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 static struct proc *bufdaemonproc;
103 static int inmem(struct vnode *vp, daddr_t blkno);
104 static void vm_hold_free_pages(struct buf *bp, int newbsize);
105 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
107 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
108 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
110 static void vfs_clean_pages_dirty_buf(struct buf *bp);
111 static void vfs_setdirty_locked_object(struct buf *bp);
112 static void vfs_vmio_release(struct buf *bp);
113 static int vfs_bio_clcheck(struct vnode *vp, int size,
114 daddr_t lblkno, daddr_t blkno);
115 static int buf_flush(int);
116 static int flushbufqueues(int, int);
117 static void buf_daemon(void);
118 static void bremfreel(struct buf *bp);
119 static __inline void bd_wakeup(void);
120 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS);
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_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
171 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L",
172 "Minimum preferred space used for in-progress I/O");
173 static long hirunningspace;
174 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE |
175 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L",
176 "Maximum amount of space to use for in-progress I/O");
177 int dirtybufferflushes;
178 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
179 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
181 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
182 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
183 int altbufferflushes;
184 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
185 0, "Number of fsync flushes to limit dirty buffers");
186 static int recursiveflushes;
187 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
188 0, "Number of flushes skipped due to being recursive");
189 static int numdirtybuffers;
190 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
191 "Number of buffers that are dirty (has unwritten changes) at the moment");
192 static int lodirtybuffers;
193 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
194 "How many buffers we want to have free before bufdaemon can sleep");
195 static int hidirtybuffers;
196 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
197 "When the number of dirty buffers is considered severe");
199 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
200 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
201 static int numfreebuffers;
202 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
203 "Number of free buffers");
204 static int lofreebuffers;
205 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
207 static int hifreebuffers;
208 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
209 "XXX Complicatedly unused");
210 static int getnewbufcalls;
211 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
212 "Number of calls to getnewbuf");
213 static int getnewbufrestarts;
214 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
215 "Number of times getnewbuf has had to restart a buffer aquisition");
216 static int mappingrestarts;
217 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0,
218 "Number of times getblk has had to restart a buffer mapping for "
220 static int flushbufqtarget = 100;
221 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
222 "Amount of work to do in flushbufqueues when helping bufdaemon");
223 static long notbufdflushes;
224 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0,
225 "Number of dirty buffer flushes done by the bufdaemon helpers");
226 static long barrierwrites;
227 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0,
228 "Number of barrier writes");
229 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD,
230 &unmapped_buf_allowed, 0,
231 "Permit the use of the unmapped i/o");
234 * Lock for the non-dirty bufqueues
236 static struct mtx_padalign bqclean;
239 * Lock for the dirty queue.
241 static struct mtx_padalign bqdirty;
244 * This lock synchronizes access to bd_request.
246 static struct mtx_padalign bdlock;
249 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
250 * waitrunningbufspace().
252 static struct mtx_padalign rbreqlock;
255 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
257 static struct rwlock_padalign nblock;
260 * Lock that protects bdirtywait.
262 static struct mtx_padalign bdirtylock;
265 * Wakeup point for bufdaemon, as well as indicator of whether it is already
266 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
269 static int bd_request;
272 * Request for the buf daemon to write more buffers than is indicated by
273 * lodirtybuf. This may be necessary to push out excess dependencies or
274 * defragment the address space where a simple count of the number of dirty
275 * buffers is insufficient to characterize the demand for flushing them.
277 static int bd_speedupreq;
280 * bogus page -- for I/O to/from partially complete buffers
281 * this is a temporary solution to the problem, but it is not
282 * really that bad. it would be better to split the buffer
283 * for input in the case of buffers partially already in memory,
284 * but the code is intricate enough already.
286 vm_page_t bogus_page;
289 * Synchronization (sleep/wakeup) variable for active buffer space requests.
290 * Set when wait starts, cleared prior to wakeup().
291 * Used in runningbufwakeup() and waitrunningbufspace().
293 static int runningbufreq;
296 * Synchronization (sleep/wakeup) variable for buffer requests.
297 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
299 * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(),
300 * getnewbuf(), and getblk().
302 static volatile int needsbuffer;
305 * Synchronization for bwillwrite() waiters.
307 static int bdirtywait;
310 * Definitions for the buffer free lists.
312 #define BUFFER_QUEUES 5 /* number of free buffer queues */
314 #define QUEUE_NONE 0 /* on no queue */
315 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
316 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
317 #define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */
318 #define QUEUE_EMPTY 4 /* empty buffer headers */
319 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */
321 /* Queues for free buffers with various properties */
322 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
324 static int bq_len[BUFFER_QUEUES];
328 * Single global constant for BUF_WMESG, to avoid getting multiple references.
329 * buf_wmesg is referred from macros.
331 const char *buf_wmesg = BUF_WMESG;
333 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
334 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
335 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
338 sysctl_runningspace(SYSCTL_HANDLER_ARGS)
343 value = *(long *)arg1;
344 error = sysctl_handle_long(oidp, &value, 0, req);
345 if (error != 0 || req->newptr == NULL)
347 mtx_lock(&rbreqlock);
348 if (arg1 == &hirunningspace) {
349 if (value < lorunningspace)
352 hirunningspace = value;
354 KASSERT(arg1 == &lorunningspace,
355 ("%s: unknown arg1", __func__));
356 if (value > hirunningspace)
359 lorunningspace = value;
361 mtx_unlock(&rbreqlock);
365 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
366 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
368 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
373 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
374 return (sysctl_handle_long(oidp, arg1, arg2, req));
375 lvalue = *(long *)arg1;
376 if (lvalue > INT_MAX)
377 /* On overflow, still write out a long to trigger ENOMEM. */
378 return (sysctl_handle_long(oidp, &lvalue, 0, req));
380 return (sysctl_handle_int(oidp, &ivalue, 0, req));
387 * Return the appropriate queue lock based on the index.
389 static inline struct mtx *
393 if (qindex == QUEUE_DIRTY)
394 return (struct mtx *)(&bqdirty);
395 return (struct mtx *)(&bqclean);
401 * Wakeup any bwillwrite() waiters.
406 mtx_lock(&bdirtylock);
411 mtx_unlock(&bdirtylock);
417 * Decrement the numdirtybuffers count by one and wakeup any
418 * threads blocked in bwillwrite().
424 if (atomic_fetchadd_int(&numdirtybuffers, -1) ==
425 (lodirtybuffers + hidirtybuffers) / 2)
432 * Increment the numdirtybuffers count by one and wakeup the buf
440 * Only do the wakeup once as we cross the boundary. The
441 * buf daemon will keep running until the condition clears.
443 if (atomic_fetchadd_int(&numdirtybuffers, 1) ==
444 (lodirtybuffers + hidirtybuffers) / 2)
451 * Called when buffer space is potentially available for recovery.
452 * getnewbuf() will block on this flag when it is unable to free
453 * sufficient buffer space. Buffer space becomes recoverable when
454 * bp's get placed back in the queues.
463 * If someone is waiting for BUF space, wake them up. Even
464 * though we haven't freed the kva space yet, the waiting
465 * process will be able to now.
471 if ((on & VFS_BIO_NEED_BUFSPACE) == 0)
474 if (atomic_cmpset_rel_int(&needsbuffer, on,
475 on & ~VFS_BIO_NEED_BUFSPACE))
479 wakeup(__DEVOLATILE(void *, &needsbuffer));
486 * Wake up processes that are waiting on asynchronous writes to fall
487 * below lorunningspace.
493 mtx_lock(&rbreqlock);
496 wakeup(&runningbufreq);
498 mtx_unlock(&rbreqlock);
504 * Decrement the outstanding write count according.
507 runningbufwakeup(struct buf *bp)
511 bspace = bp->b_runningbufspace;
514 space = atomic_fetchadd_long(&runningbufspace, -bspace);
515 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld",
517 bp->b_runningbufspace = 0;
519 * Only acquire the lock and wakeup on the transition from exceeding
520 * the threshold to falling below it.
522 if (space < lorunningspace)
524 if (space - bspace > lorunningspace)
532 * Called when a buffer has been added to one of the free queues to
533 * account for the buffer and to wakeup anyone waiting for free buffers.
534 * This typically occurs when large amounts of metadata are being handled
535 * by the buffer cache ( else buffer space runs out first, usually ).
538 bufcountadd(struct buf *bp)
540 int mask, need_wakeup, old, on;
542 KASSERT((bp->b_flags & B_INFREECNT) == 0,
543 ("buf %p already counted as free", bp));
544 bp->b_flags |= B_INFREECNT;
545 old = atomic_fetchadd_int(&numfreebuffers, 1);
546 KASSERT(old >= 0 && old < nbuf,
547 ("numfreebuffers climbed to %d", old + 1));
548 mask = VFS_BIO_NEED_ANY;
549 if (numfreebuffers >= hifreebuffers)
550 mask |= VFS_BIO_NEED_FREE;
558 if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask))
562 wakeup(__DEVOLATILE(void *, &needsbuffer));
569 * Decrement the numfreebuffers count as needed.
572 bufcountsub(struct buf *bp)
577 * Fixup numfreebuffers count. If the buffer is invalid or not
578 * delayed-write, the buffer was free and we must decrement
581 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
582 KASSERT((bp->b_flags & B_INFREECNT) != 0,
583 ("buf %p not counted in numfreebuffers", bp));
584 bp->b_flags &= ~B_INFREECNT;
585 old = atomic_fetchadd_int(&numfreebuffers, -1);
586 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
591 * waitrunningbufspace()
593 * runningbufspace is a measure of the amount of I/O currently
594 * running. This routine is used in async-write situations to
595 * prevent creating huge backups of pending writes to a device.
596 * Only asynchronous writes are governed by this function.
598 * This does NOT turn an async write into a sync write. It waits
599 * for earlier writes to complete and generally returns before the
600 * caller's write has reached the device.
603 waitrunningbufspace(void)
606 mtx_lock(&rbreqlock);
607 while (runningbufspace > hirunningspace) {
609 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
611 mtx_unlock(&rbreqlock);
616 * vfs_buf_test_cache:
618 * Called when a buffer is extended. This function clears the B_CACHE
619 * bit if the newly extended portion of the buffer does not contain
624 vfs_buf_test_cache(struct buf *bp,
625 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
629 VM_OBJECT_ASSERT_LOCKED(m->object);
630 if (bp->b_flags & B_CACHE) {
631 int base = (foff + off) & PAGE_MASK;
632 if (vm_page_is_valid(m, base, size) == 0)
633 bp->b_flags &= ~B_CACHE;
637 /* Wake up the buffer daemon if necessary */
643 if (bd_request == 0) {
651 * bd_speedup - speedup the buffer cache flushing code
660 if (bd_speedupreq == 0 || bd_request == 0)
670 #define TRANSIENT_DENOM 5
672 #define TRANSIENT_DENOM 10
676 * Calculating buffer cache scaling values and reserve space for buffer
677 * headers. This is called during low level kernel initialization and
678 * may be called more then once. We CANNOT write to the memory area
679 * being reserved at this time.
682 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
685 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz;
688 * physmem_est is in pages. Convert it to kilobytes (assumes
689 * PAGE_SIZE is >= 1K)
691 physmem_est = physmem_est * (PAGE_SIZE / 1024);
694 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
695 * For the first 64MB of ram nominally allocate sufficient buffers to
696 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
697 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
698 * the buffer cache we limit the eventual kva reservation to
701 * factor represents the 1/4 x ram conversion.
704 int factor = 4 * BKVASIZE / 1024;
707 if (physmem_est > 4096)
708 nbuf += min((physmem_est - 4096) / factor,
710 if (physmem_est > 65536)
711 nbuf += min((physmem_est - 65536) * 2 / (factor * 5),
712 32 * 1024 * 1024 / (factor * 5));
714 if (maxbcache && nbuf > maxbcache / BKVASIZE)
715 nbuf = maxbcache / BKVASIZE;
720 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
721 maxbuf = (LONG_MAX / 3) / BKVASIZE;
724 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
730 * Ideal allocation size for the transient bio submap is 10%
731 * of the maximal space buffer map. This roughly corresponds
732 * to the amount of the buffer mapped for typical UFS load.
734 * Clip the buffer map to reserve space for the transient
735 * BIOs, if its extent is bigger than 90% (80% on i386) of the
736 * maximum buffer map extent on the platform.
738 * The fall-back to the maxbuf in case of maxbcache unset,
739 * allows to not trim the buffer KVA for the architectures
740 * with ample KVA space.
742 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) {
743 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE;
744 buf_sz = (long)nbuf * BKVASIZE;
745 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM *
746 (TRANSIENT_DENOM - 1)) {
748 * There is more KVA than memory. Do not
749 * adjust buffer map size, and assign the rest
750 * of maxbuf to transient map.
752 biotmap_sz = maxbuf_sz - buf_sz;
755 * Buffer map spans all KVA we could afford on
756 * this platform. Give 10% (20% on i386) of
757 * the buffer map to the transient bio map.
759 biotmap_sz = buf_sz / TRANSIENT_DENOM;
760 buf_sz -= biotmap_sz;
762 if (biotmap_sz / INT_MAX > MAXPHYS)
763 bio_transient_maxcnt = INT_MAX;
765 bio_transient_maxcnt = biotmap_sz / MAXPHYS;
767 * Artifically limit to 1024 simultaneous in-flight I/Os
768 * using the transient mapping.
770 if (bio_transient_maxcnt > 1024)
771 bio_transient_maxcnt = 1024;
773 nbuf = buf_sz / BKVASIZE;
777 * swbufs are used as temporary holders for I/O, such as paging I/O.
778 * We have no less then 16 and no more then 256.
780 nswbuf = max(min(nbuf/4, 256), 16);
782 if (nswbuf < NSWBUF_MIN)
787 * Reserve space for the buffer cache buffers
790 v = (caddr_t)(swbuf + nswbuf);
792 v = (caddr_t)(buf + nbuf);
797 /* Initialize the buffer subsystem. Called before use of any buffers. */
804 mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF);
805 mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF);
806 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
807 rw_init(&nblock, "needsbuffer lock");
808 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
809 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF);
811 /* next, make a null set of free lists */
812 for (i = 0; i < BUFFER_QUEUES; i++)
813 TAILQ_INIT(&bufqueues[i]);
815 /* finally, initialize each buffer header and stick on empty q */
816 for (i = 0; i < nbuf; i++) {
818 bzero(bp, sizeof *bp);
819 bp->b_flags = B_INVAL | B_INFREECNT;
820 bp->b_rcred = NOCRED;
821 bp->b_wcred = NOCRED;
822 bp->b_qindex = QUEUE_EMPTY;
824 LIST_INIT(&bp->b_dep);
826 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
828 bq_len[QUEUE_EMPTY]++;
833 * maxbufspace is the absolute maximum amount of buffer space we are
834 * allowed to reserve in KVM and in real terms. The absolute maximum
835 * is nominally used by buf_daemon. hibufspace is the nominal maximum
836 * used by most other processes. The differential is required to
837 * ensure that buf_daemon is able to run when other processes might
838 * be blocked waiting for buffer space.
840 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
841 * this may result in KVM fragmentation which is not handled optimally
844 maxbufspace = (long)nbuf * BKVASIZE;
845 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
846 lobufspace = hibufspace - MAXBSIZE;
849 * Note: The 16 MiB upper limit for hirunningspace was chosen
850 * arbitrarily and may need further tuning. It corresponds to
851 * 128 outstanding write IO requests (if IO size is 128 KiB),
852 * which fits with many RAID controllers' tagged queuing limits.
853 * The lower 1 MiB limit is the historical upper limit for
856 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
857 16 * 1024 * 1024), 1024 * 1024);
858 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
861 * Limit the amount of malloc memory since it is wired permanently into
862 * the kernel space. Even though this is accounted for in the buffer
863 * allocation, we don't want the malloced region to grow uncontrolled.
864 * The malloc scheme improves memory utilization significantly on average
865 * (small) directories.
867 maxbufmallocspace = hibufspace / 20;
870 * Reduce the chance of a deadlock occuring by limiting the number
871 * of delayed-write dirty buffers we allow to stack up.
873 hidirtybuffers = nbuf / 4 + 20;
874 dirtybufthresh = hidirtybuffers * 9 / 10;
877 * To support extreme low-memory systems, make sure hidirtybuffers cannot
878 * eat up all available buffer space. This occurs when our minimum cannot
879 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
880 * BKVASIZE'd buffers.
882 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
883 hidirtybuffers >>= 1;
885 lodirtybuffers = hidirtybuffers / 2;
888 * Try to keep the number of free buffers in the specified range,
889 * and give special processes (e.g. like buf_daemon) access to an
892 lofreebuffers = nbuf / 18 + 5;
893 hifreebuffers = 2 * lofreebuffers;
894 numfreebuffers = nbuf;
896 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
897 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
898 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS);
903 vfs_buf_check_mapped(struct buf *bp)
906 KASSERT((bp->b_flags & B_UNMAPPED) == 0,
907 ("mapped buf %p %x", bp, bp->b_flags));
908 KASSERT(bp->b_kvabase != unmapped_buf,
909 ("mapped buf: b_kvabase was not updated %p", bp));
910 KASSERT(bp->b_data != unmapped_buf,
911 ("mapped buf: b_data was not updated %p", bp));
915 vfs_buf_check_unmapped(struct buf *bp)
918 KASSERT((bp->b_flags & B_UNMAPPED) == B_UNMAPPED,
919 ("unmapped buf %p %x", bp, bp->b_flags));
920 KASSERT(bp->b_kvabase == unmapped_buf,
921 ("unmapped buf: corrupted b_kvabase %p", bp));
922 KASSERT(bp->b_data == unmapped_buf,
923 ("unmapped buf: corrupted b_data %p", bp));
926 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp)
927 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp)
929 #define BUF_CHECK_MAPPED(bp) do {} while (0)
930 #define BUF_CHECK_UNMAPPED(bp) do {} while (0)
934 bpmap_qenter(struct buf *bp)
937 BUF_CHECK_MAPPED(bp);
940 * bp->b_data is relative to bp->b_offset, but
941 * bp->b_offset may be offset into the first page.
943 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
944 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages);
945 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
946 (vm_offset_t)(bp->b_offset & PAGE_MASK));
950 * bfreekva() - free the kva allocation for a buffer.
952 * Since this call frees up buffer space, we call bufspacewakeup().
955 bfreekva(struct buf *bp)
958 if (bp->b_kvasize == 0)
961 atomic_add_int(&buffreekvacnt, 1);
962 atomic_subtract_long(&bufspace, bp->b_kvasize);
963 if ((bp->b_flags & B_UNMAPPED) == 0) {
964 BUF_CHECK_MAPPED(bp);
965 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase,
968 BUF_CHECK_UNMAPPED(bp);
969 if ((bp->b_flags & B_KVAALLOC) != 0) {
970 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvaalloc,
973 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
974 bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
983 * Insert the buffer into the appropriate free list.
986 binsfree(struct buf *bp, int qindex)
988 struct mtx *olock, *nlock;
990 BUF_ASSERT_XLOCKED(bp);
992 olock = bqlock(bp->b_qindex);
993 nlock = bqlock(qindex);
995 /* Handle delayed bremfree() processing. */
996 if (bp->b_flags & B_REMFREE)
999 if (bp->b_qindex != QUEUE_NONE)
1000 panic("binsfree: free buffer onto another queue???");
1002 bp->b_qindex = qindex;
1003 if (olock != nlock) {
1007 if (bp->b_flags & B_AGE)
1008 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1010 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1012 bq_len[bp->b_qindex]++;
1017 * Something we can maybe free or reuse.
1019 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1022 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1029 * Mark the buffer for removal from the appropriate free list.
1033 bremfree(struct buf *bp)
1036 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1037 KASSERT((bp->b_flags & B_REMFREE) == 0,
1038 ("bremfree: buffer %p already marked for delayed removal.", bp));
1039 KASSERT(bp->b_qindex != QUEUE_NONE,
1040 ("bremfree: buffer %p not on a queue.", bp));
1041 BUF_ASSERT_XLOCKED(bp);
1043 bp->b_flags |= B_REMFREE;
1050 * Force an immediate removal from a free list. Used only in nfs when
1051 * it abuses the b_freelist pointer.
1054 bremfreef(struct buf *bp)
1058 qlock = bqlock(bp->b_qindex);
1067 * Removes a buffer from the free list, must be called with the
1068 * correct qlock held.
1071 bremfreel(struct buf *bp)
1074 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
1075 bp, bp->b_vp, bp->b_flags);
1076 KASSERT(bp->b_qindex != QUEUE_NONE,
1077 ("bremfreel: buffer %p not on a queue.", bp));
1078 BUF_ASSERT_XLOCKED(bp);
1079 mtx_assert(bqlock(bp->b_qindex), MA_OWNED);
1081 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
1083 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow",
1085 bq_len[bp->b_qindex]--;
1087 bp->b_qindex = QUEUE_NONE;
1089 * If this was a delayed bremfree() we only need to remove the buffer
1090 * from the queue and return the stats are already done.
1092 if (bp->b_flags & B_REMFREE) {
1093 bp->b_flags &= ~B_REMFREE;
1100 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
1101 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
1102 * the buffer is valid and we do not have to do anything.
1105 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
1106 int cnt, struct ucred * cred)
1111 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
1112 if (inmem(vp, *rablkno))
1114 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
1116 if ((rabp->b_flags & B_CACHE) == 0) {
1117 if (!TD_IS_IDLETHREAD(curthread))
1118 curthread->td_ru.ru_inblock++;
1119 rabp->b_flags |= B_ASYNC;
1120 rabp->b_flags &= ~B_INVAL;
1121 rabp->b_ioflags &= ~BIO_ERROR;
1122 rabp->b_iocmd = BIO_READ;
1123 if (rabp->b_rcred == NOCRED && cred != NOCRED)
1124 rabp->b_rcred = crhold(cred);
1125 vfs_busy_pages(rabp, 0);
1127 rabp->b_iooffset = dbtob(rabp->b_blkno);
1136 * Entry point for bread() and breadn() via #defines in sys/buf.h.
1138 * Get a buffer with the specified data. Look in the cache first. We
1139 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
1140 * is set, the buffer is valid and we do not have to do anything, see
1141 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
1144 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno,
1145 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp)
1148 int rv = 0, readwait = 0;
1150 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
1152 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
1154 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
1158 /* if not found in cache, do some I/O */
1159 if ((bp->b_flags & B_CACHE) == 0) {
1160 if (!TD_IS_IDLETHREAD(curthread))
1161 curthread->td_ru.ru_inblock++;
1162 bp->b_iocmd = BIO_READ;
1163 bp->b_flags &= ~B_INVAL;
1164 bp->b_ioflags &= ~BIO_ERROR;
1165 if (bp->b_rcred == NOCRED && cred != NOCRED)
1166 bp->b_rcred = crhold(cred);
1167 vfs_busy_pages(bp, 0);
1168 bp->b_iooffset = dbtob(bp->b_blkno);
1173 breada(vp, rablkno, rabsize, cnt, cred);
1182 * Write, release buffer on completion. (Done by iodone
1183 * if async). Do not bother writing anything if the buffer
1186 * Note that we set B_CACHE here, indicating that buffer is
1187 * fully valid and thus cacheable. This is true even of NFS
1188 * now so we set it generally. This could be set either here
1189 * or in biodone() since the I/O is synchronous. We put it
1193 bufwrite(struct buf *bp)
1200 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1201 if (bp->b_flags & B_INVAL) {
1206 if (bp->b_flags & B_BARRIER)
1209 oldflags = bp->b_flags;
1211 BUF_ASSERT_HELD(bp);
1213 if (bp->b_pin_count > 0)
1216 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
1217 ("FFS background buffer should not get here %p", bp));
1221 vp_md = vp->v_vflag & VV_MD;
1226 * Mark the buffer clean. Increment the bufobj write count
1227 * before bundirty() call, to prevent other thread from seeing
1228 * empty dirty list and zero counter for writes in progress,
1229 * falsely indicating that the bufobj is clean.
1231 bufobj_wref(bp->b_bufobj);
1234 bp->b_flags &= ~B_DONE;
1235 bp->b_ioflags &= ~BIO_ERROR;
1236 bp->b_flags |= B_CACHE;
1237 bp->b_iocmd = BIO_WRITE;
1239 vfs_busy_pages(bp, 1);
1242 * Normal bwrites pipeline writes
1244 bp->b_runningbufspace = bp->b_bufsize;
1245 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);
1247 if (!TD_IS_IDLETHREAD(curthread))
1248 curthread->td_ru.ru_oublock++;
1249 if (oldflags & B_ASYNC)
1251 bp->b_iooffset = dbtob(bp->b_blkno);
1254 if ((oldflags & B_ASYNC) == 0) {
1255 int rtval = bufwait(bp);
1258 } else if (space > hirunningspace) {
1260 * don't allow the async write to saturate the I/O
1261 * system. We will not deadlock here because
1262 * we are blocking waiting for I/O that is already in-progress
1263 * to complete. We do not block here if it is the update
1264 * or syncer daemon trying to clean up as that can lead
1267 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
1268 waitrunningbufspace();
1275 bufbdflush(struct bufobj *bo, struct buf *bp)
1279 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
1280 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
1282 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
1285 * Try to find a buffer to flush.
1287 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
1288 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
1290 LK_EXCLUSIVE | LK_NOWAIT, NULL))
1293 panic("bdwrite: found ourselves");
1295 /* Don't countdeps with the bo lock held. */
1296 if (buf_countdeps(nbp, 0)) {
1301 if (nbp->b_flags & B_CLUSTEROK) {
1302 vfs_bio_awrite(nbp);
1307 dirtybufferflushes++;
1316 * Delayed write. (Buffer is marked dirty). Do not bother writing
1317 * anything if the buffer is marked invalid.
1319 * Note that since the buffer must be completely valid, we can safely
1320 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1321 * biodone() in order to prevent getblk from writing the buffer
1322 * out synchronously.
1325 bdwrite(struct buf *bp)
1327 struct thread *td = curthread;
1331 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1332 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1333 KASSERT((bp->b_flags & B_BARRIER) == 0,
1334 ("Barrier request in delayed write %p", bp));
1335 BUF_ASSERT_HELD(bp);
1337 if (bp->b_flags & B_INVAL) {
1343 * If we have too many dirty buffers, don't create any more.
1344 * If we are wildly over our limit, then force a complete
1345 * cleanup. Otherwise, just keep the situation from getting
1346 * out of control. Note that we have to avoid a recursive
1347 * disaster and not try to clean up after our own cleanup!
1351 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1352 td->td_pflags |= TDP_INBDFLUSH;
1354 td->td_pflags &= ~TDP_INBDFLUSH;
1360 * Set B_CACHE, indicating that the buffer is fully valid. This is
1361 * true even of NFS now.
1363 bp->b_flags |= B_CACHE;
1366 * This bmap keeps the system from needing to do the bmap later,
1367 * perhaps when the system is attempting to do a sync. Since it
1368 * is likely that the indirect block -- or whatever other datastructure
1369 * that the filesystem needs is still in memory now, it is a good
1370 * thing to do this. Note also, that if the pageout daemon is
1371 * requesting a sync -- there might not be enough memory to do
1372 * the bmap then... So, this is important to do.
1374 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1375 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1379 * Set the *dirty* buffer range based upon the VM system dirty
1382 * Mark the buffer pages as clean. We need to do this here to
1383 * satisfy the vnode_pager and the pageout daemon, so that it
1384 * thinks that the pages have been "cleaned". Note that since
1385 * the pages are in a delayed write buffer -- the VFS layer
1386 * "will" see that the pages get written out on the next sync,
1387 * or perhaps the cluster will be completed.
1389 vfs_clean_pages_dirty_buf(bp);
1393 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1394 * due to the softdep code.
1401 * Turn buffer into delayed write request. We must clear BIO_READ and
1402 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1403 * itself to properly update it in the dirty/clean lists. We mark it
1404 * B_DONE to ensure that any asynchronization of the buffer properly
1405 * clears B_DONE ( else a panic will occur later ).
1407 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1408 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1409 * should only be called if the buffer is known-good.
1411 * Since the buffer is not on a queue, we do not update the numfreebuffers
1414 * The buffer must be on QUEUE_NONE.
1417 bdirty(struct buf *bp)
1420 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1421 bp, bp->b_vp, bp->b_flags);
1422 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1423 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1424 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1425 BUF_ASSERT_HELD(bp);
1426 bp->b_flags &= ~(B_RELBUF);
1427 bp->b_iocmd = BIO_WRITE;
1429 if ((bp->b_flags & B_DELWRI) == 0) {
1430 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1439 * Clear B_DELWRI for buffer.
1441 * Since the buffer is not on a queue, we do not update the numfreebuffers
1444 * The buffer must be on QUEUE_NONE.
1448 bundirty(struct buf *bp)
1451 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1452 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1453 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1454 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1455 BUF_ASSERT_HELD(bp);
1457 if (bp->b_flags & B_DELWRI) {
1458 bp->b_flags &= ~B_DELWRI;
1463 * Since it is now being written, we can clear its deferred write flag.
1465 bp->b_flags &= ~B_DEFERRED;
1471 * Asynchronous write. Start output on a buffer, but do not wait for
1472 * it to complete. The buffer is released when the output completes.
1474 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1475 * B_INVAL buffers. Not us.
1478 bawrite(struct buf *bp)
1481 bp->b_flags |= B_ASYNC;
1488 * Asynchronous barrier write. Start output on a buffer, but do not
1489 * wait for it to complete. Place a write barrier after this write so
1490 * that this buffer and all buffers written before it are committed to
1491 * the disk before any buffers written after this write are committed
1492 * to the disk. The buffer is released when the output completes.
1495 babarrierwrite(struct buf *bp)
1498 bp->b_flags |= B_ASYNC | B_BARRIER;
1505 * Synchronous barrier write. Start output on a buffer and wait for
1506 * it to complete. Place a write barrier after this write so that
1507 * this buffer and all buffers written before it are committed to
1508 * the disk before any buffers written after this write are committed
1509 * to the disk. The buffer is released when the output completes.
1512 bbarrierwrite(struct buf *bp)
1515 bp->b_flags |= B_BARRIER;
1516 return (bwrite(bp));
1522 * Called prior to the locking of any vnodes when we are expecting to
1523 * write. We do not want to starve the buffer cache with too many
1524 * dirty buffers so we block here. By blocking prior to the locking
1525 * of any vnodes we attempt to avoid the situation where a locked vnode
1526 * prevents the various system daemons from flushing related buffers.
1532 if (numdirtybuffers >= hidirtybuffers) {
1533 mtx_lock(&bdirtylock);
1534 while (numdirtybuffers >= hidirtybuffers) {
1536 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4),
1539 mtx_unlock(&bdirtylock);
1544 * Return true if we have too many dirty buffers.
1547 buf_dirty_count_severe(void)
1550 return(numdirtybuffers >= hidirtybuffers);
1553 static __noinline int
1554 buf_vm_page_count_severe(void)
1557 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1559 return vm_page_count_severe();
1565 * Release a busy buffer and, if requested, free its resources. The
1566 * buffer will be stashed in the appropriate bufqueue[] allowing it
1567 * to be accessed later as a cache entity or reused for other purposes.
1570 brelse(struct buf *bp)
1574 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1575 bp, bp->b_vp, bp->b_flags);
1576 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1577 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1579 if (BUF_LOCKRECURSED(bp)) {
1581 * Do not process, in particular, do not handle the
1582 * B_INVAL/B_RELBUF and do not release to free list.
1588 if (bp->b_flags & B_MANAGED) {
1593 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1594 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1596 * Failed write, redirty. Must clear BIO_ERROR to prevent
1597 * pages from being scrapped. If the error is anything
1598 * other than an I/O error (EIO), assume that retrying
1601 bp->b_ioflags &= ~BIO_ERROR;
1603 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1604 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1606 * Either a failed I/O or we were asked to free or not
1609 bp->b_flags |= B_INVAL;
1610 if (!LIST_EMPTY(&bp->b_dep))
1612 if (bp->b_flags & B_DELWRI)
1614 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1615 if ((bp->b_flags & B_VMIO) == 0) {
1624 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1625 * is called with B_DELWRI set, the underlying pages may wind up
1626 * getting freed causing a previous write (bdwrite()) to get 'lost'
1627 * because pages associated with a B_DELWRI bp are marked clean.
1629 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1630 * if B_DELWRI is set.
1632 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1633 * on pages to return pages to the VM page queues.
1635 if (bp->b_flags & B_DELWRI)
1636 bp->b_flags &= ~B_RELBUF;
1637 else if (buf_vm_page_count_severe()) {
1639 * BKGRDINPROG can only be set with the buf and bufobj
1640 * locks both held. We tolerate a race to clear it here.
1642 if (!(bp->b_vflags & BV_BKGRDINPROG))
1643 bp->b_flags |= B_RELBUF;
1647 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1648 * constituted, not even NFS buffers now. Two flags effect this. If
1649 * B_INVAL, the struct buf is invalidated but the VM object is kept
1650 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1652 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1653 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1654 * buffer is also B_INVAL because it hits the re-dirtying code above.
1656 * Normally we can do this whether a buffer is B_DELWRI or not. If
1657 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1658 * the commit state and we cannot afford to lose the buffer. If the
1659 * buffer has a background write in progress, we need to keep it
1660 * around to prevent it from being reconstituted and starting a second
1663 if ((bp->b_flags & B_VMIO)
1664 && !(bp->b_vp->v_mount != NULL &&
1665 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1666 !vn_isdisk(bp->b_vp, NULL) &&
1667 (bp->b_flags & B_DELWRI))
1676 obj = bp->b_bufobj->bo_object;
1679 * Get the base offset and length of the buffer. Note that
1680 * in the VMIO case if the buffer block size is not
1681 * page-aligned then b_data pointer may not be page-aligned.
1682 * But our b_pages[] array *IS* page aligned.
1684 * block sizes less then DEV_BSIZE (usually 512) are not
1685 * supported due to the page granularity bits (m->valid,
1686 * m->dirty, etc...).
1688 * See man buf(9) for more information
1690 resid = bp->b_bufsize;
1691 foff = bp->b_offset;
1692 for (i = 0; i < bp->b_npages; i++) {
1698 * If we hit a bogus page, fixup *all* the bogus pages
1701 if (m == bogus_page) {
1702 poff = OFF_TO_IDX(bp->b_offset);
1705 VM_OBJECT_RLOCK(obj);
1706 for (j = i; j < bp->b_npages; j++) {
1708 mtmp = bp->b_pages[j];
1709 if (mtmp == bogus_page) {
1710 mtmp = vm_page_lookup(obj, poff + j);
1712 panic("brelse: page missing\n");
1714 bp->b_pages[j] = mtmp;
1717 VM_OBJECT_RUNLOCK(obj);
1719 if ((bp->b_flags & (B_INVAL | B_UNMAPPED)) == 0) {
1720 BUF_CHECK_MAPPED(bp);
1722 trunc_page((vm_offset_t)bp->b_data),
1723 bp->b_pages, bp->b_npages);
1727 if ((bp->b_flags & B_NOCACHE) ||
1728 (bp->b_ioflags & BIO_ERROR &&
1729 bp->b_iocmd == BIO_READ)) {
1730 int poffset = foff & PAGE_MASK;
1731 int presid = resid > (PAGE_SIZE - poffset) ?
1732 (PAGE_SIZE - poffset) : resid;
1734 KASSERT(presid >= 0, ("brelse: extra page"));
1735 VM_OBJECT_WLOCK(obj);
1736 while (vm_page_xbusied(m)) {
1738 VM_OBJECT_WUNLOCK(obj);
1739 vm_page_busy_sleep(m, "mbncsh");
1740 VM_OBJECT_WLOCK(obj);
1742 if (pmap_page_wired_mappings(m) == 0)
1743 vm_page_set_invalid(m, poffset, presid);
1744 VM_OBJECT_WUNLOCK(obj);
1746 printf("avoided corruption bug in bogus_page/brelse code\n");
1748 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1749 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1751 if (bp->b_flags & (B_INVAL | B_RELBUF))
1752 vfs_vmio_release(bp);
1754 } else if (bp->b_flags & B_VMIO) {
1756 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1757 vfs_vmio_release(bp);
1760 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1761 if (bp->b_bufsize != 0)
1763 if (bp->b_vp != NULL)
1768 * If the buffer has junk contents signal it and eventually
1769 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1772 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1773 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1774 bp->b_flags |= B_INVAL;
1775 if (bp->b_flags & B_INVAL) {
1776 if (bp->b_flags & B_DELWRI)
1782 /* buffers with no memory */
1783 if (bp->b_bufsize == 0) {
1784 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1785 if (bp->b_vflags & BV_BKGRDINPROG)
1786 panic("losing buffer 1");
1788 qindex = QUEUE_EMPTYKVA;
1790 qindex = QUEUE_EMPTY;
1791 bp->b_flags |= B_AGE;
1792 /* buffers with junk contents */
1793 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1794 (bp->b_ioflags & BIO_ERROR)) {
1795 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1796 if (bp->b_vflags & BV_BKGRDINPROG)
1797 panic("losing buffer 2");
1798 qindex = QUEUE_CLEAN;
1799 bp->b_flags |= B_AGE;
1800 /* remaining buffers */
1801 } else if (bp->b_flags & B_DELWRI)
1802 qindex = QUEUE_DIRTY;
1804 qindex = QUEUE_CLEAN;
1806 binsfree(bp, qindex);
1808 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1809 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1810 panic("brelse: not dirty");
1816 * Release a buffer back to the appropriate queue but do not try to free
1817 * it. The buffer is expected to be used again soon.
1819 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1820 * biodone() to requeue an async I/O on completion. It is also used when
1821 * known good buffers need to be requeued but we think we may need the data
1824 * XXX we should be able to leave the B_RELBUF hint set on completion.
1827 bqrelse(struct buf *bp)
1831 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1832 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1833 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1835 if (BUF_LOCKRECURSED(bp)) {
1836 /* do not release to free list */
1840 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1842 if (bp->b_flags & B_MANAGED) {
1843 if (bp->b_flags & B_REMFREE)
1848 /* buffers with stale but valid contents */
1849 if (bp->b_flags & B_DELWRI) {
1850 qindex = QUEUE_DIRTY;
1852 if ((bp->b_flags & B_DELWRI) == 0 &&
1853 (bp->b_xflags & BX_VNDIRTY))
1854 panic("bqrelse: not dirty");
1856 * BKGRDINPROG can only be set with the buf and bufobj
1857 * locks both held. We tolerate a race to clear it here.
1859 if (buf_vm_page_count_severe() &&
1860 (bp->b_vflags & BV_BKGRDINPROG) == 0) {
1862 * We are too low on memory, we have to try to free
1863 * the buffer (most importantly: the wired pages
1864 * making up its backing store) *now*.
1869 qindex = QUEUE_CLEAN;
1871 binsfree(bp, qindex);
1878 /* Give pages used by the bp back to the VM system (where possible) */
1880 vfs_vmio_release(struct buf *bp)
1885 if ((bp->b_flags & B_UNMAPPED) == 0) {
1886 BUF_CHECK_MAPPED(bp);
1887 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1889 BUF_CHECK_UNMAPPED(bp);
1890 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
1891 for (i = 0; i < bp->b_npages; i++) {
1893 bp->b_pages[i] = NULL;
1895 * In order to keep page LRU ordering consistent, put
1896 * everything on the inactive queue.
1899 vm_page_unwire(m, PQ_INACTIVE);
1902 * Might as well free the page if we can and it has
1903 * no valid data. We also free the page if the
1904 * buffer was used for direct I/O
1906 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1907 if (m->wire_count == 0 && !vm_page_busied(m))
1909 } else if (bp->b_flags & B_DIRECT)
1910 vm_page_try_to_free(m);
1911 else if (buf_vm_page_count_severe())
1912 vm_page_try_to_cache(m);
1915 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
1917 if (bp->b_bufsize) {
1922 bp->b_flags &= ~B_VMIO;
1928 * Check to see if a block at a particular lbn is available for a clustered
1932 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1939 /* If the buf isn't in core skip it */
1940 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1943 /* If the buf is busy we don't want to wait for it */
1944 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1947 /* Only cluster with valid clusterable delayed write buffers */
1948 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1949 (B_DELWRI | B_CLUSTEROK))
1952 if (bpa->b_bufsize != size)
1956 * Check to see if it is in the expected place on disk and that the
1957 * block has been mapped.
1959 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1969 * Implement clustered async writes for clearing out B_DELWRI buffers.
1970 * This is much better then the old way of writing only one buffer at
1971 * a time. Note that we may not be presented with the buffers in the
1972 * correct order, so we search for the cluster in both directions.
1975 vfs_bio_awrite(struct buf *bp)
1980 daddr_t lblkno = bp->b_lblkno;
1981 struct vnode *vp = bp->b_vp;
1989 gbflags = (bp->b_flags & B_UNMAPPED) != 0 ? GB_UNMAPPED : 0;
1991 * right now we support clustered writing only to regular files. If
1992 * we find a clusterable block we could be in the middle of a cluster
1993 * rather then at the beginning.
1995 if ((vp->v_type == VREG) &&
1996 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1997 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1999 size = vp->v_mount->mnt_stat.f_iosize;
2000 maxcl = MAXPHYS / size;
2003 for (i = 1; i < maxcl; i++)
2004 if (vfs_bio_clcheck(vp, size, lblkno + i,
2005 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
2008 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
2009 if (vfs_bio_clcheck(vp, size, lblkno - j,
2010 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
2016 * this is a possible cluster write
2020 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl,
2026 bp->b_flags |= B_ASYNC;
2028 * default (old) behavior, writing out only one block
2030 * XXX returns b_bufsize instead of b_bcount for nwritten?
2032 nwritten = bp->b_bufsize;
2039 setbufkva(struct buf *bp, vm_offset_t addr, int maxsize, int gbflags)
2042 KASSERT((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2043 bp->b_kvasize == 0, ("call bfreekva(%p)", bp));
2044 if ((gbflags & GB_UNMAPPED) == 0) {
2045 bp->b_kvabase = (caddr_t)addr;
2046 } else if ((gbflags & GB_KVAALLOC) != 0) {
2047 KASSERT((gbflags & GB_UNMAPPED) != 0,
2048 ("GB_KVAALLOC without GB_UNMAPPED"));
2049 bp->b_kvaalloc = (caddr_t)addr;
2050 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2051 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2053 bp->b_kvasize = maxsize;
2057 * Allocate the buffer KVA and set b_kvasize. Also set b_kvabase if
2061 allocbufkva(struct buf *bp, int maxsize, int gbflags)
2068 if (vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr)) {
2070 * Buffer map is too fragmented. Request the caller
2071 * to defragment the map.
2073 atomic_add_int(&bufdefragcnt, 1);
2076 setbufkva(bp, addr, maxsize, gbflags);
2077 atomic_add_long(&bufspace, bp->b_kvasize);
2082 * Ask the bufdaemon for help, or act as bufdaemon itself, when a
2083 * locked vnode is supplied.
2086 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo,
2091 int cnt, error, flags, norunbuf, wait;
2093 mtx_assert(&bqclean, MA_OWNED);
2096 flags = VFS_BIO_NEED_BUFSPACE;
2098 } else if (bufspace >= hibufspace) {
2100 flags = VFS_BIO_NEED_BUFSPACE;
2103 flags = VFS_BIO_NEED_ANY;
2105 atomic_set_int(&needsbuffer, flags);
2106 mtx_unlock(&bqclean);
2108 bd_speedup(); /* heeeelp */
2109 if ((gbflags & GB_NOWAIT_BD) != 0)
2116 while ((needsbuffer & flags) != 0) {
2117 if (vp != NULL && vp->v_type != VCHR &&
2118 (td->td_pflags & TDP_BUFNEED) == 0) {
2119 rw_wunlock(&nblock);
2121 * getblk() is called with a vnode locked, and
2122 * some majority of the dirty buffers may as
2123 * well belong to the vnode. Flushing the
2124 * buffers there would make a progress that
2125 * cannot be achieved by the buf_daemon, that
2126 * cannot lock the vnode.
2130 ASSERT_VOP_LOCKED(vp, "bufd_helper");
2131 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 :
2132 vn_lock(vp, LK_TRYUPGRADE);
2134 /* play bufdaemon */
2135 norunbuf = curthread_pflags_set(TDP_BUFNEED |
2137 VOP_FSYNC(vp, wait, td);
2138 atomic_add_long(¬bufdflushes, 1);
2139 curthread_pflags_restore(norunbuf);
2142 if ((needsbuffer & flags) == 0)
2145 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock,
2146 (PRIBIO + 4) | slpflag, waitmsg, slptimeo);
2150 rw_wunlock(&nblock);
2154 getnewbuf_reuse_bp(struct buf *bp, int qindex)
2157 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
2158 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
2159 bp->b_kvasize, bp->b_bufsize, qindex);
2160 mtx_assert(&bqclean, MA_NOTOWNED);
2163 * Note: we no longer distinguish between VMIO and non-VMIO
2166 KASSERT((bp->b_flags & B_DELWRI) == 0,
2167 ("delwri buffer %p found in queue %d", bp, qindex));
2169 if (qindex == QUEUE_CLEAN) {
2170 if (bp->b_flags & B_VMIO) {
2171 bp->b_flags &= ~B_ASYNC;
2172 vfs_vmio_release(bp);
2174 if (bp->b_vp != NULL)
2179 * Get the rest of the buffer freed up. b_kva* is still valid
2180 * after this operation.
2183 if (bp->b_rcred != NOCRED) {
2184 crfree(bp->b_rcred);
2185 bp->b_rcred = NOCRED;
2187 if (bp->b_wcred != NOCRED) {
2188 crfree(bp->b_wcred);
2189 bp->b_wcred = NOCRED;
2191 if (!LIST_EMPTY(&bp->b_dep))
2193 if (bp->b_vflags & BV_BKGRDINPROG)
2194 panic("losing buffer 3");
2195 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d",
2196 bp, bp->b_vp, qindex));
2197 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
2198 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags));
2203 bp->b_flags &= B_UNMAPPED | B_KVAALLOC;
2206 KASSERT((bp->b_flags & B_INFREECNT) == 0,
2207 ("buf %p still counted as free?", bp));
2210 bp->b_blkno = bp->b_lblkno = 0;
2211 bp->b_offset = NOOFFSET;
2217 bp->b_dirtyoff = bp->b_dirtyend = 0;
2218 bp->b_bufobj = NULL;
2219 bp->b_pin_count = 0;
2220 bp->b_fsprivate1 = NULL;
2221 bp->b_fsprivate2 = NULL;
2222 bp->b_fsprivate3 = NULL;
2224 LIST_INIT(&bp->b_dep);
2227 static int flushingbufs;
2230 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata)
2232 struct buf *bp, *nbp;
2233 int nqindex, qindex, pass;
2235 KASSERT(!unmapped || !defrag, ("both unmapped and defrag"));
2239 atomic_add_int(&getnewbufrestarts, 1);
2242 * Setup for scan. If we do not have enough free buffers,
2243 * we setup a degenerate case that immediately fails. Note
2244 * that if we are specially marked process, we are allowed to
2245 * dip into our reserves.
2247 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2248 * for the allocation of the mapped buffer. For unmapped, the
2249 * easiest is to start with EMPTY outright.
2251 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
2252 * However, there are a number of cases (defragging, reusing, ...)
2253 * where we cannot backup.
2257 if (!defrag && unmapped) {
2258 nqindex = QUEUE_EMPTY;
2259 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2262 nqindex = QUEUE_EMPTYKVA;
2263 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2267 * If no EMPTYKVA buffers and we are either defragging or
2268 * reusing, locate a CLEAN buffer to free or reuse. If
2269 * bufspace useage is low skip this step so we can allocate a
2272 if (nbp == NULL && (defrag || bufspace >= lobufspace)) {
2273 nqindex = QUEUE_CLEAN;
2274 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2278 * If we could not find or were not allowed to reuse a CLEAN
2279 * buffer, check to see if it is ok to use an EMPTY buffer.
2280 * We can only use an EMPTY buffer if allocating its KVA would
2281 * not otherwise run us out of buffer space. No KVA is needed
2282 * for the unmapped allocation.
2284 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace ||
2286 nqindex = QUEUE_EMPTY;
2287 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
2291 * All available buffers might be clean, retry ignoring the
2292 * lobufspace as the last resort.
2294 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) {
2295 nqindex = QUEUE_CLEAN;
2296 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2300 * Run scan, possibly freeing data and/or kva mappings on the fly
2303 while ((bp = nbp) != NULL) {
2307 * Calculate next bp (we can only use it if we do not
2308 * block or do other fancy things).
2310 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
2313 nqindex = QUEUE_EMPTYKVA;
2314 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
2318 case QUEUE_EMPTYKVA:
2319 nqindex = QUEUE_CLEAN;
2320 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
2325 if (metadata && pass == 1) {
2327 nqindex = QUEUE_EMPTY;
2329 &bufqueues[QUEUE_EMPTY]);
2338 * If we are defragging then we need a buffer with
2339 * b_kvasize != 0. XXX this situation should no longer
2340 * occur, if defrag is non-zero the buffer's b_kvasize
2341 * should also be non-zero at this point. XXX
2343 if (defrag && bp->b_kvasize == 0) {
2344 printf("Warning: defrag empty buffer %p\n", bp);
2349 * Start freeing the bp. This is somewhat involved. nbp
2350 * remains valid only for QUEUE_EMPTY[KVA] bp's.
2352 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2355 * BKGRDINPROG can only be set with the buf and bufobj
2356 * locks both held. We tolerate a race to clear it here.
2358 if (bp->b_vflags & BV_BKGRDINPROG) {
2363 KASSERT(bp->b_qindex == qindex,
2364 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2367 mtx_unlock(&bqclean);
2369 * NOTE: nbp is now entirely invalid. We can only restart
2370 * the scan from this point on.
2373 getnewbuf_reuse_bp(bp, qindex);
2374 mtx_assert(&bqclean, MA_NOTOWNED);
2377 * If we are defragging then free the buffer.
2380 bp->b_flags |= B_INVAL;
2388 * Notify any waiters for the buffer lock about
2389 * identity change by freeing the buffer.
2391 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2392 bp->b_flags |= B_INVAL;
2402 * If we are overcomitted then recover the buffer and its
2403 * KVM space. This occurs in rare situations when multiple
2404 * processes are blocked in getnewbuf() or allocbuf().
2406 if (bufspace >= hibufspace)
2408 if (flushingbufs && bp->b_kvasize != 0) {
2409 bp->b_flags |= B_INVAL;
2414 if (bufspace < lobufspace)
2424 * Find and initialize a new buffer header, freeing up existing buffers
2425 * in the bufqueues as necessary. The new buffer is returned locked.
2427 * Important: B_INVAL is not set. If the caller wishes to throw the
2428 * buffer away, the caller must set B_INVAL prior to calling brelse().
2431 * We have insufficient buffer headers
2432 * We have insufficient buffer space
2433 * buffer_arena is too fragmented ( space reservation fails )
2434 * If we have to flush dirty buffers ( but we try to avoid this )
2437 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
2441 int defrag, metadata;
2443 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
2444 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
2445 if (!unmapped_buf_allowed)
2446 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC);
2449 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 ||
2455 * We can't afford to block since we might be holding a vnode lock,
2456 * which may prevent system daemons from running. We deal with
2457 * low-memory situations by proactively returning memory and running
2458 * async I/O rather then sync I/O.
2460 atomic_add_int(&getnewbufcalls, 1);
2461 atomic_subtract_int(&getnewbufrestarts, 1);
2463 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED |
2464 GB_KVAALLOC)) == GB_UNMAPPED, metadata);
2469 * If we exhausted our list, sleep as appropriate. We may have to
2470 * wakeup various daemons and write out some dirty buffers.
2472 * Generally we are sleeping due to insufficient buffer space.
2475 mtx_assert(&bqclean, MA_OWNED);
2476 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag);
2477 mtx_assert(&bqclean, MA_NOTOWNED);
2478 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) {
2479 mtx_assert(&bqclean, MA_NOTOWNED);
2482 bp->b_flags |= B_UNMAPPED;
2483 bp->b_kvabase = bp->b_data = unmapped_buf;
2484 bp->b_kvasize = maxsize;
2485 atomic_add_long(&bufspace, bp->b_kvasize);
2486 atomic_add_long(&unmapped_bufspace, bp->b_kvasize);
2487 atomic_add_int(&bufreusecnt, 1);
2489 mtx_assert(&bqclean, MA_NOTOWNED);
2492 * We finally have a valid bp. We aren't quite out of the
2493 * woods, we still have to reserve kva space. In order
2494 * to keep fragmentation sane we only allocate kva in
2497 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2499 if (maxsize != bp->b_kvasize || (bp->b_flags & (B_UNMAPPED |
2500 B_KVAALLOC)) == B_UNMAPPED) {
2501 if (allocbufkva(bp, maxsize, gbflags)) {
2503 bp->b_flags |= B_INVAL;
2507 atomic_add_int(&bufreusecnt, 1);
2508 } else if ((bp->b_flags & B_KVAALLOC) != 0 &&
2509 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == 0) {
2511 * If the reused buffer has KVA allocated,
2512 * reassign b_kvaalloc to b_kvabase.
2514 bp->b_kvabase = bp->b_kvaalloc;
2515 bp->b_flags &= ~B_KVAALLOC;
2516 atomic_subtract_long(&unmapped_bufspace,
2518 atomic_add_int(&bufreusecnt, 1);
2519 } else if ((bp->b_flags & (B_UNMAPPED | B_KVAALLOC)) == 0 &&
2520 (gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == (GB_UNMAPPED |
2523 * The case of reused buffer already have KVA
2524 * mapped, but the request is for unmapped
2525 * buffer with KVA allocated.
2527 bp->b_kvaalloc = bp->b_kvabase;
2528 bp->b_data = bp->b_kvabase = unmapped_buf;
2529 bp->b_flags |= B_UNMAPPED | B_KVAALLOC;
2530 atomic_add_long(&unmapped_bufspace,
2532 atomic_add_int(&bufreusecnt, 1);
2534 if ((gbflags & GB_UNMAPPED) == 0) {
2535 bp->b_saveaddr = bp->b_kvabase;
2536 bp->b_data = bp->b_saveaddr;
2537 bp->b_flags &= ~B_UNMAPPED;
2538 BUF_CHECK_MAPPED(bp);
2547 * buffer flushing daemon. Buffers are normally flushed by the
2548 * update daemon but if it cannot keep up this process starts to
2549 * take the load in an attempt to prevent getnewbuf() from blocking.
2552 static struct kproc_desc buf_kp = {
2557 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2560 buf_flush(int target)
2564 flushed = flushbufqueues(target, 0);
2567 * Could not find any buffers without rollback
2568 * dependencies, so just write the first one
2569 * in the hopes of eventually making progress.
2571 flushed = flushbufqueues(target, 1);
2582 * This process needs to be suspended prior to shutdown sync.
2584 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2588 * This process is allowed to take the buffer cache to the limit
2590 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2594 mtx_unlock(&bdlock);
2596 kproc_suspend_check(bufdaemonproc);
2597 lodirty = lodirtybuffers;
2598 if (bd_speedupreq) {
2599 lodirty = numdirtybuffers / 2;
2603 * Do the flush. Limit the amount of in-transit I/O we
2604 * allow to build up, otherwise we would completely saturate
2607 while (numdirtybuffers > lodirty) {
2608 if (buf_flush(numdirtybuffers - lodirty) == 0)
2610 kern_yield(PRI_USER);
2614 * Only clear bd_request if we have reached our low water
2615 * mark. The buf_daemon normally waits 1 second and
2616 * then incrementally flushes any dirty buffers that have
2617 * built up, within reason.
2619 * If we were unable to hit our low water mark and couldn't
2620 * find any flushable buffers, we sleep for a short period
2621 * to avoid endless loops on unlockable buffers.
2624 if (numdirtybuffers <= lodirtybuffers) {
2626 * We reached our low water mark, reset the
2627 * request and sleep until we are needed again.
2628 * The sleep is just so the suspend code works.
2632 * Do an extra wakeup in case dirty threshold
2633 * changed via sysctl and the explicit transition
2634 * out of shortfall was missed.
2637 if (runningbufspace <= lorunningspace)
2639 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2642 * We couldn't find any flushable dirty buffers but
2643 * still have too many dirty buffers, we
2644 * have to sleep and try again. (rare)
2646 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2654 * Try to flush a buffer in the dirty queue. We must be careful to
2655 * free up B_INVAL buffers instead of write them, which NFS is
2656 * particularly sensitive to.
2658 static int flushwithdeps = 0;
2659 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2660 0, "Number of buffers flushed with dependecies that require rollbacks");
2663 flushbufqueues(int target, int flushdeps)
2665 struct buf *sentinel;
2675 queue = QUEUE_DIRTY;
2677 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2678 sentinel->b_qindex = QUEUE_SENTINEL;
2680 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2681 mtx_unlock(&bqdirty);
2682 while (flushed != target) {
2685 bp = TAILQ_NEXT(sentinel, b_freelist);
2687 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2688 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2691 mtx_unlock(&bqdirty);
2694 KASSERT(bp->b_qindex != QUEUE_SENTINEL,
2695 ("parallel calls to flushbufqueues() bp %p", bp));
2696 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL);
2697 mtx_unlock(&bqdirty);
2700 if (bp->b_pin_count > 0) {
2705 * BKGRDINPROG can only be set with the buf and bufobj
2706 * locks both held. We tolerate a race to clear it here.
2708 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2709 (bp->b_flags & B_DELWRI) == 0) {
2713 if (bp->b_flags & B_INVAL) {
2720 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2721 if (flushdeps == 0) {
2729 * We must hold the lock on a vnode before writing
2730 * one of its buffers. Otherwise we may confuse, or
2731 * in the case of a snapshot vnode, deadlock the
2734 * The lock order here is the reverse of the normal
2735 * of vnode followed by buf lock. This is ok because
2736 * the NOWAIT will prevent deadlock.
2739 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2743 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT);
2745 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2746 bp, bp->b_vp, bp->b_flags);
2748 vn_finished_write(mp);
2750 flushwithdeps += hasdeps;
2752 if (runningbufspace > hirunningspace)
2753 waitrunningbufspace();
2756 vn_finished_write(mp);
2760 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2761 mtx_unlock(&bqdirty);
2762 free(sentinel, M_TEMP);
2767 * Check to see if a block is currently memory resident.
2770 incore(struct bufobj *bo, daddr_t blkno)
2775 bp = gbincore(bo, blkno);
2781 * Returns true if no I/O is needed to access the
2782 * associated VM object. This is like incore except
2783 * it also hunts around in the VM system for the data.
2787 inmem(struct vnode * vp, daddr_t blkno)
2790 vm_offset_t toff, tinc, size;
2794 ASSERT_VOP_LOCKED(vp, "inmem");
2796 if (incore(&vp->v_bufobj, blkno))
2798 if (vp->v_mount == NULL)
2805 if (size > vp->v_mount->mnt_stat.f_iosize)
2806 size = vp->v_mount->mnt_stat.f_iosize;
2807 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2809 VM_OBJECT_RLOCK(obj);
2810 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2811 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2815 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2816 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2817 if (vm_page_is_valid(m,
2818 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2821 VM_OBJECT_RUNLOCK(obj);
2825 VM_OBJECT_RUNLOCK(obj);
2830 * Set the dirty range for a buffer based on the status of the dirty
2831 * bits in the pages comprising the buffer. The range is limited
2832 * to the size of the buffer.
2834 * Tell the VM system that the pages associated with this buffer
2835 * are clean. This is used for delayed writes where the data is
2836 * going to go to disk eventually without additional VM intevention.
2838 * Note that while we only really need to clean through to b_bcount, we
2839 * just go ahead and clean through to b_bufsize.
2842 vfs_clean_pages_dirty_buf(struct buf *bp)
2844 vm_ooffset_t foff, noff, eoff;
2848 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2851 foff = bp->b_offset;
2852 KASSERT(bp->b_offset != NOOFFSET,
2853 ("vfs_clean_pages_dirty_buf: no buffer offset"));
2855 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
2856 vfs_drain_busy_pages(bp);
2857 vfs_setdirty_locked_object(bp);
2858 for (i = 0; i < bp->b_npages; i++) {
2859 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2861 if (eoff > bp->b_offset + bp->b_bufsize)
2862 eoff = bp->b_offset + bp->b_bufsize;
2864 vfs_page_set_validclean(bp, foff, m);
2865 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2868 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
2872 vfs_setdirty_locked_object(struct buf *bp)
2877 object = bp->b_bufobj->bo_object;
2878 VM_OBJECT_ASSERT_WLOCKED(object);
2881 * We qualify the scan for modified pages on whether the
2882 * object has been flushed yet.
2884 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2885 vm_offset_t boffset;
2886 vm_offset_t eoffset;
2889 * test the pages to see if they have been modified directly
2890 * by users through the VM system.
2892 for (i = 0; i < bp->b_npages; i++)
2893 vm_page_test_dirty(bp->b_pages[i]);
2896 * Calculate the encompassing dirty range, boffset and eoffset,
2897 * (eoffset - boffset) bytes.
2900 for (i = 0; i < bp->b_npages; i++) {
2901 if (bp->b_pages[i]->dirty)
2904 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2906 for (i = bp->b_npages - 1; i >= 0; --i) {
2907 if (bp->b_pages[i]->dirty) {
2911 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2914 * Fit it to the buffer.
2917 if (eoffset > bp->b_bcount)
2918 eoffset = bp->b_bcount;
2921 * If we have a good dirty range, merge with the existing
2925 if (boffset < eoffset) {
2926 if (bp->b_dirtyoff > boffset)
2927 bp->b_dirtyoff = boffset;
2928 if (bp->b_dirtyend < eoffset)
2929 bp->b_dirtyend = eoffset;
2935 * Allocate the KVA mapping for an existing buffer. It handles the
2936 * cases of both B_UNMAPPED buffer, and buffer with the preallocated
2937 * KVA which is not mapped (B_KVAALLOC).
2940 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags)
2942 struct buf *scratch_bp;
2943 int bsize, maxsize, need_mapping, need_kva;
2946 need_mapping = (bp->b_flags & B_UNMAPPED) != 0 &&
2947 (gbflags & GB_UNMAPPED) == 0;
2948 need_kva = (bp->b_flags & (B_KVAALLOC | B_UNMAPPED)) == B_UNMAPPED &&
2949 (gbflags & GB_KVAALLOC) != 0;
2950 if (!need_mapping && !need_kva)
2953 BUF_CHECK_UNMAPPED(bp);
2955 if (need_mapping && (bp->b_flags & B_KVAALLOC) != 0) {
2957 * Buffer is not mapped, but the KVA was already
2958 * reserved at the time of the instantiation. Use the
2961 bp->b_flags &= ~B_KVAALLOC;
2962 KASSERT(bp->b_kvaalloc != 0, ("kvaalloc == 0"));
2963 bp->b_kvabase = bp->b_kvaalloc;
2964 atomic_subtract_long(&unmapped_bufspace, bp->b_kvasize);
2969 * Calculate the amount of the address space we would reserve
2970 * if the buffer was mapped.
2972 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize;
2973 offset = blkno * bsize;
2974 maxsize = size + (offset & PAGE_MASK);
2975 maxsize = imax(maxsize, bsize);
2978 if (allocbufkva(bp, maxsize, gbflags)) {
2980 * Request defragmentation. getnewbuf() returns us the
2981 * allocated space by the scratch buffer KVA.
2983 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags |
2984 (GB_UNMAPPED | GB_KVAALLOC));
2985 if (scratch_bp == NULL) {
2986 if ((gbflags & GB_NOWAIT_BD) != 0) {
2988 * XXXKIB: defragmentation cannot
2989 * succeed, not sure what else to do.
2991 panic("GB_NOWAIT_BD and B_UNMAPPED %p", bp);
2993 atomic_add_int(&mappingrestarts, 1);
2996 KASSERT((scratch_bp->b_flags & B_KVAALLOC) != 0,
2997 ("scratch bp !B_KVAALLOC %p", scratch_bp));
2998 setbufkva(bp, (vm_offset_t)scratch_bp->b_kvaalloc,
2999 scratch_bp->b_kvasize, gbflags);
3001 /* Get rid of the scratch buffer. */
3002 scratch_bp->b_kvasize = 0;
3003 scratch_bp->b_flags |= B_INVAL;
3004 scratch_bp->b_flags &= ~(B_UNMAPPED | B_KVAALLOC);
3011 bp->b_saveaddr = bp->b_kvabase;
3012 bp->b_data = bp->b_saveaddr; /* b_offset is handled by bpmap_qenter */
3013 bp->b_flags &= ~B_UNMAPPED;
3014 BUF_CHECK_MAPPED(bp);
3021 * Get a block given a specified block and offset into a file/device.
3022 * The buffers B_DONE bit will be cleared on return, making it almost
3023 * ready for an I/O initiation. B_INVAL may or may not be set on
3024 * return. The caller should clear B_INVAL prior to initiating a
3027 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
3028 * an existing buffer.
3030 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
3031 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
3032 * and then cleared based on the backing VM. If the previous buffer is
3033 * non-0-sized but invalid, B_CACHE will be cleared.
3035 * If getblk() must create a new buffer, the new buffer is returned with
3036 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
3037 * case it is returned with B_INVAL clear and B_CACHE set based on the
3040 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
3041 * B_CACHE bit is clear.
3043 * What this means, basically, is that the caller should use B_CACHE to
3044 * determine whether the buffer is fully valid or not and should clear
3045 * B_INVAL prior to issuing a read. If the caller intends to validate
3046 * the buffer by loading its data area with something, the caller needs
3047 * to clear B_INVAL. If the caller does this without issuing an I/O,
3048 * the caller should set B_CACHE ( as an optimization ), else the caller
3049 * should issue the I/O and biodone() will set B_CACHE if the I/O was
3050 * a write attempt or if it was a successfull read. If the caller
3051 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
3052 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
3055 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo,
3060 int bsize, error, maxsize, vmio;
3063 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
3064 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC,
3065 ("GB_KVAALLOC only makes sense with GB_UNMAPPED"));
3066 ASSERT_VOP_LOCKED(vp, "getblk");
3067 if (size > MAXBSIZE)
3068 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
3069 if (!unmapped_buf_allowed)
3070 flags &= ~(GB_UNMAPPED | GB_KVAALLOC);
3075 bp = gbincore(bo, blkno);
3079 * Buffer is in-core. If the buffer is not busy nor managed,
3080 * it must be on a queue.
3082 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
3084 if (flags & GB_LOCK_NOWAIT)
3085 lockflags |= LK_NOWAIT;
3087 error = BUF_TIMELOCK(bp, lockflags,
3088 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo);
3091 * If we slept and got the lock we have to restart in case
3092 * the buffer changed identities.
3094 if (error == ENOLCK)
3096 /* We timed out or were interrupted. */
3099 /* If recursed, assume caller knows the rules. */
3100 else if (BUF_LOCKRECURSED(bp))
3104 * The buffer is locked. B_CACHE is cleared if the buffer is
3105 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
3106 * and for a VMIO buffer B_CACHE is adjusted according to the
3109 if (bp->b_flags & B_INVAL)
3110 bp->b_flags &= ~B_CACHE;
3111 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
3112 bp->b_flags |= B_CACHE;
3113 if (bp->b_flags & B_MANAGED)
3114 MPASS(bp->b_qindex == QUEUE_NONE);
3119 * check for size inconsistencies for non-VMIO case.
3121 if (bp->b_bcount != size) {
3122 if ((bp->b_flags & B_VMIO) == 0 ||
3123 (size > bp->b_kvasize)) {
3124 if (bp->b_flags & B_DELWRI) {
3126 * If buffer is pinned and caller does
3127 * not want sleep waiting for it to be
3128 * unpinned, bail out
3130 if (bp->b_pin_count > 0) {
3131 if (flags & GB_LOCK_NOWAIT) {
3138 bp->b_flags |= B_NOCACHE;
3141 if (LIST_EMPTY(&bp->b_dep)) {
3142 bp->b_flags |= B_RELBUF;
3145 bp->b_flags |= B_NOCACHE;
3154 * Handle the case of unmapped buffer which should
3155 * become mapped, or the buffer for which KVA
3156 * reservation is requested.
3158 bp_unmapped_get_kva(bp, blkno, size, flags);
3161 * If the size is inconsistant in the VMIO case, we can resize
3162 * the buffer. This might lead to B_CACHE getting set or
3163 * cleared. If the size has not changed, B_CACHE remains
3164 * unchanged from its previous state.
3166 if (bp->b_bcount != size)
3169 KASSERT(bp->b_offset != NOOFFSET,
3170 ("getblk: no buffer offset"));
3173 * A buffer with B_DELWRI set and B_CACHE clear must
3174 * be committed before we can return the buffer in
3175 * order to prevent the caller from issuing a read
3176 * ( due to B_CACHE not being set ) and overwriting
3179 * Most callers, including NFS and FFS, need this to
3180 * operate properly either because they assume they
3181 * can issue a read if B_CACHE is not set, or because
3182 * ( for example ) an uncached B_DELWRI might loop due
3183 * to softupdates re-dirtying the buffer. In the latter
3184 * case, B_CACHE is set after the first write completes,
3185 * preventing further loops.
3186 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3187 * above while extending the buffer, we cannot allow the
3188 * buffer to remain with B_CACHE set after the write
3189 * completes or it will represent a corrupt state. To
3190 * deal with this we set B_NOCACHE to scrap the buffer
3193 * We might be able to do something fancy, like setting
3194 * B_CACHE in bwrite() except if B_DELWRI is already set,
3195 * so the below call doesn't set B_CACHE, but that gets real
3196 * confusing. This is much easier.
3199 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3200 bp->b_flags |= B_NOCACHE;
3204 bp->b_flags &= ~B_DONE;
3207 * Buffer is not in-core, create new buffer. The buffer
3208 * returned by getnewbuf() is locked. Note that the returned
3209 * buffer is also considered valid (not marked B_INVAL).
3213 * If the user does not want us to create the buffer, bail out
3216 if (flags & GB_NOCREAT)
3218 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread))
3221 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
3222 offset = blkno * bsize;
3223 vmio = vp->v_object != NULL;
3225 maxsize = size + (offset & PAGE_MASK);
3228 /* Do not allow non-VMIO notmapped buffers. */
3229 flags &= ~GB_UNMAPPED;
3231 maxsize = imax(maxsize, bsize);
3233 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
3235 if (slpflag || slptimeo)
3241 * This code is used to make sure that a buffer is not
3242 * created while the getnewbuf routine is blocked.
3243 * This can be a problem whether the vnode is locked or not.
3244 * If the buffer is created out from under us, we have to
3245 * throw away the one we just created.
3247 * Note: this must occur before we associate the buffer
3248 * with the vp especially considering limitations in
3249 * the splay tree implementation when dealing with duplicate
3253 if (gbincore(bo, blkno)) {
3255 bp->b_flags |= B_INVAL;
3261 * Insert the buffer into the hash, so that it can
3262 * be found by incore.
3264 bp->b_blkno = bp->b_lblkno = blkno;
3265 bp->b_offset = offset;
3270 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
3271 * buffer size starts out as 0, B_CACHE will be set by
3272 * allocbuf() for the VMIO case prior to it testing the
3273 * backing store for validity.
3277 bp->b_flags |= B_VMIO;
3278 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
3279 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
3280 bp, vp->v_object, bp->b_bufobj->bo_object));
3282 bp->b_flags &= ~B_VMIO;
3283 KASSERT(bp->b_bufobj->bo_object == NULL,
3284 ("ARGH! has b_bufobj->bo_object %p %p\n",
3285 bp, bp->b_bufobj->bo_object));
3286 BUF_CHECK_MAPPED(bp);
3290 bp->b_flags &= ~B_DONE;
3292 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
3293 BUF_ASSERT_HELD(bp);
3295 KASSERT(bp->b_bufobj == bo,
3296 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3301 * Get an empty, disassociated buffer of given size. The buffer is initially
3305 geteblk(int size, int flags)
3310 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3311 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
3312 if ((flags & GB_NOWAIT_BD) &&
3313 (curthread->td_pflags & TDP_BUFNEED) != 0)
3317 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3318 BUF_ASSERT_HELD(bp);
3324 * This code constitutes the buffer memory from either anonymous system
3325 * memory (in the case of non-VMIO operations) or from an associated
3326 * VM object (in the case of VMIO operations). This code is able to
3327 * resize a buffer up or down.
3329 * Note that this code is tricky, and has many complications to resolve
3330 * deadlock or inconsistant data situations. Tread lightly!!!
3331 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3332 * the caller. Calling this code willy nilly can result in the loss of data.
3334 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3335 * B_CACHE for the non-VMIO case.
3339 allocbuf(struct buf *bp, int size)
3341 int newbsize, mbsize;
3344 BUF_ASSERT_HELD(bp);
3346 if (bp->b_kvasize < size)
3347 panic("allocbuf: buffer too small");
3349 if ((bp->b_flags & B_VMIO) == 0) {
3353 * Just get anonymous memory from the kernel. Don't
3354 * mess with B_CACHE.
3356 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3357 if (bp->b_flags & B_MALLOC)
3360 newbsize = round_page(size);
3362 if (newbsize < bp->b_bufsize) {
3364 * malloced buffers are not shrunk
3366 if (bp->b_flags & B_MALLOC) {
3368 bp->b_bcount = size;
3370 free(bp->b_data, M_BIOBUF);
3371 if (bp->b_bufsize) {
3372 atomic_subtract_long(
3378 bp->b_saveaddr = bp->b_kvabase;
3379 bp->b_data = bp->b_saveaddr;
3381 bp->b_flags &= ~B_MALLOC;
3385 vm_hold_free_pages(bp, newbsize);
3386 } else if (newbsize > bp->b_bufsize) {
3388 * We only use malloced memory on the first allocation.
3389 * and revert to page-allocated memory when the buffer
3393 * There is a potential smp race here that could lead
3394 * to bufmallocspace slightly passing the max. It
3395 * is probably extremely rare and not worth worrying
3398 if ( (bufmallocspace < maxbufmallocspace) &&
3399 (bp->b_bufsize == 0) &&
3400 (mbsize <= PAGE_SIZE/2)) {
3402 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
3403 bp->b_bufsize = mbsize;
3404 bp->b_bcount = size;
3405 bp->b_flags |= B_MALLOC;
3406 atomic_add_long(&bufmallocspace, mbsize);
3412 * If the buffer is growing on its other-than-first allocation,
3413 * then we revert to the page-allocation scheme.
3415 if (bp->b_flags & B_MALLOC) {
3416 origbuf = bp->b_data;
3417 origbufsize = bp->b_bufsize;
3418 bp->b_data = bp->b_kvabase;
3419 if (bp->b_bufsize) {
3420 atomic_subtract_long(&bufmallocspace,
3425 bp->b_flags &= ~B_MALLOC;
3426 newbsize = round_page(newbsize);
3430 (vm_offset_t) bp->b_data + bp->b_bufsize,
3431 (vm_offset_t) bp->b_data + newbsize);
3433 bcopy(origbuf, bp->b_data, origbufsize);
3434 free(origbuf, M_BIOBUF);
3440 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3441 desiredpages = (size == 0) ? 0 :
3442 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
3444 if (bp->b_flags & B_MALLOC)
3445 panic("allocbuf: VMIO buffer can't be malloced");
3447 * Set B_CACHE initially if buffer is 0 length or will become
3450 if (size == 0 || bp->b_bufsize == 0)
3451 bp->b_flags |= B_CACHE;
3453 if (newbsize < bp->b_bufsize) {
3455 * DEV_BSIZE aligned new buffer size is less then the
3456 * DEV_BSIZE aligned existing buffer size. Figure out
3457 * if we have to remove any pages.
3459 if (desiredpages < bp->b_npages) {
3462 if ((bp->b_flags & B_UNMAPPED) == 0) {
3463 BUF_CHECK_MAPPED(bp);
3464 pmap_qremove((vm_offset_t)trunc_page(
3465 (vm_offset_t)bp->b_data) +
3466 (desiredpages << PAGE_SHIFT),
3467 (bp->b_npages - desiredpages));
3469 BUF_CHECK_UNMAPPED(bp);
3470 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
3471 for (i = desiredpages; i < bp->b_npages; i++) {
3473 * the page is not freed here -- it
3474 * is the responsibility of
3475 * vnode_pager_setsize
3478 KASSERT(m != bogus_page,
3479 ("allocbuf: bogus page found"));
3480 while (vm_page_sleep_if_busy(m,
3484 bp->b_pages[i] = NULL;
3486 vm_page_unwire(m, PQ_INACTIVE);
3489 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
3490 bp->b_npages = desiredpages;
3492 } else if (size > bp->b_bcount) {
3494 * We are growing the buffer, possibly in a
3495 * byte-granular fashion.
3502 * Step 1, bring in the VM pages from the object,
3503 * allocating them if necessary. We must clear
3504 * B_CACHE if these pages are not valid for the
3505 * range covered by the buffer.
3508 obj = bp->b_bufobj->bo_object;
3510 VM_OBJECT_WLOCK(obj);
3511 while (bp->b_npages < desiredpages) {
3515 * We must allocate system pages since blocking
3516 * here could interfere with paging I/O, no
3517 * matter which process we are.
3519 * Only exclusive busy can be tested here.
3520 * Blocking on shared busy might lead to
3521 * deadlocks once allocbuf() is called after
3522 * pages are vfs_busy_pages().
3524 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3525 bp->b_npages, VM_ALLOC_NOBUSY |
3526 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3527 VM_ALLOC_IGN_SBUSY |
3528 VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3530 bp->b_flags &= ~B_CACHE;
3531 bp->b_pages[bp->b_npages] = m;
3536 * Step 2. We've loaded the pages into the buffer,
3537 * we have to figure out if we can still have B_CACHE
3538 * set. Note that B_CACHE is set according to the
3539 * byte-granular range ( bcount and size ), new the
3540 * aligned range ( newbsize ).
3542 * The VM test is against m->valid, which is DEV_BSIZE
3543 * aligned. Needless to say, the validity of the data
3544 * needs to also be DEV_BSIZE aligned. Note that this
3545 * fails with NFS if the server or some other client
3546 * extends the file's EOF. If our buffer is resized,
3547 * B_CACHE may remain set! XXX
3550 toff = bp->b_bcount;
3551 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3553 while ((bp->b_flags & B_CACHE) && toff < size) {
3556 if (tinc > (size - toff))
3559 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3572 VM_OBJECT_WUNLOCK(obj);
3575 * Step 3, fixup the KVM pmap.
3577 if ((bp->b_flags & B_UNMAPPED) == 0)
3580 BUF_CHECK_UNMAPPED(bp);
3583 if (newbsize < bp->b_bufsize)
3585 bp->b_bufsize = newbsize; /* actual buffer allocation */
3586 bp->b_bcount = size; /* requested buffer size */
3590 extern int inflight_transient_maps;
3593 biodone(struct bio *bp)
3596 void (*done)(struct bio *);
3597 vm_offset_t start, end;
3599 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) {
3600 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING;
3601 bp->bio_flags |= BIO_UNMAPPED;
3602 start = trunc_page((vm_offset_t)bp->bio_data);
3603 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length);
3604 pmap_qremove(start, OFF_TO_IDX(end - start));
3605 vmem_free(transient_arena, start, end - start);
3606 atomic_add_int(&inflight_transient_maps, -1);
3608 done = bp->bio_done;
3610 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3612 bp->bio_flags |= BIO_DONE;
3616 bp->bio_flags |= BIO_DONE;
3622 * Wait for a BIO to finish.
3625 biowait(struct bio *bp, const char *wchan)
3629 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3631 while ((bp->bio_flags & BIO_DONE) == 0)
3632 msleep(bp, mtxp, PRIBIO, wchan, 0);
3634 if (bp->bio_error != 0)
3635 return (bp->bio_error);
3636 if (!(bp->bio_flags & BIO_ERROR))
3642 biofinish(struct bio *bp, struct devstat *stat, int error)
3646 bp->bio_error = error;
3647 bp->bio_flags |= BIO_ERROR;
3650 devstat_end_transaction_bio(stat, bp);
3657 * Wait for buffer I/O completion, returning error status. The buffer
3658 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3659 * error and cleared.
3662 bufwait(struct buf *bp)
3664 if (bp->b_iocmd == BIO_READ)
3665 bwait(bp, PRIBIO, "biord");
3667 bwait(bp, PRIBIO, "biowr");
3668 if (bp->b_flags & B_EINTR) {
3669 bp->b_flags &= ~B_EINTR;
3672 if (bp->b_ioflags & BIO_ERROR) {
3673 return (bp->b_error ? bp->b_error : EIO);
3680 * Call back function from struct bio back up to struct buf.
3683 bufdonebio(struct bio *bip)
3687 bp = bip->bio_caller2;
3688 bp->b_resid = bip->bio_resid;
3689 bp->b_ioflags = bip->bio_flags;
3690 bp->b_error = bip->bio_error;
3692 bp->b_ioflags |= BIO_ERROR;
3698 dev_strategy(struct cdev *dev, struct buf *bp)
3703 KASSERT(dev->si_refcount > 0,
3704 ("dev_strategy on un-referenced struct cdev *(%s) %p",
3705 devtoname(dev), dev));
3707 csw = dev_refthread(dev, &ref);
3708 dev_strategy_csw(dev, csw, bp);
3709 dev_relthread(dev, ref);
3713 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp)
3717 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE,
3719 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) ||
3720 dev->si_threadcount > 0,
3721 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev),
3724 bp->b_error = ENXIO;
3725 bp->b_ioflags = BIO_ERROR;
3733 /* Try again later */
3734 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3736 bip->bio_cmd = bp->b_iocmd;
3737 bip->bio_offset = bp->b_iooffset;
3738 bip->bio_length = bp->b_bcount;
3739 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3741 bip->bio_done = bufdonebio;
3742 bip->bio_caller2 = bp;
3744 (*csw->d_strategy)(bip);
3750 * Finish I/O on a buffer, optionally calling a completion function.
3751 * This is usually called from an interrupt so process blocking is
3754 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3755 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3756 * assuming B_INVAL is clear.
3758 * For the VMIO case, we set B_CACHE if the op was a read and no
3759 * read error occured, or if the op was a write. B_CACHE is never
3760 * set if the buffer is invalid or otherwise uncacheable.
3762 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3763 * initiator to leave B_INVAL set to brelse the buffer out of existance
3764 * in the biodone routine.
3767 bufdone(struct buf *bp)
3769 struct bufobj *dropobj;
3770 void (*biodone)(struct buf *);
3772 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3775 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3776 BUF_ASSERT_HELD(bp);
3778 runningbufwakeup(bp);
3779 if (bp->b_iocmd == BIO_WRITE)
3780 dropobj = bp->b_bufobj;
3781 /* call optional completion function if requested */
3782 if (bp->b_iodone != NULL) {
3783 biodone = bp->b_iodone;
3784 bp->b_iodone = NULL;
3787 bufobj_wdrop(dropobj);
3794 bufobj_wdrop(dropobj);
3798 bufdone_finish(struct buf *bp)
3800 BUF_ASSERT_HELD(bp);
3802 if (!LIST_EMPTY(&bp->b_dep))
3805 if (bp->b_flags & B_VMIO) {
3810 int bogus, i, iosize;
3812 obj = bp->b_bufobj->bo_object;
3813 KASSERT(obj->paging_in_progress >= bp->b_npages,
3814 ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3815 obj->paging_in_progress, bp->b_npages));
3818 KASSERT(vp->v_holdcnt > 0,
3819 ("biodone_finish: vnode %p has zero hold count", vp));
3820 KASSERT(vp->v_object != NULL,
3821 ("biodone_finish: vnode %p has no vm_object", vp));
3823 foff = bp->b_offset;
3824 KASSERT(bp->b_offset != NOOFFSET,
3825 ("biodone_finish: bp %p has no buffer offset", bp));
3828 * Set B_CACHE if the op was a normal read and no error
3829 * occured. B_CACHE is set for writes in the b*write()
3832 iosize = bp->b_bcount - bp->b_resid;
3833 if (bp->b_iocmd == BIO_READ &&
3834 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3835 !(bp->b_ioflags & BIO_ERROR)) {
3836 bp->b_flags |= B_CACHE;
3839 VM_OBJECT_WLOCK(obj);
3840 for (i = 0; i < bp->b_npages; i++) {
3844 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3849 * cleanup bogus pages, restoring the originals
3852 if (m == bogus_page) {
3853 bogus = bogusflag = 1;
3854 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3856 panic("biodone: page disappeared!");
3859 KASSERT(OFF_TO_IDX(foff) == m->pindex,
3860 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3861 (intmax_t)foff, (uintmax_t)m->pindex));
3864 * In the write case, the valid and clean bits are
3865 * already changed correctly ( see bdwrite() ), so we
3866 * only need to do this here in the read case.
3868 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3869 KASSERT((m->dirty & vm_page_bits(foff &
3870 PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3871 " page %p has unexpected dirty bits", m));
3872 vfs_page_set_valid(bp, foff, m);
3876 vm_object_pip_subtract(obj, 1);
3877 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3880 vm_object_pip_wakeupn(obj, 0);
3881 VM_OBJECT_WUNLOCK(obj);
3882 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
3883 BUF_CHECK_MAPPED(bp);
3884 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3885 bp->b_pages, bp->b_npages);
3890 * For asynchronous completions, release the buffer now. The brelse
3891 * will do a wakeup there if necessary - so no need to do a wakeup
3892 * here in the async case. The sync case always needs to do a wakeup.
3895 if (bp->b_flags & B_ASYNC) {
3896 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3905 * This routine is called in lieu of iodone in the case of
3906 * incomplete I/O. This keeps the busy status for pages
3910 vfs_unbusy_pages(struct buf *bp)
3916 runningbufwakeup(bp);
3917 if (!(bp->b_flags & B_VMIO))
3920 obj = bp->b_bufobj->bo_object;
3921 VM_OBJECT_WLOCK(obj);
3922 for (i = 0; i < bp->b_npages; i++) {
3924 if (m == bogus_page) {
3925 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3927 panic("vfs_unbusy_pages: page missing\n");
3929 if ((bp->b_flags & B_UNMAPPED) == 0) {
3930 BUF_CHECK_MAPPED(bp);
3931 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3932 bp->b_pages, bp->b_npages);
3934 BUF_CHECK_UNMAPPED(bp);
3936 vm_object_pip_subtract(obj, 1);
3939 vm_object_pip_wakeupn(obj, 0);
3940 VM_OBJECT_WUNLOCK(obj);
3944 * vfs_page_set_valid:
3946 * Set the valid bits in a page based on the supplied offset. The
3947 * range is restricted to the buffer's size.
3949 * This routine is typically called after a read completes.
3952 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3957 * Compute the end offset, eoff, such that [off, eoff) does not span a
3958 * page boundary and eoff is not greater than the end of the buffer.
3959 * The end of the buffer, in this case, is our file EOF, not the
3960 * allocation size of the buffer.
3962 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3963 if (eoff > bp->b_offset + bp->b_bcount)
3964 eoff = bp->b_offset + bp->b_bcount;
3967 * Set valid range. This is typically the entire buffer and thus the
3971 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3975 * vfs_page_set_validclean:
3977 * Set the valid bits and clear the dirty bits in a page based on the
3978 * supplied offset. The range is restricted to the buffer's size.
3981 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3983 vm_ooffset_t soff, eoff;
3986 * Start and end offsets in buffer. eoff - soff may not cross a
3987 * page boundry or cross the end of the buffer. The end of the
3988 * buffer, in this case, is our file EOF, not the allocation size
3992 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3993 if (eoff > bp->b_offset + bp->b_bcount)
3994 eoff = bp->b_offset + bp->b_bcount;
3997 * Set valid range. This is typically the entire buffer and thus the
4001 vm_page_set_validclean(
4003 (vm_offset_t) (soff & PAGE_MASK),
4004 (vm_offset_t) (eoff - soff)
4010 * Ensure that all buffer pages are not exclusive busied. If any page is
4011 * exclusive busy, drain it.
4014 vfs_drain_busy_pages(struct buf *bp)
4019 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object);
4021 for (i = 0; i < bp->b_npages; i++) {
4023 if (vm_page_xbusied(m)) {
4024 for (; last_busied < i; last_busied++)
4025 vm_page_sbusy(bp->b_pages[last_busied]);
4026 while (vm_page_xbusied(m)) {
4028 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4029 vm_page_busy_sleep(m, "vbpage");
4030 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4034 for (i = 0; i < last_busied; i++)
4035 vm_page_sunbusy(bp->b_pages[i]);
4039 * This routine is called before a device strategy routine.
4040 * It is used to tell the VM system that paging I/O is in
4041 * progress, and treat the pages associated with the buffer
4042 * almost as being exclusive busy. Also the object paging_in_progress
4043 * flag is handled to make sure that the object doesn't become
4046 * Since I/O has not been initiated yet, certain buffer flags
4047 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
4048 * and should be ignored.
4051 vfs_busy_pages(struct buf *bp, int clear_modify)
4058 if (!(bp->b_flags & B_VMIO))
4061 obj = bp->b_bufobj->bo_object;
4062 foff = bp->b_offset;
4063 KASSERT(bp->b_offset != NOOFFSET,
4064 ("vfs_busy_pages: no buffer offset"));
4065 VM_OBJECT_WLOCK(obj);
4066 vfs_drain_busy_pages(bp);
4067 if (bp->b_bufsize != 0)
4068 vfs_setdirty_locked_object(bp);
4070 for (i = 0; i < bp->b_npages; i++) {
4073 if ((bp->b_flags & B_CLUSTER) == 0) {
4074 vm_object_pip_add(obj, 1);
4078 * When readying a buffer for a read ( i.e
4079 * clear_modify == 0 ), it is important to do
4080 * bogus_page replacement for valid pages in
4081 * partially instantiated buffers. Partially
4082 * instantiated buffers can, in turn, occur when
4083 * reconstituting a buffer from its VM backing store
4084 * base. We only have to do this if B_CACHE is
4085 * clear ( which causes the I/O to occur in the
4086 * first place ). The replacement prevents the read
4087 * I/O from overwriting potentially dirty VM-backed
4088 * pages. XXX bogus page replacement is, uh, bogus.
4089 * It may not work properly with small-block devices.
4090 * We need to find a better way.
4093 pmap_remove_write(m);
4094 vfs_page_set_validclean(bp, foff, m);
4095 } else if (m->valid == VM_PAGE_BITS_ALL &&
4096 (bp->b_flags & B_CACHE) == 0) {
4097 bp->b_pages[i] = bogus_page;
4100 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
4102 VM_OBJECT_WUNLOCK(obj);
4103 if (bogus && (bp->b_flags & B_UNMAPPED) == 0) {
4104 BUF_CHECK_MAPPED(bp);
4105 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4106 bp->b_pages, bp->b_npages);
4111 * vfs_bio_set_valid:
4113 * Set the range within the buffer to valid. The range is
4114 * relative to the beginning of the buffer, b_offset. Note that
4115 * b_offset itself may be offset from the beginning of the first
4119 vfs_bio_set_valid(struct buf *bp, int base, int size)
4124 if (!(bp->b_flags & B_VMIO))
4128 * Fixup base to be relative to beginning of first page.
4129 * Set initial n to be the maximum number of bytes in the
4130 * first page that can be validated.
4132 base += (bp->b_offset & PAGE_MASK);
4133 n = PAGE_SIZE - (base & PAGE_MASK);
4135 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4136 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4140 vm_page_set_valid_range(m, base & PAGE_MASK, n);
4145 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4151 * If the specified buffer is a non-VMIO buffer, clear the entire
4152 * buffer. If the specified buffer is a VMIO buffer, clear and
4153 * validate only the previously invalid portions of the buffer.
4154 * This routine essentially fakes an I/O, so we need to clear
4155 * BIO_ERROR and B_INVAL.
4157 * Note that while we only theoretically need to clear through b_bcount,
4158 * we go ahead and clear through b_bufsize.
4161 vfs_bio_clrbuf(struct buf *bp)
4163 int i, j, mask, sa, ea, slide;
4165 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
4169 bp->b_flags &= ~B_INVAL;
4170 bp->b_ioflags &= ~BIO_ERROR;
4171 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
4172 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4173 (bp->b_offset & PAGE_MASK) == 0) {
4174 if (bp->b_pages[0] == bogus_page)
4176 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4177 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object);
4178 if ((bp->b_pages[0]->valid & mask) == mask)
4180 if ((bp->b_pages[0]->valid & mask) == 0) {
4181 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize);
4182 bp->b_pages[0]->valid |= mask;
4186 sa = bp->b_offset & PAGE_MASK;
4188 for (i = 0; i < bp->b_npages; i++, sa = 0) {
4189 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize);
4190 ea = slide & PAGE_MASK;
4193 if (bp->b_pages[i] == bogus_page)
4196 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4197 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object);
4198 if ((bp->b_pages[i]->valid & mask) == mask)
4200 if ((bp->b_pages[i]->valid & mask) == 0)
4201 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa);
4203 for (; sa < ea; sa += DEV_BSIZE, j++) {
4204 if ((bp->b_pages[i]->valid & (1 << j)) == 0) {
4205 pmap_zero_page_area(bp->b_pages[i],
4210 bp->b_pages[i]->valid |= mask;
4213 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
4218 vfs_bio_bzero_buf(struct buf *bp, int base, int size)
4223 if ((bp->b_flags & B_UNMAPPED) == 0) {
4224 BUF_CHECK_MAPPED(bp);
4225 bzero(bp->b_data + base, size);
4227 BUF_CHECK_UNMAPPED(bp);
4228 n = PAGE_SIZE - (base & PAGE_MASK);
4229 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
4233 pmap_zero_page_area(m, base & PAGE_MASK, n);
4242 * vm_hold_load_pages and vm_hold_free_pages get pages into
4243 * a buffers address space. The pages are anonymous and are
4244 * not associated with a file object.
4247 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4253 BUF_CHECK_MAPPED(bp);
4255 to = round_page(to);
4256 from = round_page(from);
4257 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4259 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4262 * note: must allocate system pages since blocking here
4263 * could interfere with paging I/O, no matter which
4266 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
4267 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
4272 pmap_qenter(pg, &p, 1);
4273 bp->b_pages[index] = p;
4275 bp->b_npages = index;
4278 /* Return pages associated with this buf to the vm system */
4280 vm_hold_free_pages(struct buf *bp, int newbsize)
4284 int index, newnpages;
4286 BUF_CHECK_MAPPED(bp);
4288 from = round_page((vm_offset_t)bp->b_data + newbsize);
4289 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4290 if (bp->b_npages > newnpages)
4291 pmap_qremove(from, bp->b_npages - newnpages);
4292 for (index = newnpages; index < bp->b_npages; index++) {
4293 p = bp->b_pages[index];
4294 bp->b_pages[index] = NULL;
4295 if (vm_page_sbusied(p))
4296 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
4297 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
4300 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
4302 bp->b_npages = newnpages;
4306 * Map an IO request into kernel virtual address space.
4308 * All requests are (re)mapped into kernel VA space.
4309 * Notice that we use b_bufsize for the size of the buffer
4310 * to be mapped. b_bcount might be modified by the driver.
4312 * Note that even if the caller determines that the address space should
4313 * be valid, a race or a smaller-file mapped into a larger space may
4314 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
4315 * check the return value.
4318 vmapbuf(struct buf *bp, int mapbuf)
4324 if (bp->b_bufsize < 0)
4326 prot = VM_PROT_READ;
4327 if (bp->b_iocmd == BIO_READ)
4328 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
4329 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
4330 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
4331 btoc(MAXPHYS))) < 0)
4333 bp->b_npages = pidx;
4334 if (mapbuf || !unmapped_buf_allowed) {
4335 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
4336 kva = bp->b_saveaddr;
4337 bp->b_saveaddr = bp->b_data;
4338 bp->b_data = kva + (((vm_offset_t)bp->b_data) & PAGE_MASK);
4339 bp->b_flags &= ~B_UNMAPPED;
4341 bp->b_flags |= B_UNMAPPED;
4342 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK;
4343 bp->b_saveaddr = bp->b_data;
4344 bp->b_data = unmapped_buf;
4350 * Free the io map PTEs associated with this IO operation.
4351 * We also invalidate the TLB entries and restore the original b_addr.
4354 vunmapbuf(struct buf *bp)
4358 npages = bp->b_npages;
4359 if (bp->b_flags & B_UNMAPPED)
4360 bp->b_flags &= ~B_UNMAPPED;
4362 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4363 vm_page_unhold_pages(bp->b_pages, npages);
4365 bp->b_data = bp->b_saveaddr;
4369 bdone(struct buf *bp)
4373 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4375 bp->b_flags |= B_DONE;
4381 bwait(struct buf *bp, u_char pri, const char *wchan)
4385 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4387 while ((bp->b_flags & B_DONE) == 0)
4388 msleep(bp, mtxp, pri, wchan, 0);
4393 bufsync(struct bufobj *bo, int waitfor)
4396 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
4400 bufstrategy(struct bufobj *bo, struct buf *bp)
4406 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
4407 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
4408 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
4409 i = VOP_STRATEGY(vp, bp);
4410 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
4414 bufobj_wrefl(struct bufobj *bo)
4417 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4418 ASSERT_BO_WLOCKED(bo);
4423 bufobj_wref(struct bufobj *bo)
4426 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4433 bufobj_wdrop(struct bufobj *bo)
4436 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4438 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4439 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4440 bo->bo_flag &= ~BO_WWAIT;
4441 wakeup(&bo->bo_numoutput);
4447 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4451 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4452 ASSERT_BO_WLOCKED(bo);
4454 while (bo->bo_numoutput) {
4455 bo->bo_flag |= BO_WWAIT;
4456 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo),
4457 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4465 bpin(struct buf *bp)
4469 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4476 bunpin(struct buf *bp)
4480 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4482 if (--bp->b_pin_count == 0)
4488 bunpin_wait(struct buf *bp)
4492 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4494 while (bp->b_pin_count > 0)
4495 msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4500 * Set bio_data or bio_ma for struct bio from the struct buf.
4503 bdata2bio(struct buf *bp, struct bio *bip)
4506 if ((bp->b_flags & B_UNMAPPED) != 0) {
4507 KASSERT(unmapped_buf_allowed, ("unmapped"));
4508 bip->bio_ma = bp->b_pages;
4509 bip->bio_ma_n = bp->b_npages;
4510 bip->bio_data = unmapped_buf;
4511 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
4512 bip->bio_flags |= BIO_UNMAPPED;
4513 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) /
4514 PAGE_SIZE == bp->b_npages,
4515 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset,
4516 (long long)bip->bio_length, bip->bio_ma_n));
4518 bip->bio_data = bp->b_data;
4523 #include "opt_ddb.h"
4525 #include <ddb/ddb.h>
4527 /* DDB command to show buffer data */
4528 DB_SHOW_COMMAND(buffer, db_show_buffer)
4531 struct buf *bp = (struct buf *)addr;
4534 db_printf("usage: show buffer <addr>\n");
4538 db_printf("buf at %p\n", bp);
4539 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
4540 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
4541 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
4543 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4544 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
4546 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4547 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4548 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
4551 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4552 for (i = 0; i < bp->b_npages; i++) {
4555 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4556 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4557 if ((i + 1) < bp->b_npages)
4563 BUF_LOCKPRINTINFO(bp);
4566 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4571 for (i = 0; i < nbuf; i++) {
4573 if (BUF_ISLOCKED(bp)) {
4574 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4580 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4586 db_printf("usage: show vnodebufs <addr>\n");
4589 vp = (struct vnode *)addr;
4590 db_printf("Clean buffers:\n");
4591 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4592 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4595 db_printf("Dirty buffers:\n");
4596 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4597 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4602 DB_COMMAND(countfreebufs, db_coundfreebufs)
4605 int i, used = 0, nfree = 0;
4608 db_printf("usage: countfreebufs\n");
4612 for (i = 0; i < nbuf; i++) {
4614 if ((bp->b_flags & B_INFREECNT) != 0)
4620 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4622 db_printf("numfreebuffers is %d\n", numfreebuffers);