B_CACHE;
2318 bp->b_iocmd = BIO_WRITE;
2320 vfs_busy_pages(bp, 1);
2323 * Normal bwrites pipeline writes
2325 bp->b_runningbufspace = bp->b_bufsize;
2326 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
2331 racct_add_buf(curproc, bp, 1);
2332 PROC_UNLOCK(curproc);
2335 curthread->td_ru.ru_oublock++;
2336 if (oldflags & B_ASYNC)
2338 bp->b_iooffset = dbtob(bp->b_blkno);
2339 buf_track(bp, __func__);
2342 if ((oldflags & B_ASYNC) == 0) {
2343 int rtval = bufwait(bp);
2346 } else if (space > hirunningspace) {
2348 * don't allow the async write to saturate the I/O
2349 * system. We will not deadlock here because
2350 * we are blocking waiting for I/O that is already in-progress
2351 * to complete. We do not block here if it is the update
2352 * or syncer daemon trying to clean up as that can lead
2355 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
2356 waitrunningbufspace();
2363 bufbdflush(struct bufobj *bo, struct buf *bp)
2366 struct bufdomain *bd;
2368 bd = &bdomain[bo->bo_domain];
2369 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) {
2370 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
2372 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) {
2375 * Try to find a buffer to flush.
2377 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2378 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2380 LK_EXCLUSIVE | LK_NOWAIT, NULL))
2383 panic("bdwrite: found ourselves");
2385 /* Don't countdeps with the bo lock held. */
2386 if (buf_countdeps(nbp, 0)) {
2391 if (nbp->b_flags & B_CLUSTEROK) {
2392 vfs_bio_awrite(nbp);
2397 dirtybufferflushes++;
2406 * Delayed write. (Buffer is marked dirty). Do not bother writing
2407 * anything if the buffer is marked invalid.
2409 * Note that since the buffer must be completely valid, we can safely
2410 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
2411 * biodone() in order to prevent getblk from writing the buffer
2412 * out synchronously.
2415 bdwrite(struct buf *bp)
2417 struct thread *td = curthread;
2421 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2422 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2423 KASSERT((bp->b_flags & B_BARRIER) == 0,
2424 ("Barrier request in delayed write %p", bp));
2426 if (bp->b_flags & B_INVAL) {
2432 * If we have too many dirty buffers, don't create any more.
2433 * If we are wildly over our limit, then force a complete
2434 * cleanup. Otherwise, just keep the situation from getting
2435 * out of control. Note that we have to avoid a recursive
2436 * disaster and not try to clean up after our own cleanup!
2440 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
2441 td->td_pflags |= TDP_INBDFLUSH;
2443 td->td_pflags &= ~TDP_INBDFLUSH;
2449 * Set B_CACHE, indicating that the buffer is fully valid. This is
2450 * true even of NFS now.
2452 bp->b_flags |= B_CACHE;
2455 * This bmap keeps the system from needing to do the bmap later,
2456 * perhaps when the system is attempting to do a sync. Since it
2457 * is likely that the indirect block -- or whatever other datastructure
2458 * that the filesystem needs is still in memory now, it is a good
2459 * thing to do this. Note also, that if the pageout daemon is
2460 * requesting a sync -- there might not be enough memory to do
2461 * the bmap then... So, this is important to do.
2463 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
2464 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
2467 buf_track(bp, __func__);
2470 * Set the *dirty* buffer range based upon the VM system dirty
2473 * Mark the buffer pages as clean. We need to do this here to
2474 * satisfy the vnode_pager and the pageout daemon, so that it
2475 * thinks that the pages have been "cleaned". Note that since
2476 * the pages are in a delayed write buffer -- the VFS layer
2477 * "will" see that the pages get written out on the next sync,
2478 * or perhaps the cluster will be completed.
2480 vfs_clean_pages_dirty_buf(bp);
2484 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
2485 * due to the softdep code.
2492 * Turn buffer into delayed write request. We must clear BIO_READ and
2493 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
2494 * itself to properly update it in the dirty/clean lists. We mark it
2495 * B_DONE to ensure that any asynchronization of the buffer properly
2496 * clears B_DONE ( else a panic will occur later ).
2498 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
2499 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
2500 * should only be called if the buffer is known-good.
2502 * Since the buffer is not on a queue, we do not update the numfreebuffers
2505 * The buffer must be on QUEUE_NONE.
2508 bdirty(struct buf *bp)
2511 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
2512 bp, bp->b_vp, bp->b_flags);
2513 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2514 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2515 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
2516 bp->b_flags &= ~(B_RELBUF);
2517 bp->b_iocmd = BIO_WRITE;
2519 if ((bp->b_flags & B_DELWRI) == 0) {
2520 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
2529 * Clear B_DELWRI for buffer.
2531 * Since the buffer is not on a queue, we do not update the numfreebuffers
2534 * The buffer must be on QUEUE_NONE.
2538 bundirty(struct buf *bp)
2541 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2542 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2543 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
2544 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
2546 if (bp->b_flags & B_DELWRI) {
2547 bp->b_flags &= ~B_DELWRI;
2552 * Since it is now being written, we can clear its deferred write flag.
2554 bp->b_flags &= ~B_DEFERRED;
2560 * Asynchronous write. Start output on a buffer, but do not wait for
2561 * it to complete. The buffer is released when the output completes.
2563 * bwrite() ( or the VOP routine anyway ) is responsible for handling
2564 * B_INVAL buffers. Not us.
2567 bawrite(struct buf *bp)
2570 bp->b_flags |= B_ASYNC;
2577 * Asynchronous barrier write. Start output on a buffer, but do not
2578 * wait for it to complete. Place a write barrier after this write so
2579 * that this buffer and all buffers written before it are committed to
2580 * the disk before any buffers written after this write are committed
2581 * to the disk. The buffer is released when the output completes.
2584 babarrierwrite(struct buf *bp)
2587 bp->b_flags |= B_ASYNC | B_BARRIER;
2594 * Synchronous barrier write. Start output on a buffer and wait for
2595 * it to complete. Place a write barrier after this write so that
2596 * this buffer and all buffers written before it are committed to
2597 * the disk before any buffers written after this write are committed
2598 * to the disk. The buffer is released when the output completes.
2601 bbarrierwrite(struct buf *bp)
2604 bp->b_flags |= B_BARRIER;
2605 return (bwrite(bp));
2611 * Called prior to the locking of any vnodes when we are expecting to
2612 * write. We do not want to starve the buffer cache with too many
2613 * dirty buffers so we block here. By blocking prior to the locking
2614 * of any vnodes we attempt to avoid the situation where a locked vnode
2615 * prevents the various system daemons from flushing related buffers.
2621 if (buf_dirty_count_severe()) {
2622 mtx_lock(&bdirtylock);
2623 while (buf_dirty_count_severe()) {
2625 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
2628 mtx_unlock(&bdirtylock);
2633 * Return true if we have too many dirty buffers.
2636 buf_dirty_count_severe(void)
2639 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty));
2645 * Release a busy buffer and, if requested, free its resources. The
2646 * buffer will be stashed in the appropriate bufqueue[] allowing it
2647 * to be accessed later as a cache entity or reused for other purposes.
2650 brelse(struct buf *bp)
2652 struct mount *v_mnt;
2656 * Many functions erroneously call brelse with a NULL bp under rare
2657 * error conditions. Simply return when called with a NULL bp.
2661 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
2662 bp, bp->b_vp, bp->b_flags);
2663 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2664 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2665 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0,
2666 ("brelse: non-VMIO buffer marked NOREUSE"));
2668 if (BUF_LOCKRECURSED(bp)) {
2670 * Do not process, in particular, do not handle the
2671 * B_INVAL/B_RELBUF and do not release to free list.
2677 if (bp->b_flags & B_MANAGED) {
2682 if (LIST_EMPTY(&bp->b_dep)) {
2683 bp->b_flags &= ~B_IOSTARTED;
2685 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2686 ("brelse: SU io not finished bp %p", bp));
2689 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
2690 BO_LOCK(bp->b_bufobj);
2691 bp->b_vflags &= ~BV_BKGRDERR;
2692 BO_UNLOCK(bp->b_bufobj);
2696 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2697 (bp->b_flags & B_INVALONERR)) {
2699 * Forced invalidation of dirty buffer contents, to be used
2700 * after a failed write in the rare case that the loss of the
2701 * contents is acceptable. The buffer is invalidated and
2704 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE;
2705 bp->b_flags &= ~(B_ASYNC | B_CACHE);
2708 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
2709 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) &&
2710 !(bp->b_flags & B_INVAL)) {
2712 * Failed write, redirty. All errors except ENXIO (which
2713 * means the device is gone) are treated as being
2716 * XXX Treating EIO as transient is not correct; the
2717 * contract with the local storage device drivers is that
2718 * they will only return EIO once the I/O is no longer
2719 * retriable. Network I/O also respects this through the
2720 * guarantees of TCP and/or the internal retries of NFS.
2721 * ENOMEM might be transient, but we also have no way of
2722 * knowing when its ok to retry/reschedule. In general,
2723 * this entire case should be made obsolete through better
2724 * error handling/recovery and resource scheduling.
2726 * Do this also for buffers that failed with ENXIO, but have
2727 * non-empty dependencies - the soft updates code might need
2728 * to access the buffer to untangle them.
2730 * Must clear BIO_ERROR to prevent pages from being scrapped.
2732 bp->b_ioflags &= ~BIO_ERROR;
2734 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
2735 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
2737 * Either a failed read I/O, or we were asked to free or not
2738 * cache the buffer, or we failed to write to a device that's
2739 * no longer present.
2741 bp->b_flags |= B_INVAL;
2742 if (!LIST_EMPTY(&bp->b_dep))
2744 if (bp->b_flags & B_DELWRI)
2746 bp->b_flags &= ~(B_DELWRI | B_CACHE);
2747 if ((bp->b_flags & B_VMIO) == 0) {
2755 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate()
2756 * is called with B_DELWRI set, the underlying pages may wind up
2757 * getting freed causing a previous write (bdwrite()) to get 'lost'
2758 * because pages associated with a B_DELWRI bp are marked clean.
2760 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even
2761 * if B_DELWRI is set.
2763 if (bp->b_flags & B_DELWRI)
2764 bp->b_flags &= ~B_RELBUF;
2767 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
2768 * constituted, not even NFS buffers now. Two flags effect this. If
2769 * B_INVAL, the struct buf is invalidated but the VM object is kept
2770 * around ( i.e. so it is trivial to reconstitute the buffer later ).
2772 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
2773 * invalidated. BIO_ERROR cannot be set for a failed write unless the
2774 * buffer is also B_INVAL because it hits the re-dirtying code above.
2776 * Normally we can do this whether a buffer is B_DELWRI or not. If
2777 * the buffer is an NFS buffer, it is tracking piecemeal writes or
2778 * the commit state and we cannot afford to lose the buffer. If the
2779 * buffer has a background write in progress, we need to keep it
2780 * around to prevent it from being reconstituted and starting a second
2784 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL;
2786 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
2787 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
2788 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
2789 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
2790 vfs_vmio_invalidate(bp);
2794 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 ||
2795 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) {
2797 bp->b_flags &= ~B_NOREUSE;
2798 if (bp->b_vp != NULL)
2803 * If the buffer has junk contents signal it and eventually
2804 * clean up B_DELWRI and diassociate the vnode so that gbincore()
2807 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
2808 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
2809 bp->b_flags |= B_INVAL;
2810 if (bp->b_flags & B_INVAL) {
2811 if (bp->b_flags & B_DELWRI)
2817 buf_track(bp, __func__);
2819 /* buffers with no memory */
2820 if (bp->b_bufsize == 0) {
2824 /* buffers with junk contents */
2825 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
2826 (bp->b_ioflags & BIO_ERROR)) {
2827 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
2828 if (bp->b_vflags & BV_BKGRDINPROG)
2829 panic("losing buffer 2");
2830 qindex = QUEUE_CLEAN;
2831 bp->b_flags |= B_AGE;
2832 /* remaining buffers */
2833 } else if (bp->b_flags & B_DELWRI)
2834 qindex = QUEUE_DIRTY;
2836 qindex = QUEUE_CLEAN;
2838 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
2839 panic("brelse: not dirty");
2841 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
2842 bp->b_xflags &= ~(BX_CVTENXIO);
2843 /* binsfree unlocks bp. */
2844 binsfree(bp, qindex);
2848 * Release a buffer back to the appropriate queue but do not try to free
2849 * it. The buffer is expected to be used again soon.
2851 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
2852 * biodone() to requeue an async I/O on completion. It is also used when
2853 * known good buffers need to be requeued but we think we may need the data
2856 * XXX we should be able to leave the B_RELBUF hint set on completion.
2859 bqrelse(struct buf *bp)
2863 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2864 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
2865 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
2867 qindex = QUEUE_NONE;
2868 if (BUF_LOCKRECURSED(bp)) {
2869 /* do not release to free list */
2873 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
2874 bp->b_xflags &= ~(BX_CVTENXIO);
2876 if (LIST_EMPTY(&bp->b_dep)) {
2877 bp->b_flags &= ~B_IOSTARTED;
2879 KASSERT((bp->b_flags & B_IOSTARTED) == 0,
2880 ("bqrelse: SU io not finished bp %p", bp));
2883 if (bp->b_flags & B_MANAGED) {
2884 if (bp->b_flags & B_REMFREE)
2889 /* buffers with stale but valid contents */
2890 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
2891 BV_BKGRDERR)) == BV_BKGRDERR) {
2892 BO_LOCK(bp->b_bufobj);
2893 bp->b_vflags &= ~BV_BKGRDERR;
2894 BO_UNLOCK(bp->b_bufobj);
2895 qindex = QUEUE_DIRTY;
2897 if ((bp->b_flags & B_DELWRI) == 0 &&
2898 (bp->b_xflags & BX_VNDIRTY))
2899 panic("bqrelse: not dirty");
2900 if ((bp->b_flags & B_NOREUSE) != 0) {
2904 qindex = QUEUE_CLEAN;
2906 buf_track(bp, __func__);
2907 /* binsfree unlocks bp. */
2908 binsfree(bp, qindex);
2912 buf_track(bp, __func__);
2918 * Complete I/O to a VMIO backed page. Validate the pages as appropriate,
2919 * restore bogus pages.
2922 vfs_vmio_iodone(struct buf *bp)
2927 struct vnode *vp __unused;
2928 int i, iosize, resid;
2931 obj = bp->b_bufobj->bo_object;
2932 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages,
2933 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)",
2934 blockcount_read(&obj->paging_in_progress), bp->b_npages));
2937 VNPASS(vp->v_holdcnt > 0, vp);
2938 VNPASS(vp->v_object != NULL, vp);
2940 foff = bp->b_offset;
2941 KASSERT(bp->b_offset != NOOFFSET,
2942 ("vfs_vmio_iodone: bp %p has no buffer offset", bp));
2945 iosize = bp->b_bcount - bp->b_resid;
2946 for (i = 0; i < bp->b_npages; i++) {
2947 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2952 * cleanup bogus pages, restoring the originals
2955 if (m == bogus_page) {
2957 m = vm_page_relookup(obj, OFF_TO_IDX(foff));
2959 panic("biodone: page disappeared!");
2961 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) {
2963 * In the write case, the valid and clean bits are
2964 * already changed correctly ( see bdwrite() ), so we
2965 * only need to do this here in the read case.
2967 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK,
2968 resid)) == 0, ("vfs_vmio_iodone: page %p "
2969 "has unexpected dirty bits", m));
2970 vfs_page_set_valid(bp, foff, m);
2972 KASSERT(OFF_TO_IDX(foff) == m->pindex,
2973 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch",
2974 (intmax_t)foff, (uintmax_t)m->pindex));
2977 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2980 vm_object_pip_wakeupn(obj, bp->b_npages);
2981 if (bogus && buf_mapped(bp)) {
2982 BUF_CHECK_MAPPED(bp);
2983 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2984 bp->b_pages, bp->b_npages);
2989 * Perform page invalidation when a buffer is released. The fully invalid
2990 * pages will be reclaimed later in vfs_vmio_truncate().
2993 vfs_vmio_invalidate(struct buf *bp)
2997 int flags, i, resid, poffset, presid;
2999 if (buf_mapped(bp)) {
3000 BUF_CHECK_MAPPED(bp);
3001 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
3003 BUF_CHECK_UNMAPPED(bp);
3005 * Get the base offset and length of the buffer. Note that
3006 * in the VMIO case if the buffer block size is not
3007 * page-aligned then b_data pointer may not be page-aligned.
3008 * But our b_pages[] array *IS* page aligned.
3010 * block sizes less then DEV_BSIZE (usually 512) are not
3011 * supported due to the page granularity bits (m->valid,
3012 * m->dirty, etc...).
3014 * See man buf(9) for more information
3016 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3017 obj = bp->b_bufobj->bo_object;
3018 resid = bp->b_bufsize;
3019 poffset = bp->b_offset & PAGE_MASK;
3020 VM_OBJECT_WLOCK(obj);
3021 for (i = 0; i < bp->b_npages; i++) {
3023 if (m == bogus_page)
3024 panic("vfs_vmio_invalidate: Unexpected bogus page.");
3025 bp->b_pages[i] = NULL;
3027 presid = resid > (PAGE_SIZE - poffset) ?
3028 (PAGE_SIZE - poffset) : resid;
3029 KASSERT(presid >= 0, ("brelse: extra page"));
3030 vm_page_busy_acquire(m, VM_ALLOC_SBUSY);
3031 if (pmap_page_wired_mappings(m) == 0)
3032 vm_page_set_invalid(m, poffset, presid);
3034 vm_page_release_locked(m, flags);
3038 VM_OBJECT_WUNLOCK(obj);
3043 * Page-granular truncation of an existing VMIO buffer.
3046 vfs_vmio_truncate(struct buf *bp, int desiredpages)
3052 if (bp->b_npages == desiredpages)
3055 if (buf_mapped(bp)) {
3056 BUF_CHECK_MAPPED(bp);
3057 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) +
3058 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
3060 BUF_CHECK_UNMAPPED(bp);
3063 * The object lock is needed only if we will attempt to free pages.
3065 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0;
3066 if ((bp->b_flags & B_DIRECT) != 0) {
3067 flags |= VPR_TRYFREE;
3068 obj = bp->b_bufobj->bo_object;
3069 VM_OBJECT_WLOCK(obj);
3073 for (i = desiredpages; i < bp->b_npages; i++) {
3075 KASSERT(m != bogus_page, ("allocbuf: bogus page found"));
3076 bp->b_pages[i] = NULL;
3078 vm_page_release_locked(m, flags);
3080 vm_page_release(m, flags);
3083 VM_OBJECT_WUNLOCK(obj);
3084 bp->b_npages = desiredpages;
3088 * Byte granular extension of VMIO buffers.
3091 vfs_vmio_extend(struct buf *bp, int desiredpages, int size)
3094 * We are growing the buffer, possibly in a
3095 * byte-granular fashion.
3103 * Step 1, bring in the VM pages from the object, allocating
3104 * them if necessary. We must clear B_CACHE if these pages
3105 * are not valid for the range covered by the buffer.
3107 obj = bp->b_bufobj->bo_object;
3108 if (bp->b_npages < desiredpages) {
3109 KASSERT(desiredpages <= atop(maxbcachebuf),
3110 ("vfs_vmio_extend past maxbcachebuf %p %d %u",
3111 bp, desiredpages, maxbcachebuf));
3114 * We must allocate system pages since blocking
3115 * here could interfere with paging I/O, no
3116 * matter which process we are.
3118 * Only exclusive busy can be tested here.
3119 * Blocking on shared busy might lead to
3120 * deadlocks once allocbuf() is called after
3121 * pages are vfs_busy_pages().
3123 (void)vm_page_grab_pages_unlocked(obj,
3124 OFF_TO_IDX(bp->b_offset) + bp->b_npages,
3125 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY |
3126 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED,
3127 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages);
3128 bp->b_npages = desiredpages;
3132 * Step 2. We've loaded the pages into the buffer,
3133 * we have to figure out if we can still have B_CACHE
3134 * set. Note that B_CACHE is set according to the
3135 * byte-granular range ( bcount and size ), not the
3136 * aligned range ( newbsize ).
3138 * The VM test is against m->valid, which is DEV_BSIZE
3139 * aligned. Needless to say, the validity of the data
3140 * needs to also be DEV_BSIZE aligned. Note that this
3141 * fails with NFS if the server or some other client
3142 * extends the file's EOF. If our buffer is resized,
3143 * B_CACHE may remain set! XXX
3145 toff = bp->b_bcount;
3146 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3147 while ((bp->b_flags & B_CACHE) && toff < size) {
3150 if (tinc > (size - toff))
3152 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT;
3153 m = bp->b_pages[pi];
3154 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m);
3160 * Step 3, fixup the KVA pmap.
3165 BUF_CHECK_UNMAPPED(bp);
3169 * Check to see if a block at a particular lbn is available for a clustered
3173 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
3180 /* If the buf isn't in core skip it */
3181 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
3184 /* If the buf is busy we don't want to wait for it */
3185 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
3188 /* Only cluster with valid clusterable delayed write buffers */
3189 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
3190 (B_DELWRI | B_CLUSTEROK))
3193 if (bpa->b_bufsize != size)
3197 * Check to see if it is in the expected place on disk and that the
3198 * block has been mapped.
3200 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
3210 * Implement clustered async writes for clearing out B_DELWRI buffers.
3211 * This is much better then the old way of writing only one buffer at
3212 * a time. Note that we may not be presented with the buffers in the
3213 * correct order, so we search for the cluster in both directions.
3216 vfs_bio_awrite(struct buf *bp)
3221 daddr_t lblkno = bp->b_lblkno;
3222 struct vnode *vp = bp->b_vp;
3230 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0;
3232 * right now we support clustered writing only to regular files. If
3233 * we find a clusterable block we could be in the middle of a cluster
3234 * rather then at the beginning.
3236 if ((vp->v_type == VREG) &&
3237 (vp->v_mount != 0) && /* Only on nodes that have the size info */
3238 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
3239 size = vp->v_mount->mnt_stat.f_iosize;
3240 maxcl = maxphys / size;
3243 for (i = 1; i < maxcl; i++)
3244 if (vfs_bio_clcheck(vp, size, lblkno + i,
3245 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
3248 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
3249 if (vfs_bio_clcheck(vp, size, lblkno - j,
3250 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
3256 * this is a possible cluster write
3260 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
3266 bp->b_flags |= B_ASYNC;
3268 * default (old) behavior, writing out only one block
3270 * XXX returns b_bufsize instead of b_bcount for nwritten?
3272 nwritten = bp->b_bufsize;
3281 * Allocate KVA for an empty buf header according to gbflags.
3284 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize)
3287 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) {
3289 * In order to keep fragmentation sane we only allocate kva
3290 * in BKVASIZE chunks. XXX with vmem we can do page size.
3292 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
3294 if (maxsize != bp->b_kvasize &&
3295 bufkva_alloc(bp, maxsize, gbflags))
3304 * Find and initialize a new buffer header, freeing up existing buffers
3305 * in the bufqueues as necessary. The new buffer is returned locked.
3308 * We have insufficient buffer headers
3309 * We have insufficient buffer space
3310 * buffer_arena is too fragmented ( space reservation fails )
3311 * If we have to flush dirty buffers ( but we try to avoid this )
3313 * The caller is responsible for releasing the reserved bufspace after
3314 * allocbuf() is called.
3317 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags)
3319 struct bufdomain *bd;
3321 bool metadata, reserved;
3324 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3325 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3326 if (!unmapped_buf_allowed)
3327 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3329 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
3337 bd = &bdomain[vp->v_bufobj.bo_domain];
3339 counter_u64_add(getnewbufcalls, 1);
3342 if (reserved == false &&
3343 bufspace_reserve(bd, maxsize, metadata) != 0) {
3344 counter_u64_add(getnewbufrestarts, 1);
3348 if ((bp = buf_alloc(bd)) == NULL) {
3349 counter_u64_add(getnewbufrestarts, 1);
3352 if (getnewbuf_kva(bp, gbflags, maxsize) == 0)
3355 } while (buf_recycle(bd, false) == 0);
3358 bufspace_release(bd, maxsize);
3360 bp->b_flags |= B_INVAL;
3363 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo);
3371 * buffer flushing daemon. Buffers are normally flushed by the
3372 * update daemon but if it cannot keep up this process starts to
3373 * take the load in an attempt to prevent getnewbuf() from blocking.
3375 static struct kproc_desc buf_kp = {
3380 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
3383 buf_flush(struct vnode *vp, struct bufdomain *bd, int target)
3387 flushed = flushbufqueues(vp, bd, target, 0);
3390 * Could not find any buffers without rollback
3391 * dependencies, so just write the first one
3392 * in the hopes of eventually making progress.
3394 if (vp != NULL && target > 2)
3396 flushbufqueues(vp, bd, target, 1);
3402 buf_daemon_shutdown(void *arg __unused, int howto __unused)
3408 wakeup(&bd_request);
3409 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown",
3411 mtx_unlock(&bdlock);
3413 printf("bufdaemon wait error: %d\n", error);
3419 struct bufdomain *bd;
3425 * This process needs to be suspended prior to shutdown sync.
3427 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL,
3428 SHUTDOWN_PRI_LAST + 100);
3431 * Start the buf clean daemons as children threads.
3433 for (i = 0 ; i < buf_domains; i++) {
3436 error = kthread_add((void (*)(void *))bufspace_daemon,
3437 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i);
3439 panic("error %d spawning bufspace daemon", error);
3443 * This process is allowed to take the buffer cache to the limit
3445 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
3447 while (!bd_shutdown) {
3449 mtx_unlock(&bdlock);
3452 * Save speedupreq for this pass and reset to capture new
3455 speedupreq = bd_speedupreq;
3459 * Flush each domain sequentially according to its level and
3460 * the speedup request.
3462 for (i = 0; i < buf_domains; i++) {
3465 lodirty = bd->bd_numdirtybuffers / 2;
3467 lodirty = bd->bd_lodirtybuffers;
3468 while (bd->bd_numdirtybuffers > lodirty) {
3469 if (buf_flush(NULL, bd,
3470 bd->bd_numdirtybuffers - lodirty) == 0)
3472 kern_yield(PRI_USER);
3477 * Only clear bd_request if we have reached our low water
3478 * mark. The buf_daemon normally waits 1 second and
3479 * then incrementally flushes any dirty buffers that have
3480 * built up, within reason.
3482 * If we were unable to hit our low water mark and couldn't
3483 * find any flushable buffers, we sleep for a short period
3484 * to avoid endless loops on unlockable buffers.
3489 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) {
3491 * We reached our low water mark, reset the
3492 * request and sleep until we are needed again.
3493 * The sleep is just so the suspend code works.
3497 * Do an extra wakeup in case dirty threshold
3498 * changed via sysctl and the explicit transition
3499 * out of shortfall was missed.
3502 if (runningbufspace <= lorunningspace)
3504 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
3507 * We couldn't find any flushable dirty buffers but
3508 * still have too many dirty buffers, we
3509 * have to sleep and try again. (rare)
3511 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
3514 wakeup(&bd_shutdown);
3515 mtx_unlock(&bdlock);
3522 * Try to flush a buffer in the dirty queue. We must be careful to
3523 * free up B_INVAL buffers instead of write them, which NFS is
3524 * particularly sensitive to.
3526 static int flushwithdeps = 0;
3527 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS,
3529 "Number of buffers flushed with dependencies that require rollbacks");
3532 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target,
3535 struct bufqueue *bq;
3536 struct buf *sentinel;
3546 bq = &bd->bd_dirtyq;
3548 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
3549 sentinel->b_qindex = QUEUE_SENTINEL;
3551 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist);
3553 while (flushed != target) {
3556 bp = TAILQ_NEXT(sentinel, b_freelist);
3558 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3559 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel,
3566 * Skip sentinels inserted by other invocations of the
3567 * flushbufqueues(), taking care to not reorder them.
3569 * Only flush the buffers that belong to the
3570 * vnode locked by the curthread.
3572 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
3577 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
3583 * BKGRDINPROG can only be set with the buf and bufobj
3584 * locks both held. We tolerate a race to clear it here.
3586 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
3587 (bp->b_flags & B_DELWRI) == 0) {
3591 if (bp->b_flags & B_INVAL) {
3598 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
3599 if (flushdeps == 0) {
3607 * We must hold the lock on a vnode before writing
3608 * one of its buffers. Otherwise we may confuse, or
3609 * in the case of a snapshot vnode, deadlock the
3612 * The lock order here is the reverse of the normal
3613 * of vnode followed by buf lock. This is ok because
3614 * the NOWAIT will prevent deadlock.
3617 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
3623 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
3625 ASSERT_VOP_LOCKED(vp, "getbuf");
3627 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
3628 vn_lock(vp, LK_TRYUPGRADE);
3631 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
3632 bp, bp->b_vp, bp->b_flags);
3633 if (curproc == bufdaemonproc) {
3638 counter_u64_add(notbufdflushes, 1);
3640 vn_finished_write(mp);
3643 flushwithdeps += hasdeps;
3647 * Sleeping on runningbufspace while holding
3648 * vnode lock leads to deadlock.
3650 if (curproc == bufdaemonproc &&
3651 runningbufspace > hirunningspace)
3652 waitrunningbufspace();
3655 vn_finished_write(mp);
3659 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist);
3661 free(sentinel, M_TEMP);
3666 * Check to see if a block is currently memory resident.
3669 incore(struct bufobj *bo, daddr_t blkno)
3671 return (gbincore_unlocked(bo, blkno));
3675 * Returns true if no I/O is needed to access the
3676 * associated VM object. This is like incore except
3677 * it also hunts around in the VM system for the data.
3680 inmem(struct vnode * vp, daddr_t blkno)
3683 vm_offset_t toff, tinc, size;
3688 ASSERT_VOP_LOCKED(vp, "inmem");
3690 if (incore(&vp->v_bufobj, blkno))
3692 if (vp->v_mount == NULL)
3699 if (size > vp->v_mount->mnt_stat.f_iosize)
3700 size = vp->v_mount->mnt_stat.f_iosize;
3701 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
3703 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
3704 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3710 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
3711 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
3713 * Consider page validity only if page mapping didn't change
3716 valid = vm_page_is_valid(m,
3717 (vm_offset_t)((toff + off) & PAGE_MASK), tinc);
3718 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff));
3730 * Set the dirty range for a buffer based on the status of the dirty
3731 * bits in the pages comprising the buffer. The range is limited
3732 * to the size of the buffer.
3734 * Tell the VM system that the pages associated with this buffer
3735 * are clean. This is used for delayed writes where the data is
3736 * going to go to disk eventually without additional VM intevention.
3738 * Note that while we only really need to clean through to b_bcount, we
3739 * just go ahead and clean through to b_bufsize.
3742 vfs_clean_pages_dirty_buf(struct buf *bp)
3744 vm_ooffset_t foff, noff, eoff;
3748 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
3751 foff = bp->b_offset;
3752 KASSERT(bp->b_offset != NOOFFSET,
3753 ("vfs_clean_pages_dirty_buf: no buffer offset"));
3755 vfs_busy_pages_acquire(bp);
3756 vfs_setdirty_range(bp);
3757 for (i = 0; i < bp->b_npages; i++) {
3758 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3760 if (eoff > bp->b_offset + bp->b_bufsize)
3761 eoff = bp->b_offset + bp->b_bufsize;
3763 vfs_page_set_validclean(bp, foff, m);
3764 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3767 vfs_busy_pages_release(bp);
3771 vfs_setdirty_range(struct buf *bp)
3773 vm_offset_t boffset;
3774 vm_offset_t eoffset;
3778 * test the pages to see if they have been modified directly
3779 * by users through the VM system.
3781 for (i = 0; i < bp->b_npages; i++)
3782 vm_page_test_dirty(bp->b_pages[i]);
3785 * Calculate the encompassing dirty range, boffset and eoffset,
3786 * (eoffset - boffset) bytes.
3789 for (i = 0; i < bp->b_npages; i++) {
3790 if (bp->b_pages[i]->dirty)
3793 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3795 for (i = bp->b_npages - 1; i >= 0; --i) {
3796 if (bp->b_pages[i]->dirty) {
3800 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
3803 * Fit it to the buffer.
3806 if (eoffset > bp->b_bcount)
3807 eoffset = bp->b_bcount;
3810 * If we have a good dirty range, merge with the existing
3814 if (boffset < eoffset) {
3815 if (bp->b_dirtyoff > boffset)
3816 bp->b_dirtyoff = boffset;
3817 if (bp->b_dirtyend < eoffset)
3818 bp->b_dirtyend = eoffset;
3823 * Allocate the KVA mapping for an existing buffer.
3824 * If an unmapped buffer is provided but a mapped buffer is requested, take
3825 * also care to properly setup mappings between pages and KVA.
3828 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
3830 int bsize, maxsize, need_mapping, need_kva;
3833 need_mapping = bp->b_data == unmapped_buf &&
3834 (gbflags & GB_UNMAPPED) == 0;
3835 need_kva = bp->b_kvabase == unmapped_buf &&
3836 bp->b_data == unmapped_buf &&
3837 (gbflags & GB_KVAALLOC) != 0;
3838 if (!need_mapping && !need_kva)
3841 BUF_CHECK_UNMAPPED(bp);
3843 if (need_mapping && bp->b_kvabase != unmapped_buf) {
3845 * Buffer is not mapped, but the KVA was already
3846 * reserved at the time of the instantiation. Use the
3853 * Calculate the amount of the address space we would reserve
3854 * if the buffer was mapped.
3856 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3857 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
3858 offset = blkno * bsize;
3859 maxsize = size + (offset & PAGE_MASK);
3860 maxsize = imax(maxsize, bsize);
3862 while (bufkva_alloc(bp, maxsize, gbflags) != 0) {
3863 if ((gbflags & GB_NOWAIT_BD) != 0) {
3865 * XXXKIB: defragmentation cannot
3866 * succeed, not sure what else to do.
3868 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp);
3870 counter_u64_add(mappingrestarts, 1);
3871 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0);
3875 /* b_offset is handled by bpmap_qenter. */
3876 bp->b_data = bp->b_kvabase;
3877 BUF_CHECK_MAPPED(bp);
3883 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3889 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp);
3898 * Get a block given a specified block and offset into a file/device.
3899 * The buffers B_DONE bit will be cleared on return, making it almost
3900 * ready for an I/O initiation. B_INVAL may or may not be set on
3901 * return. The caller should clear B_INVAL prior to initiating a
3904 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3905 * an existing buffer.
3907 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3908 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3909 * and then cleared based on the backing VM. If the previous buffer is
3910 * non-0-sized but invalid, B_CACHE will be cleared.
3912 * If getblk() must create a new buffer, the new buffer is returned with
3913 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3914 * case it is returned with B_INVAL clear and B_CACHE set based on the
3917 * getblk() also forces a bwrite() for any B_DELWRI buffer whose
3918 * B_CACHE bit is clear.
3920 * What this means, basically, is that the caller should use B_CACHE to
3921 * determine whether the buffer is fully valid or not and should clear
3922 * B_INVAL prior to issuing a read. If the caller intends to validate
3923 * the buffer by loading its data area with something, the caller needs
3924 * to clear B_INVAL. If the caller does this without issuing an I/O,
3925 * the caller should set B_CACHE ( as an optimization ), else the caller
3926 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3927 * a write attempt or if it was a successful read. If the caller
3928 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3929 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3931 * The blkno parameter is the logical block being requested. Normally
3932 * the mapping of logical block number to disk block address is done
3933 * by calling VOP_BMAP(). However, if the mapping is already known, the
3934 * disk block address can be passed using the dblkno parameter. If the
3935 * disk block address is not known, then the same value should be passed
3936 * for blkno and dblkno.
3939 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag,
3940 int slptimeo, int flags, struct buf **bpp)
3945 int bsize, error, maxsize, vmio;
3948 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3949 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3950 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3951 if (vp->v_type != VCHR)
3952 ASSERT_VOP_LOCKED(vp, "getblk");
3953 if (size > maxbcachebuf)
3954 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size,
3956 if (!unmapped_buf_allowed)
3957 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3962 /* Attempt lockless lookup first. */
3963 bp = gbincore_unlocked(bo, blkno);
3965 goto newbuf_unlocked;
3967 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0,
3972 /* Verify buf identify has not changed since lookup. */
3973 if (bp->b_bufobj == bo && bp->b_lblkno == blkno)
3974 goto foundbuf_fastpath;
3976 /* It changed, fallback to locked lookup. */
3981 bp = gbincore(bo, blkno);
3986 * Buffer is in-core. If the buffer is not busy nor managed,
3987 * it must be on a queue.
3989 lockflags = LK_EXCLUSIVE | LK_INTERLOCK |
3990 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL);
3992 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0;
3995 error = BUF_TIMELOCK(bp, lockflags,
3996 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3999 * If we slept and got the lock we have to restart in case
4000 * the buffer changed identities.
4002 if (error == ENOLCK)
4004 /* We timed out or were interrupted. */
4005 else if (error != 0)
4009 /* If recursed, assume caller knows the rules. */
4010 if (BUF_LOCKRECURSED(bp))
4014 * The buffer is locked. B_CACHE is cleared if the buffer is
4015 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
4016 * and for a VMIO buffer B_CACHE is adjusted according to the
4019 if (bp->b_flags & B_INVAL)
4020 bp->b_flags &= ~B_CACHE;
4021 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
4022 bp->b_flags |= B_CACHE;
4023 if (bp->b_flags & B_MANAGED)
4024 MPASS(bp->b_qindex == QUEUE_NONE);
4029 * check for size inconsistencies for non-VMIO case.
4031 if (bp->b_bcount != size) {
4032 if ((bp->b_flags & B_VMIO) == 0 ||
4033 (size > bp->b_kvasize)) {
4034 if (bp->b_flags & B_DELWRI) {
4035 bp->b_flags |= B_NOCACHE;
4038 if (LIST_EMPTY(&bp->b_dep)) {
4039 bp->b_flags |= B_RELBUF;
4042 bp->b_flags |= B_NOCACHE;
4051 * Handle the case of unmapped buffer which should
4052 * become mapped, or the buffer for which KVA
4053 * reservation is requested.
4055 bp_unmapped_get_kva(bp, blkno, size, flags);
4058 * If the size is inconsistent in the VMIO case, we can resize
4059 * the buffer. This might lead to B_CACHE getting set or
4060 * cleared. If the size has not changed, B_CACHE remains
4061 * unchanged from its previous state.
4065 KASSERT(bp->b_offset != NOOFFSET,
4066 ("getblk: no buffer offset"));
4069 * A buffer with B_DELWRI set and B_CACHE clear must
4070 * be committed before we can return the buffer in
4071 * order to prevent the caller from issuing a read
4072 * ( due to B_CACHE not being set ) and overwriting
4075 * Most callers, including NFS and FFS, need this to
4076 * operate properly either because they assume they
4077 * can issue a read if B_CACHE is not set, or because
4078 * ( for example ) an uncached B_DELWRI might loop due
4079 * to softupdates re-dirtying the buffer. In the latter
4080 * case, B_CACHE is set after the first write completes,
4081 * preventing further loops.
4082 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
4083 * above while extending the buffer, we cannot allow the
4084 * buffer to remain with B_CACHE set after the write
4085 * completes or it will represent a corrupt state. To
4086 * deal with this we set B_NOCACHE to scrap the buffer
4089 * We might be able to do something fancy, like setting
4090 * B_CACHE in bwrite() except if B_DELWRI is already set,
4091 * so the below call doesn't set B_CACHE, but that gets real
4092 * confusing. This is much easier.
4095 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
4096 bp->b_flags |= B_NOCACHE;
4100 bp->b_flags &= ~B_DONE;
4103 * Buffer is not in-core, create new buffer. The buffer
4104 * returned by getnewbuf() is locked. Note that the returned
4105 * buffer is also considered valid (not marked B_INVAL).
4110 * If the user does not want us to create the buffer, bail out
4113 if (flags & GB_NOCREAT)
4116 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize;
4117 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize"));
4118 offset = blkno * bsize;
4119 vmio = vp->v_object != NULL;
4121 maxsize = size + (offset & PAGE_MASK);
4124 /* Do not allow non-VMIO notmapped buffers. */
4125 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
4127 maxsize = imax(maxsize, bsize);
4128 if ((flags & GB_NOSPARSE) != 0 && vmio &&
4130 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0);
4131 KASSERT(error != EOPNOTSUPP,
4132 ("GB_NOSPARSE from fs not supporting bmap, vp %p",
4137 return (EJUSTRETURN);
4140 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags);
4142 if (slpflag || slptimeo)
4145 * XXX This is here until the sleep path is diagnosed
4146 * enough to work under very low memory conditions.
4148 * There's an issue on low memory, 4BSD+non-preempt
4149 * systems (eg MIPS routers with 32MB RAM) where buffer
4150 * exhaustion occurs without sleeping for buffer
4151 * reclaimation. This just sticks in a loop and
4152 * constantly attempts to allocate a buffer, which
4153 * hits exhaustion and tries to wakeup bufdaemon.
4154 * This never happens because we never yield.
4156 * The real solution is to identify and fix these cases
4157 * so we aren't effectively busy-waiting in a loop
4158 * until the reclaimation path has cycles to run.
4160 kern_yield(PRI_USER);
4165 * This code is used to make sure that a buffer is not
4166 * created while the getnewbuf routine is blocked.
4167 * This can be a problem whether the vnode is locked or not.
4168 * If the buffer is created out from under us, we have to
4169 * throw away the one we just created.
4171 * Note: this must occur before we associate the buffer
4172 * with the vp especially considering limitations in
4173 * the splay tree implementation when dealing with duplicate
4177 if (gbincore(bo, blkno)) {
4179 bp->b_flags |= B_INVAL;
4180 bufspace_release(bufdomain(bp), maxsize);
4186 * Insert the buffer into the hash, so that it can
4187 * be found by incore.
4189 bp->b_lblkno = blkno;
4190 bp->b_blkno = d_blkno;
4191 bp->b_offset = offset;
4196 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
4197 * buffer size starts out as 0, B_CACHE will be set by
4198 * allocbuf() for the VMIO case prior to it testing the
4199 * backing store for validity.
4203 bp->b_flags |= B_VMIO;
4204 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
4205 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
4206 bp, vp->v_object, bp->b_bufobj->bo_object));
4208 bp->b_flags &= ~B_VMIO;
4209 KASSERT(bp->b_bufobj->bo_object == NULL,
4210 ("ARGH! has b_bufobj->bo_object %p %p\n",
4211 bp, bp->b_bufobj->bo_object));
4212 BUF_CHECK_MAPPED(bp);
4216 bufspace_release(bufdomain(bp), maxsize);
4217 bp->b_flags &= ~B_DONE;
4219 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
4221 buf_track(bp, __func__);
4222 KASSERT(bp->b_bufobj == bo,
4223 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
4229 * Get an empty, disassociated buffer of given size. The buffer is initially
4233 geteblk(int size, int flags)
4238 maxsize = (size + BKVAMASK) & ~BKVAMASK;
4239 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) {
4240 if ((flags & GB_NOWAIT_BD) &&
4241 (curthread->td_pflags & TDP_BUFNEED) != 0)
4245 bufspace_release(bufdomain(bp), maxsize);
4246 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
4251 * Truncate the backing store for a non-vmio buffer.
4254 vfs_nonvmio_truncate(struct buf *bp, int newbsize)
4257 if (bp->b_flags & B_MALLOC) {
4259 * malloced buffers are not shrunk
4261 if (newbsize == 0) {
4262 bufmallocadjust(bp, 0);
4263 free(bp->b_data, M_BIOBUF);
4264 bp->b_data = bp->b_kvabase;
4265 bp->b_flags &= ~B_MALLOC;
4269 vm_hold_free_pages(bp, newbsize);
4270 bufspace_adjust(bp, newbsize);
4274 * Extend the backing for a non-VMIO buffer.
4277 vfs_nonvmio_extend(struct buf *bp, int newbsize)
4283 * We only use malloced memory on the first allocation.
4284 * and revert to page-allocated memory when the buffer
4287 * There is a potential smp race here that could lead
4288 * to bufmallocspace slightly passing the max. It
4289 * is probably extremely rare and not worth worrying
4292 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 &&
4293 bufmallocspace < maxbufmallocspace) {
4294 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK);
4295 bp->b_flags |= B_MALLOC;
4296 bufmallocadjust(bp, newbsize);
4301 * If the buffer is growing on its other-than-first
4302 * allocation then we revert to the page-allocation
4307 if (bp->b_flags & B_MALLOC) {
4308 origbuf = bp->b_data;
4309 origbufsize = bp->b_bufsize;
4310 bp->b_data = bp->b_kvabase;
4311 bufmallocadjust(bp, 0);
4312 bp->b_flags &= ~B_MALLOC;
4313 newbsize = round_page(newbsize);
4315 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize,
4316 (vm_offset_t) bp->b_data + newbsize);
4317 if (origbuf != NULL) {
4318 bcopy(origbuf, bp->b_data, origbufsize);
4319 free(origbuf, M_BIOBUF);
4321 bufspace_adjust(bp, newbsize);
4325 * This code constitutes the buffer memory from either anonymous system
4326 * memory (in the case of non-VMIO operations) or from an associated
4327 * VM object (in the case of VMIO operations). This code is able to
4328 * resize a buffer up or down.
4330 * Note that this code is tricky, and has many complications to resolve
4331 * deadlock or inconsistent data situations. Tread lightly!!!
4332 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
4333 * the caller. Calling this code willy nilly can result in the loss of data.
4335 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
4336 * B_CACHE for the non-VMIO case.
4339 allocbuf(struct buf *bp, int size)
4343 if (bp->b_bcount == size)
4346 if (bp->b_kvasize != 0 && bp->b_kvasize < size)
4347 panic("allocbuf: buffer too small");
4349 newbsize = roundup2(size, DEV_BSIZE);
4350 if ((bp->b_flags & B_VMIO) == 0) {
4351 if ((bp->b_flags & B_MALLOC) == 0)
4352 newbsize = round_page(newbsize);
4354 * Just get anonymous memory from the kernel. Don't
4355 * mess with B_CACHE.
4357 if (newbsize < bp->b_bufsize)
4358 vfs_nonvmio_truncate(bp, newbsize);
4359 else if (newbsize > bp->b_bufsize)
4360 vfs_nonvmio_extend(bp, newbsize);
4364 desiredpages = (size == 0) ? 0 :
4365 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
4367 if (bp->b_flags & B_MALLOC)
4368 panic("allocbuf: VMIO buffer can't be malloced");
4370 * Set B_CACHE initially if buffer is 0 length or will become
4373 if (size == 0 || bp->b_bufsize == 0)
4374 bp->b_flags |= B_CACHE;
4376 if (newbsize < bp->b_bufsize)
4377 vfs_vmio_truncate(bp, desiredpages);
4378 /* XXX This looks as if it should be newbsize > b_bufsize */
4379 else if (size > bp->b_bcount)
4380 vfs_vmio_extend(bp, desiredpages, size);
4381 bufspace_adjust(bp, newbsize);
4383 bp->b_bcount = size; /* requested buffer size. */
4387 extern int inflight_transient_maps;
4389 static struct bio_queue nondump_bios;
4392 biodone(struct bio *bp)
4395 void (*done)(struct bio *);
4396 vm_offset_t start, end;
4398 biotrack(bp, __func__);
4401 * Avoid completing I/O when dumping after a panic since that may
4402 * result in a deadlock in the filesystem or pager code. Note that
4403 * this doesn't affect dumps that were started manually since we aim
4404 * to keep the system usable after it has been resumed.
4406 if (__predict_false(dumping && SCHEDULER_STOPPED())) {
4407 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue);
4410 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
4411 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
4412 bp->bio_flags |= BIO_UNMAPPED;
4413 start = trunc_page((vm_offset_t)bp->bio_data);
4414 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
4415 bp->bio_data = unmapped_buf;
4416 pmap_qremove(start, atop(end - start));
4417 vmem_free(transient_arena, start, end - start);
4418 atomic_add_int(&inflight_transient_maps, -1);
4420 done = bp->bio_done;
4422 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4424 bp->bio_flags |= BIO_DONE;
4432 * Wait for a BIO to finish.
4435 biowait(struct bio *bp, const char *wmesg)
4439 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4441 while ((bp->bio_flags & BIO_DONE) == 0)
4442 msleep(bp, mtxp, PRIBIO, wmesg, 0);
4444 if (bp->bio_error != 0)
4445 return (bp->bio_error);
4446 if (!(bp->bio_flags & BIO_ERROR))
4452 biofinish(struct bio *bp, struct devstat *stat, int error)
4456 bp->bio_error = error;
4457 bp->bio_flags |= BIO_ERROR;
4460 devstat_end_transaction_bio(stat, bp);
4464 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
4466 biotrack_buf(struct bio *bp, const char *location)
4469 buf_track(bp->bio_track_bp, location);
4476 * Wait for buffer I/O completion, returning error status. The buffer
4477 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
4478 * error and cleared.
4481 bufwait(struct buf *bp)
4483 if (bp->b_iocmd == BIO_READ)
4484 bwait(bp, PRIBIO, "biord");
4486 bwait(bp, PRIBIO, "biowr");
4487 if (bp->b_flags & B_EINTR) {
4488 bp->b_flags &= ~B_EINTR;
4491 if (bp->b_ioflags & BIO_ERROR) {
4492 return (bp->b_error ? bp->b_error : EIO);
4501 * Finish I/O on a buffer, optionally calling a completion function.
4502 * This is usually called from an interrupt so process blocking is
4505 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
4506 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
4507 * assuming B_INVAL is clear.
4509 * For the VMIO case, we set B_CACHE if the op was a read and no
4510 * read error occurred, or if the op was a write. B_CACHE is never
4511 * set if the buffer is invalid or otherwise uncacheable.
4513 * bufdone does not mess with B_INVAL, allowing the I/O routine or the
4514 * initiator to leave B_INVAL set to brelse the buffer out of existence
4515 * in the biodone routine.
4518 bufdone(struct buf *bp)
4520 struct bufobj *dropobj;
4521 void (*biodone)(struct buf *);
4523 buf_track(bp, __func__);
4524 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
4527 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
4529 runningbufwakeup(bp);
4530 if (bp->b_iocmd == BIO_WRITE)
4531 dropobj = bp->b_bufobj;
4532 /* call optional completion function if requested */
4533 if (bp->b_iodone != NULL) {
4534 biodone = bp->b_iodone;
4535 bp->b_iodone = NULL;
4538 bufobj_wdrop(dropobj);
4541 if (bp->b_flags & B_VMIO) {
4543 * Set B_CACHE if the op was a normal read and no error
4544 * occurred. B_CACHE is set for writes in the b*write()
4547 if (bp->b_iocmd == BIO_READ &&
4548 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
4549 !(bp->b_ioflags & BIO_ERROR))
4550 bp->b_flags |= B_CACHE;
4551 vfs_vmio_iodone(bp);
4553 if (!LIST_EMPTY(&bp->b_dep))
4555 if ((bp->b_flags & B_CKHASH) != 0) {
4556 KASSERT(bp->b_iocmd == BIO_READ,
4557 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd));
4558 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp));
4559 (*bp->b_ckhashcalc)(bp);
4562 * For asynchronous completions, release the buffer now. The brelse
4563 * will do a wakeup there if necessary - so no need to do a wakeup
4564 * here in the async case. The sync case always needs to do a wakeup.
4566 if (bp->b_flags & B_ASYNC) {
4567 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) ||
4568 (bp->b_ioflags & BIO_ERROR))
4575 bufobj_wdrop(dropobj);
4579 * This routine is called in lieu of iodone in the case of
4580 * incomplete I/O. This keeps the busy status for pages
4584 vfs_unbusy_pages(struct buf *bp)
4590 runningbufwakeup(bp);
4591 if (!(bp->b_flags & B_VMIO))
4594 obj = bp->b_bufobj->bo_object;
4595 for (i = 0; i < bp->b_npages; i++) {
4597 if (m == bogus_page) {
4598 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i);
4600 panic("vfs_unbusy_pages: page missing\n");
4602 if (buf_mapped(bp)) {
4603 BUF_CHECK_MAPPED(bp);
4604 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4605 bp->b_pages, bp->b_npages);
4607 BUF_CHECK_UNMAPPED(bp);
4611 vm_object_pip_wakeupn(obj, bp->b_npages);
4615 * vfs_page_set_valid:
4617 * Set the valid bits in a page based on the supplied offset. The
4618 * range is restricted to the buffer's size.
4620 * This routine is typically called after a read completes.
4623 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4628 * Compute the end offset, eoff, such that [off, eoff) does not span a
4629 * page boundary and eoff is not greater than the end of the buffer.
4630 * The end of the buffer, in this case, is our file EOF, not the
4631 * allocation size of the buffer.
4633 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4634 if (eoff > bp->b_offset + bp->b_bcount)
4635 eoff = bp->b_offset + bp->b_bcount;
4638 * Set valid range. This is typically the entire buffer and thus the
4642 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4646 * vfs_page_set_validclean:
4648 * Set the valid bits and clear the dirty bits in a page based on the
4649 * supplied offset. The range is restricted to the buffer's size.
4652 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4654 vm_ooffset_t soff, eoff;
4657 * Start and end offsets in buffer. eoff - soff may not cross a
4658 * page boundary or cross the end of the buffer. The end of the
4659 * buffer, in this case, is our file EOF, not the allocation size
4663 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4664 if (eoff > bp->b_offset + bp->b_bcount)
4665 eoff = bp->b_offset + bp->b_bcount;
4668 * Set valid range. This is typically the entire buffer and thus the
4672 vm_page_set_validclean(
4674 (vm_offset_t) (soff & PAGE_MASK),
4675 (vm_offset_t) (eoff - soff)
4681 * Acquire a shared busy on all pages in the buf.
4684 vfs_busy_pages_acquire(struct buf *bp)
4688 for (i = 0; i < bp->b_npages; i++)
4689 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY);
4693 vfs_busy_pages_release(struct buf *bp)
4697 for (i = 0; i < bp->b_npages; i++)
4698 vm_page_sunbusy(bp->b_pages[i]);
4702 * This routine is called before a device strategy routine.
4703 * It is used to tell the VM system that paging I/O is in
4704 * progress, and treat the pages associated with the buffer
4705 * almost as being exclusive busy. Also the object paging_in_progress
4706 * flag is handled to make sure that the object doesn't become
4709 * Since I/O has not been initiated yet, certain buffer flags
4710 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4711 * and should be ignored.
4714 vfs_busy_pages(struct buf *bp, int clear_modify)
4722 if (!(bp->b_flags & B_VMIO))
4725 obj = bp->b_bufobj->bo_object;
4726 foff = bp->b_offset;
4727 KASSERT(bp->b_offset != NOOFFSET,
4728 ("vfs_busy_pages: no buffer offset"));
4729 if ((bp->b_flags & B_CLUSTER) == 0) {
4730 vm_object_pip_add(obj, bp->b_npages);
4731 vfs_busy_pages_acquire(bp);
4733 if (bp->b_bufsize != 0)
4734 vfs_setdirty_range(bp);
4736 for (i = 0; i < bp->b_npages; i++) {
4738 vm_page_assert_sbusied(m);
4741 * When readying a buffer for a read ( i.e
4742 * clear_modify == 0 ), it is important to do
4743 * bogus_page replacement for valid pages in
4744 * partially instantiated buffers. Partially
4745 * instantiated buffers can, in turn, occur when
4746 * reconstituting a buffer from its VM backing store
4747 * base. We only have to do this if B_CACHE is
4748 * clear ( which causes the I/O to occur in the
4749 * first place ). The replacement prevents the read
4750 * I/O from overwriting potentially dirty VM-backed
4751 * pages. XXX bogus page replacement is, uh, bogus.
4752 * It may not work properly with small-block devices.
4753 * We need to find a better way.
4756 pmap_remove_write(m);
4757 vfs_page_set_validclean(bp, foff, m);
4758 } else if (vm_page_all_valid(m) &&
4759 (bp->b_flags & B_CACHE) == 0) {
4760 bp->b_pages[i] = bogus_page;
4763 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4765 if (bogus && buf_mapped(bp)) {
4766 BUF_CHECK_MAPPED(bp);
4767 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4768 bp->b_pages, bp->b_npages);
4773 * vfs_bio_set_valid:
4775 * Set the range within the buffer to valid. The range is
4776 * relative to the beginning of the buffer, b_offset. Note that
4777 * b_offset itself may be offset from the beginning of the first
4781 vfs_bio_set_valid(struct buf *bp, int base, int size)
4786 if (!(bp->b_flags & B_VMIO))
4790 * Fixup base to be relative to beginning of first page.
4791 * Set initial n to be the maximum number of bytes in the
4792 * first page that can be validated.
4794 base += (bp->b_offset & PAGE_MASK);
4795 n = PAGE_SIZE - (base & PAGE_MASK);
4798 * Busy may not be strictly necessary here because the pages are
4799 * unlikely to be fully valid and the vnode lock will synchronize
4800 * their access via getpages. It is grabbed for consistency with
4801 * other page validation.
4803 vfs_busy_pages_acquire(bp);
4804 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4808 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4813 vfs_busy_pages_release(bp);
4819 * If the specified buffer is a non-VMIO buffer, clear the entire
4820 * buffer. If the specified buffer is a VMIO buffer, clear and
4821 * validate only the previously invalid portions of the buffer.
4822 * This routine essentially fakes an I/O, so we need to clear
4823 * BIO_ERROR and B_INVAL.
4825 * Note that while we only theoretically need to clear through b_bcount,
4826 * we go ahead and clear through b_bufsize.
4829 vfs_bio_clrbuf(struct buf *bp)
4831 int i, j, mask, sa, ea, slide;
4833 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4837 bp->b_flags &= ~B_INVAL;
4838 bp->b_ioflags &= ~BIO_ERROR;
4839 vfs_busy_pages_acquire(bp);
4840 sa = bp->b_offset & PAGE_MASK;
4842 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4843 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4844 ea = slide & PAGE_MASK;
4847 if (bp->b_pages[i] == bogus_page)
4850 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4851 if ((bp->b_pages[i]->valid & mask) == mask)
4853 if ((bp->b_pages[i]->valid & mask) == 0)
4854 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4856 for (; sa < ea; sa += DEV_BSIZE, j++) {
4857 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4858 pmap_zero_page_area(bp->b_pages[i],
4863 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE,
4864 roundup2(ea - sa, DEV_BSIZE));
4866 vfs_busy_pages_release(bp);
4871 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4876 if (buf_mapped(bp)) {
4877 BUF_CHECK_MAPPED(bp);
4878 bzero(bp->b_data + base, size);
4880 BUF_CHECK_UNMAPPED(bp);
4881 n = PAGE_SIZE - (base & PAGE_MASK);
4882 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4886 pmap_zero_page_area(m, base & PAGE_MASK, n);
4895 * Update buffer flags based on I/O request parameters, optionally releasing the
4896 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM,
4897 * where they may be placed on a page queue (VMIO) or freed immediately (direct
4898 * I/O). Otherwise the buffer is released to the cache.
4901 b_io_dismiss(struct buf *bp, int ioflag, bool release)
4904 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0,
4905 ("buf %p non-VMIO noreuse", bp));
4907 if ((ioflag & IO_DIRECT) != 0)
4908 bp->b_flags |= B_DIRECT;
4909 if ((ioflag & IO_EXT) != 0)
4910 bp->b_xflags |= BX_ALTDATA;
4911 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) {
4912 bp->b_flags |= B_RELBUF;
4913 if ((ioflag & IO_NOREUSE) != 0)
4914 bp->b_flags |= B_NOREUSE;
4922 vfs_bio_brelse(struct buf *bp, int ioflag)
4925 b_io_dismiss(bp, ioflag, true);
4929 vfs_bio_set_flags(struct buf *bp, int ioflag)
4932 b_io_dismiss(bp, ioflag, false);
4936 * vm_hold_load_pages and vm_hold_free_pages get pages into
4937 * a buffers address space. The pages are anonymous and are
4938 * not associated with a file object.
4941 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4947 BUF_CHECK_MAPPED(bp);
4949 to = round_page(to);
4950 from = round_page(from);
4951 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4952 MPASS((bp->b_flags & B_MAXPHYS) == 0);
4953 KASSERT(to - from <= maxbcachebuf,
4954 ("vm_hold_load_pages too large %p %#jx %#jx %u",
4955 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf));
4957 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4959 * note: must allocate system pages since blocking here
4960 * could interfere with paging I/O, no matter which
4963 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
4964 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK);
4965 pmap_qenter(pg, &p, 1);
4966 bp->b_pages[index] = p;
4968 bp->b_npages = index;
4971 /* Return pages associated with this buf to the vm system */
4973 vm_hold_free_pages(struct buf *bp, int newbsize)
4977 int index, newnpages;
4979 BUF_CHECK_MAPPED(bp);
4981 from = round_page((vm_offset_t)bp->b_data + newbsize);
4982 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4983 if (bp->b_npages > newnpages)
4984 pmap_qremove(from, bp->b_npages - newnpages);
4985 for (index = newnpages; index < bp->b_npages; index++) {
4986 p = bp->b_pages[index];
4987 bp->b_pages[index] = NULL;
4988 vm_page_unwire_noq(p);
4991 bp->b_npages = newnpages;
4995 * Map an IO request into kernel virtual address space.
4997 * All requests are (re)mapped into kernel VA space.
4998 * Notice that we use b_bufsize for the size of the buffer
4999 * to be mapped. b_bcount might be modified by the driver.
5001 * Note that even if the caller determines that the address space should
5002 * be valid, a race or a smaller-file mapped into a larger space may
5003 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
5004 * check the return value.
5006 * This function only works with pager buffers.
5009 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf)
5014 MPASS((bp->b_flags & B_MAXPHYS) != 0);
5015 prot = VM_PROT_READ;
5016 if (bp->b_iocmd == BIO_READ)
5017 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
5018 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
5019 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES);
5022 bp->b_bufsize = len;
5023 bp->b_npages = pidx;
5024 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK;
5025 if (mapbuf || !unmapped_buf_allowed) {
5026 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx);
5027 bp->b_data = bp->b_kvabase + bp->b_offset;
5029 bp->b_data = unmapped_buf;
5034 * Free the io map PTEs associated with this IO operation.
5035 * We also invalidate the TLB entries and restore the original b_addr.
5037 * This function only works with pager buffers.
5040 vunmapbuf(struct buf *bp)
5044 npages = bp->b_npages;
5046 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
5047 vm_page_unhold_pages(bp->b_pages, npages);
5049 bp->b_data = unmapped_buf;
5053 bdone(struct buf *bp)
5057 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5059 bp->b_flags |= B_DONE;
5065 bwait(struct buf *bp, u_char pri, const char *wchan)
5069 mtxp = mtx_pool_find(mtxpool_sleep, bp);
5071 while ((bp->b_flags & B_DONE) == 0)
5072 msleep(bp, mtxp, pri, wchan, 0);
5077 bufsync(struct bufobj *bo, int waitfor)
5080 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread));
5084 bufstrategy(struct bufobj *bo, struct buf *bp)
5090 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
5091 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
5092 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
5093 i = VOP_STRATEGY(vp, bp);
5094 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
5098 * Initialize a struct bufobj before use. Memory is assumed zero filled.
5101 bufobj_init(struct bufobj *bo, void *private)
5103 static volatile int bufobj_cleanq;
5106 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains;
5107 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
5108 bo->bo_private = private;
5109 TAILQ_INIT(&bo->bo_clean.bv_hd);
5110 TAILQ_INIT(&bo->bo_dirty.bv_hd);
5114 bufobj_wrefl(struct bufobj *bo)
5117 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5118 ASSERT_BO_WLOCKED(bo);
5123 bufobj_wref(struct bufobj *bo)
5126 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
5133 bufobj_wdrop(struct bufobj *bo)
5136 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
5138 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
5139 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
5140 bo->bo_flag &= ~BO_WWAIT;
5141 wakeup(&bo->bo_numoutput);
5147 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
5151 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
5152 ASSERT_BO_WLOCKED(bo);
5154 while (bo->bo_numoutput) {
5155 bo->bo_flag |= BO_WWAIT;
5156 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
5157 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
5165 * Set bio_data or bio_ma for struct bio from the struct buf.
5168 bdata2bio(struct buf *bp, struct bio *bip)
5171 if (!buf_mapped(bp)) {
5172 KASSERT(unmapped_buf_allowed, ("unmapped"));
5173 bip->bio_ma = bp->b_pages;
5174 bip->bio_ma_n = bp->b_npages;
5175 bip->bio_data = unmapped_buf;
5176 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
5177 bip->bio_flags |= BIO_UNMAPPED;
5178 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
5179 PAGE_SIZE == bp->b_npages,
5180 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
5181 (long long)bip->bio_length, bip->bio_ma_n));
5183 bip->bio_data = bp->b_data;
5189 * The MIPS pmap code currently doesn't handle aliased pages.
5190 * The VIPT caches may not handle page aliasing themselves, leading
5191 * to data corruption.
5193 * As such, this code makes a system extremely unhappy if said
5194 * system doesn't support unaliasing the above situation in hardware.
5195 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable
5196 * this feature at build time, so it has to be handled in software.
5198 * Once the MIPS pmap/cache code grows to support this function on
5199 * earlier chips, it should be flipped back off.
5202 static int buf_pager_relbuf = 1;
5204 static int buf_pager_relbuf = 0;
5206 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN,
5207 &buf_pager_relbuf, 0,
5208 "Make buffer pager release buffers after reading");
5211 * The buffer pager. It uses buffer reads to validate pages.
5213 * In contrast to the generic local pager from vm/vnode_pager.c, this
5214 * pager correctly and easily handles volumes where the underlying
5215 * device block size is greater than the machine page size. The
5216 * buffer cache transparently extends the requested page run to be
5217 * aligned at the block boundary, and does the necessary bogus page
5218 * replacements in the addends to avoid obliterating already valid
5221 * The only non-trivial issue is that the exclusive busy state for
5222 * pages, which is assumed by the vm_pager_getpages() interface, is
5223 * incompatible with the VMIO buffer cache's desire to share-busy the
5224 * pages. This function performs a trivial downgrade of the pages'
5225 * state before reading buffers, and a less trivial upgrade from the
5226 * shared-busy to excl-busy state after the read.
5229 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count,
5230 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno,
5231 vbg_get_blksize_t get_blksize)
5238 vm_ooffset_t la, lb, poff, poffe;
5240 int br_flags, error, i, pgsin, pgsin_a, pgsin_b;
5243 object = vp->v_object;
5246 la = IDX_TO_OFF(ma[count - 1]->pindex);
5247 if (la >= object->un_pager.vnp.vnp_size)
5248 return (VM_PAGER_BAD);
5251 * Change the meaning of la from where the last requested page starts
5252 * to where it ends, because that's the end of the requested region
5253 * and the start of the potential read-ahead region.
5256 lpart = la > object->un_pager.vnp.vnp_size;
5257 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)),
5260 return (VM_PAGER_ERROR);
5263 * Calculate read-ahead, behind and total pages.
5266 lb = IDX_TO_OFF(ma[0]->pindex);
5267 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs));
5269 if (rbehind != NULL)
5271 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la);
5272 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size)
5273 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size,
5278 VM_CNT_INC(v_vnodein);
5279 VM_CNT_ADD(v_vnodepgsin, pgsin);
5281 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS)
5282 != 0) ? GB_UNMAPPED : 0;
5284 for (i = 0; i < count; i++) {
5285 if (ma[i] != bogus_page)
5286 vm_page_busy_downgrade(ma[i]);
5290 for (i = 0; i < count; i++) {
5292 if (m == bogus_page)
5296 * Pages are shared busy and the object lock is not
5297 * owned, which together allow for the pages'
5298 * invalidation. The racy test for validity avoids
5299 * useless creation of the buffer for the most typical
5300 * case when invalidation is not used in redo or for
5301 * parallel read. The shared->excl upgrade loop at
5302 * the end of the function catches the race in a
5303 * reliable way (protected by the object lock).
5305 if (vm_page_all_valid(m))
5308 poff = IDX_TO_OFF(m->pindex);
5309 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size);
5310 for (; poff < poffe; poff += bsize) {
5311 lbn = get_lblkno(vp, poff);
5316 error = get_blksize(vp, lbn, &bsize);
5318 error = bread_gb(vp, lbn, bsize,
5319 curthread->td_ucred, br_flags, &bp);
5322 if (bp->b_rcred == curthread->td_ucred) {
5323 crfree(bp->b_rcred);
5324 bp->b_rcred = NOCRED;
5326 if (LIST_EMPTY(&bp->b_dep)) {
5328 * Invalidation clears m->valid, but
5329 * may leave B_CACHE flag if the
5330 * buffer existed at the invalidation
5331 * time. In this case, recycle the
5332 * buffer to do real read on next
5333 * bread() after redo.
5335 * Otherwise B_RELBUF is not strictly
5336 * necessary, enable to reduce buf
5339 if (buf_pager_relbuf ||
5340 !vm_page_all_valid(m))
5341 bp->b_flags |= B_RELBUF;
5343 bp->b_flags &= ~B_NOCACHE;
5349 KASSERT(1 /* racy, enable for debugging */ ||
5350 vm_page_all_valid(m) || i == count - 1,
5351 ("buf %d %p invalid", i, m));
5352 if (i == count - 1 && lpart) {
5353 if (!vm_page_none_valid(m) &&
5354 !vm_page_all_valid(m))
5355 vm_page_zero_invalid(m, TRUE);
5362 for (i = 0; i < count; i++) {
5363 if (ma[i] == bogus_page)
5365 if (vm_page_busy_tryupgrade(ma[i]) == 0) {
5366 vm_page_sunbusy(ma[i]);
5367 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex,
5372 * Since the pages were only sbusy while neither the
5373 * buffer nor the object lock was held by us, or
5374 * reallocated while vm_page_grab() slept for busy
5375 * relinguish, they could have been invalidated.
5376 * Recheck the valid bits and re-read as needed.
5378 * Note that the last page is made fully valid in the
5379 * read loop, and partial validity for the page at
5380 * index count - 1 could mean that the page was
5381 * invalidated or removed, so we must restart for
5384 if (!vm_page_all_valid(ma[i]))
5387 if (redo && error == 0)
5389 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK);
5392 #include "opt_ddb.h"
5394 #include <ddb/ddb.h>
5396 /* DDB command to show buffer data */
5397 DB_SHOW_COMMAND(buffer, db_show_buffer)
5400 struct buf *bp = (struct buf *)addr;
5401 #ifdef FULL_BUF_TRACKING
5406 db_printf("usage: show buffer <addr>\n");
5410 db_printf("buf at %p\n", bp);
5411 db_printf("b_flags = 0x%b, b_xflags=0x%b\n",
5412 (u_int)bp->b_flags, PRINT_BUF_FLAGS,
5413 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS);
5414 db_printf("b_vflags=0x%b b_ioflags0x%b\n",
5415 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS,
5416 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS);
5418 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
5419 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, "
5420 "b_vp = %p, b_dep = %p\n",
5421 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5422 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
5423 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first);
5424 db_printf("b_kvabase = %p, b_kvasize = %d\n",
5425 bp->b_kvabase, bp->b_kvasize);
5428 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
5429 for (i = 0; i < bp->b_npages; i++) {
5433 db_printf("(%p, 0x%lx, 0x%lx)", m->object,
5435 (u_long)VM_PAGE_TO_PHYS(m));
5437 db_printf("( ??? )");
5438 if ((i + 1) < bp->b_npages)
5443 BUF_LOCKPRINTINFO(bp);
5444 #if defined(FULL_BUF_TRACKING)
5445 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt);
5447 i = bp->b_io_tcnt % BUF_TRACKING_SIZE;
5448 for (j = 1; j <= BUF_TRACKING_SIZE; j++) {
5449 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL)
5451 db_printf(" %2u: %s\n", j,
5452 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]);
5454 #elif defined(BUF_TRACKING)
5455 db_printf("b_io_tracking: %s\n", bp->b_io_tracking);
5460 DB_SHOW_COMMAND(bufqueues, bufqueues)
5462 struct bufdomain *bd;
5467 db_printf("bqempty: %d\n", bqempty.bq_len);
5469 for (i = 0; i < buf_domains; i++) {
5471 db_printf("Buf domain %d\n", i);
5472 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers);
5473 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers);
5474 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers);
5476 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace);
5477 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace);
5478 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace);
5479 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace);
5480 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh);
5482 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers);
5483 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers);
5484 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers);
5485 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh);
5488 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist)
5489 total += bp->b_bufsize;
5490 db_printf("\tcleanq count\t%d (%ld)\n",
5491 bd->bd_cleanq->bq_len, total);
5493 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist)
5494 total += bp->b_bufsize;
5495 db_printf("\tdirtyq count\t%d (%ld)\n",
5496 bd->bd_dirtyq.bq_len, total);
5497 db_printf("\twakeup\t\t%d\n", bd->bd_wanted);
5498 db_printf("\tlim\t\t%d\n", bd->bd_lim);
5499 db_printf("\tCPU ");
5500 for (j = 0; j <= mp_maxid; j++)
5501 db_printf("%d, ", bd->bd_subq[j].bq_len);
5505 for (j = 0; j < nbuf; j++) {
5507 if (bp->b_domain == i && BUF_ISLOCKED(bp)) {
5509 total += bp->b_bufsize;
5512 db_printf("\tLocked buffers: %d space %ld\n", cnt, total);
5515 for (j = 0; j < nbuf; j++) {
5517 if (bp->b_domain == i) {
5519 total += bp->b_bufsize;
5522 db_printf("\tTotal buffers: %d space %ld\n", cnt, total);
5526 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
5531 for (i = 0; i < nbuf; i++) {
5533 if (BUF_ISLOCKED(bp)) {
5534 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5542 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
5548 db_printf("usage: show vnodebufs <addr>\n");
5551 vp = (struct vnode *)addr;
5552 db_printf("Clean buffers:\n");
5553 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
5554 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5557 db_printf("Dirty buffers:\n");
5558 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
5559 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
5564 DB_COMMAND(countfreebufs, db_coundfreebufs)
5567 int i, used = 0, nfree = 0;
5570 db_printf("usage: countfreebufs\n");
5574 for (i = 0; i < nbuf; i++) {
5576 if (bp->b_qindex == QUEUE_EMPTY)
5582 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
5584 db_printf("numfreebuffers is %d\n", numfreebuffers);