2 * Copyright (c) 2004 Poul-Henning Kamp
3 * Copyright (c) 1994,1997 John S. Dyson
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * this file contains a new buffer I/O scheme implementing a coherent
30 * VM object and buffer cache scheme. Pains have been taken to make
31 * sure that the performance degradation associated with schemes such
32 * as this is not realized.
34 * Author: John S. Dyson
35 * Significant help during the development and debugging phases
36 * had been provided by David Greenman, also of the FreeBSD core team.
38 * see man buf(9) for more info.
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
44 #include <sys/param.h>
45 #include <sys/systm.h>
49 #include <sys/devicestat.h>
50 #include <sys/eventhandler.h>
52 #include <sys/limits.h>
54 #include <sys/malloc.h>
55 #include <sys/mount.h>
56 #include <sys/mutex.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
60 #include <sys/resourcevar.h>
61 #include <sys/sysctl.h>
62 #include <sys/vmmeter.h>
63 #include <sys/vnode.h>
64 #include <geom/geom.h>
66 #include <vm/vm_param.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_pageout.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_extern.h>
72 #include <vm/vm_map.h>
73 #include "opt_compat.h"
74 #include "opt_directio.h"
77 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer");
79 struct bio_ops bioops; /* I/O operation notification */
81 struct buf_ops buf_ops_bio = {
82 .bop_name = "buf_ops_bio",
83 .bop_write = bufwrite,
84 .bop_strategy = bufstrategy,
86 .bop_bdflush = bufbdflush,
90 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has
91 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c.
93 struct buf *buf; /* buffer header pool */
95 static struct proc *bufdaemonproc;
97 static int inmem(struct vnode *vp, daddr_t blkno);
98 static void vm_hold_free_pages(struct buf *bp, int newbsize);
99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
104 static void vfs_drain_busy_pages(struct buf *bp);
105 static void vfs_clean_pages_dirty_buf(struct buf *bp);
106 static void vfs_setdirty_locked_object(struct buf *bp);
107 static void vfs_vmio_release(struct buf *bp);
108 static int vfs_bio_clcheck(struct vnode *vp, int size,
109 daddr_t lblkno, daddr_t blkno);
110 static int buf_do_flush(struct vnode *vp);
111 static int flushbufqueues(struct vnode *, int, int);
112 static void buf_daemon(void);
113 static void bremfreel(struct buf *bp);
114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
115 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
119 int vmiodirenable = TRUE;
120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
121 "Use the VM system for directory writes");
122 long runningbufspace;
123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
124 "Amount of presently outstanding async buffer io");
125 static long bufspace;
126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
127 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
129 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
132 "Virtual memory used for buffers");
134 static long maxbufspace;
135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
136 "Maximum allowed value of bufspace (including buf_daemon)");
137 static long bufmallocspace;
138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
139 "Amount of malloced memory for buffers");
140 static long maxbufmallocspace;
141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
142 "Maximum amount of malloced memory for buffers");
143 static long lobufspace;
144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
145 "Minimum amount of buffers we want to have");
147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
148 "Maximum allowed value of bufspace (excluding buf_daemon)");
149 static int bufreusecnt;
150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
151 "Number of times we have reused a buffer");
152 static int buffreekvacnt;
153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
154 "Number of times we have freed the KVA space from some buffer");
155 static int bufdefragcnt;
156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
157 "Number of times we have had to repeat buffer allocation to defragment");
158 static long lorunningspace;
159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
160 "Minimum preferred space used for in-progress I/O");
161 static long hirunningspace;
162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
163 "Maximum amount of space to use for in-progress I/O");
164 int dirtybufferflushes;
165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
166 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
169 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
170 int altbufferflushes;
171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
172 0, "Number of fsync flushes to limit dirty buffers");
173 static int recursiveflushes;
174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
175 0, "Number of flushes skipped due to being recursive");
176 static int numdirtybuffers;
177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
178 "Number of buffers that are dirty (has unwritten changes) at the moment");
179 static int lodirtybuffers;
180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
181 "How many buffers we want to have free before bufdaemon can sleep");
182 static int hidirtybuffers;
183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
184 "When the number of dirty buffers is considered severe");
186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
187 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
188 static int numfreebuffers;
189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
190 "Number of free buffers");
191 static int lofreebuffers;
192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
194 static int hifreebuffers;
195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
196 "XXX Complicatedly unused");
197 static int getnewbufcalls;
198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
199 "Number of calls to getnewbuf");
200 static int getnewbufrestarts;
201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
202 "Number of times getnewbuf has had to restart a buffer aquisition");
203 static int flushbufqtarget = 100;
204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
205 "Amount of work to do in flushbufqueues when helping bufdaemon");
206 static long notbufdflashes;
207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0,
208 "Number of dirty buffer flushes done by the bufdaemon helpers");
211 * Wakeup point for bufdaemon, as well as indicator of whether it is already
212 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
215 static int bd_request;
218 * Request for the buf daemon to write more buffers than is indicated by
219 * lodirtybuf. This may be necessary to push out excess dependencies or
220 * defragment the address space where a simple count of the number of dirty
221 * buffers is insufficient to characterize the demand for flushing them.
223 static int bd_speedupreq;
226 * This lock synchronizes access to bd_request.
228 static struct mtx bdlock;
231 * bogus page -- for I/O to/from partially complete buffers
232 * this is a temporary solution to the problem, but it is not
233 * really that bad. it would be better to split the buffer
234 * for input in the case of buffers partially already in memory,
235 * but the code is intricate enough already.
237 vm_page_t bogus_page;
240 * Synchronization (sleep/wakeup) variable for active buffer space requests.
241 * Set when wait starts, cleared prior to wakeup().
242 * Used in runningbufwakeup() and waitrunningbufspace().
244 static int runningbufreq;
247 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
248 * waitrunningbufspace().
250 static struct mtx rbreqlock;
253 * Synchronization (sleep/wakeup) variable for buffer requests.
254 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
256 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
257 * getnewbuf(), and getblk().
259 static int needsbuffer;
262 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
264 static struct mtx nblock;
267 * Definitions for the buffer free lists.
269 #define BUFFER_QUEUES 5 /* number of free buffer queues */
271 #define QUEUE_NONE 0 /* on no queue */
272 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
273 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
274 #define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */
275 #define QUEUE_EMPTY 4 /* empty buffer headers */
276 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */
278 /* Queues for free buffers with various properties */
279 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
281 /* Lock for the bufqueues */
282 static struct mtx bqlock;
285 * Single global constant for BUF_WMESG, to avoid getting multiple references.
286 * buf_wmesg is referred from macros.
288 const char *buf_wmesg = BUF_WMESG;
290 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
291 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
292 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
293 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
295 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
296 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
298 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
303 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
304 return (sysctl_handle_long(oidp, arg1, arg2, req));
305 lvalue = *(long *)arg1;
306 if (lvalue > INT_MAX)
307 /* On overflow, still write out a long to trigger ENOMEM. */
308 return (sysctl_handle_long(oidp, &lvalue, 0, req));
310 return (sysctl_handle_int(oidp, &ivalue, 0, req));
315 extern void ffs_rawread_setup(void);
316 #endif /* DIRECTIO */
320 * If someone is blocked due to there being too many dirty buffers,
321 * and numdirtybuffers is now reasonable, wake them up.
325 numdirtywakeup(int level)
328 if (numdirtybuffers <= level) {
330 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
331 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
332 wakeup(&needsbuffer);
341 * Called when buffer space is potentially available for recovery.
342 * getnewbuf() will block on this flag when it is unable to free
343 * sufficient buffer space. Buffer space becomes recoverable when
344 * bp's get placed back in the queues.
352 * If someone is waiting for BUF space, wake them up. Even
353 * though we haven't freed the kva space yet, the waiting
354 * process will be able to now.
357 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
358 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
359 wakeup(&needsbuffer);
365 * runningbufwakeup() - in-progress I/O accounting.
369 runningbufwakeup(struct buf *bp)
372 if (bp->b_runningbufspace) {
373 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
374 bp->b_runningbufspace = 0;
375 mtx_lock(&rbreqlock);
376 if (runningbufreq && runningbufspace <= lorunningspace) {
378 wakeup(&runningbufreq);
380 mtx_unlock(&rbreqlock);
387 * Called when a buffer has been added to one of the free queues to
388 * account for the buffer and to wakeup anyone waiting for free buffers.
389 * This typically occurs when large amounts of metadata are being handled
390 * by the buffer cache ( else buffer space runs out first, usually ).
394 bufcountwakeup(struct buf *bp)
398 KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
399 ("buf %p already counted as free", bp));
400 if (bp->b_bufobj != NULL)
401 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
402 bp->b_vflags |= BV_INFREECNT;
403 old = atomic_fetchadd_int(&numfreebuffers, 1);
404 KASSERT(old >= 0 && old < nbuf,
405 ("numfreebuffers climbed to %d", old + 1));
408 needsbuffer &= ~VFS_BIO_NEED_ANY;
409 if (numfreebuffers >= hifreebuffers)
410 needsbuffer &= ~VFS_BIO_NEED_FREE;
411 wakeup(&needsbuffer);
417 * waitrunningbufspace()
419 * runningbufspace is a measure of the amount of I/O currently
420 * running. This routine is used in async-write situations to
421 * prevent creating huge backups of pending writes to a device.
422 * Only asynchronous writes are governed by this function.
424 * Reads will adjust runningbufspace, but will not block based on it.
425 * The read load has a side effect of reducing the allowed write load.
427 * This does NOT turn an async write into a sync write. It waits
428 * for earlier writes to complete and generally returns before the
429 * caller's write has reached the device.
432 waitrunningbufspace(void)
435 mtx_lock(&rbreqlock);
436 while (runningbufspace > hirunningspace) {
438 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
440 mtx_unlock(&rbreqlock);
445 * vfs_buf_test_cache:
447 * Called when a buffer is extended. This function clears the B_CACHE
448 * bit if the newly extended portion of the buffer does not contain
453 vfs_buf_test_cache(struct buf *bp,
454 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
458 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
459 if (bp->b_flags & B_CACHE) {
460 int base = (foff + off) & PAGE_MASK;
461 if (vm_page_is_valid(m, base, size) == 0)
462 bp->b_flags &= ~B_CACHE;
466 /* Wake up the buffer daemon if necessary */
469 bd_wakeup(int dirtybuflevel)
473 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
481 * bd_speedup - speedup the buffer cache flushing code
491 if (bd_speedupreq == 0 || bd_request == 0)
501 * Calculating buffer cache scaling values and reserve space for buffer
502 * headers. This is called during low level kernel initialization and
503 * may be called more then once. We CANNOT write to the memory area
504 * being reserved at this time.
507 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
513 * physmem_est is in pages. Convert it to kilobytes (assumes
514 * PAGE_SIZE is >= 1K)
516 physmem_est = physmem_est * (PAGE_SIZE / 1024);
519 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
520 * For the first 64MB of ram nominally allocate sufficient buffers to
521 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
522 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
523 * the buffer cache we limit the eventual kva reservation to
526 * factor represents the 1/4 x ram conversion.
529 int factor = 4 * BKVASIZE / 1024;
532 if (physmem_est > 4096)
533 nbuf += min((physmem_est - 4096) / factor,
535 if (physmem_est > 65536)
536 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
538 if (maxbcache && nbuf > maxbcache / BKVASIZE)
539 nbuf = maxbcache / BKVASIZE;
544 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
545 maxbuf = (LONG_MAX / 3) / BKVASIZE;
548 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
554 * swbufs are used as temporary holders for I/O, such as paging I/O.
555 * We have no less then 16 and no more then 256.
557 nswbuf = max(min(nbuf/4, 256), 16);
559 if (nswbuf < NSWBUF_MIN)
567 * Reserve space for the buffer cache buffers
570 v = (caddr_t)(swbuf + nswbuf);
572 v = (caddr_t)(buf + nbuf);
577 /* Initialize the buffer subsystem. Called before use of any buffers. */
584 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
585 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
586 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
587 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
589 /* next, make a null set of free lists */
590 for (i = 0; i < BUFFER_QUEUES; i++)
591 TAILQ_INIT(&bufqueues[i]);
593 /* finally, initialize each buffer header and stick on empty q */
594 for (i = 0; i < nbuf; i++) {
596 bzero(bp, sizeof *bp);
597 bp->b_flags = B_INVAL; /* we're just an empty header */
598 bp->b_rcred = NOCRED;
599 bp->b_wcred = NOCRED;
600 bp->b_qindex = QUEUE_EMPTY;
601 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */
603 LIST_INIT(&bp->b_dep);
605 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
609 * maxbufspace is the absolute maximum amount of buffer space we are
610 * allowed to reserve in KVM and in real terms. The absolute maximum
611 * is nominally used by buf_daemon. hibufspace is the nominal maximum
612 * used by most other processes. The differential is required to
613 * ensure that buf_daemon is able to run when other processes might
614 * be blocked waiting for buffer space.
616 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
617 * this may result in KVM fragmentation which is not handled optimally
620 maxbufspace = (long)nbuf * BKVASIZE;
621 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
622 lobufspace = hibufspace - MAXBSIZE;
625 * Note: The 16 MiB upper limit for hirunningspace was chosen
626 * arbitrarily and may need further tuning. It corresponds to
627 * 128 outstanding write IO requests (if IO size is 128 KiB),
628 * which fits with many RAID controllers' tagged queuing limits.
629 * The lower 1 MiB limit is the historical upper limit for
632 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE),
633 16 * 1024 * 1024), 1024 * 1024);
634 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE);
637 * Limit the amount of malloc memory since it is wired permanently into
638 * the kernel space. Even though this is accounted for in the buffer
639 * allocation, we don't want the malloced region to grow uncontrolled.
640 * The malloc scheme improves memory utilization significantly on average
641 * (small) directories.
643 maxbufmallocspace = hibufspace / 20;
646 * Reduce the chance of a deadlock occuring by limiting the number
647 * of delayed-write dirty buffers we allow to stack up.
649 hidirtybuffers = nbuf / 4 + 20;
650 dirtybufthresh = hidirtybuffers * 9 / 10;
653 * To support extreme low-memory systems, make sure hidirtybuffers cannot
654 * eat up all available buffer space. This occurs when our minimum cannot
655 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
656 * BKVASIZE'd buffers.
658 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
659 hidirtybuffers >>= 1;
661 lodirtybuffers = hidirtybuffers / 2;
664 * Try to keep the number of free buffers in the specified range,
665 * and give special processes (e.g. like buf_daemon) access to an
668 lofreebuffers = nbuf / 18 + 5;
669 hifreebuffers = 2 * lofreebuffers;
670 numfreebuffers = nbuf;
672 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
673 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
677 * bfreekva() - free the kva allocation for a buffer.
679 * Since this call frees up buffer space, we call bufspacewakeup().
682 bfreekva(struct buf *bp)
686 atomic_add_int(&buffreekvacnt, 1);
687 atomic_subtract_long(&bufspace, bp->b_kvasize);
688 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
689 (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
698 * Mark the buffer for removal from the appropriate free list in brelse.
702 bremfree(struct buf *bp)
706 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
707 KASSERT((bp->b_flags & B_REMFREE) == 0,
708 ("bremfree: buffer %p already marked for delayed removal.", bp));
709 KASSERT(bp->b_qindex != QUEUE_NONE,
710 ("bremfree: buffer %p not on a queue.", bp));
713 bp->b_flags |= B_REMFREE;
714 /* Fixup numfreebuffers count. */
715 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
716 KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
717 ("buf %p not counted in numfreebuffers", bp));
718 if (bp->b_bufobj != NULL)
719 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
720 bp->b_vflags &= ~BV_INFREECNT;
721 old = atomic_fetchadd_int(&numfreebuffers, -1);
722 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
729 * Force an immediate removal from a free list. Used only in nfs when
730 * it abuses the b_freelist pointer.
733 bremfreef(struct buf *bp)
743 * Removes a buffer from the free list, must be called with the
747 bremfreel(struct buf *bp)
751 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
752 bp, bp->b_vp, bp->b_flags);
753 KASSERT(bp->b_qindex != QUEUE_NONE,
754 ("bremfreel: buffer %p not on a queue.", bp));
756 mtx_assert(&bqlock, MA_OWNED);
758 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
759 bp->b_qindex = QUEUE_NONE;
761 * If this was a delayed bremfree() we only need to remove the buffer
762 * from the queue and return the stats are already done.
764 if (bp->b_flags & B_REMFREE) {
765 bp->b_flags &= ~B_REMFREE;
769 * Fixup numfreebuffers count. If the buffer is invalid or not
770 * delayed-write, the buffer was free and we must decrement
773 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
774 KASSERT((bp->b_vflags & BV_INFREECNT) != 0,
775 ("buf %p not counted in numfreebuffers", bp));
776 if (bp->b_bufobj != NULL)
777 mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED);
778 bp->b_vflags &= ~BV_INFREECNT;
779 old = atomic_fetchadd_int(&numfreebuffers, -1);
780 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1));
785 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
786 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
787 * the buffer is valid and we do not have to do anything.
790 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
791 int cnt, struct ucred * cred)
796 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
797 if (inmem(vp, *rablkno))
799 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
801 if ((rabp->b_flags & B_CACHE) == 0) {
802 if (!TD_IS_IDLETHREAD(curthread))
803 curthread->td_ru.ru_inblock++;
804 rabp->b_flags |= B_ASYNC;
805 rabp->b_flags &= ~B_INVAL;
806 rabp->b_ioflags &= ~BIO_ERROR;
807 rabp->b_iocmd = BIO_READ;
808 if (rabp->b_rcred == NOCRED && cred != NOCRED)
809 rabp->b_rcred = crhold(cred);
810 vfs_busy_pages(rabp, 0);
812 rabp->b_iooffset = dbtob(rabp->b_blkno);
821 * Entry point for bread() and breadn() via #defines in sys/buf.h.
823 * Get a buffer with the specified data. Look in the cache first. We
824 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
825 * is set, the buffer is valid and we do not have to do anything, see
826 * getblk(). Also starts asynchronous I/O on read-ahead blocks.
829 breadn_flags(struct vnode * vp, daddr_t blkno, int size,
830 daddr_t * rablkno, int *rabsize, int cnt,
831 struct ucred * cred, int flags, struct buf **bpp)
834 int rv = 0, readwait = 0;
836 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
838 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
840 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
844 /* if not found in cache, do some I/O */
845 if ((bp->b_flags & B_CACHE) == 0) {
846 if (!TD_IS_IDLETHREAD(curthread))
847 curthread->td_ru.ru_inblock++;
848 bp->b_iocmd = BIO_READ;
849 bp->b_flags &= ~B_INVAL;
850 bp->b_ioflags &= ~BIO_ERROR;
851 if (bp->b_rcred == NOCRED && cred != NOCRED)
852 bp->b_rcred = crhold(cred);
853 vfs_busy_pages(bp, 0);
854 bp->b_iooffset = dbtob(bp->b_blkno);
859 breada(vp, rablkno, rabsize, cnt, cred);
868 * Write, release buffer on completion. (Done by iodone
869 * if async). Do not bother writing anything if the buffer
872 * Note that we set B_CACHE here, indicating that buffer is
873 * fully valid and thus cacheable. This is true even of NFS
874 * now so we set it generally. This could be set either here
875 * or in biodone() since the I/O is synchronous. We put it
879 bufwrite(struct buf *bp)
885 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
886 if (bp->b_flags & B_INVAL) {
891 oldflags = bp->b_flags;
895 if (bp->b_pin_count > 0)
898 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
899 ("FFS background buffer should not get here %p", bp));
903 vp_md = vp->v_vflag & VV_MD;
907 /* Mark the buffer clean */
910 bp->b_flags &= ~B_DONE;
911 bp->b_ioflags &= ~BIO_ERROR;
912 bp->b_flags |= B_CACHE;
913 bp->b_iocmd = BIO_WRITE;
915 bufobj_wref(bp->b_bufobj);
916 vfs_busy_pages(bp, 1);
919 * Normal bwrites pipeline writes
921 bp->b_runningbufspace = bp->b_bufsize;
922 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
924 if (!TD_IS_IDLETHREAD(curthread))
925 curthread->td_ru.ru_oublock++;
926 if (oldflags & B_ASYNC)
928 bp->b_iooffset = dbtob(bp->b_blkno);
931 if ((oldflags & B_ASYNC) == 0) {
932 int rtval = bufwait(bp);
937 * don't allow the async write to saturate the I/O
938 * system. We will not deadlock here because
939 * we are blocking waiting for I/O that is already in-progress
940 * to complete. We do not block here if it is the update
941 * or syncer daemon trying to clean up as that can lead
944 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
945 waitrunningbufspace();
952 bufbdflush(struct bufobj *bo, struct buf *bp)
956 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
957 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
959 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
962 * Try to find a buffer to flush.
964 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
965 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
967 LK_EXCLUSIVE | LK_NOWAIT, NULL))
970 panic("bdwrite: found ourselves");
972 /* Don't countdeps with the bo lock held. */
973 if (buf_countdeps(nbp, 0)) {
978 if (nbp->b_flags & B_CLUSTEROK) {
984 dirtybufferflushes++;
993 * Delayed write. (Buffer is marked dirty). Do not bother writing
994 * anything if the buffer is marked invalid.
996 * Note that since the buffer must be completely valid, we can safely
997 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
998 * biodone() in order to prevent getblk from writing the buffer
1002 bdwrite(struct buf *bp)
1004 struct thread *td = curthread;
1008 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1009 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1010 BUF_ASSERT_HELD(bp);
1012 if (bp->b_flags & B_INVAL) {
1018 * If we have too many dirty buffers, don't create any more.
1019 * If we are wildly over our limit, then force a complete
1020 * cleanup. Otherwise, just keep the situation from getting
1021 * out of control. Note that we have to avoid a recursive
1022 * disaster and not try to clean up after our own cleanup!
1026 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
1027 td->td_pflags |= TDP_INBDFLUSH;
1029 td->td_pflags &= ~TDP_INBDFLUSH;
1035 * Set B_CACHE, indicating that the buffer is fully valid. This is
1036 * true even of NFS now.
1038 bp->b_flags |= B_CACHE;
1041 * This bmap keeps the system from needing to do the bmap later,
1042 * perhaps when the system is attempting to do a sync. Since it
1043 * is likely that the indirect block -- or whatever other datastructure
1044 * that the filesystem needs is still in memory now, it is a good
1045 * thing to do this. Note also, that if the pageout daemon is
1046 * requesting a sync -- there might not be enough memory to do
1047 * the bmap then... So, this is important to do.
1049 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1050 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1054 * Set the *dirty* buffer range based upon the VM system dirty
1057 * Mark the buffer pages as clean. We need to do this here to
1058 * satisfy the vnode_pager and the pageout daemon, so that it
1059 * thinks that the pages have been "cleaned". Note that since
1060 * the pages are in a delayed write buffer -- the VFS layer
1061 * "will" see that the pages get written out on the next sync,
1062 * or perhaps the cluster will be completed.
1064 vfs_clean_pages_dirty_buf(bp);
1068 * Wakeup the buffer flushing daemon if we have a lot of dirty
1069 * buffers (midpoint between our recovery point and our stall
1072 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1075 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1076 * due to the softdep code.
1083 * Turn buffer into delayed write request. We must clear BIO_READ and
1084 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1085 * itself to properly update it in the dirty/clean lists. We mark it
1086 * B_DONE to ensure that any asynchronization of the buffer properly
1087 * clears B_DONE ( else a panic will occur later ).
1089 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1090 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1091 * should only be called if the buffer is known-good.
1093 * Since the buffer is not on a queue, we do not update the numfreebuffers
1096 * The buffer must be on QUEUE_NONE.
1099 bdirty(struct buf *bp)
1102 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1103 bp, bp->b_vp, bp->b_flags);
1104 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1105 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1106 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1107 BUF_ASSERT_HELD(bp);
1108 bp->b_flags &= ~(B_RELBUF);
1109 bp->b_iocmd = BIO_WRITE;
1111 if ((bp->b_flags & B_DELWRI) == 0) {
1112 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1114 atomic_add_int(&numdirtybuffers, 1);
1115 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1122 * Clear B_DELWRI for buffer.
1124 * Since the buffer is not on a queue, we do not update the numfreebuffers
1127 * The buffer must be on QUEUE_NONE.
1131 bundirty(struct buf *bp)
1134 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1135 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1136 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1137 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1138 BUF_ASSERT_HELD(bp);
1140 if (bp->b_flags & B_DELWRI) {
1141 bp->b_flags &= ~B_DELWRI;
1143 atomic_subtract_int(&numdirtybuffers, 1);
1144 numdirtywakeup(lodirtybuffers);
1147 * Since it is now being written, we can clear its deferred write flag.
1149 bp->b_flags &= ~B_DEFERRED;
1155 * Asynchronous write. Start output on a buffer, but do not wait for
1156 * it to complete. The buffer is released when the output completes.
1158 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1159 * B_INVAL buffers. Not us.
1162 bawrite(struct buf *bp)
1165 bp->b_flags |= B_ASYNC;
1172 * Called prior to the locking of any vnodes when we are expecting to
1173 * write. We do not want to starve the buffer cache with too many
1174 * dirty buffers so we block here. By blocking prior to the locking
1175 * of any vnodes we attempt to avoid the situation where a locked vnode
1176 * prevents the various system daemons from flushing related buffers.
1183 if (numdirtybuffers >= hidirtybuffers) {
1185 while (numdirtybuffers >= hidirtybuffers) {
1187 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1188 msleep(&needsbuffer, &nblock,
1189 (PRIBIO + 4), "flswai", 0);
1191 mtx_unlock(&nblock);
1196 * Return true if we have too many dirty buffers.
1199 buf_dirty_count_severe(void)
1202 return(numdirtybuffers >= hidirtybuffers);
1205 static __noinline int
1206 buf_vm_page_count_severe(void)
1209 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1211 return vm_page_count_severe();
1217 * Release a busy buffer and, if requested, free its resources. The
1218 * buffer will be stashed in the appropriate bufqueue[] allowing it
1219 * to be accessed later as a cache entity or reused for other purposes.
1222 brelse(struct buf *bp)
1224 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1225 bp, bp->b_vp, bp->b_flags);
1226 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1227 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1229 if (bp->b_flags & B_MANAGED) {
1234 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1235 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1237 * Failed write, redirty. Must clear BIO_ERROR to prevent
1238 * pages from being scrapped. If the error is anything
1239 * other than an I/O error (EIO), assume that retrying
1242 bp->b_ioflags &= ~BIO_ERROR;
1244 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1245 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1247 * Either a failed I/O or we were asked to free or not
1250 bp->b_flags |= B_INVAL;
1251 if (!LIST_EMPTY(&bp->b_dep))
1253 if (bp->b_flags & B_DELWRI) {
1254 atomic_subtract_int(&numdirtybuffers, 1);
1255 numdirtywakeup(lodirtybuffers);
1257 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1258 if ((bp->b_flags & B_VMIO) == 0) {
1267 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1268 * is called with B_DELWRI set, the underlying pages may wind up
1269 * getting freed causing a previous write (bdwrite()) to get 'lost'
1270 * because pages associated with a B_DELWRI bp are marked clean.
1272 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1273 * if B_DELWRI is set.
1275 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1276 * on pages to return pages to the VM page queues.
1278 if (bp->b_flags & B_DELWRI)
1279 bp->b_flags &= ~B_RELBUF;
1280 else if (buf_vm_page_count_severe()) {
1282 * The locking of the BO_LOCK is not necessary since
1283 * BKGRDINPROG cannot be set while we hold the buf
1284 * lock, it can only be cleared if it is already
1288 if (!(bp->b_vflags & BV_BKGRDINPROG))
1289 bp->b_flags |= B_RELBUF;
1291 bp->b_flags |= B_RELBUF;
1295 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1296 * constituted, not even NFS buffers now. Two flags effect this. If
1297 * B_INVAL, the struct buf is invalidated but the VM object is kept
1298 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1300 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1301 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1302 * buffer is also B_INVAL because it hits the re-dirtying code above.
1304 * Normally we can do this whether a buffer is B_DELWRI or not. If
1305 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1306 * the commit state and we cannot afford to lose the buffer. If the
1307 * buffer has a background write in progress, we need to keep it
1308 * around to prevent it from being reconstituted and starting a second
1311 if ((bp->b_flags & B_VMIO)
1312 && !(bp->b_vp->v_mount != NULL &&
1313 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1314 !vn_isdisk(bp->b_vp, NULL) &&
1315 (bp->b_flags & B_DELWRI))
1324 obj = bp->b_bufobj->bo_object;
1327 * Get the base offset and length of the buffer. Note that
1328 * in the VMIO case if the buffer block size is not
1329 * page-aligned then b_data pointer may not be page-aligned.
1330 * But our b_pages[] array *IS* page aligned.
1332 * block sizes less then DEV_BSIZE (usually 512) are not
1333 * supported due to the page granularity bits (m->valid,
1334 * m->dirty, etc...).
1336 * See man buf(9) for more information
1338 resid = bp->b_bufsize;
1339 foff = bp->b_offset;
1340 VM_OBJECT_LOCK(obj);
1341 for (i = 0; i < bp->b_npages; i++) {
1347 * If we hit a bogus page, fixup *all* the bogus pages
1350 if (m == bogus_page) {
1351 poff = OFF_TO_IDX(bp->b_offset);
1354 for (j = i; j < bp->b_npages; j++) {
1356 mtmp = bp->b_pages[j];
1357 if (mtmp == bogus_page) {
1358 mtmp = vm_page_lookup(obj, poff + j);
1360 panic("brelse: page missing\n");
1362 bp->b_pages[j] = mtmp;
1366 if ((bp->b_flags & B_INVAL) == 0) {
1368 trunc_page((vm_offset_t)bp->b_data),
1369 bp->b_pages, bp->b_npages);
1373 if ((bp->b_flags & B_NOCACHE) ||
1374 (bp->b_ioflags & BIO_ERROR &&
1375 bp->b_iocmd == BIO_READ)) {
1376 int poffset = foff & PAGE_MASK;
1377 int presid = resid > (PAGE_SIZE - poffset) ?
1378 (PAGE_SIZE - poffset) : resid;
1380 KASSERT(presid >= 0, ("brelse: extra page"));
1381 vm_page_set_invalid(m, poffset, presid);
1383 printf("avoided corruption bug in bogus_page/brelse code\n");
1385 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1386 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1388 VM_OBJECT_UNLOCK(obj);
1389 if (bp->b_flags & (B_INVAL | B_RELBUF))
1390 vfs_vmio_release(bp);
1392 } else if (bp->b_flags & B_VMIO) {
1394 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1395 vfs_vmio_release(bp);
1398 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1399 if (bp->b_bufsize != 0)
1401 if (bp->b_vp != NULL)
1405 if (BUF_LOCKRECURSED(bp)) {
1406 /* do not release to free list */
1413 /* Handle delayed bremfree() processing. */
1414 if (bp->b_flags & B_REMFREE) {
1424 if (bp->b_qindex != QUEUE_NONE)
1425 panic("brelse: free buffer onto another queue???");
1428 * If the buffer has junk contents signal it and eventually
1429 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1432 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1433 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1434 bp->b_flags |= B_INVAL;
1435 if (bp->b_flags & B_INVAL) {
1436 if (bp->b_flags & B_DELWRI)
1442 /* buffers with no memory */
1443 if (bp->b_bufsize == 0) {
1444 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1445 if (bp->b_vflags & BV_BKGRDINPROG)
1446 panic("losing buffer 1");
1447 if (bp->b_kvasize) {
1448 bp->b_qindex = QUEUE_EMPTYKVA;
1450 bp->b_qindex = QUEUE_EMPTY;
1452 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1453 /* buffers with junk contents */
1454 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1455 (bp->b_ioflags & BIO_ERROR)) {
1456 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1457 if (bp->b_vflags & BV_BKGRDINPROG)
1458 panic("losing buffer 2");
1459 bp->b_qindex = QUEUE_CLEAN;
1460 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1461 /* remaining buffers */
1463 if (bp->b_flags & B_DELWRI)
1464 bp->b_qindex = QUEUE_DIRTY;
1466 bp->b_qindex = QUEUE_CLEAN;
1467 if (bp->b_flags & B_AGE)
1468 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1470 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1472 mtx_unlock(&bqlock);
1475 * Fixup numfreebuffers count. The bp is on an appropriate queue
1476 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1477 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1478 * if B_INVAL is set ).
1481 if (!(bp->b_flags & B_DELWRI)) {
1493 * Something we can maybe free or reuse
1495 if (bp->b_bufsize || bp->b_kvasize)
1498 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1499 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1500 panic("brelse: not dirty");
1506 * Release a buffer back to the appropriate queue but do not try to free
1507 * it. The buffer is expected to be used again soon.
1509 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1510 * biodone() to requeue an async I/O on completion. It is also used when
1511 * known good buffers need to be requeued but we think we may need the data
1514 * XXX we should be able to leave the B_RELBUF hint set on completion.
1517 bqrelse(struct buf *bp)
1521 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1522 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1523 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1525 if (BUF_LOCKRECURSED(bp)) {
1526 /* do not release to free list */
1532 if (bp->b_flags & B_MANAGED) {
1533 if (bp->b_flags & B_REMFREE) {
1540 mtx_unlock(&bqlock);
1542 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1548 /* Handle delayed bremfree() processing. */
1549 if (bp->b_flags & B_REMFREE) {
1556 if (bp->b_qindex != QUEUE_NONE)
1557 panic("bqrelse: free buffer onto another queue???");
1558 /* buffers with stale but valid contents */
1559 if (bp->b_flags & B_DELWRI) {
1560 bp->b_qindex = QUEUE_DIRTY;
1561 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1564 * The locking of the BO_LOCK for checking of the
1565 * BV_BKGRDINPROG is not necessary since the
1566 * BV_BKGRDINPROG cannot be set while we hold the buf
1567 * lock, it can only be cleared if it is already
1570 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1571 bp->b_qindex = QUEUE_CLEAN;
1572 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1576 * We are too low on memory, we have to try to free
1577 * the buffer (most importantly: the wired pages
1578 * making up its backing store) *now*.
1580 mtx_unlock(&bqlock);
1585 mtx_unlock(&bqlock);
1587 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) {
1596 * Something we can maybe free or reuse.
1598 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1601 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1602 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1603 panic("bqrelse: not dirty");
1608 /* Give pages used by the bp back to the VM system (where possible) */
1610 vfs_vmio_release(struct buf *bp)
1615 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages);
1616 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1617 for (i = 0; i < bp->b_npages; i++) {
1619 bp->b_pages[i] = NULL;
1621 * In order to keep page LRU ordering consistent, put
1622 * everything on the inactive queue.
1625 vm_page_unwire(m, 0);
1627 * We don't mess with busy pages, it is
1628 * the responsibility of the process that
1629 * busied the pages to deal with them.
1631 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 &&
1632 m->wire_count == 0) {
1634 * Might as well free the page if we can and it has
1635 * no valid data. We also free the page if the
1636 * buffer was used for direct I/O
1638 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) {
1640 } else if (bp->b_flags & B_DIRECT) {
1641 vm_page_try_to_free(m);
1642 } else if (buf_vm_page_count_severe()) {
1643 vm_page_try_to_cache(m);
1648 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1650 if (bp->b_bufsize) {
1655 bp->b_flags &= ~B_VMIO;
1661 * Check to see if a block at a particular lbn is available for a clustered
1665 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1672 /* If the buf isn't in core skip it */
1673 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1676 /* If the buf is busy we don't want to wait for it */
1677 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1680 /* Only cluster with valid clusterable delayed write buffers */
1681 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1682 (B_DELWRI | B_CLUSTEROK))
1685 if (bpa->b_bufsize != size)
1689 * Check to see if it is in the expected place on disk and that the
1690 * block has been mapped.
1692 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1702 * Implement clustered async writes for clearing out B_DELWRI buffers.
1703 * This is much better then the old way of writing only one buffer at
1704 * a time. Note that we may not be presented with the buffers in the
1705 * correct order, so we search for the cluster in both directions.
1708 vfs_bio_awrite(struct buf *bp)
1713 daddr_t lblkno = bp->b_lblkno;
1714 struct vnode *vp = bp->b_vp;
1722 * right now we support clustered writing only to regular files. If
1723 * we find a clusterable block we could be in the middle of a cluster
1724 * rather then at the beginning.
1726 if ((vp->v_type == VREG) &&
1727 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1728 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1730 size = vp->v_mount->mnt_stat.f_iosize;
1731 maxcl = MAXPHYS / size;
1734 for (i = 1; i < maxcl; i++)
1735 if (vfs_bio_clcheck(vp, size, lblkno + i,
1736 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1739 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1740 if (vfs_bio_clcheck(vp, size, lblkno - j,
1741 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1747 * this is a possible cluster write
1751 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1756 bp->b_flags |= B_ASYNC;
1758 * default (old) behavior, writing out only one block
1760 * XXX returns b_bufsize instead of b_bcount for nwritten?
1762 nwritten = bp->b_bufsize;
1771 * Find and initialize a new buffer header, freeing up existing buffers
1772 * in the bufqueues as necessary. The new buffer is returned locked.
1774 * Important: B_INVAL is not set. If the caller wishes to throw the
1775 * buffer away, the caller must set B_INVAL prior to calling brelse().
1778 * We have insufficient buffer headers
1779 * We have insufficient buffer space
1780 * buffer_map is too fragmented ( space reservation fails )
1781 * If we have to flush dirty buffers ( but we try to avoid this )
1783 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1784 * Instead we ask the buf daemon to do it for us. We attempt to
1785 * avoid piecemeal wakeups of the pageout daemon.
1789 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1797 static int flushingbufs;
1801 * We can't afford to block since we might be holding a vnode lock,
1802 * which may prevent system daemons from running. We deal with
1803 * low-memory situations by proactively returning memory and running
1804 * async I/O rather then sync I/O.
1806 atomic_add_int(&getnewbufcalls, 1);
1807 atomic_subtract_int(&getnewbufrestarts, 1);
1809 atomic_add_int(&getnewbufrestarts, 1);
1812 * Setup for scan. If we do not have enough free buffers,
1813 * we setup a degenerate case that immediately fails. Note
1814 * that if we are specially marked process, we are allowed to
1815 * dip into our reserves.
1817 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1819 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1820 * However, there are a number of cases (defragging, reusing, ...)
1821 * where we cannot backup.
1824 nqindex = QUEUE_EMPTYKVA;
1825 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1829 * If no EMPTYKVA buffers and we are either
1830 * defragging or reusing, locate a CLEAN buffer
1831 * to free or reuse. If bufspace useage is low
1832 * skip this step so we can allocate a new buffer.
1834 if (defrag || bufspace >= lobufspace) {
1835 nqindex = QUEUE_CLEAN;
1836 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1840 * If we could not find or were not allowed to reuse a
1841 * CLEAN buffer, check to see if it is ok to use an EMPTY
1842 * buffer. We can only use an EMPTY buffer if allocating
1843 * its KVA would not otherwise run us out of buffer space.
1845 if (nbp == NULL && defrag == 0 &&
1846 bufspace + maxsize < hibufspace) {
1847 nqindex = QUEUE_EMPTY;
1848 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1853 * Run scan, possibly freeing data and/or kva mappings on the fly
1857 while ((bp = nbp) != NULL) {
1858 int qindex = nqindex;
1861 * Calculate next bp ( we can only use it if we do not block
1862 * or do other fancy things ).
1864 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1867 nqindex = QUEUE_EMPTYKVA;
1868 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1871 case QUEUE_EMPTYKVA:
1872 nqindex = QUEUE_CLEAN;
1873 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1884 * If we are defragging then we need a buffer with
1885 * b_kvasize != 0. XXX this situation should no longer
1886 * occur, if defrag is non-zero the buffer's b_kvasize
1887 * should also be non-zero at this point. XXX
1889 if (defrag && bp->b_kvasize == 0) {
1890 printf("Warning: defrag empty buffer %p\n", bp);
1895 * Start freeing the bp. This is somewhat involved. nbp
1896 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1898 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1901 BO_LOCK(bp->b_bufobj);
1902 if (bp->b_vflags & BV_BKGRDINPROG) {
1903 BO_UNLOCK(bp->b_bufobj);
1907 BO_UNLOCK(bp->b_bufobj);
1910 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1911 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1912 bp->b_kvasize, bp->b_bufsize, qindex);
1917 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1920 * Note: we no longer distinguish between VMIO and non-VMIO
1924 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1926 if (bp->b_bufobj != NULL)
1927 BO_LOCK(bp->b_bufobj);
1929 if (bp->b_bufobj != NULL)
1930 BO_UNLOCK(bp->b_bufobj);
1931 mtx_unlock(&bqlock);
1933 if (qindex == QUEUE_CLEAN) {
1934 if (bp->b_flags & B_VMIO) {
1935 bp->b_flags &= ~B_ASYNC;
1936 vfs_vmio_release(bp);
1943 * NOTE: nbp is now entirely invalid. We can only restart
1944 * the scan from this point on.
1946 * Get the rest of the buffer freed up. b_kva* is still
1947 * valid after this operation.
1950 if (bp->b_rcred != NOCRED) {
1951 crfree(bp->b_rcred);
1952 bp->b_rcred = NOCRED;
1954 if (bp->b_wcred != NOCRED) {
1955 crfree(bp->b_wcred);
1956 bp->b_wcred = NOCRED;
1958 if (!LIST_EMPTY(&bp->b_dep))
1960 if (bp->b_vflags & BV_BKGRDINPROG)
1961 panic("losing buffer 3");
1962 KASSERT(bp->b_vp == NULL,
1963 ("bp: %p still has vnode %p. qindex: %d",
1964 bp, bp->b_vp, qindex));
1965 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1966 ("bp: %p still on a buffer list. xflags %X",
1975 KASSERT((bp->b_vflags & BV_INFREECNT) == 0,
1976 ("buf %p still counted as free?", bp));
1979 bp->b_blkno = bp->b_lblkno = 0;
1980 bp->b_offset = NOOFFSET;
1986 bp->b_dirtyoff = bp->b_dirtyend = 0;
1987 bp->b_bufobj = NULL;
1988 bp->b_pin_count = 0;
1989 bp->b_fsprivate1 = NULL;
1990 bp->b_fsprivate2 = NULL;
1991 bp->b_fsprivate3 = NULL;
1993 LIST_INIT(&bp->b_dep);
1996 * If we are defragging then free the buffer.
1999 bp->b_flags |= B_INVAL;
2007 * Notify any waiters for the buffer lock about
2008 * identity change by freeing the buffer.
2010 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
2011 bp->b_flags |= B_INVAL;
2018 * If we are overcomitted then recover the buffer and its
2019 * KVM space. This occurs in rare situations when multiple
2020 * processes are blocked in getnewbuf() or allocbuf().
2022 if (bufspace >= hibufspace)
2024 if (flushingbufs && bp->b_kvasize != 0) {
2025 bp->b_flags |= B_INVAL;
2030 if (bufspace < lobufspace)
2036 * If we exhausted our list, sleep as appropriate. We may have to
2037 * wakeup various daemons and write out some dirty buffers.
2039 * Generally we are sleeping due to insufficient buffer space.
2043 int flags, norunbuf;
2048 flags = VFS_BIO_NEED_BUFSPACE;
2050 } else if (bufspace >= hibufspace) {
2052 flags = VFS_BIO_NEED_BUFSPACE;
2055 flags = VFS_BIO_NEED_ANY;
2058 needsbuffer |= flags;
2059 mtx_unlock(&nblock);
2060 mtx_unlock(&bqlock);
2062 bd_speedup(); /* heeeelp */
2063 if (gbflags & GB_NOWAIT_BD)
2067 while (needsbuffer & flags) {
2068 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2069 mtx_unlock(&nblock);
2071 * getblk() is called with a vnode
2072 * locked, and some majority of the
2073 * dirty buffers may as well belong to
2074 * the vnode. Flushing the buffers
2075 * there would make a progress that
2076 * cannot be achieved by the
2077 * buf_daemon, that cannot lock the
2080 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2081 (td->td_pflags & TDP_NORUNNINGBUF);
2082 /* play bufdaemon */
2083 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2084 fl = buf_do_flush(vp);
2085 td->td_pflags &= norunbuf;
2089 if ((needsbuffer & flags) == 0)
2092 if (msleep(&needsbuffer, &nblock,
2093 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2094 mtx_unlock(&nblock);
2098 mtx_unlock(&nblock);
2101 * We finally have a valid bp. We aren't quite out of the
2102 * woods, we still have to reserve kva space. In order
2103 * to keep fragmentation sane we only allocate kva in
2106 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2108 if (maxsize != bp->b_kvasize) {
2109 vm_offset_t addr = 0;
2113 vm_map_lock(buffer_map);
2114 if (vm_map_findspace(buffer_map,
2115 vm_map_min(buffer_map), maxsize, &addr)) {
2117 * Uh oh. Buffer map is to fragmented. We
2118 * must defragment the map.
2120 atomic_add_int(&bufdefragcnt, 1);
2121 vm_map_unlock(buffer_map);
2123 bp->b_flags |= B_INVAL;
2128 vm_map_insert(buffer_map, NULL, 0,
2129 addr, addr + maxsize,
2130 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2132 bp->b_kvabase = (caddr_t) addr;
2133 bp->b_kvasize = maxsize;
2134 atomic_add_long(&bufspace, bp->b_kvasize);
2135 atomic_add_int(&bufreusecnt, 1);
2137 vm_map_unlock(buffer_map);
2139 bp->b_saveaddr = bp->b_kvabase;
2140 bp->b_data = bp->b_saveaddr;
2148 * buffer flushing daemon. Buffers are normally flushed by the
2149 * update daemon but if it cannot keep up this process starts to
2150 * take the load in an attempt to prevent getnewbuf() from blocking.
2153 static struct kproc_desc buf_kp = {
2158 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2161 buf_do_flush(struct vnode *vp)
2165 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2168 * Could not find any buffers without rollback
2169 * dependencies, so just write the first one
2170 * in the hopes of eventually making progress.
2172 flushbufqueues(vp, QUEUE_DIRTY, 1);
2183 * This process needs to be suspended prior to shutdown sync.
2185 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2189 * This process is allowed to take the buffer cache to the limit
2191 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2195 mtx_unlock(&bdlock);
2197 kproc_suspend_check(bufdaemonproc);
2198 lodirtysave = lodirtybuffers;
2199 if (bd_speedupreq) {
2200 lodirtybuffers = numdirtybuffers / 2;
2204 * Do the flush. Limit the amount of in-transit I/O we
2205 * allow to build up, otherwise we would completely saturate
2206 * the I/O system. Wakeup any waiting processes before we
2207 * normally would so they can run in parallel with our drain.
2209 while (numdirtybuffers > lodirtybuffers) {
2210 if (buf_do_flush(NULL) == 0)
2212 kern_yield(PRI_UNCHANGED);
2214 lodirtybuffers = lodirtysave;
2217 * Only clear bd_request if we have reached our low water
2218 * mark. The buf_daemon normally waits 1 second and
2219 * then incrementally flushes any dirty buffers that have
2220 * built up, within reason.
2222 * If we were unable to hit our low water mark and couldn't
2223 * find any flushable buffers, we sleep half a second.
2224 * Otherwise we loop immediately.
2227 if (numdirtybuffers <= lodirtybuffers) {
2229 * We reached our low water mark, reset the
2230 * request and sleep until we are needed again.
2231 * The sleep is just so the suspend code works.
2234 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2237 * We couldn't find any flushable dirty buffers but
2238 * still have too many dirty buffers, we
2239 * have to sleep and try again. (rare)
2241 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2249 * Try to flush a buffer in the dirty queue. We must be careful to
2250 * free up B_INVAL buffers instead of write them, which NFS is
2251 * particularly sensitive to.
2253 static int flushwithdeps = 0;
2254 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2255 0, "Number of buffers flushed with dependecies that require rollbacks");
2258 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2260 struct buf *sentinel;
2269 target = numdirtybuffers - lodirtybuffers;
2270 if (flushdeps && target > 2)
2273 target = flushbufqtarget;
2276 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2277 sentinel->b_qindex = QUEUE_SENTINEL;
2279 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2280 while (flushed != target) {
2281 bp = TAILQ_NEXT(sentinel, b_freelist);
2283 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2284 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2289 * Skip sentinels inserted by other invocations of the
2290 * flushbufqueues(), taking care to not reorder them.
2292 if (bp->b_qindex == QUEUE_SENTINEL)
2295 * Only flush the buffers that belong to the
2296 * vnode locked by the curthread.
2298 if (lvp != NULL && bp->b_vp != lvp)
2300 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2302 if (bp->b_pin_count > 0) {
2306 BO_LOCK(bp->b_bufobj);
2307 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2308 (bp->b_flags & B_DELWRI) == 0) {
2309 BO_UNLOCK(bp->b_bufobj);
2313 BO_UNLOCK(bp->b_bufobj);
2314 if (bp->b_flags & B_INVAL) {
2316 mtx_unlock(&bqlock);
2319 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2324 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2325 if (flushdeps == 0) {
2333 * We must hold the lock on a vnode before writing
2334 * one of its buffers. Otherwise we may confuse, or
2335 * in the case of a snapshot vnode, deadlock the
2338 * The lock order here is the reverse of the normal
2339 * of vnode followed by buf lock. This is ok because
2340 * the NOWAIT will prevent deadlock.
2343 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2347 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2348 mtx_unlock(&bqlock);
2349 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2350 bp, bp->b_vp, bp->b_flags);
2351 if (curproc == bufdaemonproc)
2358 vn_finished_write(mp);
2360 flushwithdeps += hasdeps;
2364 * Sleeping on runningbufspace while holding
2365 * vnode lock leads to deadlock.
2367 if (curproc == bufdaemonproc)
2368 waitrunningbufspace();
2369 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2373 vn_finished_write(mp);
2376 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2377 mtx_unlock(&bqlock);
2378 free(sentinel, M_TEMP);
2383 * Check to see if a block is currently memory resident.
2386 incore(struct bufobj *bo, daddr_t blkno)
2391 bp = gbincore(bo, blkno);
2397 * Returns true if no I/O is needed to access the
2398 * associated VM object. This is like incore except
2399 * it also hunts around in the VM system for the data.
2403 inmem(struct vnode * vp, daddr_t blkno)
2406 vm_offset_t toff, tinc, size;
2410 ASSERT_VOP_LOCKED(vp, "inmem");
2412 if (incore(&vp->v_bufobj, blkno))
2414 if (vp->v_mount == NULL)
2421 if (size > vp->v_mount->mnt_stat.f_iosize)
2422 size = vp->v_mount->mnt_stat.f_iosize;
2423 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2425 VM_OBJECT_LOCK(obj);
2426 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2427 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2431 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2432 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2433 if (vm_page_is_valid(m,
2434 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2437 VM_OBJECT_UNLOCK(obj);
2441 VM_OBJECT_UNLOCK(obj);
2446 * Set the dirty range for a buffer based on the status of the dirty
2447 * bits in the pages comprising the buffer. The range is limited
2448 * to the size of the buffer.
2450 * Tell the VM system that the pages associated with this buffer
2451 * are clean. This is used for delayed writes where the data is
2452 * going to go to disk eventually without additional VM intevention.
2454 * Note that while we only really need to clean through to b_bcount, we
2455 * just go ahead and clean through to b_bufsize.
2458 vfs_clean_pages_dirty_buf(struct buf *bp)
2460 vm_ooffset_t foff, noff, eoff;
2464 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0)
2467 foff = bp->b_offset;
2468 KASSERT(bp->b_offset != NOOFFSET,
2469 ("vfs_clean_pages_dirty_buf: no buffer offset"));
2471 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2472 vfs_drain_busy_pages(bp);
2473 vfs_setdirty_locked_object(bp);
2474 for (i = 0; i < bp->b_npages; i++) {
2475 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2477 if (eoff > bp->b_offset + bp->b_bufsize)
2478 eoff = bp->b_offset + bp->b_bufsize;
2480 vfs_page_set_validclean(bp, foff, m);
2481 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
2484 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2488 vfs_setdirty_locked_object(struct buf *bp)
2493 object = bp->b_bufobj->bo_object;
2494 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2497 * We qualify the scan for modified pages on whether the
2498 * object has been flushed yet.
2500 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) {
2501 vm_offset_t boffset;
2502 vm_offset_t eoffset;
2505 * test the pages to see if they have been modified directly
2506 * by users through the VM system.
2508 for (i = 0; i < bp->b_npages; i++)
2509 vm_page_test_dirty(bp->b_pages[i]);
2512 * Calculate the encompassing dirty range, boffset and eoffset,
2513 * (eoffset - boffset) bytes.
2516 for (i = 0; i < bp->b_npages; i++) {
2517 if (bp->b_pages[i]->dirty)
2520 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2522 for (i = bp->b_npages - 1; i >= 0; --i) {
2523 if (bp->b_pages[i]->dirty) {
2527 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2530 * Fit it to the buffer.
2533 if (eoffset > bp->b_bcount)
2534 eoffset = bp->b_bcount;
2537 * If we have a good dirty range, merge with the existing
2541 if (boffset < eoffset) {
2542 if (bp->b_dirtyoff > boffset)
2543 bp->b_dirtyoff = boffset;
2544 if (bp->b_dirtyend < eoffset)
2545 bp->b_dirtyend = eoffset;
2553 * Get a block given a specified block and offset into a file/device.
2554 * The buffers B_DONE bit will be cleared on return, making it almost
2555 * ready for an I/O initiation. B_INVAL may or may not be set on
2556 * return. The caller should clear B_INVAL prior to initiating a
2559 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2560 * an existing buffer.
2562 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2563 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2564 * and then cleared based on the backing VM. If the previous buffer is
2565 * non-0-sized but invalid, B_CACHE will be cleared.
2567 * If getblk() must create a new buffer, the new buffer is returned with
2568 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2569 * case it is returned with B_INVAL clear and B_CACHE set based on the
2572 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2573 * B_CACHE bit is clear.
2575 * What this means, basically, is that the caller should use B_CACHE to
2576 * determine whether the buffer is fully valid or not and should clear
2577 * B_INVAL prior to issuing a read. If the caller intends to validate
2578 * the buffer by loading its data area with something, the caller needs
2579 * to clear B_INVAL. If the caller does this without issuing an I/O,
2580 * the caller should set B_CACHE ( as an optimization ), else the caller
2581 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2582 * a write attempt or if it was a successfull read. If the caller
2583 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2584 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2587 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2594 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2595 ASSERT_VOP_LOCKED(vp, "getblk");
2596 if (size > MAXBSIZE)
2597 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2602 * Block if we are low on buffers. Certain processes are allowed
2603 * to completely exhaust the buffer cache.
2605 * If this check ever becomes a bottleneck it may be better to
2606 * move it into the else, when gbincore() fails. At the moment
2607 * it isn't a problem.
2609 * XXX remove if 0 sections (clean this up after its proven)
2611 if (numfreebuffers == 0) {
2612 if (TD_IS_IDLETHREAD(curthread))
2615 needsbuffer |= VFS_BIO_NEED_ANY;
2616 mtx_unlock(&nblock);
2620 bp = gbincore(bo, blkno);
2624 * Buffer is in-core. If the buffer is not busy nor managed,
2625 * it must be on a queue.
2627 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2629 if (flags & GB_LOCK_NOWAIT)
2630 lockflags |= LK_NOWAIT;
2632 error = BUF_TIMELOCK(bp, lockflags,
2633 BO_MTX(bo), "getblk", slpflag, slptimeo);
2636 * If we slept and got the lock we have to restart in case
2637 * the buffer changed identities.
2639 if (error == ENOLCK)
2641 /* We timed out or were interrupted. */
2646 * The buffer is locked. B_CACHE is cleared if the buffer is
2647 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2648 * and for a VMIO buffer B_CACHE is adjusted according to the
2651 if (bp->b_flags & B_INVAL)
2652 bp->b_flags &= ~B_CACHE;
2653 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2654 bp->b_flags |= B_CACHE;
2655 if (bp->b_flags & B_MANAGED)
2656 MPASS(bp->b_qindex == QUEUE_NONE);
2664 * check for size inconsistancies for non-VMIO case.
2667 if (bp->b_bcount != size) {
2668 if ((bp->b_flags & B_VMIO) == 0 ||
2669 (size > bp->b_kvasize)) {
2670 if (bp->b_flags & B_DELWRI) {
2672 * If buffer is pinned and caller does
2673 * not want sleep waiting for it to be
2674 * unpinned, bail out
2676 if (bp->b_pin_count > 0) {
2677 if (flags & GB_LOCK_NOWAIT) {
2684 bp->b_flags |= B_NOCACHE;
2687 if (LIST_EMPTY(&bp->b_dep)) {
2688 bp->b_flags |= B_RELBUF;
2691 bp->b_flags |= B_NOCACHE;
2700 * If the size is inconsistant in the VMIO case, we can resize
2701 * the buffer. This might lead to B_CACHE getting set or
2702 * cleared. If the size has not changed, B_CACHE remains
2703 * unchanged from its previous state.
2706 if (bp->b_bcount != size)
2709 KASSERT(bp->b_offset != NOOFFSET,
2710 ("getblk: no buffer offset"));
2713 * A buffer with B_DELWRI set and B_CACHE clear must
2714 * be committed before we can return the buffer in
2715 * order to prevent the caller from issuing a read
2716 * ( due to B_CACHE not being set ) and overwriting
2719 * Most callers, including NFS and FFS, need this to
2720 * operate properly either because they assume they
2721 * can issue a read if B_CACHE is not set, or because
2722 * ( for example ) an uncached B_DELWRI might loop due
2723 * to softupdates re-dirtying the buffer. In the latter
2724 * case, B_CACHE is set after the first write completes,
2725 * preventing further loops.
2726 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2727 * above while extending the buffer, we cannot allow the
2728 * buffer to remain with B_CACHE set after the write
2729 * completes or it will represent a corrupt state. To
2730 * deal with this we set B_NOCACHE to scrap the buffer
2733 * We might be able to do something fancy, like setting
2734 * B_CACHE in bwrite() except if B_DELWRI is already set,
2735 * so the below call doesn't set B_CACHE, but that gets real
2736 * confusing. This is much easier.
2739 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2740 bp->b_flags |= B_NOCACHE;
2744 bp->b_flags &= ~B_DONE;
2746 int bsize, maxsize, vmio;
2750 * Buffer is not in-core, create new buffer. The buffer
2751 * returned by getnewbuf() is locked. Note that the returned
2752 * buffer is also considered valid (not marked B_INVAL).
2756 * If the user does not want us to create the buffer, bail out
2759 if (flags & GB_NOCREAT)
2761 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize;
2762 offset = blkno * bsize;
2763 vmio = vp->v_object != NULL;
2764 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2765 maxsize = imax(maxsize, bsize);
2767 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2769 if (slpflag || slptimeo)
2775 * This code is used to make sure that a buffer is not
2776 * created while the getnewbuf routine is blocked.
2777 * This can be a problem whether the vnode is locked or not.
2778 * If the buffer is created out from under us, we have to
2779 * throw away the one we just created.
2781 * Note: this must occur before we associate the buffer
2782 * with the vp especially considering limitations in
2783 * the splay tree implementation when dealing with duplicate
2787 if (gbincore(bo, blkno)) {
2789 bp->b_flags |= B_INVAL;
2795 * Insert the buffer into the hash, so that it can
2796 * be found by incore.
2798 bp->b_blkno = bp->b_lblkno = blkno;
2799 bp->b_offset = offset;
2804 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2805 * buffer size starts out as 0, B_CACHE will be set by
2806 * allocbuf() for the VMIO case prior to it testing the
2807 * backing store for validity.
2811 bp->b_flags |= B_VMIO;
2812 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2813 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2814 bp, vp->v_object, bp->b_bufobj->bo_object));
2816 bp->b_flags &= ~B_VMIO;
2817 KASSERT(bp->b_bufobj->bo_object == NULL,
2818 ("ARGH! has b_bufobj->bo_object %p %p\n",
2819 bp, bp->b_bufobj->bo_object));
2823 bp->b_flags &= ~B_DONE;
2825 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2826 BUF_ASSERT_HELD(bp);
2827 KASSERT(bp->b_bufobj == bo,
2828 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2833 * Get an empty, disassociated buffer of given size. The buffer is initially
2837 geteblk(int size, int flags)
2842 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2843 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2844 if ((flags & GB_NOWAIT_BD) &&
2845 (curthread->td_pflags & TDP_BUFNEED) != 0)
2849 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2850 BUF_ASSERT_HELD(bp);
2856 * This code constitutes the buffer memory from either anonymous system
2857 * memory (in the case of non-VMIO operations) or from an associated
2858 * VM object (in the case of VMIO operations). This code is able to
2859 * resize a buffer up or down.
2861 * Note that this code is tricky, and has many complications to resolve
2862 * deadlock or inconsistant data situations. Tread lightly!!!
2863 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2864 * the caller. Calling this code willy nilly can result in the loss of data.
2866 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2867 * B_CACHE for the non-VMIO case.
2871 allocbuf(struct buf *bp, int size)
2873 int newbsize, mbsize;
2876 BUF_ASSERT_HELD(bp);
2878 if (bp->b_kvasize < size)
2879 panic("allocbuf: buffer too small");
2881 if ((bp->b_flags & B_VMIO) == 0) {
2885 * Just get anonymous memory from the kernel. Don't
2886 * mess with B_CACHE.
2888 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2889 if (bp->b_flags & B_MALLOC)
2892 newbsize = round_page(size);
2894 if (newbsize < bp->b_bufsize) {
2896 * malloced buffers are not shrunk
2898 if (bp->b_flags & B_MALLOC) {
2900 bp->b_bcount = size;
2902 free(bp->b_data, M_BIOBUF);
2903 if (bp->b_bufsize) {
2904 atomic_subtract_long(
2910 bp->b_saveaddr = bp->b_kvabase;
2911 bp->b_data = bp->b_saveaddr;
2913 bp->b_flags &= ~B_MALLOC;
2917 vm_hold_free_pages(bp, newbsize);
2918 } else if (newbsize > bp->b_bufsize) {
2920 * We only use malloced memory on the first allocation.
2921 * and revert to page-allocated memory when the buffer
2925 * There is a potential smp race here that could lead
2926 * to bufmallocspace slightly passing the max. It
2927 * is probably extremely rare and not worth worrying
2930 if ( (bufmallocspace < maxbufmallocspace) &&
2931 (bp->b_bufsize == 0) &&
2932 (mbsize <= PAGE_SIZE/2)) {
2934 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2935 bp->b_bufsize = mbsize;
2936 bp->b_bcount = size;
2937 bp->b_flags |= B_MALLOC;
2938 atomic_add_long(&bufmallocspace, mbsize);
2944 * If the buffer is growing on its other-than-first allocation,
2945 * then we revert to the page-allocation scheme.
2947 if (bp->b_flags & B_MALLOC) {
2948 origbuf = bp->b_data;
2949 origbufsize = bp->b_bufsize;
2950 bp->b_data = bp->b_kvabase;
2951 if (bp->b_bufsize) {
2952 atomic_subtract_long(&bufmallocspace,
2957 bp->b_flags &= ~B_MALLOC;
2958 newbsize = round_page(newbsize);
2962 (vm_offset_t) bp->b_data + bp->b_bufsize,
2963 (vm_offset_t) bp->b_data + newbsize);
2965 bcopy(origbuf, bp->b_data, origbufsize);
2966 free(origbuf, M_BIOBUF);
2972 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2973 desiredpages = (size == 0) ? 0 :
2974 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2976 if (bp->b_flags & B_MALLOC)
2977 panic("allocbuf: VMIO buffer can't be malloced");
2979 * Set B_CACHE initially if buffer is 0 length or will become
2982 if (size == 0 || bp->b_bufsize == 0)
2983 bp->b_flags |= B_CACHE;
2985 if (newbsize < bp->b_bufsize) {
2987 * DEV_BSIZE aligned new buffer size is less then the
2988 * DEV_BSIZE aligned existing buffer size. Figure out
2989 * if we have to remove any pages.
2991 if (desiredpages < bp->b_npages) {
2994 pmap_qremove((vm_offset_t)trunc_page(
2995 (vm_offset_t)bp->b_data) +
2996 (desiredpages << PAGE_SHIFT),
2997 (bp->b_npages - desiredpages));
2998 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2999 for (i = desiredpages; i < bp->b_npages; i++) {
3001 * the page is not freed here -- it
3002 * is the responsibility of
3003 * vnode_pager_setsize
3006 KASSERT(m != bogus_page,
3007 ("allocbuf: bogus page found"));
3008 while (vm_page_sleep_if_busy(m, TRUE,
3012 bp->b_pages[i] = NULL;
3014 vm_page_unwire(m, 0);
3017 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3018 bp->b_npages = desiredpages;
3020 } else if (size > bp->b_bcount) {
3022 * We are growing the buffer, possibly in a
3023 * byte-granular fashion.
3030 * Step 1, bring in the VM pages from the object,
3031 * allocating them if necessary. We must clear
3032 * B_CACHE if these pages are not valid for the
3033 * range covered by the buffer.
3036 obj = bp->b_bufobj->bo_object;
3038 VM_OBJECT_LOCK(obj);
3039 while (bp->b_npages < desiredpages) {
3043 * We must allocate system pages since blocking
3044 * here could interfere with paging I/O, no
3045 * matter which process we are.
3047 * We can only test VPO_BUSY here. Blocking on
3048 * m->busy might lead to a deadlock:
3049 * vm_fault->getpages->cluster_read->allocbuf
3050 * Thus, we specify VM_ALLOC_IGN_SBUSY.
3052 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) +
3053 bp->b_npages, VM_ALLOC_NOBUSY |
3054 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED |
3055 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY |
3056 VM_ALLOC_COUNT(desiredpages - bp->b_npages));
3058 bp->b_flags &= ~B_CACHE;
3059 bp->b_pages[bp->b_npages] = m;
3064 * Step 2. We've loaded the pages into the buffer,
3065 * we have to figure out if we can still have B_CACHE
3066 * set. Note that B_CACHE is set according to the
3067 * byte-granular range ( bcount and size ), new the
3068 * aligned range ( newbsize ).
3070 * The VM test is against m->valid, which is DEV_BSIZE
3071 * aligned. Needless to say, the validity of the data
3072 * needs to also be DEV_BSIZE aligned. Note that this
3073 * fails with NFS if the server or some other client
3074 * extends the file's EOF. If our buffer is resized,
3075 * B_CACHE may remain set! XXX
3078 toff = bp->b_bcount;
3079 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3081 while ((bp->b_flags & B_CACHE) && toff < size) {
3084 if (tinc > (size - toff))
3087 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3100 VM_OBJECT_UNLOCK(obj);
3103 * Step 3, fixup the KVM pmap. Remember that
3104 * bp->b_data is relative to bp->b_offset, but
3105 * bp->b_offset may be offset into the first page.
3108 bp->b_data = (caddr_t)
3109 trunc_page((vm_offset_t)bp->b_data);
3111 (vm_offset_t)bp->b_data,
3116 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3117 (vm_offset_t)(bp->b_offset & PAGE_MASK));
3120 if (newbsize < bp->b_bufsize)
3122 bp->b_bufsize = newbsize; /* actual buffer allocation */
3123 bp->b_bcount = size; /* requested buffer size */
3128 biodone(struct bio *bp)
3131 void (*done)(struct bio *);
3133 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3135 bp->bio_flags |= BIO_DONE;
3136 done = bp->bio_done;
3145 * Wait for a BIO to finish.
3147 * XXX: resort to a timeout for now. The optimal locking (if any) for this
3148 * case is not yet clear.
3151 biowait(struct bio *bp, const char *wchan)
3155 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3157 while ((bp->bio_flags & BIO_DONE) == 0)
3158 msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3160 if (bp->bio_error != 0)
3161 return (bp->bio_error);
3162 if (!(bp->bio_flags & BIO_ERROR))
3168 biofinish(struct bio *bp, struct devstat *stat, int error)
3172 bp->bio_error = error;
3173 bp->bio_flags |= BIO_ERROR;
3176 devstat_end_transaction_bio(stat, bp);
3183 * Wait for buffer I/O completion, returning error status. The buffer
3184 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3185 * error and cleared.
3188 bufwait(struct buf *bp)
3190 if (bp->b_iocmd == BIO_READ)
3191 bwait(bp, PRIBIO, "biord");
3193 bwait(bp, PRIBIO, "biowr");
3194 if (bp->b_flags & B_EINTR) {
3195 bp->b_flags &= ~B_EINTR;
3198 if (bp->b_ioflags & BIO_ERROR) {
3199 return (bp->b_error ? bp->b_error : EIO);
3206 * Call back function from struct bio back up to struct buf.
3209 bufdonebio(struct bio *bip)
3213 bp = bip->bio_caller2;
3214 bp->b_resid = bp->b_bcount - bip->bio_completed;
3215 bp->b_resid = bip->bio_resid; /* XXX: remove */
3216 bp->b_ioflags = bip->bio_flags;
3217 bp->b_error = bip->bio_error;
3219 bp->b_ioflags |= BIO_ERROR;
3225 dev_strategy(struct cdev *dev, struct buf *bp)
3231 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3232 panic("b_iocmd botch");
3237 /* Try again later */
3238 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3240 bip->bio_cmd = bp->b_iocmd;
3241 bip->bio_offset = bp->b_iooffset;
3242 bip->bio_length = bp->b_bcount;
3243 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3244 bip->bio_data = bp->b_data;
3245 bip->bio_done = bufdonebio;
3246 bip->bio_caller2 = bp;
3248 KASSERT(dev->si_refcount > 0,
3249 ("dev_strategy on un-referenced struct cdev *(%s)",
3251 csw = dev_refthread(dev, &ref);
3254 bp->b_error = ENXIO;
3255 bp->b_ioflags = BIO_ERROR;
3259 (*csw->d_strategy)(bip);
3260 dev_relthread(dev, ref);
3266 * Finish I/O on a buffer, optionally calling a completion function.
3267 * This is usually called from an interrupt so process blocking is
3270 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3271 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3272 * assuming B_INVAL is clear.
3274 * For the VMIO case, we set B_CACHE if the op was a read and no
3275 * read error occured, or if the op was a write. B_CACHE is never
3276 * set if the buffer is invalid or otherwise uncacheable.
3278 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3279 * initiator to leave B_INVAL set to brelse the buffer out of existance
3280 * in the biodone routine.
3283 bufdone(struct buf *bp)
3285 struct bufobj *dropobj;
3286 void (*biodone)(struct buf *);
3288 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3291 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3292 BUF_ASSERT_HELD(bp);
3294 runningbufwakeup(bp);
3295 if (bp->b_iocmd == BIO_WRITE)
3296 dropobj = bp->b_bufobj;
3297 /* call optional completion function if requested */
3298 if (bp->b_iodone != NULL) {
3299 biodone = bp->b_iodone;
3300 bp->b_iodone = NULL;
3303 bufobj_wdrop(dropobj);
3310 bufobj_wdrop(dropobj);
3314 bufdone_finish(struct buf *bp)
3316 BUF_ASSERT_HELD(bp);
3318 if (!LIST_EMPTY(&bp->b_dep))
3321 if (bp->b_flags & B_VMIO) {
3326 int bogus, i, iosize;
3328 obj = bp->b_bufobj->bo_object;
3329 KASSERT(obj->paging_in_progress >= bp->b_npages,
3330 ("biodone_finish: paging in progress(%d) < b_npages(%d)",
3331 obj->paging_in_progress, bp->b_npages));
3334 KASSERT(vp->v_holdcnt > 0,
3335 ("biodone_finish: vnode %p has zero hold count", vp));
3336 KASSERT(vp->v_object != NULL,
3337 ("biodone_finish: vnode %p has no vm_object", vp));
3339 foff = bp->b_offset;
3340 KASSERT(bp->b_offset != NOOFFSET,
3341 ("biodone_finish: bp %p has no buffer offset", bp));
3344 * Set B_CACHE if the op was a normal read and no error
3345 * occured. B_CACHE is set for writes in the b*write()
3348 iosize = bp->b_bcount - bp->b_resid;
3349 if (bp->b_iocmd == BIO_READ &&
3350 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3351 !(bp->b_ioflags & BIO_ERROR)) {
3352 bp->b_flags |= B_CACHE;
3355 VM_OBJECT_LOCK(obj);
3356 for (i = 0; i < bp->b_npages; i++) {
3360 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3365 * cleanup bogus pages, restoring the originals
3368 if (m == bogus_page) {
3369 bogus = bogusflag = 1;
3370 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3372 panic("biodone: page disappeared!");
3375 KASSERT(OFF_TO_IDX(foff) == m->pindex,
3376 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch",
3377 (intmax_t)foff, (uintmax_t)m->pindex));
3380 * In the write case, the valid and clean bits are
3381 * already changed correctly ( see bdwrite() ), so we
3382 * only need to do this here in the read case.
3384 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3385 KASSERT((m->dirty & vm_page_bits(foff &
3386 PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3387 " page %p has unexpected dirty bits", m));
3388 vfs_page_set_valid(bp, foff, m);
3391 vm_page_io_finish(m);
3392 vm_object_pip_subtract(obj, 1);
3393 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3396 vm_object_pip_wakeupn(obj, 0);
3397 VM_OBJECT_UNLOCK(obj);
3399 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3400 bp->b_pages, bp->b_npages);
3404 * For asynchronous completions, release the buffer now. The brelse
3405 * will do a wakeup there if necessary - so no need to do a wakeup
3406 * here in the async case. The sync case always needs to do a wakeup.
3409 if (bp->b_flags & B_ASYNC) {
3410 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3419 * This routine is called in lieu of iodone in the case of
3420 * incomplete I/O. This keeps the busy status for pages
3424 vfs_unbusy_pages(struct buf *bp)
3430 runningbufwakeup(bp);
3431 if (!(bp->b_flags & B_VMIO))
3434 obj = bp->b_bufobj->bo_object;
3435 VM_OBJECT_LOCK(obj);
3436 for (i = 0; i < bp->b_npages; i++) {
3438 if (m == bogus_page) {
3439 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3441 panic("vfs_unbusy_pages: page missing\n");
3443 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3444 bp->b_pages, bp->b_npages);
3446 vm_object_pip_subtract(obj, 1);
3447 vm_page_io_finish(m);
3449 vm_object_pip_wakeupn(obj, 0);
3450 VM_OBJECT_UNLOCK(obj);
3454 * vfs_page_set_valid:
3456 * Set the valid bits in a page based on the supplied offset. The
3457 * range is restricted to the buffer's size.
3459 * This routine is typically called after a read completes.
3462 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3467 * Compute the end offset, eoff, such that [off, eoff) does not span a
3468 * page boundary and eoff is not greater than the end of the buffer.
3469 * The end of the buffer, in this case, is our file EOF, not the
3470 * allocation size of the buffer.
3472 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3473 if (eoff > bp->b_offset + bp->b_bcount)
3474 eoff = bp->b_offset + bp->b_bcount;
3477 * Set valid range. This is typically the entire buffer and thus the
3481 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off);
3485 * vfs_page_set_validclean:
3487 * Set the valid bits and clear the dirty bits in a page based on the
3488 * supplied offset. The range is restricted to the buffer's size.
3491 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3493 vm_ooffset_t soff, eoff;
3496 * Start and end offsets in buffer. eoff - soff may not cross a
3497 * page boundry or cross the end of the buffer. The end of the
3498 * buffer, in this case, is our file EOF, not the allocation size
3502 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3503 if (eoff > bp->b_offset + bp->b_bcount)
3504 eoff = bp->b_offset + bp->b_bcount;
3507 * Set valid range. This is typically the entire buffer and thus the
3511 vm_page_set_validclean(
3513 (vm_offset_t) (soff & PAGE_MASK),
3514 (vm_offset_t) (eoff - soff)
3520 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If
3521 * any page is busy, drain the flag.
3524 vfs_drain_busy_pages(struct buf *bp)
3529 VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED);
3531 for (i = 0; i < bp->b_npages; i++) {
3533 if ((m->oflags & VPO_BUSY) != 0) {
3534 for (; last_busied < i; last_busied++)
3535 vm_page_busy(bp->b_pages[last_busied]);
3536 while ((m->oflags & VPO_BUSY) != 0)
3537 vm_page_sleep(m, "vbpage");
3540 for (i = 0; i < last_busied; i++)
3541 vm_page_wakeup(bp->b_pages[i]);
3545 * This routine is called before a device strategy routine.
3546 * It is used to tell the VM system that paging I/O is in
3547 * progress, and treat the pages associated with the buffer
3548 * almost as being VPO_BUSY. Also the object paging_in_progress
3549 * flag is handled to make sure that the object doesn't become
3552 * Since I/O has not been initiated yet, certain buffer flags
3553 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3554 * and should be ignored.
3557 vfs_busy_pages(struct buf *bp, int clear_modify)
3564 if (!(bp->b_flags & B_VMIO))
3567 obj = bp->b_bufobj->bo_object;
3568 foff = bp->b_offset;
3569 KASSERT(bp->b_offset != NOOFFSET,
3570 ("vfs_busy_pages: no buffer offset"));
3571 VM_OBJECT_LOCK(obj);
3572 vfs_drain_busy_pages(bp);
3573 if (bp->b_bufsize != 0)
3574 vfs_setdirty_locked_object(bp);
3576 for (i = 0; i < bp->b_npages; i++) {
3579 if ((bp->b_flags & B_CLUSTER) == 0) {
3580 vm_object_pip_add(obj, 1);
3581 vm_page_io_start(m);
3584 * When readying a buffer for a read ( i.e
3585 * clear_modify == 0 ), it is important to do
3586 * bogus_page replacement for valid pages in
3587 * partially instantiated buffers. Partially
3588 * instantiated buffers can, in turn, occur when
3589 * reconstituting a buffer from its VM backing store
3590 * base. We only have to do this if B_CACHE is
3591 * clear ( which causes the I/O to occur in the
3592 * first place ). The replacement prevents the read
3593 * I/O from overwriting potentially dirty VM-backed
3594 * pages. XXX bogus page replacement is, uh, bogus.
3595 * It may not work properly with small-block devices.
3596 * We need to find a better way.
3599 pmap_remove_write(m);
3600 vfs_page_set_validclean(bp, foff, m);
3601 } else if (m->valid == VM_PAGE_BITS_ALL &&
3602 (bp->b_flags & B_CACHE) == 0) {
3603 bp->b_pages[i] = bogus_page;
3606 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3608 VM_OBJECT_UNLOCK(obj);
3610 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3611 bp->b_pages, bp->b_npages);
3615 * vfs_bio_set_valid:
3617 * Set the range within the buffer to valid. The range is
3618 * relative to the beginning of the buffer, b_offset. Note that
3619 * b_offset itself may be offset from the beginning of the first
3623 vfs_bio_set_valid(struct buf *bp, int base, int size)
3628 if (!(bp->b_flags & B_VMIO))
3632 * Fixup base to be relative to beginning of first page.
3633 * Set initial n to be the maximum number of bytes in the
3634 * first page that can be validated.
3636 base += (bp->b_offset & PAGE_MASK);
3637 n = PAGE_SIZE - (base & PAGE_MASK);
3639 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3640 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3644 vm_page_set_valid_range(m, base & PAGE_MASK, n);
3649 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3655 * If the specified buffer is a non-VMIO buffer, clear the entire
3656 * buffer. If the specified buffer is a VMIO buffer, clear and
3657 * validate only the previously invalid portions of the buffer.
3658 * This routine essentially fakes an I/O, so we need to clear
3659 * BIO_ERROR and B_INVAL.
3661 * Note that while we only theoretically need to clear through b_bcount,
3662 * we go ahead and clear through b_bufsize.
3665 vfs_bio_clrbuf(struct buf *bp)
3670 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3674 bp->b_flags &= ~B_INVAL;
3675 bp->b_ioflags &= ~BIO_ERROR;
3676 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3677 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3678 (bp->b_offset & PAGE_MASK) == 0) {
3679 if (bp->b_pages[0] == bogus_page)
3681 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3682 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3683 if ((bp->b_pages[0]->valid & mask) == mask)
3685 if ((bp->b_pages[0]->valid & mask) == 0) {
3686 bzero(bp->b_data, bp->b_bufsize);
3687 bp->b_pages[0]->valid |= mask;
3691 ea = sa = bp->b_data;
3692 for(i = 0; i < bp->b_npages; i++, sa = ea) {
3693 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3694 ea = (caddr_t)(vm_offset_t)ulmin(
3695 (u_long)(vm_offset_t)ea,
3696 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3697 if (bp->b_pages[i] == bogus_page)
3699 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3700 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3701 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3702 if ((bp->b_pages[i]->valid & mask) == mask)
3704 if ((bp->b_pages[i]->valid & mask) == 0)
3707 for (; sa < ea; sa += DEV_BSIZE, j++) {
3708 if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3709 bzero(sa, DEV_BSIZE);
3712 bp->b_pages[i]->valid |= mask;
3715 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3720 * vm_hold_load_pages and vm_hold_free_pages get pages into
3721 * a buffers address space. The pages are anonymous and are
3722 * not associated with a file object.
3725 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3731 to = round_page(to);
3732 from = round_page(from);
3733 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3735 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3738 * note: must allocate system pages since blocking here
3739 * could interfere with paging I/O, no matter which
3742 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ |
3743 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT));
3748 pmap_qenter(pg, &p, 1);
3749 bp->b_pages[index] = p;
3751 bp->b_npages = index;
3754 /* Return pages associated with this buf to the vm system */
3756 vm_hold_free_pages(struct buf *bp, int newbsize)
3760 int index, newnpages;
3762 from = round_page((vm_offset_t)bp->b_data + newbsize);
3763 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3764 if (bp->b_npages > newnpages)
3765 pmap_qremove(from, bp->b_npages - newnpages);
3766 for (index = newnpages; index < bp->b_npages; index++) {
3767 p = bp->b_pages[index];
3768 bp->b_pages[index] = NULL;
3770 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3771 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno);
3774 atomic_subtract_int(&cnt.v_wire_count, 1);
3776 bp->b_npages = newnpages;
3780 * Map an IO request into kernel virtual address space.
3782 * All requests are (re)mapped into kernel VA space.
3783 * Notice that we use b_bufsize for the size of the buffer
3784 * to be mapped. b_bcount might be modified by the driver.
3786 * Note that even if the caller determines that the address space should
3787 * be valid, a race or a smaller-file mapped into a larger space may
3788 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3789 * check the return value.
3792 vmapbuf(struct buf *bp)
3798 if (bp->b_bufsize < 0)
3800 prot = VM_PROT_READ;
3801 if (bp->b_iocmd == BIO_READ)
3802 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
3803 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
3804 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages,
3805 btoc(MAXPHYS))) < 0)
3807 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3809 kva = bp->b_saveaddr;
3810 bp->b_npages = pidx;
3811 bp->b_saveaddr = bp->b_data;
3812 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3817 * Free the io map PTEs associated with this IO operation.
3818 * We also invalidate the TLB entries and restore the original b_addr.
3821 vunmapbuf(struct buf *bp)
3825 npages = bp->b_npages;
3826 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3827 vm_page_unhold_pages(bp->b_pages, npages);
3829 bp->b_data = bp->b_saveaddr;
3833 bdone(struct buf *bp)
3837 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3839 bp->b_flags |= B_DONE;
3845 bwait(struct buf *bp, u_char pri, const char *wchan)
3849 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3851 while ((bp->b_flags & B_DONE) == 0)
3852 msleep(bp, mtxp, pri, wchan, 0);
3857 bufsync(struct bufobj *bo, int waitfor)
3860 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3864 bufstrategy(struct bufobj *bo, struct buf *bp)
3870 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3871 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3872 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3873 i = VOP_STRATEGY(vp, bp);
3874 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3878 bufobj_wrefl(struct bufobj *bo)
3881 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3882 ASSERT_BO_LOCKED(bo);
3887 bufobj_wref(struct bufobj *bo)
3890 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3897 bufobj_wdrop(struct bufobj *bo)
3900 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
3902 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
3903 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
3904 bo->bo_flag &= ~BO_WWAIT;
3905 wakeup(&bo->bo_numoutput);
3911 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
3915 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
3916 ASSERT_BO_LOCKED(bo);
3918 while (bo->bo_numoutput) {
3919 bo->bo_flag |= BO_WWAIT;
3920 error = msleep(&bo->bo_numoutput, BO_MTX(bo),
3921 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
3929 bpin(struct buf *bp)
3933 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3940 bunpin(struct buf *bp)
3944 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3946 if (--bp->b_pin_count == 0)
3952 bunpin_wait(struct buf *bp)
3956 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3958 while (bp->b_pin_count > 0)
3959 msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
3963 #include "opt_ddb.h"
3965 #include <ddb/ddb.h>
3967 /* DDB command to show buffer data */
3968 DB_SHOW_COMMAND(buffer, db_show_buffer)
3971 struct buf *bp = (struct buf *)addr;
3974 db_printf("usage: show buffer <addr>\n");
3978 db_printf("buf at %p\n", bp);
3979 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n",
3980 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags,
3981 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS);
3983 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
3984 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, "
3986 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3987 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
3988 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first);
3991 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3992 for (i = 0; i < bp->b_npages; i++) {
3995 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3996 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3997 if ((i + 1) < bp->b_npages)
4003 BUF_LOCKPRINTINFO(bp);
4006 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4011 for (i = 0; i < nbuf; i++) {
4013 if (BUF_ISLOCKED(bp)) {
4014 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4020 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4026 db_printf("usage: show vnodebufs <addr>\n");
4029 vp = (struct vnode *)addr;
4030 db_printf("Clean buffers:\n");
4031 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4032 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4035 db_printf("Dirty buffers:\n");
4036 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4037 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4042 DB_COMMAND(countfreebufs, db_coundfreebufs)
4045 int i, used = 0, nfree = 0;
4048 db_printf("usage: countfreebufs\n");
4052 for (i = 0; i < nbuf; i++) {
4054 if ((bp->b_vflags & BV_INFREECNT) != 0)
4060 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used,
4062 db_printf("numfreebuffers is %d\n", numfreebuffers);