2 * Copyright (c) 2004 Poul-Henning Kamp
3 * Copyright (c) 1994,1997 John S. Dyson
4 * Copyright (c) 2013 The FreeBSD Foundation
7 * Portions of this software were developed by Konstantin Belousov
8 * under sponsorship from the FreeBSD Foundation.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * this file contains a new buffer I/O scheme implementing a coherent
34 * VM object and buffer cache scheme. Pains have been taken to make
35 * sure that the performance degradation associated with schemes such
36 * as this is not realized.
38 * Author: John S. Dyson
39 * Significant help during the development and debugging phases
40 * had been provided by David Greenman, also of the FreeBSD core team.
42 * see man buf(9) for more info.
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include <sys/param.h>
49 #include <sys/systm.h>
53 #include <sys/devicestat.h>
54 #include <sys/eventhandler.h>
56 #include <sys/limits.h>
58 #include <sys/malloc.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/kernel.h>
62 #include <sys/kthread.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
66 #include <sys/sysctl.h>
68 #include <sys/vmmeter.h>
69 #include <sys/vnode.h>
70 #include <geom/geom.h>
72 #include <vm/vm_param.h>
73 #include <vm/vm_kern.h>
74 #include <vm/vm_pageout.h>
75 #include <vm/vm_page.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/vm_map.h>
79 #include "opt_compat.h"
82 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
84 struct bio_ops bioops; /* I/O operation notification */
86 struct buf_ops buf_ops_bio = {
87 .bop_name = "buf_ops_bio",
88 .bop_write = bufwrite,
89 .bop_strategy = bufstrategy,
91 .bop_bdflush = bufbdflush,
95 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
96 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
98 struct buf *buf; /* buffer header pool */
101 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */
102 struct proc *bufdaemonproc;
104 static int inmem(struct vnode *vp, daddr_t blkno);
105 static void vm_hold_free_pages(struct buf *bp, int newbsize);
106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
111 static void vfs_clean_pages_dirty_buf(struct buf *bp);
112 static void vfs_setdirty_locked_object(struct buf *bp);
113 static void vfs_vmio_release(struct buf *bp);
114 static int vfs_bio_clcheck(struct vnode *vp, int size,
115 daddr_t lblkno, daddr_t blkno);
116 static int buf_flush(struct vnode *vp, int);
117 static int flushbufqueues(struct vnode *, int, int);
118 static void buf_daemon(void);
119 static void bremfreel(struct buf *bp);
120 static __inline void bd_wakeup(void);
121 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
122 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
123 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
126 int vmiodirenable = TRUE;
127 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
128 "Use the VM system for directory writes");
129 long runningbufspace;
130 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
131 "Amount of presently outstanding async buffer io");
132 static long bufspace;
133 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
134 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
135 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
136 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
138 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
139 "Virtual memory used for buffers");
141 static long unmapped_bufspace;
142 SYSCTL_LONG(_vfs, OID_AUTO, unmapped_bufspace, CTLFLAG_RD,
143 &unmapped_bufspace, 0,
144 "Amount of unmapped buffers, inclusive in the bufspace");
145 static long maxbufspace;
146 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
147 "Maximum allowed value of bufspace (including buf_daemon)");
148 static long bufmallocspace;
149 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
150 "Amount of malloced memory for buffers");
151 static long maxbufmallocspace;
152 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
153 "Maximum amount of malloced memory for buffers");
154 static long lobufspace;
155 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
156 "Minimum amount of buffers we want to have");
158 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
159 "Maximum allowed value of bufspace (excluding buf_daemon)");
160 static int bufreusecnt;
161 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
162 "Number of times we have reused a buffer");
163 static int buffreekvacnt;
164 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
165 "Number of times we have freed the KVA space from some buffer");
166 static int bufdefragcnt;
167 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
168 "Number of times we have had to repeat buffer allocation to defragment");
169 static long lorunningspace;
170 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
171 "Minimum preferred space used for in-progress I/O");
172 static long hirunningspace;
173 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
174 "Maximum amount of space to use for in-progress I/O");
175 int dirtybufferflushes;
176 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
177 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
179 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
180 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
181 int altbufferflushes;
182 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
183 0, "Number of fsync flushes to limit dirty buffers");
184 static int recursiveflushes;
185 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
186 0, "Number of flushes skipped due to being recursive");
187 static int numdirtybuffers;
188 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
189 "Number of buffers that are dirty (has unwritten changes) at the moment");
190 static int lodirtybuffers;
191 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
192 "How many buffers we want to have free before bufdaemon can sleep");
193 static int hidirtybuffers;
194 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
195 "When the number of dirty buffers is considered severe");
197 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
198 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
199 static int numfreebuffers;
200 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
201 "Number of free buffers");
202 static int lofreebuffers;
203 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
205 static int hifreebuffers;
206 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
207 "XXX Complicatedly unused");
208 static int getnewbufcalls;
209 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
210 "Number of calls to getnewbuf");
211 static int getnewbufrestarts;
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
213 "Number of times getnewbuf has had to restart a buffer aquisition");
214 static int mappingrestarts;
215 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
216 "Number of times getblk has had to restart a buffer mapping for "
218 static int flushbufqtarget = 100;
219 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
220 "Amount of work to do in flushbufqueues when helping bufdaemon");
221 static long notbufdflushes;
222 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0,
223 "Number of dirty buffer flushes done by the bufdaemon helpers");
224 static long barrierwrites;
225 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
226 "Number of barrier writes");
227 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
228 &unmapped_buf_allowed, 0,
229 "Permit the use of the unmapped i/o");
232 * Lock for the non-dirty bufqueues
234 static struct mtx_padalign bqclean;
237 * Lock for the dirty queue.
239 static struct mtx_padalign bqdirty;
242 * This lock synchronizes access to bd_request.
244 static struct mtx_padalign bdlock;
247 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
248 * waitrunningbufspace().
250 static struct mtx_padalign rbreqlock;
253 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
255 static struct rwlock_padalign nblock;
258 * Lock that protects bdirtywait.
260 static struct mtx_padalign bdirtylock;
263 * Wakeup point for bufdaemon, as well as indicator of whether it is already
264 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
267 static int bd_request;
270 * Request for the buf daemon to write more buffers than is indicated by
271 * lodirtybuf. This may be necessary to push out excess dependencies or
272 * defragment the address space where a simple count of the number of dirty
273 * buffers is insufficient to characterize the demand for flushing them.
275 static int bd_speedupreq;
278 * bogus page -- for I/O to/from partially complete buffers
279 * this is a temporary solution to the problem, but it is not
280 * really that bad. it would be better to split the buffer
281 * for input in the case of buffers partially already in memory,
282 * but the code is intricate enough already.
284 vm_page_t bogus_page;
287 * Synchronization (sleep/wakeup) variable for active buffer space requests.
288 * Set when wait starts, cleared prior to wakeup().
289 * Used in runningbufwakeup() and waitrunningbufspace().
291 static int runningbufreq;
294 * Synchronization (sleep/wakeup) variable for buffer requests.
295 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
297 * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
298 * getnewbuf(), and getblk().
300 static volatile int needsbuffer;
303 * Synchronization for bwillwrite() waiters.
305 static int bdirtywait;
308 * Definitions for the buffer free lists.
310 #define BUFFER_QUEUES 5 /* number of free buffer queues */
312 #define QUEUE_NONE 0 /* on no queue */
313 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
314 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
315 #define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */
316 #define QUEUE_EMPTY 4 /* empty buffer headers */
317 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */
319 /* Queues for free buffers with various properties */
320 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
322 static int bq_len[BUFFER_QUEUES];
326 * Single global constant for BUF_WMESG, to avoid getting multiple references.
327 * buf_wmesg is referred from macros.
329 const char *buf_wmesg = BUF_WMESG;
331 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
332 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
333 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
335 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
336 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
338 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
343 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
344 return (sysctl_handle_long(oidp, arg1, arg2, req));
345 lvalue = *(long *)arg1;
346 if (lvalue > INT_MAX)
347 /* On overflow, still write out a long to trigger ENOMEM. */
348 return (sysctl_handle_long(oidp, &lvalue, 0, req));
350 return (sysctl_handle_int(oidp, &ivalue, 0, req));
357 * Return the appropriate queue lock based on the index.
359 static inline struct mtx *
363 if (qindex == QUEUE_DIRTY)
364 return (struct mtx *)(&bqdirty);
365 return (struct mtx *)(&bqclean);
371 * Wakeup any bwillwrite() waiters.
376 mtx_lock(&bdirtylock);
381 mtx_unlock(&bdirtylock);
387 * Decrement the numdirtybuffers count by one and wakeup any
388 * threads blocked in bwillwrite().
394 if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
395 (lodirtybuffers + hidirtybuffers) / 2)
402 * Increment the numdirtybuffers count by one and wakeup the buf
410 * Only do the wakeup once as we cross the boundary. The
411 * buf daemon will keep running until the condition clears.
413 if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
414 (lodirtybuffers + hidirtybuffers) / 2)
421 * Called when buffer space is potentially available for recovery.
422 * getnewbuf() will block on this flag when it is unable to free
423 * sufficient buffer space. Buffer space becomes recoverable when
424 * bp's get placed back in the queues.
433 * If someone is waiting for BUF space, wake them up. Even
434 * though we haven't freed the kva space yet, the waiting
435 * process will be able to now.
441 if ((on & VFS_BIO_NEED_BUFSPACE) == 0)
444 if (atomic_cmpset_rel_int(&needsbuffer, on,
445 on & ~VFS_BIO_NEED_BUFSPACE))
449 wakeup(__DEVOLATILE(void *, &needsbuffer));
456 * Wake up processes that are waiting on asynchronous writes to fall
457 * below lorunningspace.
463 mtx_lock(&rbreqlock);
466 wakeup(&runningbufreq);
468 mtx_unlock(&rbreqlock);
474 * Decrement the outstanding write count according.
477 runningbufwakeup(struct buf *bp)
481 bspace = bp->b_runningbufspace;
484 space = atomic_fetchadd_long(&runningbufspace, -bspace);
485 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
487 bp->b_runningbufspace = 0;
489 * Only acquire the lock and wakeup on the transition from exceeding
490 * the threshold to falling below it.
492 if (space < lorunningspace)
494 if (space - bspace > lorunningspace)
502 * Called when a buffer has been added to one of the free queues to
503 * account for the buffer and to wakeup anyone waiting for free buffers.
504 * This typically occurs when large amounts of metadata are being handled
505 * by the buffer cache ( else buffer space runs out first, usually ).
508 bufcountadd(struct buf *bp)
510 int mask, need_wakeup, old, on;
512 KASSERT((bp->b_flags & B_INFREECNT) == 0,
513 ("buf %p already counted as free", bp));
514 bp->b_flags |= B_INFREECNT;
515 old = atomic_fetchadd_int(&numfreebuffers, 1);
516 KASSERT(old >= 0 && old < nbuf,
517 ("numfreebuffers climbed to %d", old + 1));
518 mask = VFS_BIO_NEED_ANY;
519 if (numfreebuffers >= hifreebuffers)
520 mask |= VFS_BIO_NEED_FREE;
528 if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask))
532 wakeup(__DEVOLATILE(void *, &needsbuffer));
539 * Decrement the numfreebuffers count as needed.
542 bufcountsub(struct buf *bp)
547 * Fixup numfreebuffers count. If the buffer is invalid or not
548 * delayed-write, the buffer was free and we must decrement
551 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
552 KASSERT((bp->b_flags & B_INFREECNT) != 0,
553 ("buf %p not counted in numfreebuffers", bp));
554 bp->b_flags &= ~B_INFREECNT;
555 old = atomic_fetchadd_int(&numfreebuffers, -1);
556 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
561 * waitrunningbufspace()
563 * runningbufspace is a measure of the amount of I/O currently
564 * running. This routine is used in async-write situations to
565 * prevent creating huge backups of pending writes to a device.
566 * Only asynchronous writes are governed by this function.
568 * This does NOT turn an async write into a sync write. It waits
569 * for earlier writes to complete and generally returns before the
570 * caller's write has reached the device.
573 waitrunningbufspace(void)
576 mtx_lock(&rbreqlock);
577 while (runningbufspace > hirunningspace) {
579 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
581 mtx_unlock(&rbreqlock);
586 * vfs_buf_test_cache:
588 * Called when a buffer is extended. This function clears the B_CACHE
589 * bit if the newly extended portion of the buffer does not contain
594 vfs_buf_test_cache(struct buf *bp,
595 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
599 VM_OBJECT_ASSERT_LOCKED(m->object);
600 if (bp->b_flags & B_CACHE) {
601 int base = (foff + off) & PAGE_MASK;
602 if (vm_page_is_valid(m, base, size) == 0)
603 bp->b_flags &= ~B_CACHE;
607 /* Wake up the buffer daemon if necessary */
613 if (bd_request == 0) {
621 * bd_speedup - speedup the buffer cache flushing code
630 if (bd_speedupreq == 0 || bd_request == 0)
640 #define NSWBUF_MIN 16
644 #define TRANSIENT_DENOM 5
646 #define TRANSIENT_DENOM 10
650 * Calculating buffer cache scaling values and reserve space for buffer
651 * headers. This is called during low level kernel initialization and
652 * may be called more then once. We CANNOT write to the memory area
653 * being reserved at this time.
656 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
659 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
662 * physmem_est is in pages. Convert it to kilobytes (assumes
663 * PAGE_SIZE is >= 1K)
665 physmem_est = physmem_est * (PAGE_SIZE / 1024);
668 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
669 * For the first 64MB of ram nominally allocate sufficient buffers to
670 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
671 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
672 * the buffer cache we limit the eventual kva reservation to
675 * factor represents the 1/4 x ram conversion.
678 int factor = 4 * BKVASIZE / 1024;
681 if (physmem_est > 4096)
682 nbuf += min((physmem_est - 4096) / factor,
684 if (physmem_est > 65536)
685 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
686 32 * 1024 * 1024 / (factor * 5));
688 if (maxbcache && nbuf > maxbcache / BKVASIZE)
689 nbuf = maxbcache / BKVASIZE;
694 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
695 maxbuf = (LONG_MAX / 3) / BKVASIZE;
698 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
704 * Ideal allocation size for the transient bio submap if 10%
705 * of the maximal space buffer map. This roughly corresponds
706 * to the amount of the buffer mapped for typical UFS load.
708 * Clip the buffer map to reserve space for the transient
709 * BIOs, if its extent is bigger than 90% (80% on i386) of the
710 * maximum buffer map extent on the platform.
712 * The fall-back to the maxbuf in case of maxbcache unset,
713 * allows to not trim the buffer KVA for the architectures
714 * with ample KVA space.
716 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
717 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
718 buf_sz = (long)nbuf * BKVASIZE;
719 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
720 (TRANSIENT_DENOM - 1)) {
722 * There is more KVA than memory. Do not
723 * adjust buffer map size, and assign the rest
724 * of maxbuf to transient map.
726 biotmap_sz = maxbuf_sz - buf_sz;
729 * Buffer map spans all KVA we could afford on
730 * this platform. Give 10% (20% on i386) of
731 * the buffer map to the transient bio map.
733 biotmap_sz = buf_sz / TRANSIENT_DENOM;
734 buf_sz -= biotmap_sz;
736 if (biotmap_sz / INT_MAX > MAXPHYS)
737 bio_transient_maxcnt = INT_MAX;
739 bio_transient_maxcnt = biotmap_sz / MAXPHYS;
741 * Artifically limit to 1024 simultaneous in-flight I/Os
742 * using the transient mapping.
744 if (bio_transient_maxcnt > 1024)
745 bio_transient_maxcnt = 1024;
747 nbuf = buf_sz / BKVASIZE;
751 * swbufs are used as temporary holders for I/O, such as paging I/O.
752 * We have no less then 16 and no more then 256.
754 nswbuf = min(nbuf / 4, 256);
755 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf);
756 if (nswbuf < NSWBUF_MIN)
760 * Reserve space for the buffer cache buffers
763 v = (caddr_t)(swbuf + nswbuf);
765 v = (caddr_t)(buf + nbuf);
770 /* Initialize the buffer subsystem. Called before use of any buffers. */
777 CTASSERT(MAXBCACHEBUF >= MAXBSIZE);
778 mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
779 mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
780 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
781 rw_init(&nblock, "needsbuffer lock");
782 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
783 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
785 /* next, make a null set of free lists */
786 for (i = 0; i < BUFFER_QUEUES; i++)
787 TAILQ_INIT(&bufqueues[i]);
789 /* finally, initialize each buffer header and stick on empty q */
790 for (i = 0; i < nbuf; i++) {
792 bzero(bp, sizeof *bp);
793 bp->b_flags = B_INVAL | B_INFREECNT;
794 bp->b_rcred = NOCRED;
795 bp->b_wcred = NOCRED;
796 bp->b_qindex = QUEUE_EMPTY;
798 LIST_INIT(&bp->b_dep);
800 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
802 bq_len[QUEUE_EMPTY]++;
807 * maxbufspace is the absolute maximum amount of buffer space we are
808 * allowed to reserve in KVM and in real terms. The absolute maximum
809 * is nominally used by buf_daemon. hibufspace is the nominal maximum
810 * used by most other processes. The differential is required to
811 * ensure that buf_daemon is able to run when other processes might
812 * be blocked waiting for buffer space.
814 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
815 * this may result in KVM fragmentation which is not handled optimally
818 maxbufspace = (long)nbuf * BKVASIZE;
819 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10);
820 lobufspace = hibufspace - MAXBCACHEBUF;
823 * Note: The 16 MiB upper limit for hirunningspace was chosen
824 * arbitrarily and may need further tuning. It corresponds to
825 * 128 outstanding write IO requests (if IO size is 128 KiB),
826 * which fits with many RAID controllers' tagged queuing limits.
827 * The lower 1 MiB limit is the historical upper limit for
830 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF),
831 16 * 1024 * 1024), 1024 * 1024);
832 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF);
835 * Limit the amount of malloc memory since it is wired permanently into
836 * the kernel space. Even though this is accounted for in the buffer
837 * allocation, we don't want the malloced region to grow uncontrolled.
838 * The malloc scheme improves memory utilization significantly on average
839 * (small) directories.
841 maxbufmallocspace = hibufspace / 20;
844 * Reduce the chance of a deadlock occuring by limiting the number
845 * of delayed-write dirty buffers we allow to stack up.
847 hidirtybuffers = nbuf / 4 + 20;
848 dirtybufthresh = hidirtybuffers * 9 / 10;
851 * To support extreme low-memory systems, make sure hidirtybuffers cannot
852 * eat up all available buffer space. This occurs when our minimum cannot
853 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
854 * BKVASIZE'd buffers.
856 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
857 hidirtybuffers >>= 1;
859 lodirtybuffers = hidirtybuffers / 2;
862 * Try to keep the number of free buffers in the specified range,
863 * and give special processes (e.g. like buf_daemon) access to an
866 lofreebuffers = nbuf / 18 + 5;
867 hifreebuffers = 2 * lofreebuffers;
868 numfreebuffers = nbuf;
870 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
871 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
872 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
877 vfs_buf_check_mapped(struct buf *bp)
880 KASSERT((bp->b_flags & B_UNMAPPED) == 0,
881 ("mapped buf %p %x", bp, bp->b_flags));
882 KASSERT(bp->b_kvabase != unmapped_buf,
883 ("mapped buf: b_kvabase was not updated %p", bp));
884 KASSERT(bp->b_data != unmapped_buf,
885 ("mapped buf: b_data was not updated %p", bp));
889 vfs_buf_check_unmapped(struct buf *bp)
892 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
893 ("unmapped buf %p %x", bp, bp->b_flags));
894 KASSERT(bp->b_kvabase == unmapped_buf,
895 ("unmapped buf: corrupted b_kvabase %p", bp));
896 KASSERT(bp->b_data == unmapped_buf,
897 ("unmapped buf: corrupted b_data %p", bp));
900 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
901 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
903 #define BUF_CHECK_MAPPED(bp) do {} while (0)
904 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
908 bpmap_qenter(struct buf *bp)
911 BUF_CHECK_MAPPED(bp);
914 * bp->b_data is relative to bp->b_offset, but
915 * bp->b_offset may be offset into the first page.
917 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
918 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
919 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
920 (vm_offset_t)(bp->b_offset & PAGE_MASK));
924 * bfreekva() - free the kva allocation for a buffer.
926 * Since this call frees up buffer space, we call bufspacewakeup().
929 bfreekva(struct buf *bp)
932 if (bp->b_kvasize == 0)
935 atomic_add_int(&buffreekvacnt, 1);
936 atomic_subtract_long(&bufspace, bp->b_kvasize);
937 if ((bp->b_flags & B_UNMAPPED) == 0) {
938 BUF_CHECK_MAPPED(bp);
939 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
942 BUF_CHECK_UNMAPPED(bp);
943 if ((bp->b_flags & B_KVAALLOC) != 0) {
944 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
947 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
948 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
957 * Insert the buffer into the appropriate free list.
960 binsfree(struct buf *bp, int qindex)
962 struct mtx *olock, *nlock;
964 BUF_ASSERT_XLOCKED(bp);
966 nlock = bqlock(qindex);
967 /* Handle delayed bremfree() processing. */
968 if (bp->b_flags & B_REMFREE) {
969 olock = bqlock(bp->b_qindex);
972 if (olock != nlock) {
979 if (bp->b_qindex != QUEUE_NONE)
980 panic("binsfree: free buffer onto another queue???");
982 bp->b_qindex = qindex;
983 if (bp->b_flags & B_AGE)
984 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
986 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
988 bq_len[bp->b_qindex]++;
993 * Something we can maybe free or reuse.
995 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
998 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1005 * Mark the buffer for removal from the appropriate free list.
1009 bremfree(struct buf *bp)
1012 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1013 KASSERT((bp->b_flags & B_REMFREE) == 0,
1014 ("bremfree: buffer %p already marked for delayed removal.", bp));
1015 KASSERT(bp->b_qindex != QUEUE_NONE,
1016 ("bremfree: buffer %p not on a queue.", bp));
1017 BUF_ASSERT_XLOCKED(bp);
1019 bp->b_flags |= B_REMFREE;
1026 * Force an immediate removal from a free list. Used only in nfs when
1027 * it abuses the b_freelist pointer.
1030 bremfreef(struct buf *bp)
1034 qlock = bqlock(bp->b_qindex);
1043 * Removes a buffer from the free list, must be called with the
1044 * correct qlock held.
1047 bremfreel(struct buf *bp)
1050 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1051 bp, bp->b_vp, bp->b_flags);
1052 KASSERT(bp->b_qindex != QUEUE_NONE,
1053 ("bremfreel: buffer %p not on a queue.", bp));
1054 BUF_ASSERT_XLOCKED(bp);
1055 mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1057 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1059 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1061 bq_len[bp->b_qindex]--;
1063 bp->b_qindex = QUEUE_NONE;
1065 * If this was a delayed bremfree() we only need to remove the buffer
1066 * from the queue and return the stats are already done.
1068 if (bp->b_flags & B_REMFREE) {
1069 bp->b_flags &= ~B_REMFREE;
1076 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
1077 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1078 * the buffer is valid and we do not have to do anything.
1081 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1082 int cnt, struct ucred * cred)
1087 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1088 if (inmem(vp, *rablkno))
1090 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1092 if ((rabp->b_flags & B_CACHE) == 0) {
1093 if (!TD_IS_IDLETHREAD(curthread))
1094 curthread->td_ru.ru_inblock++;
1095 rabp->b_flags |= B_ASYNC;
1096 rabp->b_flags &= ~B_INVAL;
1097 rabp->b_ioflags &= ~BIO_ERROR;
1098 rabp->b_iocmd = BIO_READ;
1099 if (rabp->b_rcred == NOCRED && cred != NOCRED)
1100 rabp->b_rcred = crhold(cred);
1101 vfs_busy_pages(rabp, 0);
1103 rabp->b_iooffset = dbtob(rabp->b_blkno);
1112 * Entry point for bread() and breadn() via #defines in sys/buf.h.
1114 * Get a buffer with the specified data. Look in the cache first. We
1115 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
1116 * is set, the buffer is valid and we do not have to do anything, see
1117 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1120 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1121 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1124 int rv = 0, readwait = 0;
1126 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1128 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1130 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1134 /* if not found in cache, do some I/O */
1135 if ((bp->b_flags & B_CACHE) == 0) {
1136 if (!TD_IS_IDLETHREAD(curthread))
1137 curthread->td_ru.ru_inblock++;
1138 bp->b_iocmd = BIO_READ;
1139 bp->b_flags &= ~B_INVAL;
1140 bp->b_ioflags &= ~BIO_ERROR;
1141 if (bp->b_rcred == NOCRED && cred != NOCRED)
1142 bp->b_rcred = crhold(cred);
1143 vfs_busy_pages(bp, 0);
1144 bp->b_iooffset = dbtob(bp->b_blkno);
1149 breada(vp, rablkno, rabsize, cnt, cred);
1158 * Write, release buffer on completion. (Done by iodone
1159 * if async). Do not bother writing anything if the buffer
1162 * Note that we set B_CACHE here, indicating that buffer is
1163 * fully valid and thus cacheable. This is true even of NFS
1164 * now so we set it generally. This could be set either here
1165 * or in biodone() since the I/O is synchronous. We put it
1169 bufwrite(struct buf *bp)
1176 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1177 if (bp->b_flags & B_INVAL) {
1182 if (bp->b_flags & B_BARRIER)
1185 oldflags = bp->b_flags;
1187 BUF_ASSERT_HELD(bp);
1189 if (bp->b_pin_count > 0)
1192 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1193 ("FFS background buffer should not get here %p", bp));
1197 vp_md = vp->v_vflag & VV_MD;
1202 * Mark the buffer clean. Increment the bufobj write count
1203 * before bundirty() call, to prevent other thread from seeing
1204 * empty dirty list and zero counter for writes in progress,
1205 * falsely indicating that the bufobj is clean.
1207 bufobj_wref(bp->b_bufobj);
1210 bp->b_flags &= ~B_DONE;
1211 bp->b_ioflags &= ~BIO_ERROR;
1212 bp->b_flags |= B_CACHE;
1213 bp->b_iocmd = BIO_WRITE;
1215 vfs_busy_pages(bp, 1);
1218 * Normal bwrites pipeline writes
1220 bp->b_runningbufspace = bp->b_bufsize;
1221 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1223 if (!TD_IS_IDLETHREAD(curthread))
1224 curthread->td_ru.ru_oublock++;
1225 if (oldflags & B_ASYNC)
1227 bp->b_iooffset = dbtob(bp->b_blkno);
1230 if ((oldflags & B_ASYNC) == 0) {
1231 int rtval = bufwait(bp);
1234 } else if (space > hirunningspace) {
1236 * don't allow the async write to saturate the I/O
1237 * system. We will not deadlock here because
1238 * we are blocking waiting for I/O that is already in-progress
1239 * to complete. We do not block here if it is the update
1240 * or syncer daemon trying to clean up as that can lead
1243 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1244 waitrunningbufspace();
1251 bufbdflush(struct bufobj *bo, struct buf *bp)
1255 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1256 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1258 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1261 * Try to find a buffer to flush.
1263 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1264 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1266 LK_EXCLUSIVE | LK_NOWAIT, NULL))
1269 panic("bdwrite: found ourselves");
1271 /* Don't countdeps with the bo lock held. */
1272 if (buf_countdeps(nbp, 0)) {
1277 if (nbp->b_flags & B_CLUSTEROK) {
1278 vfs_bio_awrite(nbp);
1283 dirtybufferflushes++;
1292 * Delayed write. (Buffer is marked dirty). Do not bother writing
1293 * anything if the buffer is marked invalid.
1295 * Note that since the buffer must be completely valid, we can safely
1296 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1297 * biodone() in order to prevent getblk from writing the buffer
1298 * out synchronously.
1301 bdwrite(struct buf *bp)
1303 struct thread *td = curthread;
1307 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1308 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1309 KASSERT((bp->b_flags & B_BARRIER) == 0,
1310 ("Barrier request in delayed write %p", bp));
1311 BUF_ASSERT_HELD(bp);
1313 if (bp->b_flags & B_INVAL) {
1319 * If we have too many dirty buffers, don't create any more.
1320 * If we are wildly over our limit, then force a complete
1321 * cleanup. Otherwise, just keep the situation from getting
1322 * out of control. Note that we have to avoid a recursive
1323 * disaster and not try to clean up after our own cleanup!
1327 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1328 td->td_pflags |= TDP_INBDFLUSH;
1330 td->td_pflags &= ~TDP_INBDFLUSH;
1336 * Set B_CACHE, indicating that the buffer is fully valid. This is
1337 * true even of NFS now.
1339 bp->b_flags |= B_CACHE;
1342 * This bmap keeps the system from needing to do the bmap later,
1343 * perhaps when the system is attempting to do a sync. Since it
1344 * is likely that the indirect block -- or whatever other datastructure
1345 * that the filesystem needs is still in memory now, it is a good
1346 * thing to do this. Note also, that if the pageout daemon is
1347 * requesting a sync -- there might not be enough memory to do
1348 * the bmap then... So, this is important to do.
1350 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1351 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1355 * Set the *dirty* buffer range based upon the VM system dirty
1358 * Mark the buffer pages as clean. We need to do this here to
1359 * satisfy the vnode_pager and the pageout daemon, so that it
1360 * thinks that the pages have been "cleaned". Note that since
1361 * the pages are in a delayed write buffer -- the VFS layer
1362 * "will" see that the pages get written out on the next sync,
1363 * or perhaps the cluster will be completed.
1365 vfs_clean_pages_dirty_buf(bp);
1369 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1370 * due to the softdep code.
1377 * Turn buffer into delayed write request. We must clear BIO_READ and
1378 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1379 * itself to properly update it in the dirty/clean lists. We mark it
1380 * B_DONE to ensure that any asynchronization of the buffer properly
1381 * clears B_DONE ( else a panic will occur later ).
1383 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1384 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1385 * should only be called if the buffer is known-good.
1387 * Since the buffer is not on a queue, we do not update the numfreebuffers
1390 * The buffer must be on QUEUE_NONE.
1393 bdirty(struct buf *bp)
1396 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1397 bp, bp->b_vp, bp->b_flags);
1398 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1399 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1400 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1401 BUF_ASSERT_HELD(bp);
1402 bp->b_flags &= ~(B_RELBUF);
1403 bp->b_iocmd = BIO_WRITE;
1405 if ((bp->b_flags & B_DELWRI) == 0) {
1406 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1415 * Clear B_DELWRI for buffer.
1417 * Since the buffer is not on a queue, we do not update the numfreebuffers
1420 * The buffer must be on QUEUE_NONE.
1424 bundirty(struct buf *bp)
1427 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1428 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1429 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1430 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1431 BUF_ASSERT_HELD(bp);
1433 if (bp->b_flags & B_DELWRI) {
1434 bp->b_flags &= ~B_DELWRI;
1439 * Since it is now being written, we can clear its deferred write flag.
1441 bp->b_flags &= ~B_DEFERRED;
1447 * Asynchronous write. Start output on a buffer, but do not wait for
1448 * it to complete. The buffer is released when the output completes.
1450 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1451 * B_INVAL buffers. Not us.
1454 bawrite(struct buf *bp)
1457 bp->b_flags |= B_ASYNC;
1464 * Asynchronous barrier write. Start output on a buffer, but do not
1465 * wait for it to complete. Place a write barrier after this write so
1466 * that this buffer and all buffers written before it are committed to
1467 * the disk before any buffers written after this write are committed
1468 * to the disk. The buffer is released when the output completes.
1471 babarrierwrite(struct buf *bp)
1474 bp->b_flags |= B_ASYNC | B_BARRIER;
1481 * Synchronous barrier write. Start output on a buffer and wait for
1482 * it to complete. Place a write barrier after this write so that
1483 * this buffer and all buffers written before it are committed to
1484 * the disk before any buffers written after this write are committed
1485 * to the disk. The buffer is released when the output completes.
1488 bbarrierwrite(struct buf *bp)
1491 bp->b_flags |= B_BARRIER;
1492 return (bwrite(bp));
1498 * Called prior to the locking of any vnodes when we are expecting to
1499 * write. We do not want to starve the buffer cache with too many
1500 * dirty buffers so we block here. By blocking prior to the locking
1501 * of any vnodes we attempt to avoid the situation where a locked vnode
1502 * prevents the various system daemons from flushing related buffers.
1508 if (numdirtybuffers >= hidirtybuffers) {
1509 mtx_lock(&bdirtylock);
1510 while (numdirtybuffers >= hidirtybuffers) {
1512 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1515 mtx_unlock(&bdirtylock);
1520 * Return true if we have too many dirty buffers.
1523 buf_dirty_count_severe(void)
1526 return(numdirtybuffers >= hidirtybuffers);
1529 static __noinline int
1530 buf_vm_page_count_severe(void)
1533 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1535 return vm_page_count_severe();
1541 * Release a busy buffer and, if requested, free its resources. The
1542 * buffer will be stashed in the appropriate bufqueue[] allowing it
1543 * to be accessed later as a cache entity or reused for other purposes.
1546 brelse(struct buf *bp)
1550 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1551 bp, bp->b_vp, bp->b_flags);
1552 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1553 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1555 if (BUF_LOCKRECURSED(bp)) {
1557 * Do not process, in particular, do not handle the
1558 * B_INVAL/B_RELBUF and do not release to free list.
1564 if (bp->b_flags & B_MANAGED) {
1569 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
1570 BO_LOCK(bp->b_bufobj);
1571 bp->b_vflags &= ~BV_BKGRDERR;
1572 BO_UNLOCK(bp->b_bufobj);
1575 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1576 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1578 * Failed write, redirty. Must clear BIO_ERROR to prevent
1579 * pages from being scrapped. If the error is anything
1580 * other than an I/O error (EIO), assume that retrying
1583 bp->b_ioflags &= ~BIO_ERROR;
1585 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1586 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1588 * Either a failed I/O or we were asked to free or not
1591 bp->b_flags |= B_INVAL;
1592 if (!LIST_EMPTY(&bp->b_dep))
1594 if (bp->b_flags & B_DELWRI)
1596 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1597 if ((bp->b_flags & B_VMIO) == 0) {
1606 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1607 * is called with B_DELWRI set, the underlying pages may wind up
1608 * getting freed causing a previous write (bdwrite()) to get 'lost'
1609 * because pages associated with a B_DELWRI bp are marked clean.
1611 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1612 * if B_DELWRI is set.
1614 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1615 * on pages to return pages to the VM page queues.
1617 if (bp->b_flags & B_DELWRI)
1618 bp->b_flags &= ~B_RELBUF;
1619 else if (buf_vm_page_count_severe()) {
1621 * BKGRDINPROG can only be set with the buf and bufobj
1622 * locks both held. We tolerate a race to clear it here.
1624 if (!(bp->b_vflags & BV_BKGRDINPROG))
1625 bp->b_flags |= B_RELBUF;
1629 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1630 * constituted, not even NFS buffers now. Two flags effect this. If
1631 * B_INVAL, the struct buf is invalidated but the VM object is kept
1632 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1634 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1635 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1636 * buffer is also B_INVAL because it hits the re-dirtying code above.
1638 * Normally we can do this whether a buffer is B_DELWRI or not. If
1639 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1640 * the commit state and we cannot afford to lose the buffer. If the
1641 * buffer has a background write in progress, we need to keep it
1642 * around to prevent it from being reconstituted and starting a second
1645 if ((bp->b_flags & B_VMIO)
1646 && !(bp->b_vp->v_mount != NULL &&
1647 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1648 !vn_isdisk(bp->b_vp, NULL) &&
1649 (bp->b_flags & B_DELWRI))
1658 obj = bp->b_bufobj->bo_object;
1661 * Get the base offset and length of the buffer. Note that
1662 * in the VMIO case if the buffer block size is not
1663 * page-aligned then b_data pointer may not be page-aligned.
1664 * But our b_pages[] array *IS* page aligned.
1666 * block sizes less then DEV_BSIZE (usually 512) are not
1667 * supported due to the page granularity bits (m->valid,
1668 * m->dirty, etc...).
1670 * See man buf(9) for more information
1672 resid = bp->b_bufsize;
1673 foff = bp->b_offset;
1674 for (i = 0; i < bp->b_npages; i++) {
1680 * If we hit a bogus page, fixup *all* the bogus pages
1683 if (m == bogus_page) {
1684 poff = OFF_TO_IDX(bp->b_offset);
1687 VM_OBJECT_RLOCK(obj);
1688 for (j = i; j < bp->b_npages; j++) {
1690 mtmp = bp->b_pages[j];
1691 if (mtmp == bogus_page) {
1692 mtmp = vm_page_lookup(obj, poff + j);
1694 panic("brelse: page missing\n");
1696 bp->b_pages[j] = mtmp;
1699 VM_OBJECT_RUNLOCK(obj);
1701 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1702 BUF_CHECK_MAPPED(bp);
1704 trunc_page((vm_offset_t)bp->b_data),
1705 bp->b_pages, bp->b_npages);
1709 if ((bp->b_flags & B_NOCACHE) ||
1710 (bp->b_ioflags & BIO_ERROR &&
1711 bp->b_iocmd == BIO_READ)) {
1712 int poffset = foff & PAGE_MASK;
1713 int presid = resid > (PAGE_SIZE - poffset) ?
1714 (PAGE_SIZE - poffset) : resid;
1716 KASSERT(presid >= 0, ("brelse: extra page"));
1717 VM_OBJECT_WLOCK(obj);
1718 while (vm_page_xbusied(m)) {
1720 VM_OBJECT_WUNLOCK(obj);
1721 vm_page_busy_sleep(m, "mbncsh");
1722 VM_OBJECT_WLOCK(obj);
1724 if (pmap_page_wired_mappings(m) == 0)
1725 vm_page_set_invalid(m, poffset, presid);
1726 VM_OBJECT_WUNLOCK(obj);
1728 printf("avoided corruption bug in bogus_page/brelse code\n");
1730 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1731 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1733 if (bp->b_flags & (B_INVAL | B_RELBUF))
1734 vfs_vmio_release(bp);
1736 } else if (bp->b_flags & B_VMIO) {
1738 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1739 vfs_vmio_release(bp);
1742 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1743 if (bp->b_bufsize != 0)
1745 if (bp->b_vp != NULL)
1750 * If the buffer has junk contents signal it and eventually
1751 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1754 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1755 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1756 bp->b_flags |= B_INVAL;
1757 if (bp->b_flags & B_INVAL) {
1758 if (bp->b_flags & B_DELWRI)
1764 /* buffers with no memory */
1765 if (bp->b_bufsize == 0) {
1766 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1767 if (bp->b_vflags & BV_BKGRDINPROG)
1768 panic("losing buffer 1");
1770 qindex = QUEUE_EMPTYKVA;
1772 qindex = QUEUE_EMPTY;
1773 bp->b_flags |= B_AGE;
1774 /* buffers with junk contents */
1775 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1776 (bp->b_ioflags & BIO_ERROR)) {
1777 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1778 if (bp->b_vflags & BV_BKGRDINPROG)
1779 panic("losing buffer 2");
1780 qindex = QUEUE_CLEAN;
1781 bp->b_flags |= B_AGE;
1782 /* remaining buffers */
1783 } else if (bp->b_flags & B_DELWRI)
1784 qindex = QUEUE_DIRTY;
1786 qindex = QUEUE_CLEAN;
1788 binsfree(bp, qindex);
1790 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1791 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1792 panic("brelse: not dirty");
1798 * Release a buffer back to the appropriate queue but do not try to free
1799 * it. The buffer is expected to be used again soon.
1801 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1802 * biodone() to requeue an async I/O on completion. It is also used when
1803 * known good buffers need to be requeued but we think we may need the data
1806 * XXX we should be able to leave the B_RELBUF hint set on completion.
1809 bqrelse(struct buf *bp)
1813 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1814 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1815 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1817 if (BUF_LOCKRECURSED(bp)) {
1818 /* do not release to free list */
1822 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1824 if (bp->b_flags & B_MANAGED) {
1825 if (bp->b_flags & B_REMFREE)
1830 /* buffers with stale but valid contents */
1831 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
1832 BV_BKGRDERR)) == BV_BKGRDERR) {
1833 BO_LOCK(bp->b_bufobj);
1834 bp->b_vflags &= ~BV_BKGRDERR;
1835 BO_UNLOCK(bp->b_bufobj);
1836 qindex = QUEUE_DIRTY;
1838 if ((bp->b_flags & B_DELWRI) == 0 &&
1839 (bp->b_xflags & BX_VNDIRTY))
1840 panic("bqrelse: not dirty");
1842 * BKGRDINPROG can only be set with the buf and bufobj
1843 * locks both held. We tolerate a race to clear it here.
1845 if (buf_vm_page_count_severe() &&
1846 (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1848 * We are too low on memory, we have to try to free
1849 * the buffer (most importantly: the wired pages
1850 * making up its backing store) *now*.
1855 qindex = QUEUE_CLEAN;
1857 binsfree(bp, qindex);
1864 /* Give pages used by the bp back to the VM system (where possible) */
1866 vfs_vmio_release(struct buf *bp)
1872 if ((bp->b_flags & B_UNMAPPED) == 0) {
1873 BUF_CHECK_MAPPED(bp);
1874 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1876 BUF_CHECK_UNMAPPED(bp);
1877 obj = bp->b_bufobj->bo_object;
1879 VM_OBJECT_WLOCK(obj);
1880 for (i = 0; i < bp->b_npages; i++) {
1882 bp->b_pages[i] = NULL;
1884 * In order to keep page LRU ordering consistent, put
1885 * everything on the inactive queue.
1888 vm_page_unwire(m, 0);
1891 * Might as well free the page if we can and it has
1892 * no valid data. We also free the page if the
1893 * buffer was used for direct I/O
1895 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1896 if (m->wire_count == 0 && !vm_page_busied(m))
1898 } else if (bp->b_flags & B_DIRECT)
1899 vm_page_try_to_free(m);
1900 else if (buf_vm_page_count_severe())
1901 vm_page_try_to_cache(m);
1905 VM_OBJECT_WUNLOCK(obj);
1907 if (bp->b_bufsize) {
1912 bp->b_flags &= ~B_VMIO;
1918 * Check to see if a block at a particular lbn is available for a clustered
1922 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1929 /* If the buf isn't in core skip it */
1930 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1933 /* If the buf is busy we don't want to wait for it */
1934 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1937 /* Only cluster with valid clusterable delayed write buffers */
1938 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1939 (B_DELWRI | B_CLUSTEROK))
1942 if (bpa->b_bufsize != size)
1946 * Check to see if it is in the expected place on disk and that the
1947 * block has been mapped.
1949 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1959 * Implement clustered async writes for clearing out B_DELWRI buffers.
1960 * This is much better then the old way of writing only one buffer at
1961 * a time. Note that we may not be presented with the buffers in the
1962 * correct order, so we search for the cluster in both directions.
1965 vfs_bio_awrite(struct buf *bp)
1970 daddr_t lblkno = bp->b_lblkno;
1971 struct vnode *vp = bp->b_vp;
1979 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1981 * right now we support clustered writing only to regular files. If
1982 * we find a clusterable block we could be in the middle of a cluster
1983 * rather then at the beginning.
1985 if ((vp->v_type == VREG) &&
1986 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1987 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1989 size = vp->v_mount->mnt_stat.f_iosize;
1990 maxcl = MAXPHYS / size;
1993 for (i = 1; i < maxcl; i++)
1994 if (vfs_bio_clcheck(vp, size, lblkno + i,
1995 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1998 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1999 if (vfs_bio_clcheck(vp, size, lblkno - j,
2000 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2006 * this is a possible cluster write
2010 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2016 bp->b_flags |= B_ASYNC;
2018 * default (old) behavior, writing out only one block
2020 * XXX returns b_bufsize instead of b_bcount for nwritten?
2022 nwritten = bp->b_bufsize;
2029 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2032 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2033 bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2034 if ((gbflags & GB_UNMAPPED) == 0) {
2035 bp->b_kvabase = (caddr_t)addr;
2036 } else if ((gbflags & GB_KVAALLOC) != 0) {
2037 KASSERT((gbflags & GB_UNMAPPED) != 0,
2038 ("GB_KVAALLOC without GB_UNMAPPED"));
2039 bp->b_kvaalloc = (caddr_t)addr;
2040 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2041 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2043 bp->b_kvasize = maxsize;
2047 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2051 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2058 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2060 * Buffer map is too fragmented. Request the caller
2061 * to defragment the map.
2063 atomic_add_int(&bufdefragcnt, 1);
2066 setbufkva(bp, addr, maxsize, gbflags);
2067 atomic_add_long(&bufspace, bp->b_kvasize);
2072 * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2073 * locked vnode is supplied.
2076 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2081 int error, fl, flags, norunbuf;
2083 mtx_assert(&bqclean, MA_OWNED);
2086 flags = VFS_BIO_NEED_BUFSPACE;
2088 } else if (bufspace >= hibufspace) {
2090 flags = VFS_BIO_NEED_BUFSPACE;
2093 flags = VFS_BIO_NEED_ANY;
2095 atomic_set_int(&needsbuffer, flags);
2096 mtx_unlock(&bqclean);
2098 bd_speedup(); /* heeeelp */
2099 if ((gbflags & GB_NOWAIT_BD) != 0)
2104 while ((needsbuffer & flags) != 0) {
2105 if (vp != NULL && vp->v_type != VCHR &&
2106 (td->td_pflags & TDP_BUFNEED) == 0) {
2107 rw_wunlock(&nblock);
2109 * getblk() is called with a vnode locked, and
2110 * some majority of the dirty buffers may as
2111 * well belong to the vnode. Flushing the
2112 * buffers there would make a progress that
2113 * cannot be achieved by the buf_daemon, that
2114 * cannot lock the vnode.
2116 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2117 (td->td_pflags & TDP_NORUNNINGBUF);
2120 * Play bufdaemon. The getnewbuf() function
2121 * may be called while the thread owns lock
2122 * for another dirty buffer for the same
2123 * vnode, which makes it impossible to use
2124 * VOP_FSYNC() there, due to the buffer lock
2127 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2128 fl = buf_flush(vp, flushbufqtarget);
2129 td->td_pflags &= norunbuf;
2133 if ((needsbuffer & flags) == 0)
2136 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
2137 (PRIBIO + 4) | slpflag, waitmsg, slptimeo);
2141 rw_wunlock(&nblock);
2145 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2148 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2149 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2150 bp->b_kvasize, bp->b_bufsize, qindex);
2151 mtx_assert(&bqclean, MA_NOTOWNED);
2154 * Note: we no longer distinguish between VMIO and non-VMIO
2157 KASSERT((bp->b_flags & B_DELWRI) == 0,
2158 ("delwri buffer %p found in queue %d", bp, qindex));
2160 if (qindex == QUEUE_CLEAN) {
2161 if (bp->b_flags & B_VMIO) {
2162 bp->b_flags &= ~B_ASYNC;
2163 vfs_vmio_release(bp);
2165 if (bp->b_vp != NULL)
2170 * Get the rest of the buffer freed up. b_kva* is still valid
2171 * after this operation.
2174 if (bp->b_rcred != NOCRED) {
2175 crfree(bp->b_rcred);
2176 bp->b_rcred = NOCRED;
2178 if (bp->b_wcred != NOCRED) {
2179 crfree(bp->b_wcred);
2180 bp->b_wcred = NOCRED;
2182 if (!LIST_EMPTY(&bp->b_dep))
2184 if (bp->b_vflags & BV_BKGRDINPROG)
2185 panic("losing buffer 3");
2186 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d",
2187 bp, bp->b_vp, qindex));
2188 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2189 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2194 bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2197 KASSERT((bp->b_flags & B_INFREECNT) == 0,
2198 ("buf %p still counted as free?", bp));
2201 bp->b_blkno = bp->b_lblkno = 0;
2202 bp->b_offset = NOOFFSET;
2208 bp->b_dirtyoff = bp->b_dirtyend = 0;
2209 bp->b_bufobj = NULL;
2210 bp->b_pin_count = 0;
2211 bp->b_fsprivate1 = NULL;
2212 bp->b_fsprivate2 = NULL;
2213 bp->b_fsprivate3 = NULL;
2215 LIST_INIT(&bp->b_dep);
2218 static int flushingbufs;
2221 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2223 struct buf *bp, *nbp;
2224 int nqindex, qindex, pass;
2226 KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2230 atomic_add_int(&getnewbufrestarts, 1);
2233 * Setup for scan. If we do not have enough free buffers,
2234 * we setup a degenerate case that immediately fails. Note
2235 * that if we are specially marked process, we are allowed to
2236 * dip into our reserves.
2238 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2239 * for the allocation of the mapped buffer. For unmapped, the
2240 * easiest is to start with EMPTY outright.
2242 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
2243 * However, there are a number of cases (defragging, reusing, ...)
2244 * where we cannot backup.
2248 if (!defrag && unmapped) {
2249 nqindex = QUEUE_EMPTY;
2250 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2253 nqindex = QUEUE_EMPTYKVA;
2254 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2258 * If no EMPTYKVA buffers and we are either defragging or
2259 * reusing, locate a CLEAN buffer to free or reuse. If
2260 * bufspace useage is low skip this step so we can allocate a
2263 if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2264 nqindex = QUEUE_CLEAN;
2265 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2269 * If we could not find or were not allowed to reuse a CLEAN
2270 * buffer, check to see if it is ok to use an EMPTY buffer.
2271 * We can only use an EMPTY buffer if allocating its KVA would
2272 * not otherwise run us out of buffer space. No KVA is needed
2273 * for the unmapped allocation.
2275 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2277 nqindex = QUEUE_EMPTY;
2278 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2282 * All available buffers might be clean, retry ignoring the
2283 * lobufspace as the last resort.
2285 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2286 nqindex = QUEUE_CLEAN;
2287 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2291 * Run scan, possibly freeing data and/or kva mappings on the fly
2294 while ((bp = nbp) != NULL) {
2298 * Calculate next bp (we can only use it if we do not
2299 * block or do other fancy things).
2301 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2304 nqindex = QUEUE_EMPTYKVA;
2305 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2309 case QUEUE_EMPTYKVA:
2310 nqindex = QUEUE_CLEAN;
2311 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2316 if (metadata && pass == 1) {
2318 nqindex = QUEUE_EMPTY;
2320 &bufqueues[QUEUE_EMPTY]);
2329 * If we are defragging then we need a buffer with
2330 * b_kvasize != 0. XXX this situation should no longer
2331 * occur, if defrag is non-zero the buffer's b_kvasize
2332 * should also be non-zero at this point. XXX
2334 if (defrag && bp->b_kvasize == 0) {
2335 printf("Warning: defrag empty buffer %p\n", bp);
2340 * Start freeing the bp. This is somewhat involved. nbp
2341 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2343 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2346 * BKGRDINPROG can only be set with the buf and bufobj
2347 * locks both held. We tolerate a race to clear it here.
2349 if (bp->b_vflags & BV_BKGRDINPROG) {
2355 * Requeue the background write buffer with error.
2357 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
2359 mtx_unlock(&bqclean);
2364 KASSERT(bp->b_qindex == qindex,
2365 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2368 mtx_unlock(&bqclean);
2370 * NOTE: nbp is now entirely invalid. We can only restart
2371 * the scan from this point on.
2374 getnewbuf_reuse_bp(bp, qindex);
2375 mtx_assert(&bqclean, MA_NOTOWNED);
2378 * If we are defragging then free the buffer.
2381 bp->b_flags |= B_INVAL;
2389 * Notify any waiters for the buffer lock about
2390 * identity change by freeing the buffer.
2392 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2393 bp->b_flags |= B_INVAL;
2403 * If we are overcomitted then recover the buffer and its
2404 * KVM space. This occurs in rare situations when multiple
2405 * processes are blocked in getnewbuf() or allocbuf().
2407 if (bufspace >= hibufspace)
2409 if (flushingbufs && bp->b_kvasize != 0) {
2410 bp->b_flags |= B_INVAL;
2415 if (bufspace < lobufspace)
2425 * Find and initialize a new buffer header, freeing up existing buffers
2426 * in the bufqueues as necessary. The new buffer is returned locked.
2428 * Important: B_INVAL is not set. If the caller wishes to throw the
2429 * buffer away, the caller must set B_INVAL prior to calling brelse().
2432 * We have insufficient buffer headers
2433 * We have insufficient buffer space
2434 * buffer_arena is too fragmented ( space reservation fails )
2435 * If we have to flush dirty buffers ( but we try to avoid this )
2438 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2442 int defrag, metadata;
2444 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2445 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2446 if (!unmapped_buf_allowed)
2447 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2450 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2456 * We can't afford to block since we might be holding a vnode lock,
2457 * which may prevent system daemons from running. We deal with
2458 * low-memory situations by proactively returning memory and running
2459 * async I/O rather then sync I/O.
2461 atomic_add_int(&getnewbufcalls, 1);
2462 atomic_subtract_int(&getnewbufrestarts, 1);
2464 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2465 GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2470 * If we exhausted our list, sleep as appropriate. We may have to
2471 * wakeup various daemons and write out some dirty buffers.
2473 * Generally we are sleeping due to insufficient buffer space.
2476 mtx_assert(&bqclean, MA_OWNED);
2477 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2478 mtx_assert(&bqclean, MA_NOTOWNED);
2479 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2480 mtx_assert(&bqclean, MA_NOTOWNED);
2483 bp->b_flags |= B_UNMAPPED;
2484 bp->b_kvabase = bp->b_data = unmapped_buf;
2485 bp->b_kvasize = maxsize;
2486 atomic_add_long(&bufspace, bp->b_kvasize);
2487 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2488 atomic_add_int(&bufreusecnt, 1);
2490 mtx_assert(&bqclean, MA_NOTOWNED);
2493 * We finally have a valid bp. We aren't quite out of the
2494 * woods, we still have to reserve kva space. In order
2495 * to keep fragmentation sane we only allocate kva in
2498 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2500 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2501 B_KVAALLOC)) == B_UNMAPPED) {
2502 if (allocbufkva(bp, maxsize, gbflags)) {
2504 bp->b_flags |= B_INVAL;
2508 atomic_add_int(&bufreusecnt, 1);
2509 } else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2510 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2512 * If the reused buffer has KVA allocated,
2513 * reassign b_kvaalloc to b_kvabase.
2515 bp->b_kvabase = bp->b_kvaalloc;
2516 bp->b_flags &= ~B_KVAALLOC;
2517 atomic_subtract_long(&unmapped_bufspace,
2519 atomic_add_int(&bufreusecnt, 1);
2520 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2521 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2524 * The case of reused buffer already have KVA
2525 * mapped, but the request is for unmapped
2526 * buffer with KVA allocated.
2528 bp->b_kvaalloc = bp->b_kvabase;
2529 bp->b_data = bp->b_kvabase = unmapped_buf;
2530 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2531 atomic_add_long(&unmapped_bufspace,
2533 atomic_add_int(&bufreusecnt, 1);
2535 if ((gbflags & GB_UNMAPPED) == 0) {
2536 bp->b_saveaddr = bp->b_kvabase;
2537 bp->b_data = bp->b_saveaddr;
2538 bp->b_flags &= ~B_UNMAPPED;
2539 BUF_CHECK_MAPPED(bp);
2548 * buffer flushing daemon. Buffers are normally flushed by the
2549 * update daemon but if it cannot keep up this process starts to
2550 * take the load in an attempt to prevent getnewbuf() from blocking.
2553 static struct kproc_desc buf_kp = {
2558 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2561 buf_flush(struct vnode *vp, int target)
2565 flushed = flushbufqueues(vp, target, 0);
2568 * Could not find any buffers without rollback
2569 * dependencies, so just write the first one
2570 * in the hopes of eventually making progress.
2572 if (vp != NULL && target > 2)
2574 flushbufqueues(vp, target, 1);
2585 * This process needs to be suspended prior to shutdown sync.
2587 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2591 * This process is allowed to take the buffer cache to the limit
2593 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2597 mtx_unlock(&bdlock);
2599 kproc_suspend_check(bufdaemonproc);
2600 lodirty = lodirtybuffers;
2601 if (bd_speedupreq) {
2602 lodirty = numdirtybuffers / 2;
2606 * Do the flush. Limit the amount of in-transit I/O we
2607 * allow to build up, otherwise we would completely saturate
2610 while (numdirtybuffers > lodirty) {
2611 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
2613 kern_yield(PRI_USER);
2617 * Only clear bd_request if we have reached our low water
2618 * mark. The buf_daemon normally waits 1 second and
2619 * then incrementally flushes any dirty buffers that have
2620 * built up, within reason.
2622 * If we were unable to hit our low water mark and couldn't
2623 * find any flushable buffers, we sleep for a short period
2624 * to avoid endless loops on unlockable buffers.
2627 if (numdirtybuffers <= lodirtybuffers) {
2629 * We reached our low water mark, reset the
2630 * request and sleep until we are needed again.
2631 * The sleep is just so the suspend code works.
2635 * Do an extra wakeup in case dirty threshold
2636 * changed via sysctl and the explicit transition
2637 * out of shortfall was missed.
2640 if (runningbufspace <= lorunningspace)
2642 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2645 * We couldn't find any flushable dirty buffers but
2646 * still have too many dirty buffers, we
2647 * have to sleep and try again. (rare)
2649 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2657 * Try to flush a buffer in the dirty queue. We must be careful to
2658 * free up B_INVAL buffers instead of write them, which NFS is
2659 * particularly sensitive to.
2661 static int flushwithdeps = 0;
2662 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2663 0, "Number of buffers flushed with dependecies that require rollbacks");
2666 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
2668 struct buf *sentinel;
2679 queue = QUEUE_DIRTY;
2681 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2682 sentinel->b_qindex = QUEUE_SENTINEL;
2684 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2685 mtx_unlock(&bqdirty);
2686 while (flushed != target) {
2689 bp = TAILQ_NEXT(sentinel, b_freelist);
2691 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2692 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2695 mtx_unlock(&bqdirty);
2699 * Skip sentinels inserted by other invocations of the
2700 * flushbufqueues(), taking care to not reorder them.
2702 * Only flush the buffers that belong to the
2703 * vnode locked by the curthread.
2705 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
2707 mtx_unlock(&bqdirty);
2710 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2711 mtx_unlock(&bqdirty);
2714 if (bp->b_pin_count > 0) {
2719 * BKGRDINPROG can only be set with the buf and bufobj
2720 * locks both held. We tolerate a race to clear it here.
2722 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2723 (bp->b_flags & B_DELWRI) == 0) {
2727 if (bp->b_flags & B_INVAL) {
2734 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2735 if (flushdeps == 0) {
2743 * We must hold the lock on a vnode before writing
2744 * one of its buffers. Otherwise we may confuse, or
2745 * in the case of a snapshot vnode, deadlock the
2748 * The lock order here is the reverse of the normal
2749 * of vnode followed by buf lock. This is ok because
2750 * the NOWAIT will prevent deadlock.
2753 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2759 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2761 ASSERT_VOP_LOCKED(vp, "getbuf");
2763 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2764 vn_lock(vp, LK_TRYUPGRADE);
2767 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2768 bp, bp->b_vp, bp->b_flags);
2769 if (curproc == bufdaemonproc) {
2776 vn_finished_write(mp);
2779 flushwithdeps += hasdeps;
2783 * Sleeping on runningbufspace while holding
2784 * vnode lock leads to deadlock.
2786 if (curproc == bufdaemonproc &&
2787 runningbufspace > hirunningspace)
2788 waitrunningbufspace();
2791 vn_finished_write(mp);
2795 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2796 mtx_unlock(&bqdirty);
2797 free(sentinel, M_TEMP);
2802 * Check to see if a block is currently memory resident.
2805 incore(struct bufobj *bo, daddr_t blkno)
2810 bp = gbincore(bo, blkno);
2816 * Returns true if no I/O is needed to access the
2817 * associated VM object. This is like incore except
2818 * it also hunts around in the VM system for the data.
2822 inmem(struct vnode * vp, daddr_t blkno)
2825 vm_offset_t toff, tinc, size;
2829 ASSERT_VOP_LOCKED(vp, "inmem");
2831 if (incore(&vp->v_bufobj, blkno))
2833 if (vp->v_mount == NULL)
2840 if (size > vp->v_mount->mnt_stat.f_iosize)
2841 size = vp->v_mount->mnt_stat.f_iosize;
2842 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2844 VM_OBJECT_RLOCK(obj);
2845 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2846 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2850 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2851 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2852 if (vm_page_is_valid(m,
2853 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2856 VM_OBJECT_RUNLOCK(obj);
2860 VM_OBJECT_RUNLOCK(obj);
2865 * Set the dirty range for a buffer based on the status of the dirty
2866 * bits in the pages comprising the buffer. The range is limited
2867 * to the size of the buffer.
2869 * Tell the VM system that the pages associated with this buffer
2870 * are clean. This is used for delayed writes where the data is
2871 * going to go to disk eventually without additional VM intevention.
2873 * Note that while we only really need to clean through to b_bcount, we
2874 * just go ahead and clean through to b_bufsize.
2877 vfs_clean_pages_dirty_buf(struct buf *bp)
2879 vm_ooffset_t foff, noff, eoff;
2883 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2886 foff = bp->b_offset;
2887 KASSERT(bp->b_offset != NOOFFSET,
2888 ("vfs_clean_pages_dirty_buf: no buffer offset"));
2890 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2891 vfs_drain_busy_pages(bp);
2892 vfs_setdirty_locked_object(bp);
2893 for (i = 0; i < bp->b_npages; i++) {
2894 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2896 if (eoff > bp->b_offset + bp->b_bufsize)
2897 eoff = bp->b_offset + bp->b_bufsize;
2899 vfs_page_set_validclean(bp, foff, m);
2900 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2903 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2907 vfs_setdirty_locked_object(struct buf *bp)
2912 object = bp->b_bufobj->bo_object;
2913 VM_OBJECT_ASSERT_WLOCKED(object);
2916 * We qualify the scan for modified pages on whether the
2917 * object has been flushed yet.
2919 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2920 vm_offset_t boffset;
2921 vm_offset_t eoffset;
2924 * test the pages to see if they have been modified directly
2925 * by users through the VM system.
2927 for (i = 0; i < bp->b_npages; i++)
2928 vm_page_test_dirty(bp->b_pages[i]);
2931 * Calculate the encompassing dirty range, boffset and eoffset,
2932 * (eoffset - boffset) bytes.
2935 for (i = 0; i < bp->b_npages; i++) {
2936 if (bp->b_pages[i]->dirty)
2939 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2941 for (i = bp->b_npages - 1; i >= 0; --i) {
2942 if (bp->b_pages[i]->dirty) {
2946 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2949 * Fit it to the buffer.
2952 if (eoffset > bp->b_bcount)
2953 eoffset = bp->b_bcount;
2956 * If we have a good dirty range, merge with the existing
2960 if (boffset < eoffset) {
2961 if (bp->b_dirtyoff > boffset)
2962 bp->b_dirtyoff = boffset;
2963 if (bp->b_dirtyend < eoffset)
2964 bp->b_dirtyend = eoffset;
2970 * Allocate the KVA mapping for an existing buffer. It handles the
2971 * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2972 * KVA which is not mapped (B_KVAALLOC).
2975 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2977 struct buf *scratch_bp;
2978 int bsize, maxsize, need_mapping, need_kva;
2981 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2982 (gbflags & GB_UNMAPPED) == 0;
2983 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2984 (gbflags & GB_KVAALLOC) != 0;
2985 if (!need_mapping && !need_kva)
2988 BUF_CHECK_UNMAPPED(bp);
2990 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2992 * Buffer is not mapped, but the KVA was already
2993 * reserved at the time of the instantiation. Use the
2996 bp->b_flags &= ~B_KVAALLOC;
2997 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2998 bp->b_kvabase = bp->b_kvaalloc;
2999 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
3004 * Calculate the amount of the address space we would reserve
3005 * if the buffer was mapped.
3007 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3008 offset = blkno * bsize;
3009 maxsize = size + (offset & PAGE_MASK);
3010 maxsize = imax(maxsize, bsize);
3013 if (allocbufkva(bp, maxsize, gbflags)) {
3015 * Request defragmentation. getnewbuf() returns us the
3016 * allocated space by the scratch buffer KVA.
3018 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
3019 (GB_UNMAPPED | GB_KVAALLOC));
3020 if (scratch_bp == NULL) {
3021 if ((gbflags & GB_NOWAIT_BD) != 0) {
3023 * XXXKIB: defragmentation cannot
3024 * succeed, not sure what else to do.
3026 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
3028 atomic_add_int(&mappingrestarts, 1);
3031 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
3032 ("scratch bp !B_KVAALLOC %p", scratch_bp));
3033 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
3034 scratch_bp->b_kvasize, gbflags);
3036 /* Get rid of the scratch buffer. */
3037 scratch_bp->b_kvasize = 0;
3038 scratch_bp->b_flags |= B_INVAL;
3039 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
3046 bp->b_saveaddr = bp->b_kvabase;
3047 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
3048 bp->b_flags &= ~B_UNMAPPED;
3049 BUF_CHECK_MAPPED(bp);
3056 * Get a block given a specified block and offset into a file/device.
3057 * The buffers B_DONE bit will be cleared on return, making it almost
3058 * ready for an I/O initiation. B_INVAL may or may not be set on
3059 * return. The caller should clear B_INVAL prior to initiating a
3062 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3063 * an existing buffer.
3065 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3066 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3067 * and then cleared based on the backing VM. If the previous buffer is
3068 * non-0-sized but invalid, B_CACHE will be cleared.
3070 * If getblk() must create a new buffer, the new buffer is returned with
3071 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3072 * case it is returned with B_INVAL clear and B_CACHE set based on the
3075 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
3076 * B_CACHE bit is clear.
3078 * What this means, basically, is that the caller should use B_CACHE to
3079 * determine whether the buffer is fully valid or not and should clear
3080 * B_INVAL prior to issuing a read. If the caller intends to validate
3081 * the buffer by loading its data area with something, the caller needs
3082 * to clear B_INVAL. If the caller does this without issuing an I/O,
3083 * the caller should set B_CACHE ( as an optimization ), else the caller
3084 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3085 * a write attempt or if it was a successfull read. If the caller
3086 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3087 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3090 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3095 int bsize, error, maxsize, vmio;
3098 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3099 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3100 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3101 ASSERT_VOP_LOCKED(vp, "getblk");
3102 if (size > MAXBCACHEBUF)
3103 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size,
3105 if (!unmapped_buf_allowed)
3106 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3111 bp = gbincore(bo, blkno);
3115 * Buffer is in-core. If the buffer is not busy nor managed,
3116 * it must be on a queue.
3118 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3120 if (flags & GB_LOCK_NOWAIT)
3121 lockflags |= LK_NOWAIT;
3123 error = BUF_TIMELOCK(bp, lockflags,
3124 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3127 * If we slept and got the lock we have to restart in case
3128 * the buffer changed identities.
3130 if (error == ENOLCK)
3132 /* We timed out or were interrupted. */
3135 /* If recursed, assume caller knows the rules. */
3136 else if (BUF_LOCKRECURSED(bp))
3140 * The buffer is locked. B_CACHE is cleared if the buffer is
3141 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3142 * and for a VMIO buffer B_CACHE is adjusted according to the
3145 if (bp->b_flags & B_INVAL)
3146 bp->b_flags &= ~B_CACHE;
3147 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3148 bp->b_flags |= B_CACHE;
3149 if (bp->b_flags & B_MANAGED)
3150 MPASS(bp->b_qindex == QUEUE_NONE);
3155 * check for size inconsistencies for non-VMIO case.
3157 if (bp->b_bcount != size) {
3158 if ((bp->b_flags & B_VMIO) == 0 ||
3159 (size > bp->b_kvasize)) {
3160 if (bp->b_flags & B_DELWRI) {
3162 * If buffer is pinned and caller does
3163 * not want sleep waiting for it to be
3164 * unpinned, bail out
3166 if (bp->b_pin_count > 0) {
3167 if (flags & GB_LOCK_NOWAIT) {
3174 bp->b_flags |= B_NOCACHE;
3177 if (LIST_EMPTY(&bp->b_dep)) {
3178 bp->b_flags |= B_RELBUF;
3181 bp->b_flags |= B_NOCACHE;
3190 * Handle the case of unmapped buffer which should
3191 * become mapped, or the buffer for which KVA
3192 * reservation is requested.
3194 bp_unmapped_get_kva(bp, blkno, size, flags);
3197 * If the size is inconsistant in the VMIO case, we can resize
3198 * the buffer. This might lead to B_CACHE getting set or
3199 * cleared. If the size has not changed, B_CACHE remains
3200 * unchanged from its previous state.
3202 if (bp->b_bcount != size)
3205 KASSERT(bp->b_offset != NOOFFSET,
3206 ("getblk: no buffer offset"));
3209 * A buffer with B_DELWRI set and B_CACHE clear must
3210 * be committed before we can return the buffer in
3211 * order to prevent the caller from issuing a read
3212 * ( due to B_CACHE not being set ) and overwriting
3215 * Most callers, including NFS and FFS, need this to
3216 * operate properly either because they assume they
3217 * can issue a read if B_CACHE is not set, or because
3218 * ( for example ) an uncached B_DELWRI might loop due
3219 * to softupdates re-dirtying the buffer. In the latter
3220 * case, B_CACHE is set after the first write completes,
3221 * preventing further loops.
3222 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3223 * above while extending the buffer, we cannot allow the
3224 * buffer to remain with B_CACHE set after the write
3225 * completes or it will represent a corrupt state. To
3226 * deal with this we set B_NOCACHE to scrap the buffer
3229 * We might be able to do something fancy, like setting
3230 * B_CACHE in bwrite() except if B_DELWRI is already set,
3231 * so the below call doesn't set B_CACHE, but that gets real
3232 * confusing. This is much easier.
3235 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3236 bp->b_flags |= B_NOCACHE;
3240 bp->b_flags &= ~B_DONE;
3243 * Buffer is not in-core, create new buffer. The buffer
3244 * returned by getnewbuf() is locked. Note that the returned
3245 * buffer is also considered valid (not marked B_INVAL).
3249 * If the user does not want us to create the buffer, bail out
3252 if (flags & GB_NOCREAT)
3254 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3257 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3258 offset = blkno * bsize;
3259 vmio = vp->v_object != NULL;
3261 maxsize = size + (offset & PAGE_MASK);
3264 /* Do not allow non-VMIO notmapped buffers. */
3265 flags &= ~GB_UNMAPPED;
3267 maxsize = imax(maxsize, bsize);
3269 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3271 if (slpflag || slptimeo)
3277 * This code is used to make sure that a buffer is not
3278 * created while the getnewbuf routine is blocked.
3279 * This can be a problem whether the vnode is locked or not.
3280 * If the buffer is created out from under us, we have to
3281 * throw away the one we just created.
3283 * Note: this must occur before we associate the buffer
3284 * with the vp especially considering limitations in
3285 * the splay tree implementation when dealing with duplicate
3289 if (gbincore(bo, blkno)) {
3291 bp->b_flags |= B_INVAL;
3297 * Insert the buffer into the hash, so that it can
3298 * be found by incore.
3300 bp->b_blkno = bp->b_lblkno = blkno;
3301 bp->b_offset = offset;
3306 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
3307 * buffer size starts out as 0, B_CACHE will be set by
3308 * allocbuf() for the VMIO case prior to it testing the
3309 * backing store for validity.
3313 bp->b_flags |= B_VMIO;
3314 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3315 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3316 bp, vp->v_object, bp->b_bufobj->bo_object));
3318 bp->b_flags &= ~B_VMIO;
3319 KASSERT(bp->b_bufobj->bo_object == NULL,
3320 ("ARGH! has b_bufobj->bo_object %p %p\n",
3321 bp, bp->b_bufobj->bo_object));
3322 BUF_CHECK_MAPPED(bp);
3326 bp->b_flags &= ~B_DONE;
3328 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3329 BUF_ASSERT_HELD(bp);
3331 KASSERT(bp->b_bufobj == bo,
3332 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3337 * Get an empty, disassociated buffer of given size. The buffer is initially
3341 geteblk(int size, int flags)
3346 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3347 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3348 if ((flags & GB_NOWAIT_BD) &&
3349 (curthread->td_pflags & TDP_BUFNEED) != 0)
3353 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3354 BUF_ASSERT_HELD(bp);
3360 * This code constitutes the buffer memory from either anonymous system
3361 * memory (in the case of non-VMIO operations) or from an associated
3362 * VM object (in the case of VMIO operations). This code is able to
3363 * resize a buffer up or down.
3365 * Note that this code is tricky, and has many complications to resolve
3366 * deadlock or inconsistant data situations. Tread lightly!!!
3367 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3368 * the caller. Calling this code willy nilly can result in the loss of data.
3370 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3371 * B_CACHE for the non-VMIO case.
3375 allocbuf(struct buf *bp, int size)
3377 int newbsize, mbsize;
3380 BUF_ASSERT_HELD(bp);
3382 if (bp->b_kvasize < size)
3383 panic("allocbuf: buffer too small");
3385 if ((bp->b_flags & B_VMIO) == 0) {
3389 * Just get anonymous memory from the kernel. Don't
3390 * mess with B_CACHE.
3392 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3393 if (bp->b_flags & B_MALLOC)
3396 newbsize = round_page(size);
3398 if (newbsize < bp->b_bufsize) {
3400 * malloced buffers are not shrunk
3402 if (bp->b_flags & B_MALLOC) {
3404 bp->b_bcount = size;
3406 free(bp->b_data, M_BIOBUF);
3407 if (bp->b_bufsize) {
3408 atomic_subtract_long(
3414 bp->b_saveaddr = bp->b_kvabase;
3415 bp->b_data = bp->b_saveaddr;
3417 bp->b_flags &= ~B_MALLOC;
3421 vm_hold_free_pages(bp, newbsize);
3422 } else if (newbsize > bp->b_bufsize) {
3424 * We only use malloced memory on the first allocation.
3425 * and revert to page-allocated memory when the buffer
3429 * There is a potential smp race here that could lead
3430 * to bufmallocspace slightly passing the max. It
3431 * is probably extremely rare and not worth worrying
3434 if ( (bufmallocspace < maxbufmallocspace) &&
3435 (bp->b_bufsize == 0) &&
3436 (mbsize <= PAGE_SIZE/2)) {
3438 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3439 bp->b_bufsize = mbsize;
3440 bp->b_bcount = size;
3441 bp->b_flags |= B_MALLOC;
3442 atomic_add_long(&bufmallocspace, mbsize);
3448 * If the buffer is growing on its other-than-first allocation,
3449 * then we revert to the page-allocation scheme.
3451 if (bp->b_flags & B_MALLOC) {
3452 origbuf = bp->b_data;
3453 origbufsize = bp->b_bufsize;
3454 bp->b_data = bp->b_kvabase;
3455 if (bp->b_bufsize) {
3456 atomic_subtract_long(&bufmallocspace,
3461 bp->b_flags &= ~B_MALLOC;
3462 newbsize = round_page(newbsize);
3466 (vm_offset_t) bp->b_data + bp->b_bufsize,
3467 (vm_offset_t) bp->b_data + newbsize);
3469 bcopy(origbuf, bp->b_data, origbufsize);
3470 free(origbuf, M_BIOBUF);
3476 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3477 desiredpages = (size == 0) ? 0 :
3478 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3480 if (bp->b_flags & B_MALLOC)
3481 panic("allocbuf: VMIO buffer can't be malloced");
3483 * Set B_CACHE initially if buffer is 0 length or will become
3486 if (size == 0 || bp->b_bufsize == 0)
3487 bp->b_flags |= B_CACHE;
3489 if (newbsize < bp->b_bufsize) {
3491 * DEV_BSIZE aligned new buffer size is less then the
3492 * DEV_BSIZE aligned existing buffer size. Figure out
3493 * if we have to remove any pages.
3495 if (desiredpages < bp->b_npages) {
3498 if ((bp->b_flags & B_UNMAPPED) == 0) {
3499 BUF_CHECK_MAPPED(bp);
3500 pmap_qremove((vm_offset_t)trunc_page(
3501 (vm_offset_t)bp->b_data) +
3502 (desiredpages << PAGE_SHIFT),
3503 (bp->b_npages - desiredpages));
3505 BUF_CHECK_UNMAPPED(bp);
3506 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3507 for (i = desiredpages; i < bp->b_npages; i++) {
3509 * the page is not freed here -- it
3510 * is the responsibility of
3511 * vnode_pager_setsize
3514 KASSERT(m != bogus_page,
3515 ("allocbuf: bogus page found"));
3516 while (vm_page_sleep_if_busy(m,
3520 bp->b_pages[i] = NULL;
3522 vm_page_unwire(m, 0);
3525 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3526 bp->b_npages = desiredpages;
3528 } else if (size > bp->b_bcount) {
3530 * We are growing the buffer, possibly in a
3531 * byte-granular fashion.
3538 * Step 1, bring in the VM pages from the object,
3539 * allocating them if necessary. We must clear
3540 * B_CACHE if these pages are not valid for the
3541 * range covered by the buffer.
3544 obj = bp->b_bufobj->bo_object;
3546 VM_OBJECT_WLOCK(obj);
3547 while (bp->b_npages < desiredpages) {
3551 * We must allocate system pages since blocking
3552 * here could interfere with paging I/O, no
3553 * matter which process we are.
3555 * Only exclusive busy can be tested here.
3556 * Blocking on shared busy might lead to
3557 * deadlocks once allocbuf() is called after
3558 * pages are vfs_busy_pages().
3560 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3561 bp->b_npages, VM_ALLOC_NOBUSY |
3562 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3563 VM_ALLOC_IGN_SBUSY |
3564 VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3566 bp->b_flags &= ~B_CACHE;
3567 bp->b_pages[bp->b_npages] = m;
3572 * Step 2. We've loaded the pages into the buffer,
3573 * we have to figure out if we can still have B_CACHE
3574 * set. Note that B_CACHE is set according to the
3575 * byte-granular range ( bcount and size ), new the
3576 * aligned range ( newbsize ).
3578 * The VM test is against m->valid, which is DEV_BSIZE
3579 * aligned. Needless to say, the validity of the data
3580 * needs to also be DEV_BSIZE aligned. Note that this
3581 * fails with NFS if the server or some other client
3582 * extends the file's EOF. If our buffer is resized,
3583 * B_CACHE may remain set! XXX
3586 toff = bp->b_bcount;
3587 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3589 while ((bp->b_flags & B_CACHE) && toff < size) {
3592 if (tinc > (size - toff))
3595 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3608 VM_OBJECT_WUNLOCK(obj);
3611 * Step 3, fixup the KVM pmap.
3613 if ((bp->b_flags & B_UNMAPPED) == 0)
3616 BUF_CHECK_UNMAPPED(bp);
3619 if (newbsize < bp->b_bufsize)
3621 bp->b_bufsize = newbsize; /* actual buffer allocation */
3622 bp->b_bcount = size; /* requested buffer size */
3626 extern int inflight_transient_maps;
3629 biodone(struct bio *bp)
3632 void (*done)(struct bio *);
3633 vm_offset_t start, end;
3635 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3636 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3637 bp->bio_flags |= BIO_UNMAPPED;
3638 start = trunc_page((vm_offset_t)bp->bio_data);
3639 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3640 pmap_qremove(start, OFF_TO_IDX(end - start));
3641 vmem_free(transient_arena, start, end - start);
3642 atomic_add_int(&inflight_transient_maps, -1);
3644 done = bp->bio_done;
3646 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3648 bp->bio_flags |= BIO_DONE;
3652 bp->bio_flags |= BIO_DONE;
3658 * Wait for a BIO to finish.
3661 biowait(struct bio *bp, const char *wchan)
3665 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3667 while ((bp->bio_flags & BIO_DONE) == 0)
3668 msleep(bp, mtxp, PRIBIO, wchan, 0);
3670 if (bp->bio_error != 0)
3671 return (bp->bio_error);
3672 if (!(bp->bio_flags & BIO_ERROR))
3678 biofinish(struct bio *bp, struct devstat *stat, int error)
3682 bp->bio_error = error;
3683 bp->bio_flags |= BIO_ERROR;
3686 devstat_end_transaction_bio(stat, bp);
3693 * Wait for buffer I/O completion, returning error status. The buffer
3694 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3695 * error and cleared.
3698 bufwait(struct buf *bp)
3700 if (bp->b_iocmd == BIO_READ)
3701 bwait(bp, PRIBIO, "biord");
3703 bwait(bp, PRIBIO, "biowr");
3704 if (bp->b_flags & B_EINTR) {
3705 bp->b_flags &= ~B_EINTR;
3708 if (bp->b_ioflags & BIO_ERROR) {
3709 return (bp->b_error ? bp->b_error : EIO);
3716 * Call back function from struct bio back up to struct buf.
3719 bufdonebio(struct bio *bip)
3723 bp = bip->bio_caller2;
3724 bp->b_resid = bp->b_bcount - bip->bio_completed;
3725 bp->b_resid = bip->bio_resid; /* XXX: remove */
3726 bp->b_ioflags = bip->bio_flags;
3727 bp->b_error = bip->bio_error;
3729 bp->b_ioflags |= BIO_ERROR;
3735 dev_strategy(struct cdev *dev, struct buf *bp)
3740 KASSERT(dev->si_refcount > 0,
3741 ("dev_strategy on un-referenced struct cdev *(%s) %p",
3742 devtoname(dev), dev));
3744 csw = dev_refthread(dev, &ref);
3745 dev_strategy_csw(dev, csw, bp);
3746 dev_relthread(dev, ref);
3750 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3754 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3756 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3757 dev->si_threadcount > 0,
3758 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3761 bp->b_error = ENXIO;
3762 bp->b_ioflags = BIO_ERROR;
3770 /* Try again later */
3771 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3773 bip->bio_cmd = bp->b_iocmd;
3774 bip->bio_offset = bp->b_iooffset;
3775 bip->bio_length = bp->b_bcount;
3776 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3778 bip->bio_done = bufdonebio;
3779 bip->bio_caller2 = bp;
3781 (*csw->d_strategy)(bip);
3787 * Finish I/O on a buffer, optionally calling a completion function.
3788 * This is usually called from an interrupt so process blocking is
3791 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3792 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3793 * assuming B_INVAL is clear.
3795 * For the VMIO case, we set B_CACHE if the op was a read and no
3796 * read error occured, or if the op was a write. B_CACHE is never
3797 * set if the buffer is invalid or otherwise uncacheable.
3799 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3800 * initiator to leave B_INVAL set to brelse the buffer out of existance
3801 * in the biodone routine.
3804 bufdone(struct buf *bp)
3806 struct bufobj *dropobj;
3807 void (*biodone)(struct buf *);
3809 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3812 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3813 BUF_ASSERT_HELD(bp);
3815 runningbufwakeup(bp);
3816 if (bp->b_iocmd == BIO_WRITE)
3817 dropobj = bp->b_bufobj;
3818 /* call optional completion function if requested */
3819 if (bp->b_iodone != NULL) {
3820 biodone = bp->b_iodone;
3821 bp->b_iodone = NULL;
3824 bufobj_wdrop(dropobj);
3831 bufobj_wdrop(dropobj);
3835 bufdone_finish(struct buf *bp)
3837 BUF_ASSERT_HELD(bp);
3839 if (!LIST_EMPTY(&bp->b_dep))
3842 if (bp->b_flags & B_VMIO) {
3847 int bogus, i, iosize;
3849 obj = bp->b_bufobj->bo_object;
3850 KASSERT(obj->paging_in_progress >= bp->b_npages,
3851 ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3852 obj->paging_in_progress, bp->b_npages));
3855 KASSERT(vp->v_holdcnt > 0,
3856 ("biodone_finish: vnode %p has zero hold count", vp));
3857 KASSERT(vp->v_object != NULL,
3858 ("biodone_finish: vnode %p has no vm_object", vp));
3860 foff = bp->b_offset;
3861 KASSERT(bp->b_offset != NOOFFSET,
3862 ("biodone_finish: bp %p has no buffer offset", bp));
3865 * Set B_CACHE if the op was a normal read and no error
3866 * occured. B_CACHE is set for writes in the b*write()
3869 iosize = bp->b_bcount - bp->b_resid;
3870 if (bp->b_iocmd == BIO_READ &&
3871 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3872 !(bp->b_ioflags & BIO_ERROR)) {
3873 bp->b_flags |= B_CACHE;
3876 VM_OBJECT_WLOCK(obj);
3877 for (i = 0; i < bp->b_npages; i++) {
3881 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3886 * cleanup bogus pages, restoring the originals
3889 if (m == bogus_page) {
3890 bogus = bogusflag = 1;
3891 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3893 panic("biodone: page disappeared!");
3896 KASSERT(OFF_TO_IDX(foff) == m->pindex,
3897 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3898 (intmax_t)foff, (uintmax_t)m->pindex));
3901 * In the write case, the valid and clean bits are
3902 * already changed correctly ( see bdwrite() ), so we
3903 * only need to do this here in the read case.
3905 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3906 KASSERT((m->dirty & vm_page_bits(foff &
3907 PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3908 " page %p has unexpected dirty bits", m));
3909 vfs_page_set_valid(bp, foff, m);
3913 vm_object_pip_subtract(obj, 1);
3914 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3917 vm_object_pip_wakeupn(obj, 0);
3918 VM_OBJECT_WUNLOCK(obj);
3919 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3920 BUF_CHECK_MAPPED(bp);
3921 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3922 bp->b_pages, bp->b_npages);
3927 * For asynchronous completions, release the buffer now. The brelse
3928 * will do a wakeup there if necessary - so no need to do a wakeup
3929 * here in the async case. The sync case always needs to do a wakeup.
3932 if (bp->b_flags & B_ASYNC) {
3933 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3942 * This routine is called in lieu of iodone in the case of
3943 * incomplete I/O. This keeps the busy status for pages
3947 vfs_unbusy_pages(struct buf *bp)
3953 runningbufwakeup(bp);
3954 if (!(bp->b_flags & B_VMIO))
3957 obj = bp->b_bufobj->bo_object;
3958 VM_OBJECT_WLOCK(obj);
3959 for (i = 0; i < bp->b_npages; i++) {
3961 if (m == bogus_page) {
3962 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3964 panic("vfs_unbusy_pages: page missing\n");
3966 if ((bp->b_flags & B_UNMAPPED) == 0) {
3967 BUF_CHECK_MAPPED(bp);
3968 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3969 bp->b_pages, bp->b_npages);
3971 BUF_CHECK_UNMAPPED(bp);
3973 vm_object_pip_subtract(obj, 1);
3976 vm_object_pip_wakeupn(obj, 0);
3977 VM_OBJECT_WUNLOCK(obj);
3981 * vfs_page_set_valid:
3983 * Set the valid bits in a page based on the supplied offset. The
3984 * range is restricted to the buffer's size.
3986 * This routine is typically called after a read completes.
3989 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3994 * Compute the end offset, eoff, such that [off, eoff) does not span a
3995 * page boundary and eoff is not greater than the end of the buffer.
3996 * The end of the buffer, in this case, is our file EOF, not the
3997 * allocation size of the buffer.
3999 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4000 if (eoff > bp->b_offset + bp->b_bcount)
4001 eoff = bp->b_offset + bp->b_bcount;
4004 * Set valid range. This is typically the entire buffer and thus the
4008 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4012 * vfs_page_set_validclean:
4014 * Set the valid bits and clear the dirty bits in a page based on the
4015 * supplied offset. The range is restricted to the buffer's size.
4018 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4020 vm_ooffset_t soff, eoff;
4023 * Start and end offsets in buffer. eoff - soff may not cross a
4024 * page boundry or cross the end of the buffer. The end of the
4025 * buffer, in this case, is our file EOF, not the allocation size
4029 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4030 if (eoff > bp->b_offset + bp->b_bcount)
4031 eoff = bp->b_offset + bp->b_bcount;
4034 * Set valid range. This is typically the entire buffer and thus the
4038 vm_page_set_validclean(
4040 (vm_offset_t) (soff & PAGE_MASK),
4041 (vm_offset_t) (eoff - soff)
4047 * Ensure that all buffer pages are not exclusive busied. If any page is
4048 * exclusive busy, drain it.
4051 vfs_drain_busy_pages(struct buf *bp)
4056 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4058 for (i = 0; i < bp->b_npages; i++) {
4060 if (vm_page_xbusied(m)) {
4061 for (; last_busied < i; last_busied++)
4062 vm_page_sbusy(bp->b_pages[last_busied]);
4063 while (vm_page_xbusied(m)) {
4065 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4066 vm_page_busy_sleep(m, "vbpage");
4067 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4071 for (i = 0; i < last_busied; i++)
4072 vm_page_sunbusy(bp->b_pages[i]);
4076 * This routine is called before a device strategy routine.
4077 * It is used to tell the VM system that paging I/O is in
4078 * progress, and treat the pages associated with the buffer
4079 * almost as being exclusive busy. Also the object paging_in_progress
4080 * flag is handled to make sure that the object doesn't become
4083 * Since I/O has not been initiated yet, certain buffer flags
4084 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4085 * and should be ignored.
4088 vfs_busy_pages(struct buf *bp, int clear_modify)
4095 if (!(bp->b_flags & B_VMIO))
4098 obj = bp->b_bufobj->bo_object;
4099 foff = bp->b_offset;
4100 KASSERT(bp->b_offset != NOOFFSET,
4101 ("vfs_busy_pages: no buffer offset"));
4102 VM_OBJECT_WLOCK(obj);
4103 vfs_drain_busy_pages(bp);
4104 if (bp->b_bufsize != 0)
4105 vfs_setdirty_locked_object(bp);
4107 for (i = 0; i < bp->b_npages; i++) {
4110 if ((bp->b_flags & B_CLUSTER) == 0) {
4111 vm_object_pip_add(obj, 1);
4115 * When readying a buffer for a read ( i.e
4116 * clear_modify == 0 ), it is important to do
4117 * bogus_page replacement for valid pages in
4118 * partially instantiated buffers. Partially
4119 * instantiated buffers can, in turn, occur when
4120 * reconstituting a buffer from its VM backing store
4121 * base. We only have to do this if B_CACHE is
4122 * clear ( which causes the I/O to occur in the
4123 * first place ). The replacement prevents the read
4124 * I/O from overwriting potentially dirty VM-backed
4125 * pages. XXX bogus page replacement is, uh, bogus.
4126 * It may not work properly with small-block devices.
4127 * We need to find a better way.
4130 pmap_remove_write(m);
4131 vfs_page_set_validclean(bp, foff, m);
4132 } else if (m->valid == VM_PAGE_BITS_ALL &&
4133 (bp->b_flags & B_CACHE) == 0) {
4134 bp->b_pages[i] = bogus_page;
4137 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4139 VM_OBJECT_WUNLOCK(obj);
4140 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4141 BUF_CHECK_MAPPED(bp);
4142 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4143 bp->b_pages, bp->b_npages);
4148 * vfs_bio_set_valid:
4150 * Set the range within the buffer to valid. The range is
4151 * relative to the beginning of the buffer, b_offset. Note that
4152 * b_offset itself may be offset from the beginning of the first
4156 vfs_bio_set_valid(struct buf *bp, int base, int size)
4161 if (!(bp->b_flags & B_VMIO))
4165 * Fixup base to be relative to beginning of first page.
4166 * Set initial n to be the maximum number of bytes in the
4167 * first page that can be validated.
4169 base += (bp->b_offset & PAGE_MASK);
4170 n = PAGE_SIZE - (base & PAGE_MASK);
4172 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4173 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4177 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4182 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4188 * If the specified buffer is a non-VMIO buffer, clear the entire
4189 * buffer. If the specified buffer is a VMIO buffer, clear and
4190 * validate only the previously invalid portions of the buffer.
4191 * This routine essentially fakes an I/O, so we need to clear
4192 * BIO_ERROR and B_INVAL.
4194 * Note that while we only theoretically need to clear through b_bcount,
4195 * we go ahead and clear through b_bufsize.
4198 vfs_bio_clrbuf(struct buf *bp)
4200 int i, j, mask, sa, ea, slide;
4202 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4206 bp->b_flags &= ~B_INVAL;
4207 bp->b_ioflags &= ~BIO_ERROR;
4208 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4209 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4210 (bp->b_offset & PAGE_MASK) == 0) {
4211 if (bp->b_pages[0] == bogus_page)
4213 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4214 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4215 if ((bp->b_pages[0]->valid & mask) == mask)
4217 if ((bp->b_pages[0]->valid & mask) == 0) {
4218 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4219 bp->b_pages[0]->valid |= mask;
4223 sa = bp->b_offset & PAGE_MASK;
4225 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4226 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4227 ea = slide & PAGE_MASK;
4230 if (bp->b_pages[i] == bogus_page)
4233 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4234 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4235 if ((bp->b_pages[i]->valid & mask) == mask)
4237 if ((bp->b_pages[i]->valid & mask) == 0)
4238 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4240 for (; sa < ea; sa += DEV_BSIZE, j++) {
4241 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4242 pmap_zero_page_area(bp->b_pages[i],
4247 bp->b_pages[i]->valid |= mask;
4250 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4255 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4260 if ((bp->b_flags & B_UNMAPPED) == 0) {
4261 BUF_CHECK_MAPPED(bp);
4262 bzero(bp->b_data + base, size);
4264 BUF_CHECK_UNMAPPED(bp);
4265 n = PAGE_SIZE - (base & PAGE_MASK);
4266 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4270 pmap_zero_page_area(m, base & PAGE_MASK, n);
4279 * vm_hold_load_pages and vm_hold_free_pages get pages into
4280 * a buffers address space. The pages are anonymous and are
4281 * not associated with a file object.
4284 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4290 BUF_CHECK_MAPPED(bp);
4292 to = round_page(to);
4293 from = round_page(from);
4294 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4296 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4299 * note: must allocate system pages since blocking here
4300 * could interfere with paging I/O, no matter which
4303 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4304 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4309 pmap_qenter(pg, &p, 1);
4310 bp->b_pages[index] = p;
4312 bp->b_npages = index;
4315 /* Return pages associated with this buf to the vm system */
4317 vm_hold_free_pages(struct buf *bp, int newbsize)
4321 int index, newnpages;
4323 BUF_CHECK_MAPPED(bp);
4325 from = round_page((vm_offset_t)bp->b_data + newbsize);
4326 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4327 if (bp->b_npages > newnpages)
4328 pmap_qremove(from, bp->b_npages - newnpages);
4329 for (index = newnpages; index < bp->b_npages; index++) {
4330 p = bp->b_pages[index];
4331 bp->b_pages[index] = NULL;
4332 if (vm_page_sbusied(p))
4333 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4334 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4337 atomic_subtract_int(&cnt.v_wire_count, 1);
4339 bp->b_npages = newnpages;
4343 * Map an IO request into kernel virtual address space.
4345 * All requests are (re)mapped into kernel VA space.
4346 * Notice that we use b_bufsize for the size of the buffer
4347 * to be mapped. b_bcount might be modified by the driver.
4349 * Note that even if the caller determines that the address space should
4350 * be valid, a race or a smaller-file mapped into a larger space may
4351 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4352 * check the return value.
4355 vmapbuf(struct buf *bp, int mapbuf)
4361 if (bp->b_bufsize < 0)
4363 prot = VM_PROT_READ;
4364 if (bp->b_iocmd == BIO_READ)
4365 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4366 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4367 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4368 btoc(MAXPHYS))) < 0)
4370 bp->b_npages = pidx;
4371 if (mapbuf || !unmapped_buf_allowed) {
4372 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4373 kva = bp->b_saveaddr;
4374 bp->b_saveaddr = bp->b_data;
4375 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4376 bp->b_flags &= ~B_UNMAPPED;
4378 bp->b_flags |= B_UNMAPPED;
4379 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4380 bp->b_saveaddr = bp->b_data;
4381 bp->b_data = unmapped_buf;
4387 * Free the io map PTEs associated with this IO operation.
4388 * We also invalidate the TLB entries and restore the original b_addr.
4391 vunmapbuf(struct buf *bp)
4395 npages = bp->b_npages;
4396 if (bp->b_flags & B_UNMAPPED)
4397 bp->b_flags &= ~B_UNMAPPED;
4399 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4400 vm_page_unhold_pages(bp->b_pages, npages);
4402 bp->b_data = bp->b_saveaddr;
4406 bdone(struct buf *bp)
4410 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4412 bp->b_flags |= B_DONE;
4418 bwait(struct buf *bp, u_char pri, const char *wchan)
4422 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4424 while ((bp->b_flags & B_DONE) == 0)
4425 msleep(bp, mtxp, pri, wchan, 0);
4430 bufsync(struct bufobj *bo, int waitfor)
4433 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4437 bufstrategy(struct bufobj *bo, struct buf *bp)
4443 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4444 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4445 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4446 i = VOP_STRATEGY(vp, bp);
4447 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4451 bufobj_wrefl(struct bufobj *bo)
4454 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4455 ASSERT_BO_WLOCKED(bo);
4460 bufobj_wref(struct bufobj *bo)
4463 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4470 bufobj_wdrop(struct bufobj *bo)
4473 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4475 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4476 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4477 bo->bo_flag &= ~BO_WWAIT;
4478 wakeup(&bo->bo_numoutput);
4484 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4488 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4489 ASSERT_BO_WLOCKED(bo);
4491 while (bo->bo_numoutput) {
4492 bo->bo_flag |= BO_WWAIT;
4493 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4494 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4502 bpin(struct buf *bp)
4506 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4513 bunpin(struct buf *bp)
4517 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4519 if (--bp->b_pin_count == 0)
4525 bunpin_wait(struct buf *bp)
4529 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4531 while (bp->b_pin_count > 0)
4532 msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4537 * Set bio_data or bio_ma for struct bio from the struct buf.
4540 bdata2bio(struct buf *bp, struct bio *bip)
4543 if ((bp->b_flags & B_UNMAPPED) != 0) {
4544 KASSERT(unmapped_buf_allowed, ("unmapped"));
4545 bip->bio_ma = bp->b_pages;
4546 bip->bio_ma_n = bp->b_npages;
4547 bip->bio_data = unmapped_buf;
4548 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4549 bip->bio_flags |= BIO_UNMAPPED;
4550 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4551 PAGE_SIZE == bp->b_npages,
4552 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4553 (long long)bip->bio_length, bip->bio_ma_n));
4555 bip->bio_data = bp->b_data;
4560 #include "opt_ddb.h"
4562 #include <ddb/ddb.h>
4564 /* DDB command to show buffer data */
4565 DB_SHOW_COMMAND(buffer, db_show_buffer)
4568 struct buf *bp = (struct buf *)addr;
4571 db_printf("usage: show buffer <addr>\n");
4575 db_printf("buf at %p\n", bp);
4576 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4577 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4578 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4580 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4581 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4583 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4584 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4585 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4588 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4589 for (i = 0; i < bp->b_npages; i++) {
4592 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4593 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4594 if ((i + 1) < bp->b_npages)
4600 BUF_LOCKPRINTINFO(bp);
4603 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4608 for (i = 0; i < nbuf; i++) {
4610 if (BUF_ISLOCKED(bp)) {
4611 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4617 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4623 db_printf("usage: show vnodebufs <addr>\n");
4626 vp = (struct vnode *)addr;
4627 db_printf("Clean buffers:\n");
4628 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4629 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4632 db_printf("Dirty buffers:\n");
4633 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4634 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4639 DB_COMMAND(countfreebufs, db_coundfreebufs)
4642 int i, used = 0, nfree = 0;
4645 db_printf("usage: countfreebufs\n");
4649 for (i = 0; i < nbuf; i++) {
4651 if ((bp->b_flags & B_INFREECNT) != 0)
4657 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4659 db_printf("numfreebuffers is %d\n", numfreebuffers);