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 acquisition");
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 * Artificially 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_bufobj->bo_flag & BO_DEAD) != 0) {
1178 bp->b_flags |= B_INVAL | B_RELBUF;
1179 bp->b_flags &= ~B_CACHE;
1183 if (bp->b_flags & B_INVAL) {
1188 if (bp->b_flags & B_BARRIER)
1191 oldflags = bp->b_flags;
1193 BUF_ASSERT_HELD(bp);
1195 if (bp->b_pin_count > 0)
1198 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1199 ("FFS background buffer should not get here %p", bp));
1203 vp_md = vp->v_vflag & VV_MD;
1208 * Mark the buffer clean. Increment the bufobj write count
1209 * before bundirty() call, to prevent other thread from seeing
1210 * empty dirty list and zero counter for writes in progress,
1211 * falsely indicating that the bufobj is clean.
1213 bufobj_wref(bp->b_bufobj);
1216 bp->b_flags &= ~B_DONE;
1217 bp->b_ioflags &= ~BIO_ERROR;
1218 bp->b_flags |= B_CACHE;
1219 bp->b_iocmd = BIO_WRITE;
1221 vfs_busy_pages(bp, 1);
1224 * Normal bwrites pipeline writes
1226 bp->b_runningbufspace = bp->b_bufsize;
1227 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1229 if (!TD_IS_IDLETHREAD(curthread))
1230 curthread->td_ru.ru_oublock++;
1231 if (oldflags & B_ASYNC)
1233 bp->b_iooffset = dbtob(bp->b_blkno);
1236 if ((oldflags & B_ASYNC) == 0) {
1237 int rtval = bufwait(bp);
1240 } else if (space > hirunningspace) {
1242 * don't allow the async write to saturate the I/O
1243 * system. We will not deadlock here because
1244 * we are blocking waiting for I/O that is already in-progress
1245 * to complete. We do not block here if it is the update
1246 * or syncer daemon trying to clean up as that can lead
1249 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1250 waitrunningbufspace();
1257 bufbdflush(struct bufobj *bo, struct buf *bp)
1261 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1262 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1264 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1267 * Try to find a buffer to flush.
1269 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1270 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1272 LK_EXCLUSIVE | LK_NOWAIT, NULL))
1275 panic("bdwrite: found ourselves");
1277 /* Don't countdeps with the bo lock held. */
1278 if (buf_countdeps(nbp, 0)) {
1283 if (nbp->b_flags & B_CLUSTEROK) {
1284 vfs_bio_awrite(nbp);
1289 dirtybufferflushes++;
1298 * Delayed write. (Buffer is marked dirty). Do not bother writing
1299 * anything if the buffer is marked invalid.
1301 * Note that since the buffer must be completely valid, we can safely
1302 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1303 * biodone() in order to prevent getblk from writing the buffer
1304 * out synchronously.
1307 bdwrite(struct buf *bp)
1309 struct thread *td = curthread;
1313 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1314 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1315 KASSERT((bp->b_flags & B_BARRIER) == 0,
1316 ("Barrier request in delayed write %p", bp));
1317 BUF_ASSERT_HELD(bp);
1319 if (bp->b_flags & B_INVAL) {
1325 * If we have too many dirty buffers, don't create any more.
1326 * If we are wildly over our limit, then force a complete
1327 * cleanup. Otherwise, just keep the situation from getting
1328 * out of control. Note that we have to avoid a recursive
1329 * disaster and not try to clean up after our own cleanup!
1333 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1334 td->td_pflags |= TDP_INBDFLUSH;
1336 td->td_pflags &= ~TDP_INBDFLUSH;
1342 * Set B_CACHE, indicating that the buffer is fully valid. This is
1343 * true even of NFS now.
1345 bp->b_flags |= B_CACHE;
1348 * This bmap keeps the system from needing to do the bmap later,
1349 * perhaps when the system is attempting to do a sync. Since it
1350 * is likely that the indirect block -- or whatever other datastructure
1351 * that the filesystem needs is still in memory now, it is a good
1352 * thing to do this. Note also, that if the pageout daemon is
1353 * requesting a sync -- there might not be enough memory to do
1354 * the bmap then... So, this is important to do.
1356 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1357 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1361 * Set the *dirty* buffer range based upon the VM system dirty
1364 * Mark the buffer pages as clean. We need to do this here to
1365 * satisfy the vnode_pager and the pageout daemon, so that it
1366 * thinks that the pages have been "cleaned". Note that since
1367 * the pages are in a delayed write buffer -- the VFS layer
1368 * "will" see that the pages get written out on the next sync,
1369 * or perhaps the cluster will be completed.
1371 vfs_clean_pages_dirty_buf(bp);
1375 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1376 * due to the softdep code.
1383 * Turn buffer into delayed write request. We must clear BIO_READ and
1384 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1385 * itself to properly update it in the dirty/clean lists. We mark it
1386 * B_DONE to ensure that any asynchronization of the buffer properly
1387 * clears B_DONE ( else a panic will occur later ).
1389 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1390 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1391 * should only be called if the buffer is known-good.
1393 * Since the buffer is not on a queue, we do not update the numfreebuffers
1396 * The buffer must be on QUEUE_NONE.
1399 bdirty(struct buf *bp)
1402 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1403 bp, bp->b_vp, bp->b_flags);
1404 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1405 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1406 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1407 BUF_ASSERT_HELD(bp);
1408 bp->b_flags &= ~(B_RELBUF);
1409 bp->b_iocmd = BIO_WRITE;
1411 if ((bp->b_flags & B_DELWRI) == 0) {
1412 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1421 * Clear B_DELWRI for buffer.
1423 * Since the buffer is not on a queue, we do not update the numfreebuffers
1426 * The buffer must be on QUEUE_NONE.
1430 bundirty(struct buf *bp)
1433 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1434 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1435 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1436 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1437 BUF_ASSERT_HELD(bp);
1439 if (bp->b_flags & B_DELWRI) {
1440 bp->b_flags &= ~B_DELWRI;
1445 * Since it is now being written, we can clear its deferred write flag.
1447 bp->b_flags &= ~B_DEFERRED;
1453 * Asynchronous write. Start output on a buffer, but do not wait for
1454 * it to complete. The buffer is released when the output completes.
1456 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1457 * B_INVAL buffers. Not us.
1460 bawrite(struct buf *bp)
1463 bp->b_flags |= B_ASYNC;
1470 * Asynchronous barrier write. Start output on a buffer, but do not
1471 * wait for it to complete. Place a write barrier after this write so
1472 * that this buffer and all buffers written before it are committed to
1473 * the disk before any buffers written after this write are committed
1474 * to the disk. The buffer is released when the output completes.
1477 babarrierwrite(struct buf *bp)
1480 bp->b_flags |= B_ASYNC | B_BARRIER;
1487 * Synchronous barrier write. Start output on a buffer and wait for
1488 * it to complete. Place a write barrier after this write so that
1489 * this buffer and all buffers written before it are committed to
1490 * the disk before any buffers written after this write are committed
1491 * to the disk. The buffer is released when the output completes.
1494 bbarrierwrite(struct buf *bp)
1497 bp->b_flags |= B_BARRIER;
1498 return (bwrite(bp));
1504 * Called prior to the locking of any vnodes when we are expecting to
1505 * write. We do not want to starve the buffer cache with too many
1506 * dirty buffers so we block here. By blocking prior to the locking
1507 * of any vnodes we attempt to avoid the situation where a locked vnode
1508 * prevents the various system daemons from flushing related buffers.
1514 if (numdirtybuffers >= hidirtybuffers) {
1515 mtx_lock(&bdirtylock);
1516 while (numdirtybuffers >= hidirtybuffers) {
1518 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1521 mtx_unlock(&bdirtylock);
1526 * Return true if we have too many dirty buffers.
1529 buf_dirty_count_severe(void)
1532 return(numdirtybuffers >= hidirtybuffers);
1535 static __noinline int
1536 buf_vm_page_count_severe(void)
1539 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1541 return vm_page_count_severe();
1547 * Release a busy buffer and, if requested, free its resources. The
1548 * buffer will be stashed in the appropriate bufqueue[] allowing it
1549 * to be accessed later as a cache entity or reused for other purposes.
1552 brelse(struct buf *bp)
1556 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1557 bp, bp->b_vp, bp->b_flags);
1558 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1559 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1561 if (BUF_LOCKRECURSED(bp)) {
1563 * Do not process, in particular, do not handle the
1564 * B_INVAL/B_RELBUF and do not release to free list.
1570 if (bp->b_flags & B_MANAGED) {
1575 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) {
1576 BO_LOCK(bp->b_bufobj);
1577 bp->b_vflags &= ~BV_BKGRDERR;
1578 BO_UNLOCK(bp->b_bufobj);
1581 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1582 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1584 * Failed write, redirty. Must clear BIO_ERROR to prevent
1585 * pages from being scrapped. If the error is anything
1586 * other than an I/O error (EIO), assume that retrying
1589 bp->b_ioflags &= ~BIO_ERROR;
1591 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1592 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1594 * Either a failed I/O or we were asked to free or not
1597 bp->b_flags |= B_INVAL;
1598 if (!LIST_EMPTY(&bp->b_dep))
1600 if (bp->b_flags & B_DELWRI)
1602 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1603 if ((bp->b_flags & B_VMIO) == 0) {
1612 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1613 * is called with B_DELWRI set, the underlying pages may wind up
1614 * getting freed causing a previous write (bdwrite()) to get 'lost'
1615 * because pages associated with a B_DELWRI bp are marked clean.
1617 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1618 * if B_DELWRI is set.
1620 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1621 * on pages to return pages to the VM page queues.
1623 if (bp->b_flags & B_DELWRI)
1624 bp->b_flags &= ~B_RELBUF;
1625 else if (buf_vm_page_count_severe()) {
1627 * BKGRDINPROG can only be set with the buf and bufobj
1628 * locks both held. We tolerate a race to clear it here.
1630 if (!(bp->b_vflags & BV_BKGRDINPROG))
1631 bp->b_flags |= B_RELBUF;
1635 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1636 * constituted, not even NFS buffers now. Two flags effect this. If
1637 * B_INVAL, the struct buf is invalidated but the VM object is kept
1638 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1640 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1641 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1642 * buffer is also B_INVAL because it hits the re-dirtying code above.
1644 * Normally we can do this whether a buffer is B_DELWRI or not. If
1645 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1646 * the commit state and we cannot afford to lose the buffer. If the
1647 * buffer has a background write in progress, we need to keep it
1648 * around to prevent it from being reconstituted and starting a second
1651 if ((bp->b_flags & B_VMIO)
1652 && !(bp->b_vp->v_mount != NULL &&
1653 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1654 !vn_isdisk(bp->b_vp, NULL) &&
1655 (bp->b_flags & B_DELWRI))
1664 obj = bp->b_bufobj->bo_object;
1667 * Get the base offset and length of the buffer. Note that
1668 * in the VMIO case if the buffer block size is not
1669 * page-aligned then b_data pointer may not be page-aligned.
1670 * But our b_pages[] array *IS* page aligned.
1672 * block sizes less then DEV_BSIZE (usually 512) are not
1673 * supported due to the page granularity bits (m->valid,
1674 * m->dirty, etc...).
1676 * See man buf(9) for more information
1678 resid = bp->b_bufsize;
1679 foff = bp->b_offset;
1680 for (i = 0; i < bp->b_npages; i++) {
1686 * If we hit a bogus page, fixup *all* the bogus pages
1689 if (m == bogus_page) {
1690 poff = OFF_TO_IDX(bp->b_offset);
1693 VM_OBJECT_RLOCK(obj);
1694 for (j = i; j < bp->b_npages; j++) {
1696 mtmp = bp->b_pages[j];
1697 if (mtmp == bogus_page) {
1698 mtmp = vm_page_lookup(obj, poff + j);
1700 panic("brelse: page missing\n");
1702 bp->b_pages[j] = mtmp;
1705 VM_OBJECT_RUNLOCK(obj);
1707 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1708 BUF_CHECK_MAPPED(bp);
1710 trunc_page((vm_offset_t)bp->b_data),
1711 bp->b_pages, bp->b_npages);
1715 if ((bp->b_flags & B_NOCACHE) ||
1716 (bp->b_ioflags & BIO_ERROR &&
1717 bp->b_iocmd == BIO_READ)) {
1718 int poffset = foff & PAGE_MASK;
1719 int presid = resid > (PAGE_SIZE - poffset) ?
1720 (PAGE_SIZE - poffset) : resid;
1722 KASSERT(presid >= 0, ("brelse: extra page"));
1723 VM_OBJECT_WLOCK(obj);
1724 while (vm_page_xbusied(m)) {
1726 VM_OBJECT_WUNLOCK(obj);
1727 vm_page_busy_sleep(m, "mbncsh", true);
1728 VM_OBJECT_WLOCK(obj);
1730 if (pmap_page_wired_mappings(m) == 0)
1731 vm_page_set_invalid(m, poffset, presid);
1732 VM_OBJECT_WUNLOCK(obj);
1734 printf("avoided corruption bug in bogus_page/brelse code\n");
1736 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1737 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1739 if (bp->b_flags & (B_INVAL | B_RELBUF))
1740 vfs_vmio_release(bp);
1742 } else if (bp->b_flags & B_VMIO) {
1744 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1745 vfs_vmio_release(bp);
1748 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1749 if (bp->b_bufsize != 0)
1751 if (bp->b_vp != NULL)
1756 * If the buffer has junk contents signal it and eventually
1757 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1760 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1761 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1762 bp->b_flags |= B_INVAL;
1763 if (bp->b_flags & B_INVAL) {
1764 if (bp->b_flags & B_DELWRI)
1770 /* buffers with no memory */
1771 if (bp->b_bufsize == 0) {
1772 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1773 if (bp->b_vflags & BV_BKGRDINPROG)
1774 panic("losing buffer 1");
1776 qindex = QUEUE_EMPTYKVA;
1778 qindex = QUEUE_EMPTY;
1779 bp->b_flags |= B_AGE;
1780 /* buffers with junk contents */
1781 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1782 (bp->b_ioflags & BIO_ERROR)) {
1783 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1784 if (bp->b_vflags & BV_BKGRDINPROG)
1785 panic("losing buffer 2");
1786 qindex = QUEUE_CLEAN;
1787 bp->b_flags |= B_AGE;
1788 /* remaining buffers */
1789 } else if (bp->b_flags & B_DELWRI)
1790 qindex = QUEUE_DIRTY;
1792 qindex = QUEUE_CLEAN;
1794 binsfree(bp, qindex);
1796 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1797 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1798 panic("brelse: not dirty");
1804 * Release a buffer back to the appropriate queue but do not try to free
1805 * it. The buffer is expected to be used again soon.
1807 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1808 * biodone() to requeue an async I/O on completion. It is also used when
1809 * known good buffers need to be requeued but we think we may need the data
1812 * XXX we should be able to leave the B_RELBUF hint set on completion.
1815 bqrelse(struct buf *bp)
1819 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1820 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1821 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1823 if (BUF_LOCKRECURSED(bp)) {
1824 /* do not release to free list */
1828 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1830 if (bp->b_flags & B_MANAGED) {
1831 if (bp->b_flags & B_REMFREE)
1836 /* buffers with stale but valid contents */
1837 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG |
1838 BV_BKGRDERR)) == BV_BKGRDERR) {
1839 BO_LOCK(bp->b_bufobj);
1840 bp->b_vflags &= ~BV_BKGRDERR;
1841 BO_UNLOCK(bp->b_bufobj);
1842 qindex = QUEUE_DIRTY;
1844 if ((bp->b_flags & B_DELWRI) == 0 &&
1845 (bp->b_xflags & BX_VNDIRTY))
1846 panic("bqrelse: not dirty");
1848 * BKGRDINPROG can only be set with the buf and bufobj
1849 * locks both held. We tolerate a race to clear it here.
1851 if (buf_vm_page_count_severe() &&
1852 (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1854 * We are too low on memory, we have to try to free
1855 * the buffer (most importantly: the wired pages
1856 * making up its backing store) *now*.
1861 qindex = QUEUE_CLEAN;
1863 binsfree(bp, qindex);
1870 /* Give pages used by the bp back to the VM system (where possible) */
1872 vfs_vmio_release(struct buf *bp)
1878 if ((bp->b_flags & B_UNMAPPED) == 0) {
1879 BUF_CHECK_MAPPED(bp);
1880 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1882 BUF_CHECK_UNMAPPED(bp);
1883 obj = bp->b_bufobj->bo_object;
1885 VM_OBJECT_WLOCK(obj);
1886 for (i = 0; i < bp->b_npages; i++) {
1888 bp->b_pages[i] = NULL;
1890 * In order to keep page LRU ordering consistent, put
1891 * everything on the inactive queue.
1894 vm_page_unwire(m, 0);
1897 * Might as well free the page if we can and it has
1898 * no valid data. We also free the page if the
1899 * buffer was used for direct I/O
1901 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1902 if (m->wire_count == 0 && !vm_page_busied(m))
1904 } else if (bp->b_flags & B_DIRECT)
1905 vm_page_try_to_free(m);
1906 else if (buf_vm_page_count_severe())
1907 vm_page_try_to_cache(m);
1911 VM_OBJECT_WUNLOCK(obj);
1913 if (bp->b_bufsize) {
1918 bp->b_flags &= ~B_VMIO;
1924 * Check to see if a block at a particular lbn is available for a clustered
1928 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1935 /* If the buf isn't in core skip it */
1936 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1939 /* If the buf is busy we don't want to wait for it */
1940 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1943 /* Only cluster with valid clusterable delayed write buffers */
1944 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1945 (B_DELWRI | B_CLUSTEROK))
1948 if (bpa->b_bufsize != size)
1952 * Check to see if it is in the expected place on disk and that the
1953 * block has been mapped.
1955 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1965 * Implement clustered async writes for clearing out B_DELWRI buffers.
1966 * This is much better then the old way of writing only one buffer at
1967 * a time. Note that we may not be presented with the buffers in the
1968 * correct order, so we search for the cluster in both directions.
1971 vfs_bio_awrite(struct buf *bp)
1976 daddr_t lblkno = bp->b_lblkno;
1977 struct vnode *vp = bp->b_vp;
1985 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1987 * right now we support clustered writing only to regular files. If
1988 * we find a clusterable block we could be in the middle of a cluster
1989 * rather then at the beginning.
1991 if ((vp->v_type == VREG) &&
1992 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1993 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1995 size = vp->v_mount->mnt_stat.f_iosize;
1996 maxcl = MAXPHYS / size;
1999 for (i = 1; i < maxcl; i++)
2000 if (vfs_bio_clcheck(vp, size, lblkno + i,
2001 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2004 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2005 if (vfs_bio_clcheck(vp, size, lblkno - j,
2006 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2012 * this is a possible cluster write
2016 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2022 bp->b_flags |= B_ASYNC;
2024 * default (old) behavior, writing out only one block
2026 * XXX returns b_bufsize instead of b_bcount for nwritten?
2028 nwritten = bp->b_bufsize;
2035 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2038 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2039 bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2040 if ((gbflags & GB_UNMAPPED) == 0) {
2041 bp->b_kvabase = (caddr_t)addr;
2042 } else if ((gbflags & GB_KVAALLOC) != 0) {
2043 KASSERT((gbflags & GB_UNMAPPED) != 0,
2044 ("GB_KVAALLOC without GB_UNMAPPED"));
2045 bp->b_kvaalloc = (caddr_t)addr;
2046 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2047 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2049 bp->b_kvasize = maxsize;
2053 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2057 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2064 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2066 * Buffer map is too fragmented. Request the caller
2067 * to defragment the map.
2069 atomic_add_int(&bufdefragcnt, 1);
2072 setbufkva(bp, addr, maxsize, gbflags);
2073 atomic_add_long(&bufspace, bp->b_kvasize);
2078 * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2079 * locked vnode is supplied.
2082 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2087 int error, fl, flags, norunbuf;
2089 mtx_assert(&bqclean, MA_OWNED);
2092 flags = VFS_BIO_NEED_BUFSPACE;
2094 } else if (bufspace >= hibufspace) {
2096 flags = VFS_BIO_NEED_BUFSPACE;
2099 flags = VFS_BIO_NEED_ANY;
2101 atomic_set_int(&needsbuffer, flags);
2102 mtx_unlock(&bqclean);
2104 bd_speedup(); /* heeeelp */
2105 if ((gbflags & GB_NOWAIT_BD) != 0)
2110 while ((needsbuffer & flags) != 0) {
2111 if (vp != NULL && vp->v_type != VCHR &&
2112 (td->td_pflags & TDP_BUFNEED) == 0) {
2113 rw_wunlock(&nblock);
2115 * getblk() is called with a vnode locked, and
2116 * some majority of the dirty buffers may as
2117 * well belong to the vnode. Flushing the
2118 * buffers there would make a progress that
2119 * cannot be achieved by the buf_daemon, that
2120 * cannot lock the vnode.
2122 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2123 (td->td_pflags & TDP_NORUNNINGBUF);
2126 * Play bufdaemon. The getnewbuf() function
2127 * may be called while the thread owns lock
2128 * for another dirty buffer for the same
2129 * vnode, which makes it impossible to use
2130 * VOP_FSYNC() there, due to the buffer lock
2133 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2134 fl = buf_flush(vp, flushbufqtarget);
2135 td->td_pflags &= norunbuf;
2139 if ((needsbuffer & flags) == 0)
2142 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
2143 (PRIBIO + 4) | slpflag, waitmsg, slptimeo);
2147 rw_wunlock(&nblock);
2151 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2154 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2155 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2156 bp->b_kvasize, bp->b_bufsize, qindex);
2157 mtx_assert(&bqclean, MA_NOTOWNED);
2160 * Note: we no longer distinguish between VMIO and non-VMIO
2163 KASSERT((bp->b_flags & B_DELWRI) == 0,
2164 ("delwri buffer %p found in queue %d", bp, qindex));
2166 if (qindex == QUEUE_CLEAN) {
2167 if (bp->b_flags & B_VMIO) {
2168 bp->b_flags &= ~B_ASYNC;
2169 vfs_vmio_release(bp);
2171 if (bp->b_vp != NULL)
2176 * Get the rest of the buffer freed up. b_kva* is still valid
2177 * after this operation.
2180 if (bp->b_rcred != NOCRED) {
2181 crfree(bp->b_rcred);
2182 bp->b_rcred = NOCRED;
2184 if (bp->b_wcred != NOCRED) {
2185 crfree(bp->b_wcred);
2186 bp->b_wcred = NOCRED;
2188 if (!LIST_EMPTY(&bp->b_dep))
2190 if (bp->b_vflags & BV_BKGRDINPROG)
2191 panic("losing buffer 3");
2192 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d",
2193 bp, bp->b_vp, qindex));
2194 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2195 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2200 bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2203 KASSERT((bp->b_flags & B_INFREECNT) == 0,
2204 ("buf %p still counted as free?", bp));
2207 bp->b_blkno = bp->b_lblkno = 0;
2208 bp->b_offset = NOOFFSET;
2214 bp->b_dirtyoff = bp->b_dirtyend = 0;
2215 bp->b_bufobj = NULL;
2216 bp->b_pin_count = 0;
2217 bp->b_fsprivate1 = NULL;
2218 bp->b_fsprivate2 = NULL;
2219 bp->b_fsprivate3 = NULL;
2221 LIST_INIT(&bp->b_dep);
2224 static int flushingbufs;
2227 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2229 struct buf *bp, *nbp;
2230 int nqindex, qindex, pass;
2232 KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2236 atomic_add_int(&getnewbufrestarts, 1);
2239 * Setup for scan. If we do not have enough free buffers,
2240 * we setup a degenerate case that immediately fails. Note
2241 * that if we are specially marked process, we are allowed to
2242 * dip into our reserves.
2244 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2245 * for the allocation of the mapped buffer. For unmapped, the
2246 * easiest is to start with EMPTY outright.
2248 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
2249 * However, there are a number of cases (defragging, reusing, ...)
2250 * where we cannot backup.
2254 if (!defrag && unmapped) {
2255 nqindex = QUEUE_EMPTY;
2256 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2259 nqindex = QUEUE_EMPTYKVA;
2260 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2264 * If no EMPTYKVA buffers and we are either defragging or
2265 * reusing, locate a CLEAN buffer to free or reuse. If
2266 * bufspace useage is low skip this step so we can allocate a
2269 if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2270 nqindex = QUEUE_CLEAN;
2271 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2275 * If we could not find or were not allowed to reuse a CLEAN
2276 * buffer, check to see if it is ok to use an EMPTY buffer.
2277 * We can only use an EMPTY buffer if allocating its KVA would
2278 * not otherwise run us out of buffer space. No KVA is needed
2279 * for the unmapped allocation.
2281 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2283 nqindex = QUEUE_EMPTY;
2284 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2288 * All available buffers might be clean, retry ignoring the
2289 * lobufspace as the last resort.
2291 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2292 nqindex = QUEUE_CLEAN;
2293 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2297 * Run scan, possibly freeing data and/or kva mappings on the fly
2300 while ((bp = nbp) != NULL) {
2304 * Calculate next bp (we can only use it if we do not
2305 * block or do other fancy things).
2307 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2310 nqindex = QUEUE_EMPTYKVA;
2311 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2315 case QUEUE_EMPTYKVA:
2316 nqindex = QUEUE_CLEAN;
2317 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2322 if (metadata && pass == 1) {
2324 nqindex = QUEUE_EMPTY;
2326 &bufqueues[QUEUE_EMPTY]);
2335 * If we are defragging then we need a buffer with
2336 * b_kvasize != 0. XXX this situation should no longer
2337 * occur, if defrag is non-zero the buffer's b_kvasize
2338 * should also be non-zero at this point. XXX
2340 if (defrag && bp->b_kvasize == 0) {
2341 printf("Warning: defrag empty buffer %p\n", bp);
2346 * Start freeing the bp. This is somewhat involved. nbp
2347 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2349 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2352 * BKGRDINPROG can only be set with the buf and bufobj
2353 * locks both held. We tolerate a race to clear it here.
2355 if (bp->b_vflags & BV_BKGRDINPROG) {
2361 * Requeue the background write buffer with error.
2363 if ((bp->b_vflags & BV_BKGRDERR) != 0) {
2365 mtx_unlock(&bqclean);
2370 KASSERT(bp->b_qindex == qindex,
2371 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2374 mtx_unlock(&bqclean);
2376 * NOTE: nbp is now entirely invalid. We can only restart
2377 * the scan from this point on.
2380 getnewbuf_reuse_bp(bp, qindex);
2381 mtx_assert(&bqclean, MA_NOTOWNED);
2384 * If we are defragging then free the buffer.
2387 bp->b_flags |= B_INVAL;
2395 * Notify any waiters for the buffer lock about
2396 * identity change by freeing the buffer.
2398 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2399 bp->b_flags |= B_INVAL;
2409 * If we are overcomitted then recover the buffer and its
2410 * KVM space. This occurs in rare situations when multiple
2411 * processes are blocked in getnewbuf() or allocbuf().
2413 if (bufspace >= hibufspace)
2415 if (flushingbufs && bp->b_kvasize != 0) {
2416 bp->b_flags |= B_INVAL;
2421 if (bufspace < lobufspace)
2431 * Find and initialize a new buffer header, freeing up existing buffers
2432 * in the bufqueues as necessary. The new buffer is returned locked.
2434 * Important: B_INVAL is not set. If the caller wishes to throw the
2435 * buffer away, the caller must set B_INVAL prior to calling brelse().
2438 * We have insufficient buffer headers
2439 * We have insufficient buffer space
2440 * buffer_arena is too fragmented ( space reservation fails )
2441 * If we have to flush dirty buffers ( but we try to avoid this )
2444 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2448 int defrag, metadata;
2450 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2451 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2452 if (!unmapped_buf_allowed)
2453 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2456 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2462 * We can't afford to block since we might be holding a vnode lock,
2463 * which may prevent system daemons from running. We deal with
2464 * low-memory situations by proactively returning memory and running
2465 * async I/O rather then sync I/O.
2467 atomic_add_int(&getnewbufcalls, 1);
2468 atomic_subtract_int(&getnewbufrestarts, 1);
2470 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2471 GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2476 * If we exhausted our list, sleep as appropriate. We may have to
2477 * wakeup various daemons and write out some dirty buffers.
2479 * Generally we are sleeping due to insufficient buffer space.
2482 mtx_assert(&bqclean, MA_OWNED);
2483 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2484 mtx_assert(&bqclean, MA_NOTOWNED);
2485 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2486 mtx_assert(&bqclean, MA_NOTOWNED);
2489 bp->b_flags |= B_UNMAPPED;
2490 bp->b_kvabase = bp->b_data = unmapped_buf;
2491 bp->b_kvasize = maxsize;
2492 atomic_add_long(&bufspace, bp->b_kvasize);
2493 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2494 atomic_add_int(&bufreusecnt, 1);
2496 mtx_assert(&bqclean, MA_NOTOWNED);
2499 * We finally have a valid bp. We aren't quite out of the
2500 * woods, we still have to reserve kva space. In order
2501 * to keep fragmentation sane we only allocate kva in
2504 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2506 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2507 B_KVAALLOC)) == B_UNMAPPED) {
2508 if (allocbufkva(bp, maxsize, gbflags)) {
2510 bp->b_flags |= B_INVAL;
2514 atomic_add_int(&bufreusecnt, 1);
2515 } else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2516 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2518 * If the reused buffer has KVA allocated,
2519 * reassign b_kvaalloc to b_kvabase.
2521 bp->b_kvabase = bp->b_kvaalloc;
2522 bp->b_flags &= ~B_KVAALLOC;
2523 atomic_subtract_long(&unmapped_bufspace,
2525 atomic_add_int(&bufreusecnt, 1);
2526 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2527 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2530 * The case of reused buffer already have KVA
2531 * mapped, but the request is for unmapped
2532 * buffer with KVA allocated.
2534 bp->b_kvaalloc = bp->b_kvabase;
2535 bp->b_data = bp->b_kvabase = unmapped_buf;
2536 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2537 atomic_add_long(&unmapped_bufspace,
2539 atomic_add_int(&bufreusecnt, 1);
2541 if ((gbflags & GB_UNMAPPED) == 0) {
2542 bp->b_saveaddr = bp->b_kvabase;
2543 bp->b_data = bp->b_saveaddr;
2544 bp->b_flags &= ~B_UNMAPPED;
2545 BUF_CHECK_MAPPED(bp);
2554 * buffer flushing daemon. Buffers are normally flushed by the
2555 * update daemon but if it cannot keep up this process starts to
2556 * take the load in an attempt to prevent getnewbuf() from blocking.
2559 static struct kproc_desc buf_kp = {
2564 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2567 buf_flush(struct vnode *vp, int target)
2571 flushed = flushbufqueues(vp, target, 0);
2574 * Could not find any buffers without rollback
2575 * dependencies, so just write the first one
2576 * in the hopes of eventually making progress.
2578 if (vp != NULL && target > 2)
2580 flushbufqueues(vp, target, 1);
2591 * This process needs to be suspended prior to shutdown sync.
2593 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2597 * This process is allowed to take the buffer cache to the limit
2599 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2603 mtx_unlock(&bdlock);
2605 kproc_suspend_check(bufdaemonproc);
2606 lodirty = lodirtybuffers;
2607 if (bd_speedupreq) {
2608 lodirty = numdirtybuffers / 2;
2612 * Do the flush. Limit the amount of in-transit I/O we
2613 * allow to build up, otherwise we would completely saturate
2616 while (numdirtybuffers > lodirty) {
2617 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0)
2619 kern_yield(PRI_USER);
2623 * Only clear bd_request if we have reached our low water
2624 * mark. The buf_daemon normally waits 1 second and
2625 * then incrementally flushes any dirty buffers that have
2626 * built up, within reason.
2628 * If we were unable to hit our low water mark and couldn't
2629 * find any flushable buffers, we sleep for a short period
2630 * to avoid endless loops on unlockable buffers.
2633 if (numdirtybuffers <= lodirtybuffers) {
2635 * We reached our low water mark, reset the
2636 * request and sleep until we are needed again.
2637 * The sleep is just so the suspend code works.
2641 * Do an extra wakeup in case dirty threshold
2642 * changed via sysctl and the explicit transition
2643 * out of shortfall was missed.
2646 if (runningbufspace <= lorunningspace)
2648 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2651 * We couldn't find any flushable dirty buffers but
2652 * still have too many dirty buffers, we
2653 * have to sleep and try again. (rare)
2655 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2663 * Try to flush a buffer in the dirty queue. We must be careful to
2664 * free up B_INVAL buffers instead of write them, which NFS is
2665 * particularly sensitive to.
2667 static int flushwithdeps = 0;
2668 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2669 0, "Number of buffers flushed with dependecies that require rollbacks");
2672 flushbufqueues(struct vnode *lvp, int target, int flushdeps)
2674 struct buf *sentinel;
2685 queue = QUEUE_DIRTY;
2687 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2688 sentinel->b_qindex = QUEUE_SENTINEL;
2690 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2691 mtx_unlock(&bqdirty);
2692 while (flushed != target) {
2695 bp = TAILQ_NEXT(sentinel, b_freelist);
2697 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2698 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2701 mtx_unlock(&bqdirty);
2705 * Skip sentinels inserted by other invocations of the
2706 * flushbufqueues(), taking care to not reorder them.
2708 * Only flush the buffers that belong to the
2709 * vnode locked by the curthread.
2711 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL &&
2713 mtx_unlock(&bqdirty);
2716 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2717 mtx_unlock(&bqdirty);
2720 if (bp->b_pin_count > 0) {
2725 * BKGRDINPROG can only be set with the buf and bufobj
2726 * locks both held. We tolerate a race to clear it here.
2728 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2729 (bp->b_flags & B_DELWRI) == 0) {
2733 if (bp->b_flags & B_INVAL) {
2740 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2741 if (flushdeps == 0) {
2749 * We must hold the lock on a vnode before writing
2750 * one of its buffers. Otherwise we may confuse, or
2751 * in the case of a snapshot vnode, deadlock the
2754 * The lock order here is the reverse of the normal
2755 * of vnode followed by buf lock. This is ok because
2756 * the NOWAIT will prevent deadlock.
2759 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2765 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2767 ASSERT_VOP_LOCKED(vp, "getbuf");
2769 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2770 vn_lock(vp, LK_TRYUPGRADE);
2773 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2774 bp, bp->b_vp, bp->b_flags);
2775 if (curproc == bufdaemonproc) {
2782 vn_finished_write(mp);
2785 flushwithdeps += hasdeps;
2789 * Sleeping on runningbufspace while holding
2790 * vnode lock leads to deadlock.
2792 if (curproc == bufdaemonproc &&
2793 runningbufspace > hirunningspace)
2794 waitrunningbufspace();
2797 vn_finished_write(mp);
2801 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2802 mtx_unlock(&bqdirty);
2803 free(sentinel, M_TEMP);
2808 * Check to see if a block is currently memory resident.
2811 incore(struct bufobj *bo, daddr_t blkno)
2816 bp = gbincore(bo, blkno);
2822 * Returns true if no I/O is needed to access the
2823 * associated VM object. This is like incore except
2824 * it also hunts around in the VM system for the data.
2828 inmem(struct vnode * vp, daddr_t blkno)
2831 vm_offset_t toff, tinc, size;
2835 ASSERT_VOP_LOCKED(vp, "inmem");
2837 if (incore(&vp->v_bufobj, blkno))
2839 if (vp->v_mount == NULL)
2846 if (size > vp->v_mount->mnt_stat.f_iosize)
2847 size = vp->v_mount->mnt_stat.f_iosize;
2848 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2850 VM_OBJECT_RLOCK(obj);
2851 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2852 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2856 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2857 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2858 if (vm_page_is_valid(m,
2859 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2862 VM_OBJECT_RUNLOCK(obj);
2866 VM_OBJECT_RUNLOCK(obj);
2871 * Set the dirty range for a buffer based on the status of the dirty
2872 * bits in the pages comprising the buffer. The range is limited
2873 * to the size of the buffer.
2875 * Tell the VM system that the pages associated with this buffer
2876 * are clean. This is used for delayed writes where the data is
2877 * going to go to disk eventually without additional VM intevention.
2879 * Note that while we only really need to clean through to b_bcount, we
2880 * just go ahead and clean through to b_bufsize.
2883 vfs_clean_pages_dirty_buf(struct buf *bp)
2885 vm_ooffset_t foff, noff, eoff;
2889 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2892 foff = bp->b_offset;
2893 KASSERT(bp->b_offset != NOOFFSET,
2894 ("vfs_clean_pages_dirty_buf: no buffer offset"));
2896 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2897 vfs_drain_busy_pages(bp);
2898 vfs_setdirty_locked_object(bp);
2899 for (i = 0; i < bp->b_npages; i++) {
2900 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2902 if (eoff > bp->b_offset + bp->b_bufsize)
2903 eoff = bp->b_offset + bp->b_bufsize;
2905 vfs_page_set_validclean(bp, foff, m);
2906 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2909 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2913 vfs_setdirty_locked_object(struct buf *bp)
2918 object = bp->b_bufobj->bo_object;
2919 VM_OBJECT_ASSERT_WLOCKED(object);
2922 * We qualify the scan for modified pages on whether the
2923 * object has been flushed yet.
2925 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2926 vm_offset_t boffset;
2927 vm_offset_t eoffset;
2930 * test the pages to see if they have been modified directly
2931 * by users through the VM system.
2933 for (i = 0; i < bp->b_npages; i++)
2934 vm_page_test_dirty(bp->b_pages[i]);
2937 * Calculate the encompassing dirty range, boffset and eoffset,
2938 * (eoffset - boffset) bytes.
2941 for (i = 0; i < bp->b_npages; i++) {
2942 if (bp->b_pages[i]->dirty)
2945 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2947 for (i = bp->b_npages - 1; i >= 0; --i) {
2948 if (bp->b_pages[i]->dirty) {
2952 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2955 * Fit it to the buffer.
2958 if (eoffset > bp->b_bcount)
2959 eoffset = bp->b_bcount;
2962 * If we have a good dirty range, merge with the existing
2966 if (boffset < eoffset) {
2967 if (bp->b_dirtyoff > boffset)
2968 bp->b_dirtyoff = boffset;
2969 if (bp->b_dirtyend < eoffset)
2970 bp->b_dirtyend = eoffset;
2976 * Allocate the KVA mapping for an existing buffer. It handles the
2977 * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2978 * KVA which is not mapped (B_KVAALLOC).
2981 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2983 struct buf *scratch_bp;
2984 int bsize, maxsize, need_mapping, need_kva;
2987 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2988 (gbflags & GB_UNMAPPED) == 0;
2989 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2990 (gbflags & GB_KVAALLOC) != 0;
2991 if (!need_mapping && !need_kva)
2994 BUF_CHECK_UNMAPPED(bp);
2996 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2998 * Buffer is not mapped, but the KVA was already
2999 * reserved at the time of the instantiation. Use the
3002 bp->b_flags &= ~B_KVAALLOC;
3003 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
3004 bp->b_kvabase = bp->b_kvaalloc;
3005 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
3010 * Calculate the amount of the address space we would reserve
3011 * if the buffer was mapped.
3013 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
3014 offset = blkno * bsize;
3015 maxsize = size + (offset & PAGE_MASK);
3016 maxsize = imax(maxsize, bsize);
3019 if (allocbufkva(bp, maxsize, gbflags)) {
3021 * Request defragmentation. getnewbuf() returns us the
3022 * allocated space by the scratch buffer KVA.
3024 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
3025 (GB_UNMAPPED | GB_KVAALLOC));
3026 if (scratch_bp == NULL) {
3027 if ((gbflags & GB_NOWAIT_BD) != 0) {
3029 * XXXKIB: defragmentation cannot
3030 * succeed, not sure what else to do.
3032 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
3034 atomic_add_int(&mappingrestarts, 1);
3037 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
3038 ("scratch bp !B_KVAALLOC %p", scratch_bp));
3039 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
3040 scratch_bp->b_kvasize, gbflags);
3042 /* Get rid of the scratch buffer. */
3043 scratch_bp->b_kvasize = 0;
3044 scratch_bp->b_flags |= B_INVAL;
3045 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
3052 bp->b_saveaddr = bp->b_kvabase;
3053 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
3054 bp->b_flags &= ~B_UNMAPPED;
3055 BUF_CHECK_MAPPED(bp);
3062 * Get a block given a specified block and offset into a file/device.
3063 * The buffers B_DONE bit will be cleared on return, making it almost
3064 * ready for an I/O initiation. B_INVAL may or may not be set on
3065 * return. The caller should clear B_INVAL prior to initiating a
3068 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3069 * an existing buffer.
3071 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3072 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3073 * and then cleared based on the backing VM. If the previous buffer is
3074 * non-0-sized but invalid, B_CACHE will be cleared.
3076 * If getblk() must create a new buffer, the new buffer is returned with
3077 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3078 * case it is returned with B_INVAL clear and B_CACHE set based on the
3081 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
3082 * B_CACHE bit is clear.
3084 * What this means, basically, is that the caller should use B_CACHE to
3085 * determine whether the buffer is fully valid or not and should clear
3086 * B_INVAL prior to issuing a read. If the caller intends to validate
3087 * the buffer by loading its data area with something, the caller needs
3088 * to clear B_INVAL. If the caller does this without issuing an I/O,
3089 * the caller should set B_CACHE ( as an optimization ), else the caller
3090 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3091 * a write attempt or if it was a successful read. If the caller
3092 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3093 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3096 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3101 int bsize, error, maxsize, vmio;
3104 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3105 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3106 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3107 ASSERT_VOP_LOCKED(vp, "getblk");
3108 if (size > MAXBCACHEBUF)
3109 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size,
3111 if (!unmapped_buf_allowed)
3112 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3117 bp = gbincore(bo, blkno);
3121 * Buffer is in-core. If the buffer is not busy nor managed,
3122 * it must be on a queue.
3124 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3126 if (flags & GB_LOCK_NOWAIT)
3127 lockflags |= LK_NOWAIT;
3129 error = BUF_TIMELOCK(bp, lockflags,
3130 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3133 * If we slept and got the lock we have to restart in case
3134 * the buffer changed identities.
3136 if (error == ENOLCK)
3138 /* We timed out or were interrupted. */
3141 /* If recursed, assume caller knows the rules. */
3142 else if (BUF_LOCKRECURSED(bp))
3146 * The buffer is locked. B_CACHE is cleared if the buffer is
3147 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3148 * and for a VMIO buffer B_CACHE is adjusted according to the
3151 if (bp->b_flags & B_INVAL)
3152 bp->b_flags &= ~B_CACHE;
3153 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3154 bp->b_flags |= B_CACHE;
3155 if (bp->b_flags & B_MANAGED)
3156 MPASS(bp->b_qindex == QUEUE_NONE);
3161 * check for size inconsistencies for non-VMIO case.
3163 if (bp->b_bcount != size) {
3164 if ((bp->b_flags & B_VMIO) == 0 ||
3165 (size > bp->b_kvasize)) {
3166 if (bp->b_flags & B_DELWRI) {
3168 * If buffer is pinned and caller does
3169 * not want sleep waiting for it to be
3170 * unpinned, bail out
3172 if (bp->b_pin_count > 0) {
3173 if (flags & GB_LOCK_NOWAIT) {
3180 bp->b_flags |= B_NOCACHE;
3183 if (LIST_EMPTY(&bp->b_dep)) {
3184 bp->b_flags |= B_RELBUF;
3187 bp->b_flags |= B_NOCACHE;
3196 * Handle the case of unmapped buffer which should
3197 * become mapped, or the buffer for which KVA
3198 * reservation is requested.
3200 bp_unmapped_get_kva(bp, blkno, size, flags);
3203 * If the size is inconsistent in the VMIO case, we can resize
3204 * the buffer. This might lead to B_CACHE getting set or
3205 * cleared. If the size has not changed, B_CACHE remains
3206 * unchanged from its previous state.
3208 if (bp->b_bcount != size)
3211 KASSERT(bp->b_offset != NOOFFSET,
3212 ("getblk: no buffer offset"));
3215 * A buffer with B_DELWRI set and B_CACHE clear must
3216 * be committed before we can return the buffer in
3217 * order to prevent the caller from issuing a read
3218 * ( due to B_CACHE not being set ) and overwriting
3221 * Most callers, including NFS and FFS, need this to
3222 * operate properly either because they assume they
3223 * can issue a read if B_CACHE is not set, or because
3224 * ( for example ) an uncached B_DELWRI might loop due
3225 * to softupdates re-dirtying the buffer. In the latter
3226 * case, B_CACHE is set after the first write completes,
3227 * preventing further loops.
3228 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3229 * above while extending the buffer, we cannot allow the
3230 * buffer to remain with B_CACHE set after the write
3231 * completes or it will represent a corrupt state. To
3232 * deal with this we set B_NOCACHE to scrap the buffer
3235 * We might be able to do something fancy, like setting
3236 * B_CACHE in bwrite() except if B_DELWRI is already set,
3237 * so the below call doesn't set B_CACHE, but that gets real
3238 * confusing. This is much easier.
3241 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3242 bp->b_flags |= B_NOCACHE;
3246 bp->b_flags &= ~B_DONE;
3249 * Buffer is not in-core, create new buffer. The buffer
3250 * returned by getnewbuf() is locked. Note that the returned
3251 * buffer is also considered valid (not marked B_INVAL).
3255 * If the user does not want us to create the buffer, bail out
3258 if (flags & GB_NOCREAT)
3260 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3263 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3264 offset = blkno * bsize;
3265 vmio = vp->v_object != NULL;
3267 maxsize = size + (offset & PAGE_MASK);
3270 /* Do not allow non-VMIO notmapped buffers. */
3271 flags &= ~GB_UNMAPPED;
3273 maxsize = imax(maxsize, bsize);
3275 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3277 if (slpflag || slptimeo)
3283 * This code is used to make sure that a buffer is not
3284 * created while the getnewbuf routine is blocked.
3285 * This can be a problem whether the vnode is locked or not.
3286 * If the buffer is created out from under us, we have to
3287 * throw away the one we just created.
3289 * Note: this must occur before we associate the buffer
3290 * with the vp especially considering limitations in
3291 * the splay tree implementation when dealing with duplicate
3295 if (gbincore(bo, blkno)) {
3297 bp->b_flags |= B_INVAL;
3303 * Insert the buffer into the hash, so that it can
3304 * be found by incore.
3306 bp->b_blkno = bp->b_lblkno = blkno;
3307 bp->b_offset = offset;
3312 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
3313 * buffer size starts out as 0, B_CACHE will be set by
3314 * allocbuf() for the VMIO case prior to it testing the
3315 * backing store for validity.
3319 bp->b_flags |= B_VMIO;
3320 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3321 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3322 bp, vp->v_object, bp->b_bufobj->bo_object));
3324 bp->b_flags &= ~B_VMIO;
3325 KASSERT(bp->b_bufobj->bo_object == NULL,
3326 ("ARGH! has b_bufobj->bo_object %p %p\n",
3327 bp, bp->b_bufobj->bo_object));
3328 BUF_CHECK_MAPPED(bp);
3332 bp->b_flags &= ~B_DONE;
3334 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3335 BUF_ASSERT_HELD(bp);
3337 KASSERT(bp->b_bufobj == bo,
3338 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3343 * Get an empty, disassociated buffer of given size. The buffer is initially
3347 geteblk(int size, int flags)
3352 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3353 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3354 if ((flags & GB_NOWAIT_BD) &&
3355 (curthread->td_pflags & TDP_BUFNEED) != 0)
3359 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3360 BUF_ASSERT_HELD(bp);
3366 * This code constitutes the buffer memory from either anonymous system
3367 * memory (in the case of non-VMIO operations) or from an associated
3368 * VM object (in the case of VMIO operations). This code is able to
3369 * resize a buffer up or down.
3371 * Note that this code is tricky, and has many complications to resolve
3372 * deadlock or inconsistent data situations. Tread lightly!!!
3373 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3374 * the caller. Calling this code willy nilly can result in the loss of data.
3376 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3377 * B_CACHE for the non-VMIO case.
3381 allocbuf(struct buf *bp, int size)
3383 int newbsize, mbsize;
3386 BUF_ASSERT_HELD(bp);
3388 if (bp->b_kvasize < size)
3389 panic("allocbuf: buffer too small");
3391 if ((bp->b_flags & B_VMIO) == 0) {
3395 * Just get anonymous memory from the kernel. Don't
3396 * mess with B_CACHE.
3398 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3399 if (bp->b_flags & B_MALLOC)
3402 newbsize = round_page(size);
3404 if (newbsize < bp->b_bufsize) {
3406 * malloced buffers are not shrunk
3408 if (bp->b_flags & B_MALLOC) {
3410 bp->b_bcount = size;
3412 free(bp->b_data, M_BIOBUF);
3413 if (bp->b_bufsize) {
3414 atomic_subtract_long(
3420 bp->b_saveaddr = bp->b_kvabase;
3421 bp->b_data = bp->b_saveaddr;
3423 bp->b_flags &= ~B_MALLOC;
3427 vm_hold_free_pages(bp, newbsize);
3428 } else if (newbsize > bp->b_bufsize) {
3430 * We only use malloced memory on the first allocation.
3431 * and revert to page-allocated memory when the buffer
3435 * There is a potential smp race here that could lead
3436 * to bufmallocspace slightly passing the max. It
3437 * is probably extremely rare and not worth worrying
3440 if ( (bufmallocspace < maxbufmallocspace) &&
3441 (bp->b_bufsize == 0) &&
3442 (mbsize <= PAGE_SIZE/2)) {
3444 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3445 bp->b_bufsize = mbsize;
3446 bp->b_bcount = size;
3447 bp->b_flags |= B_MALLOC;
3448 atomic_add_long(&bufmallocspace, mbsize);
3454 * If the buffer is growing on its other-than-first allocation,
3455 * then we revert to the page-allocation scheme.
3457 if (bp->b_flags & B_MALLOC) {
3458 origbuf = bp->b_data;
3459 origbufsize = bp->b_bufsize;
3460 bp->b_data = bp->b_kvabase;
3461 if (bp->b_bufsize) {
3462 atomic_subtract_long(&bufmallocspace,
3467 bp->b_flags &= ~B_MALLOC;
3468 newbsize = round_page(newbsize);
3472 (vm_offset_t) bp->b_data + bp->b_bufsize,
3473 (vm_offset_t) bp->b_data + newbsize);
3475 bcopy(origbuf, bp->b_data, origbufsize);
3476 free(origbuf, M_BIOBUF);
3482 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3483 desiredpages = (size == 0) ? 0 :
3484 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3486 if (bp->b_flags & B_MALLOC)
3487 panic("allocbuf: VMIO buffer can't be malloced");
3489 * Set B_CACHE initially if buffer is 0 length or will become
3492 if (size == 0 || bp->b_bufsize == 0)
3493 bp->b_flags |= B_CACHE;
3495 if (newbsize < bp->b_bufsize) {
3497 * DEV_BSIZE aligned new buffer size is less then the
3498 * DEV_BSIZE aligned existing buffer size. Figure out
3499 * if we have to remove any pages.
3501 if (desiredpages < bp->b_npages) {
3504 if ((bp->b_flags & B_UNMAPPED) == 0) {
3505 BUF_CHECK_MAPPED(bp);
3506 pmap_qremove((vm_offset_t)trunc_page(
3507 (vm_offset_t)bp->b_data) +
3508 (desiredpages << PAGE_SHIFT),
3509 (bp->b_npages - desiredpages));
3511 BUF_CHECK_UNMAPPED(bp);
3512 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3513 for (i = desiredpages; i < bp->b_npages; i++) {
3515 * the page is not freed here -- it
3516 * is the responsibility of
3517 * vnode_pager_setsize
3520 KASSERT(m != bogus_page,
3521 ("allocbuf: bogus page found"));
3522 while (vm_page_sleep_if_busy(m,
3526 bp->b_pages[i] = NULL;
3528 vm_page_unwire(m, 0);
3531 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3532 bp->b_npages = desiredpages;
3534 } else if (size > bp->b_bcount) {
3536 * We are growing the buffer, possibly in a
3537 * byte-granular fashion.
3544 * Step 1, bring in the VM pages from the object,
3545 * allocating them if necessary. We must clear
3546 * B_CACHE if these pages are not valid for the
3547 * range covered by the buffer.
3550 obj = bp->b_bufobj->bo_object;
3552 VM_OBJECT_WLOCK(obj);
3553 while (bp->b_npages < desiredpages) {
3557 * We must allocate system pages since blocking
3558 * here could interfere with paging I/O, no
3559 * matter which process we are.
3561 * Only exclusive busy can be tested here.
3562 * Blocking on shared busy might lead to
3563 * deadlocks once allocbuf() is called after
3564 * pages are vfs_busy_pages().
3566 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3567 bp->b_npages, VM_ALLOC_NOBUSY |
3568 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3569 VM_ALLOC_IGN_SBUSY |
3570 VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3572 bp->b_flags &= ~B_CACHE;
3573 bp->b_pages[bp->b_npages] = m;
3578 * Step 2. We've loaded the pages into the buffer,
3579 * we have to figure out if we can still have B_CACHE
3580 * set. Note that B_CACHE is set according to the
3581 * byte-granular range ( bcount and size ), new the
3582 * aligned range ( newbsize ).
3584 * The VM test is against m->valid, which is DEV_BSIZE
3585 * aligned. Needless to say, the validity of the data
3586 * needs to also be DEV_BSIZE aligned. Note that this
3587 * fails with NFS if the server or some other client
3588 * extends the file's EOF. If our buffer is resized,
3589 * B_CACHE may remain set! XXX
3592 toff = bp->b_bcount;
3593 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3595 while ((bp->b_flags & B_CACHE) && toff < size) {
3598 if (tinc > (size - toff))
3601 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3614 VM_OBJECT_WUNLOCK(obj);
3617 * Step 3, fixup the KVM pmap.
3619 if ((bp->b_flags & B_UNMAPPED) == 0)
3622 BUF_CHECK_UNMAPPED(bp);
3625 if (newbsize < bp->b_bufsize)
3627 bp->b_bufsize = newbsize; /* actual buffer allocation */
3628 bp->b_bcount = size; /* requested buffer size */
3632 extern int inflight_transient_maps;
3635 biodone(struct bio *bp)
3638 void (*done)(struct bio *);
3639 vm_offset_t start, end;
3641 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3642 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3643 bp->bio_flags |= BIO_UNMAPPED;
3644 start = trunc_page((vm_offset_t)bp->bio_data);
3645 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3646 pmap_qremove(start, OFF_TO_IDX(end - start));
3647 vmem_free(transient_arena, start, end - start);
3648 atomic_add_int(&inflight_transient_maps, -1);
3650 done = bp->bio_done;
3652 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3654 bp->bio_flags |= BIO_DONE;
3658 bp->bio_flags |= BIO_DONE;
3664 * Wait for a BIO to finish.
3667 biowait(struct bio *bp, const char *wchan)
3671 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3673 while ((bp->bio_flags & BIO_DONE) == 0)
3674 msleep(bp, mtxp, PRIBIO, wchan, 0);
3676 if (bp->bio_error != 0)
3677 return (bp->bio_error);
3678 if (!(bp->bio_flags & BIO_ERROR))
3684 biofinish(struct bio *bp, struct devstat *stat, int error)
3688 bp->bio_error = error;
3689 bp->bio_flags |= BIO_ERROR;
3692 devstat_end_transaction_bio(stat, bp);
3699 * Wait for buffer I/O completion, returning error status. The buffer
3700 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3701 * error and cleared.
3704 bufwait(struct buf *bp)
3706 if (bp->b_iocmd == BIO_READ)
3707 bwait(bp, PRIBIO, "biord");
3709 bwait(bp, PRIBIO, "biowr");
3710 if (bp->b_flags & B_EINTR) {
3711 bp->b_flags &= ~B_EINTR;
3714 if (bp->b_ioflags & BIO_ERROR) {
3715 return (bp->b_error ? bp->b_error : EIO);
3722 * Call back function from struct bio back up to struct buf.
3725 bufdonebio(struct bio *bip)
3729 bp = bip->bio_caller2;
3730 bp->b_resid = bp->b_bcount - bip->bio_completed;
3731 bp->b_resid = bip->bio_resid; /* XXX: remove */
3732 bp->b_ioflags = bip->bio_flags;
3733 bp->b_error = bip->bio_error;
3735 bp->b_ioflags |= BIO_ERROR;
3741 dev_strategy(struct cdev *dev, struct buf *bp)
3746 KASSERT(dev->si_refcount > 0,
3747 ("dev_strategy on un-referenced struct cdev *(%s) %p",
3748 devtoname(dev), dev));
3750 csw = dev_refthread(dev, &ref);
3751 dev_strategy_csw(dev, csw, bp);
3752 dev_relthread(dev, ref);
3756 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3760 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3762 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3763 dev->si_threadcount > 0,
3764 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3767 bp->b_error = ENXIO;
3768 bp->b_ioflags = BIO_ERROR;
3776 /* Try again later */
3777 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3779 bip->bio_cmd = bp->b_iocmd;
3780 bip->bio_offset = bp->b_iooffset;
3781 bip->bio_length = bp->b_bcount;
3782 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3784 bip->bio_done = bufdonebio;
3785 bip->bio_caller2 = bp;
3787 (*csw->d_strategy)(bip);
3793 * Finish I/O on a buffer, optionally calling a completion function.
3794 * This is usually called from an interrupt so process blocking is
3797 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3798 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3799 * assuming B_INVAL is clear.
3801 * For the VMIO case, we set B_CACHE if the op was a read and no
3802 * read error occurred, or if the op was a write. B_CACHE is never
3803 * set if the buffer is invalid or otherwise uncacheable.
3805 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3806 * initiator to leave B_INVAL set to brelse the buffer out of existence
3807 * in the biodone routine.
3810 bufdone(struct buf *bp)
3812 struct bufobj *dropobj;
3813 void (*biodone)(struct buf *);
3815 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3818 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3819 BUF_ASSERT_HELD(bp);
3821 runningbufwakeup(bp);
3822 if (bp->b_iocmd == BIO_WRITE)
3823 dropobj = bp->b_bufobj;
3824 /* call optional completion function if requested */
3825 if (bp->b_iodone != NULL) {
3826 biodone = bp->b_iodone;
3827 bp->b_iodone = NULL;
3830 bufobj_wdrop(dropobj);
3837 bufobj_wdrop(dropobj);
3841 bufdone_finish(struct buf *bp)
3843 BUF_ASSERT_HELD(bp);
3845 if (!LIST_EMPTY(&bp->b_dep))
3848 if (bp->b_flags & B_VMIO) {
3853 int bogus, i, iosize;
3855 obj = bp->b_bufobj->bo_object;
3856 KASSERT(obj->paging_in_progress >= bp->b_npages,
3857 ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3858 obj->paging_in_progress, bp->b_npages));
3861 KASSERT(vp->v_holdcnt > 0,
3862 ("biodone_finish: vnode %p has zero hold count", vp));
3863 KASSERT(vp->v_object != NULL,
3864 ("biodone_finish: vnode %p has no vm_object", vp));
3866 foff = bp->b_offset;
3867 KASSERT(bp->b_offset != NOOFFSET,
3868 ("biodone_finish: bp %p has no buffer offset", bp));
3871 * Set B_CACHE if the op was a normal read and no error
3872 * occurred. B_CACHE is set for writes in the b*write()
3875 iosize = bp->b_bcount - bp->b_resid;
3876 if (bp->b_iocmd == BIO_READ &&
3877 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3878 !(bp->b_ioflags & BIO_ERROR)) {
3879 bp->b_flags |= B_CACHE;
3882 VM_OBJECT_WLOCK(obj);
3883 for (i = 0; i < bp->b_npages; i++) {
3887 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3892 * cleanup bogus pages, restoring the originals
3895 if (m == bogus_page) {
3896 bogus = bogusflag = 1;
3897 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3899 panic("biodone: page disappeared!");
3902 KASSERT(OFF_TO_IDX(foff) == m->pindex,
3903 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3904 (intmax_t)foff, (uintmax_t)m->pindex));
3907 * In the write case, the valid and clean bits are
3908 * already changed correctly ( see bdwrite() ), so we
3909 * only need to do this here in the read case.
3911 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3912 KASSERT((m->dirty & vm_page_bits(foff &
3913 PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3914 " page %p has unexpected dirty bits", m));
3915 vfs_page_set_valid(bp, foff, m);
3919 vm_object_pip_subtract(obj, 1);
3920 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3923 vm_object_pip_wakeupn(obj, 0);
3924 VM_OBJECT_WUNLOCK(obj);
3925 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3926 BUF_CHECK_MAPPED(bp);
3927 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3928 bp->b_pages, bp->b_npages);
3933 * For asynchronous completions, release the buffer now. The brelse
3934 * will do a wakeup there if necessary - so no need to do a wakeup
3935 * here in the async case. The sync case always needs to do a wakeup.
3938 if (bp->b_flags & B_ASYNC) {
3939 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3948 * This routine is called in lieu of iodone in the case of
3949 * incomplete I/O. This keeps the busy status for pages
3953 vfs_unbusy_pages(struct buf *bp)
3959 runningbufwakeup(bp);
3960 if (!(bp->b_flags & B_VMIO))
3963 obj = bp->b_bufobj->bo_object;
3964 VM_OBJECT_WLOCK(obj);
3965 for (i = 0; i < bp->b_npages; i++) {
3967 if (m == bogus_page) {
3968 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3970 panic("vfs_unbusy_pages: page missing\n");
3972 if ((bp->b_flags & B_UNMAPPED) == 0) {
3973 BUF_CHECK_MAPPED(bp);
3974 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3975 bp->b_pages, bp->b_npages);
3977 BUF_CHECK_UNMAPPED(bp);
3979 vm_object_pip_subtract(obj, 1);
3982 vm_object_pip_wakeupn(obj, 0);
3983 VM_OBJECT_WUNLOCK(obj);
3987 * vfs_page_set_valid:
3989 * Set the valid bits in a page based on the supplied offset. The
3990 * range is restricted to the buffer's size.
3992 * This routine is typically called after a read completes.
3995 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4000 * Compute the end offset, eoff, such that [off, eoff) does not span a
4001 * page boundary and eoff is not greater than the end of the buffer.
4002 * The end of the buffer, in this case, is our file EOF, not the
4003 * allocation size of the buffer.
4005 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
4006 if (eoff > bp->b_offset + bp->b_bcount)
4007 eoff = bp->b_offset + bp->b_bcount;
4010 * Set valid range. This is typically the entire buffer and thus the
4014 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
4018 * vfs_page_set_validclean:
4020 * Set the valid bits and clear the dirty bits in a page based on the
4021 * supplied offset. The range is restricted to the buffer's size.
4024 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
4026 vm_ooffset_t soff, eoff;
4029 * Start and end offsets in buffer. eoff - soff may not cross a
4030 * page boundary or cross the end of the buffer. The end of the
4031 * buffer, in this case, is our file EOF, not the allocation size
4035 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4036 if (eoff > bp->b_offset + bp->b_bcount)
4037 eoff = bp->b_offset + bp->b_bcount;
4040 * Set valid range. This is typically the entire buffer and thus the
4044 vm_page_set_validclean(
4046 (vm_offset_t) (soff & PAGE_MASK),
4047 (vm_offset_t) (eoff - soff)
4053 * Ensure that all buffer pages are not exclusive busied. If any page is
4054 * exclusive busy, drain it.
4057 vfs_drain_busy_pages(struct buf *bp)
4062 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4064 for (i = 0; i < bp->b_npages; i++) {
4066 if (vm_page_xbusied(m)) {
4067 for (; last_busied < i; last_busied++)
4068 vm_page_sbusy(bp->b_pages[last_busied]);
4069 while (vm_page_xbusied(m)) {
4071 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4072 vm_page_busy_sleep(m, "vbpage", true);
4073 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4077 for (i = 0; i < last_busied; i++)
4078 vm_page_sunbusy(bp->b_pages[i]);
4082 * This routine is called before a device strategy routine.
4083 * It is used to tell the VM system that paging I/O is in
4084 * progress, and treat the pages associated with the buffer
4085 * almost as being exclusive busy. Also the object paging_in_progress
4086 * flag is handled to make sure that the object doesn't become
4089 * Since I/O has not been initiated yet, certain buffer flags
4090 * such as BIO_ERROR or B_INVAL may be in an inconsistent state
4091 * and should be ignored.
4094 vfs_busy_pages(struct buf *bp, int clear_modify)
4101 if (!(bp->b_flags & B_VMIO))
4104 obj = bp->b_bufobj->bo_object;
4105 foff = bp->b_offset;
4106 KASSERT(bp->b_offset != NOOFFSET,
4107 ("vfs_busy_pages: no buffer offset"));
4108 VM_OBJECT_WLOCK(obj);
4109 vfs_drain_busy_pages(bp);
4110 if (bp->b_bufsize != 0)
4111 vfs_setdirty_locked_object(bp);
4113 for (i = 0; i < bp->b_npages; i++) {
4116 if ((bp->b_flags & B_CLUSTER) == 0) {
4117 vm_object_pip_add(obj, 1);
4121 * When readying a buffer for a read ( i.e
4122 * clear_modify == 0 ), it is important to do
4123 * bogus_page replacement for valid pages in
4124 * partially instantiated buffers. Partially
4125 * instantiated buffers can, in turn, occur when
4126 * reconstituting a buffer from its VM backing store
4127 * base. We only have to do this if B_CACHE is
4128 * clear ( which causes the I/O to occur in the
4129 * first place ). The replacement prevents the read
4130 * I/O from overwriting potentially dirty VM-backed
4131 * pages. XXX bogus page replacement is, uh, bogus.
4132 * It may not work properly with small-block devices.
4133 * We need to find a better way.
4136 pmap_remove_write(m);
4137 vfs_page_set_validclean(bp, foff, m);
4138 } else if (m->valid == VM_PAGE_BITS_ALL &&
4139 (bp->b_flags & B_CACHE) == 0) {
4140 bp->b_pages[i] = bogus_page;
4143 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4145 VM_OBJECT_WUNLOCK(obj);
4146 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4147 BUF_CHECK_MAPPED(bp);
4148 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4149 bp->b_pages, bp->b_npages);
4154 * vfs_bio_set_valid:
4156 * Set the range within the buffer to valid. The range is
4157 * relative to the beginning of the buffer, b_offset. Note that
4158 * b_offset itself may be offset from the beginning of the first
4162 vfs_bio_set_valid(struct buf *bp, int base, int size)
4167 if (!(bp->b_flags & B_VMIO))
4171 * Fixup base to be relative to beginning of first page.
4172 * Set initial n to be the maximum number of bytes in the
4173 * first page that can be validated.
4175 base += (bp->b_offset & PAGE_MASK);
4176 n = PAGE_SIZE - (base & PAGE_MASK);
4178 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4179 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4183 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4188 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4194 * If the specified buffer is a non-VMIO buffer, clear the entire
4195 * buffer. If the specified buffer is a VMIO buffer, clear and
4196 * validate only the previously invalid portions of the buffer.
4197 * This routine essentially fakes an I/O, so we need to clear
4198 * BIO_ERROR and B_INVAL.
4200 * Note that while we only theoretically need to clear through b_bcount,
4201 * we go ahead and clear through b_bufsize.
4204 vfs_bio_clrbuf(struct buf *bp)
4206 int i, j, mask, sa, ea, slide;
4208 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4212 bp->b_flags &= ~B_INVAL;
4213 bp->b_ioflags &= ~BIO_ERROR;
4214 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4215 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4216 (bp->b_offset & PAGE_MASK) == 0) {
4217 if (bp->b_pages[0] == bogus_page)
4219 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4220 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4221 if ((bp->b_pages[0]->valid & mask) == mask)
4223 if ((bp->b_pages[0]->valid & mask) == 0) {
4224 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4225 bp->b_pages[0]->valid |= mask;
4229 sa = bp->b_offset & PAGE_MASK;
4231 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4232 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4233 ea = slide & PAGE_MASK;
4236 if (bp->b_pages[i] == bogus_page)
4239 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4240 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4241 if ((bp->b_pages[i]->valid & mask) == mask)
4243 if ((bp->b_pages[i]->valid & mask) == 0)
4244 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4246 for (; sa < ea; sa += DEV_BSIZE, j++) {
4247 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4248 pmap_zero_page_area(bp->b_pages[i],
4253 bp->b_pages[i]->valid |= mask;
4256 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4261 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4266 if ((bp->b_flags & B_UNMAPPED) == 0) {
4267 BUF_CHECK_MAPPED(bp);
4268 bzero(bp->b_data + base, size);
4270 BUF_CHECK_UNMAPPED(bp);
4271 n = PAGE_SIZE - (base & PAGE_MASK);
4272 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4276 pmap_zero_page_area(m, base & PAGE_MASK, n);
4285 * vm_hold_load_pages and vm_hold_free_pages get pages into
4286 * a buffers address space. The pages are anonymous and are
4287 * not associated with a file object.
4290 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4296 BUF_CHECK_MAPPED(bp);
4298 to = round_page(to);
4299 from = round_page(from);
4300 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4302 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4305 * note: must allocate system pages since blocking here
4306 * could interfere with paging I/O, no matter which
4309 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4310 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4315 pmap_qenter(pg, &p, 1);
4316 bp->b_pages[index] = p;
4318 bp->b_npages = index;
4321 /* Return pages associated with this buf to the vm system */
4323 vm_hold_free_pages(struct buf *bp, int newbsize)
4327 int index, newnpages;
4329 BUF_CHECK_MAPPED(bp);
4331 from = round_page((vm_offset_t)bp->b_data + newbsize);
4332 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4333 if (bp->b_npages > newnpages)
4334 pmap_qremove(from, bp->b_npages - newnpages);
4335 for (index = newnpages; index < bp->b_npages; index++) {
4336 p = bp->b_pages[index];
4337 bp->b_pages[index] = NULL;
4338 if (vm_page_sbusied(p))
4339 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4340 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4343 atomic_subtract_int(&cnt.v_wire_count, 1);
4345 bp->b_npages = newnpages;
4349 * Map an IO request into kernel virtual address space.
4351 * All requests are (re)mapped into kernel VA space.
4352 * Notice that we use b_bufsize for the size of the buffer
4353 * to be mapped. b_bcount might be modified by the driver.
4355 * Note that even if the caller determines that the address space should
4356 * be valid, a race or a smaller-file mapped into a larger space may
4357 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4358 * check the return value.
4361 vmapbuf(struct buf *bp, int mapbuf)
4367 if (bp->b_bufsize < 0)
4369 prot = VM_PROT_READ;
4370 if (bp->b_iocmd == BIO_READ)
4371 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4372 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4373 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4374 btoc(MAXPHYS))) < 0)
4376 bp->b_npages = pidx;
4377 if (mapbuf || !unmapped_buf_allowed) {
4378 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4379 kva = bp->b_saveaddr;
4380 bp->b_saveaddr = bp->b_data;
4381 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4382 bp->b_flags &= ~B_UNMAPPED;
4384 bp->b_flags |= B_UNMAPPED;
4385 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4386 bp->b_saveaddr = bp->b_data;
4387 bp->b_data = unmapped_buf;
4393 * Free the io map PTEs associated with this IO operation.
4394 * We also invalidate the TLB entries and restore the original b_addr.
4397 vunmapbuf(struct buf *bp)
4401 npages = bp->b_npages;
4402 if (bp->b_flags & B_UNMAPPED)
4403 bp->b_flags &= ~B_UNMAPPED;
4405 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4406 vm_page_unhold_pages(bp->b_pages, npages);
4408 bp->b_data = bp->b_saveaddr;
4412 bdone(struct buf *bp)
4416 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4418 bp->b_flags |= B_DONE;
4424 bwait(struct buf *bp, u_char pri, const char *wchan)
4428 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4430 while ((bp->b_flags & B_DONE) == 0)
4431 msleep(bp, mtxp, pri, wchan, 0);
4436 bufsync(struct bufobj *bo, int waitfor)
4439 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4443 bufstrategy(struct bufobj *bo, struct buf *bp)
4449 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4450 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4451 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4452 i = VOP_STRATEGY(vp, bp);
4453 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4457 bufobj_wrefl(struct bufobj *bo)
4460 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4461 ASSERT_BO_WLOCKED(bo);
4466 bufobj_wref(struct bufobj *bo)
4469 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4476 bufobj_wdrop(struct bufobj *bo)
4479 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4481 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4482 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4483 bo->bo_flag &= ~BO_WWAIT;
4484 wakeup(&bo->bo_numoutput);
4490 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4494 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4495 ASSERT_BO_WLOCKED(bo);
4497 while (bo->bo_numoutput) {
4498 bo->bo_flag |= BO_WWAIT;
4499 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4500 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4508 bpin(struct buf *bp)
4512 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4519 bunpin(struct buf *bp)
4523 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4525 if (--bp->b_pin_count == 0)
4531 bunpin_wait(struct buf *bp)
4535 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4537 while (bp->b_pin_count > 0)
4538 msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4543 * Set bio_data or bio_ma for struct bio from the struct buf.
4546 bdata2bio(struct buf *bp, struct bio *bip)
4549 if ((bp->b_flags & B_UNMAPPED) != 0) {
4550 KASSERT(unmapped_buf_allowed, ("unmapped"));
4551 bip->bio_ma = bp->b_pages;
4552 bip->bio_ma_n = bp->b_npages;
4553 bip->bio_data = unmapped_buf;
4554 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4555 bip->bio_flags |= BIO_UNMAPPED;
4556 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4557 PAGE_SIZE == bp->b_npages,
4558 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4559 (long long)bip->bio_length, bip->bio_ma_n));
4561 bip->bio_data = bp->b_data;
4566 #include "opt_ddb.h"
4568 #include <ddb/ddb.h>
4570 /* DDB command to show buffer data */
4571 DB_SHOW_COMMAND(buffer, db_show_buffer)
4574 struct buf *bp = (struct buf *)addr;
4577 db_printf("usage: show buffer <addr>\n");
4581 db_printf("buf at %p\n", bp);
4582 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4583 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4584 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4586 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4587 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4589 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4590 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4591 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4594 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4595 for (i = 0; i < bp->b_npages; i++) {
4598 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4599 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4600 if ((i + 1) < bp->b_npages)
4606 BUF_LOCKPRINTINFO(bp);
4609 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4614 for (i = 0; i < nbuf; i++) {
4616 if (BUF_ISLOCKED(bp)) {
4617 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4623 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4629 db_printf("usage: show vnodebufs <addr>\n");
4632 vp = (struct vnode *)addr;
4633 db_printf("Clean buffers:\n");
4634 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4635 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4638 db_printf("Dirty buffers:\n");
4639 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4640 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4645 DB_COMMAND(countfreebufs, db_coundfreebufs)
4648 int i, used = 0, nfree = 0;
4651 db_printf("usage: countfreebufs\n");
4655 for (i = 0; i < nbuf; i++) {
4657 if ((bp->b_flags & B_INFREECNT) != 0)
4663 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4665 db_printf("numfreebuffers is %d\n", numfreebuffers);