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, vm_offset_t from,
100 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
102 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m);
103 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off,
105 static void vfs_clean_pages(struct buf *bp);
106 static void vfs_setdirty(struct buf *bp);
107 static void vfs_setdirty_locked_object(struct buf *bp);
108 static void vfs_vmio_release(struct buf *bp);
109 static int vfs_bio_clcheck(struct vnode *vp, int size,
110 daddr_t lblkno, daddr_t blkno);
111 static int buf_do_flush(struct vnode *vp);
112 static int flushbufqueues(struct vnode *, int, int);
113 static void buf_daemon(void);
114 static void bremfreel(struct buf *bp);
115 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
116 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
117 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS);
120 int vmiodirenable = TRUE;
121 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0,
122 "Use the VM system for directory writes");
123 long runningbufspace;
124 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
125 "Amount of presently outstanding async buffer io");
126 static long bufspace;
127 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
128 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
129 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD,
130 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers");
132 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
133 "Virtual memory used for buffers");
135 static long maxbufspace;
136 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
137 "Maximum allowed value of bufspace (including buf_daemon)");
138 static long bufmallocspace;
139 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
140 "Amount of malloced memory for buffers");
141 static long maxbufmallocspace;
142 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0,
143 "Maximum amount of malloced memory for buffers");
144 static long lobufspace;
145 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
146 "Minimum amount of buffers we want to have");
148 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
149 "Maximum allowed value of bufspace (excluding buf_daemon)");
150 static int bufreusecnt;
151 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0,
152 "Number of times we have reused a buffer");
153 static int buffreekvacnt;
154 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0,
155 "Number of times we have freed the KVA space from some buffer");
156 static int bufdefragcnt;
157 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0,
158 "Number of times we have had to repeat buffer allocation to defragment");
159 static long lorunningspace;
160 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
161 "Minimum preferred space used for in-progress I/O");
162 static long hirunningspace;
163 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
164 "Maximum amount of space to use for in-progress I/O");
165 int dirtybufferflushes;
166 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes,
167 0, "Number of bdwrite to bawrite conversions to limit dirty buffers");
169 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip,
170 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk");
171 int altbufferflushes;
172 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes,
173 0, "Number of fsync flushes to limit dirty buffers");
174 static int recursiveflushes;
175 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes,
176 0, "Number of flushes skipped due to being recursive");
177 static int numdirtybuffers;
178 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
179 "Number of buffers that are dirty (has unwritten changes) at the moment");
180 static int lodirtybuffers;
181 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
182 "How many buffers we want to have free before bufdaemon can sleep");
183 static int hidirtybuffers;
184 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
185 "When the number of dirty buffers is considered severe");
187 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh,
188 0, "Number of bdwrite to bawrite conversions to clear dirty buffers");
189 static int numfreebuffers;
190 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
191 "Number of free buffers");
192 static int lofreebuffers;
193 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
195 static int hifreebuffers;
196 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
197 "XXX Complicatedly unused");
198 static int getnewbufcalls;
199 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0,
200 "Number of calls to getnewbuf");
201 static int getnewbufrestarts;
202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0,
203 "Number of times getnewbuf has had to restart a buffer aquisition");
204 static int flushbufqtarget = 100;
205 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0,
206 "Amount of work to do in flushbufqueues when helping bufdaemon");
207 static long notbufdflashes;
208 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0,
209 "Number of dirty buffer flushes done by the bufdaemon helpers");
212 * Wakeup point for bufdaemon, as well as indicator of whether it is already
213 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it
216 static int bd_request;
219 * This lock synchronizes access to bd_request.
221 static struct mtx bdlock;
224 * bogus page -- for I/O to/from partially complete buffers
225 * this is a temporary solution to the problem, but it is not
226 * really that bad. it would be better to split the buffer
227 * for input in the case of buffers partially already in memory,
228 * but the code is intricate enough already.
230 vm_page_t bogus_page;
233 * Synchronization (sleep/wakeup) variable for active buffer space requests.
234 * Set when wait starts, cleared prior to wakeup().
235 * Used in runningbufwakeup() and waitrunningbufspace().
237 static int runningbufreq;
240 * This lock protects the runningbufreq and synchronizes runningbufwakeup and
241 * waitrunningbufspace().
243 static struct mtx rbreqlock;
246 * Synchronization (sleep/wakeup) variable for buffer requests.
247 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done
249 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(),
250 * getnewbuf(), and getblk().
252 static int needsbuffer;
255 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it.
257 static struct mtx nblock;
260 * Definitions for the buffer free lists.
262 #define BUFFER_QUEUES 6 /* number of free buffer queues */
264 #define QUEUE_NONE 0 /* on no queue */
265 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */
266 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */
267 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */
268 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */
269 #define QUEUE_EMPTY 5 /* empty buffer headers */
270 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */
272 /* Queues for free buffers with various properties */
273 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } };
275 /* Lock for the bufqueues */
276 static struct mtx bqlock;
279 * Single global constant for BUF_WMESG, to avoid getting multiple references.
280 * buf_wmesg is referred from macros.
282 const char *buf_wmesg = BUF_WMESG;
284 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
285 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
286 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
287 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
289 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \
290 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7)
292 sysctl_bufspace(SYSCTL_HANDLER_ARGS)
297 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long))
298 return (sysctl_handle_long(oidp, arg1, arg2, req));
299 lvalue = *(long *)arg1;
300 if (lvalue > INT_MAX)
301 /* On overflow, still write out a long to trigger ENOMEM. */
302 return (sysctl_handle_long(oidp, &lvalue, 0, req));
304 return (sysctl_handle_int(oidp, &ivalue, 0, req));
309 extern void ffs_rawread_setup(void);
310 #endif /* DIRECTIO */
314 * If someone is blocked due to there being too many dirty buffers,
315 * and numdirtybuffers is now reasonable, wake them up.
319 numdirtywakeup(int level)
322 if (numdirtybuffers <= level) {
324 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
325 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
326 wakeup(&needsbuffer);
335 * Called when buffer space is potentially available for recovery.
336 * getnewbuf() will block on this flag when it is unable to free
337 * sufficient buffer space. Buffer space becomes recoverable when
338 * bp's get placed back in the queues.
346 * If someone is waiting for BUF space, wake them up. Even
347 * though we haven't freed the kva space yet, the waiting
348 * process will be able to now.
351 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
352 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
353 wakeup(&needsbuffer);
359 * runningbufwakeup() - in-progress I/O accounting.
363 runningbufwakeup(struct buf *bp)
366 if (bp->b_runningbufspace) {
367 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace);
368 bp->b_runningbufspace = 0;
369 mtx_lock(&rbreqlock);
370 if (runningbufreq && runningbufspace <= lorunningspace) {
372 wakeup(&runningbufreq);
374 mtx_unlock(&rbreqlock);
381 * Called when a buffer has been added to one of the free queues to
382 * account for the buffer and to wakeup anyone waiting for free buffers.
383 * This typically occurs when large amounts of metadata are being handled
384 * by the buffer cache ( else buffer space runs out first, usually ).
391 atomic_add_int(&numfreebuffers, 1);
394 needsbuffer &= ~VFS_BIO_NEED_ANY;
395 if (numfreebuffers >= hifreebuffers)
396 needsbuffer &= ~VFS_BIO_NEED_FREE;
397 wakeup(&needsbuffer);
403 * waitrunningbufspace()
405 * runningbufspace is a measure of the amount of I/O currently
406 * running. This routine is used in async-write situations to
407 * prevent creating huge backups of pending writes to a device.
408 * Only asynchronous writes are governed by this function.
410 * Reads will adjust runningbufspace, but will not block based on it.
411 * The read load has a side effect of reducing the allowed write load.
413 * This does NOT turn an async write into a sync write. It waits
414 * for earlier writes to complete and generally returns before the
415 * caller's write has reached the device.
418 waitrunningbufspace(void)
421 mtx_lock(&rbreqlock);
422 while (runningbufspace > hirunningspace) {
424 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0);
426 mtx_unlock(&rbreqlock);
431 * vfs_buf_test_cache:
433 * Called when a buffer is extended. This function clears the B_CACHE
434 * bit if the newly extended portion of the buffer does not contain
439 vfs_buf_test_cache(struct buf *bp,
440 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
444 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
445 if (bp->b_flags & B_CACHE) {
446 int base = (foff + off) & PAGE_MASK;
447 if (vm_page_is_valid(m, base, size) == 0)
448 bp->b_flags &= ~B_CACHE;
452 /* Wake up the buffer daemon if necessary */
455 bd_wakeup(int dirtybuflevel)
459 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
467 * bd_speedup - speedup the buffer cache flushing code
479 * Calculating buffer cache scaling values and reserve space for buffer
480 * headers. This is called during low level kernel initialization and
481 * may be called more then once. We CANNOT write to the memory area
482 * being reserved at this time.
485 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est)
491 * physmem_est is in pages. Convert it to kilobytes (assumes
492 * PAGE_SIZE is >= 1K)
494 physmem_est = physmem_est * (PAGE_SIZE / 1024);
497 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE.
498 * For the first 64MB of ram nominally allocate sufficient buffers to
499 * cover 1/4 of our ram. Beyond the first 64MB allocate additional
500 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing
501 * the buffer cache we limit the eventual kva reservation to
504 * factor represents the 1/4 x ram conversion.
507 int factor = 4 * BKVASIZE / 1024;
510 if (physmem_est > 4096)
511 nbuf += min((physmem_est - 4096) / factor,
513 if (physmem_est > 65536)
514 nbuf += (physmem_est - 65536) * 2 / (factor * 5);
516 if (maxbcache && nbuf > maxbcache / BKVASIZE)
517 nbuf = maxbcache / BKVASIZE;
522 /* XXX Avoid unsigned long overflows later on with maxbufspace. */
523 maxbuf = (LONG_MAX / 3) / BKVASIZE;
526 printf("Warning: nbufs lowered from %d to %ld\n", nbuf,
532 * swbufs are used as temporary holders for I/O, such as paging I/O.
533 * We have no less then 16 and no more then 256.
535 nswbuf = max(min(nbuf/4, 256), 16);
537 if (nswbuf < NSWBUF_MIN)
545 * Reserve space for the buffer cache buffers
548 v = (caddr_t)(swbuf + nswbuf);
550 v = (caddr_t)(buf + nbuf);
555 /* Initialize the buffer subsystem. Called before use of any buffers. */
562 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF);
563 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF);
564 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF);
565 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF);
567 /* next, make a null set of free lists */
568 for (i = 0; i < BUFFER_QUEUES; i++)
569 TAILQ_INIT(&bufqueues[i]);
571 /* finally, initialize each buffer header and stick on empty q */
572 for (i = 0; i < nbuf; i++) {
574 bzero(bp, sizeof *bp);
575 bp->b_flags = B_INVAL; /* we're just an empty header */
576 bp->b_rcred = NOCRED;
577 bp->b_wcred = NOCRED;
578 bp->b_qindex = QUEUE_EMPTY;
581 LIST_INIT(&bp->b_dep);
583 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
587 * maxbufspace is the absolute maximum amount of buffer space we are
588 * allowed to reserve in KVM and in real terms. The absolute maximum
589 * is nominally used by buf_daemon. hibufspace is the nominal maximum
590 * used by most other processes. The differential is required to
591 * ensure that buf_daemon is able to run when other processes might
592 * be blocked waiting for buffer space.
594 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
595 * this may result in KVM fragmentation which is not handled optimally
598 maxbufspace = (long)nbuf * BKVASIZE;
599 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
600 lobufspace = hibufspace - MAXBSIZE;
602 lorunningspace = 512 * 1024;
603 hirunningspace = 1024 * 1024;
606 * Limit the amount of malloc memory since it is wired permanently into
607 * the kernel space. Even though this is accounted for in the buffer
608 * allocation, we don't want the malloced region to grow uncontrolled.
609 * The malloc scheme improves memory utilization significantly on average
610 * (small) directories.
612 maxbufmallocspace = hibufspace / 20;
615 * Reduce the chance of a deadlock occuring by limiting the number
616 * of delayed-write dirty buffers we allow to stack up.
618 hidirtybuffers = nbuf / 4 + 20;
619 dirtybufthresh = hidirtybuffers * 9 / 10;
622 * To support extreme low-memory systems, make sure hidirtybuffers cannot
623 * eat up all available buffer space. This occurs when our minimum cannot
624 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
625 * BKVASIZE'd (8K) buffers.
627 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
628 hidirtybuffers >>= 1;
630 lodirtybuffers = hidirtybuffers / 2;
633 * Try to keep the number of free buffers in the specified range,
634 * and give special processes (e.g. like buf_daemon) access to an
637 lofreebuffers = nbuf / 18 + 5;
638 hifreebuffers = 2 * lofreebuffers;
639 numfreebuffers = nbuf;
642 * Maximum number of async ops initiated per buf_daemon loop. This is
643 * somewhat of a hack at the moment, we really need to limit ourselves
644 * based on the number of bytes of I/O in-transit that were initiated
648 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
649 VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
653 * bfreekva() - free the kva allocation for a buffer.
655 * Since this call frees up buffer space, we call bufspacewakeup().
658 bfreekva(struct buf *bp)
662 atomic_add_int(&buffreekvacnt, 1);
663 atomic_subtract_long(&bufspace, bp->b_kvasize);
664 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase,
665 (vm_offset_t) bp->b_kvabase + bp->b_kvasize);
674 * Mark the buffer for removal from the appropriate free list in brelse.
678 bremfree(struct buf *bp)
681 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
682 KASSERT((bp->b_flags & B_REMFREE) == 0,
683 ("bremfree: buffer %p already marked for delayed removal.", bp));
684 KASSERT(bp->b_qindex != QUEUE_NONE,
685 ("bremfree: buffer %p not on a queue.", bp));
688 bp->b_flags |= B_REMFREE;
689 /* Fixup numfreebuffers count. */
690 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
691 atomic_subtract_int(&numfreebuffers, 1);
697 * Force an immediate removal from a free list. Used only in nfs when
698 * it abuses the b_freelist pointer.
701 bremfreef(struct buf *bp)
711 * Removes a buffer from the free list, must be called with the
715 bremfreel(struct buf *bp)
717 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X",
718 bp, bp->b_vp, bp->b_flags);
719 KASSERT(bp->b_qindex != QUEUE_NONE,
720 ("bremfreel: buffer %p not on a queue.", bp));
722 mtx_assert(&bqlock, MA_OWNED);
724 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
725 bp->b_qindex = QUEUE_NONE;
727 * If this was a delayed bremfree() we only need to remove the buffer
728 * from the queue and return the stats are already done.
730 if (bp->b_flags & B_REMFREE) {
731 bp->b_flags &= ~B_REMFREE;
735 * Fixup numfreebuffers count. If the buffer is invalid or not
736 * delayed-write, the buffer was free and we must decrement
739 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)
740 atomic_subtract_int(&numfreebuffers, 1);
745 * Get a buffer with the specified data. Look in the cache first. We
746 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
747 * is set, the buffer is valid and we do not have to do anything ( see
748 * getblk() ). This is really just a special case of breadn().
751 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
755 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
759 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must
760 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
761 * the buffer is valid and we do not have to do anything.
764 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
765 int cnt, struct ucred * cred)
770 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
771 if (inmem(vp, *rablkno))
773 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0);
775 if ((rabp->b_flags & B_CACHE) == 0) {
776 if (!TD_IS_IDLETHREAD(curthread))
777 curthread->td_ru.ru_inblock++;
778 rabp->b_flags |= B_ASYNC;
779 rabp->b_flags &= ~B_INVAL;
780 rabp->b_ioflags &= ~BIO_ERROR;
781 rabp->b_iocmd = BIO_READ;
782 if (rabp->b_rcred == NOCRED && cred != NOCRED)
783 rabp->b_rcred = crhold(cred);
784 vfs_busy_pages(rabp, 0);
786 rabp->b_iooffset = dbtob(rabp->b_blkno);
795 * Operates like bread, but also starts asynchronous I/O on
799 breadn(struct vnode * vp, daddr_t blkno, int size,
800 daddr_t * rablkno, int *rabsize,
801 int cnt, struct ucred * cred, struct buf **bpp)
804 int rv = 0, readwait = 0;
806 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
807 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
809 /* if not found in cache, do some I/O */
810 if ((bp->b_flags & B_CACHE) == 0) {
811 if (!TD_IS_IDLETHREAD(curthread))
812 curthread->td_ru.ru_inblock++;
813 bp->b_iocmd = BIO_READ;
814 bp->b_flags &= ~B_INVAL;
815 bp->b_ioflags &= ~BIO_ERROR;
816 if (bp->b_rcred == NOCRED && cred != NOCRED)
817 bp->b_rcred = crhold(cred);
818 vfs_busy_pages(bp, 0);
819 bp->b_iooffset = dbtob(bp->b_blkno);
824 breada(vp, rablkno, rabsize, cnt, cred);
833 * Write, release buffer on completion. (Done by iodone
834 * if async). Do not bother writing anything if the buffer
837 * Note that we set B_CACHE here, indicating that buffer is
838 * fully valid and thus cacheable. This is true even of NFS
839 * now so we set it generally. This could be set either here
840 * or in biodone() since the I/O is synchronous. We put it
844 bufwrite(struct buf *bp)
850 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
851 if (bp->b_flags & B_INVAL) {
856 oldflags = bp->b_flags;
860 if (bp->b_pin_count > 0)
863 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG),
864 ("FFS background buffer should not get here %p", bp));
868 vp_md = vp->v_vflag & VV_MD;
872 /* Mark the buffer clean */
875 bp->b_flags &= ~B_DONE;
876 bp->b_ioflags &= ~BIO_ERROR;
877 bp->b_flags |= B_CACHE;
878 bp->b_iocmd = BIO_WRITE;
880 bufobj_wref(bp->b_bufobj);
881 vfs_busy_pages(bp, 1);
884 * Normal bwrites pipeline writes
886 bp->b_runningbufspace = bp->b_bufsize;
887 atomic_add_long(&runningbufspace, bp->b_runningbufspace);
889 if (!TD_IS_IDLETHREAD(curthread))
890 curthread->td_ru.ru_oublock++;
891 if (oldflags & B_ASYNC)
893 bp->b_iooffset = dbtob(bp->b_blkno);
896 if ((oldflags & B_ASYNC) == 0) {
897 int rtval = bufwait(bp);
902 * don't allow the async write to saturate the I/O
903 * system. We will not deadlock here because
904 * we are blocking waiting for I/O that is already in-progress
905 * to complete. We do not block here if it is the update
906 * or syncer daemon trying to clean up as that can lead
909 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md)
910 waitrunningbufspace();
917 bufbdflush(struct bufobj *bo, struct buf *bp)
921 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) {
922 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread);
924 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) {
927 * Try to find a buffer to flush.
929 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
930 if ((nbp->b_vflags & BV_BKGRDINPROG) ||
932 LK_EXCLUSIVE | LK_NOWAIT, NULL))
935 panic("bdwrite: found ourselves");
937 /* Don't countdeps with the bo lock held. */
938 if (buf_countdeps(nbp, 0)) {
943 if (nbp->b_flags & B_CLUSTEROK) {
949 dirtybufferflushes++;
958 * Delayed write. (Buffer is marked dirty). Do not bother writing
959 * anything if the buffer is marked invalid.
961 * Note that since the buffer must be completely valid, we can safely
962 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
963 * biodone() in order to prevent getblk from writing the buffer
967 bdwrite(struct buf *bp)
969 struct thread *td = curthread;
973 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
974 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
977 if (bp->b_flags & B_INVAL) {
983 * If we have too many dirty buffers, don't create any more.
984 * If we are wildly over our limit, then force a complete
985 * cleanup. Otherwise, just keep the situation from getting
986 * out of control. Note that we have to avoid a recursive
987 * disaster and not try to clean up after our own cleanup!
991 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) {
992 td->td_pflags |= TDP_INBDFLUSH;
994 td->td_pflags &= ~TDP_INBDFLUSH;
1000 * Set B_CACHE, indicating that the buffer is fully valid. This is
1001 * true even of NFS now.
1003 bp->b_flags |= B_CACHE;
1006 * This bmap keeps the system from needing to do the bmap later,
1007 * perhaps when the system is attempting to do a sync. Since it
1008 * is likely that the indirect block -- or whatever other datastructure
1009 * that the filesystem needs is still in memory now, it is a good
1010 * thing to do this. Note also, that if the pageout daemon is
1011 * requesting a sync -- there might not be enough memory to do
1012 * the bmap then... So, this is important to do.
1014 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) {
1015 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1019 * Set the *dirty* buffer range based upon the VM system dirty pages.
1024 * We need to do this here to satisfy the vnode_pager and the
1025 * pageout daemon, so that it thinks that the pages have been
1026 * "cleaned". Note that since the pages are in a delayed write
1027 * buffer -- the VFS layer "will" see that the pages get written
1028 * out on the next sync, or perhaps the cluster will be completed.
1030 vfs_clean_pages(bp);
1034 * Wakeup the buffer flushing daemon if we have a lot of dirty
1035 * buffers (midpoint between our recovery point and our stall
1038 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1041 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1042 * due to the softdep code.
1049 * Turn buffer into delayed write request. We must clear BIO_READ and
1050 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
1051 * itself to properly update it in the dirty/clean lists. We mark it
1052 * B_DONE to ensure that any asynchronization of the buffer properly
1053 * clears B_DONE ( else a panic will occur later ).
1055 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
1056 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
1057 * should only be called if the buffer is known-good.
1059 * Since the buffer is not on a queue, we do not update the numfreebuffers
1062 * The buffer must be on QUEUE_NONE.
1065 bdirty(struct buf *bp)
1068 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X",
1069 bp, bp->b_vp, bp->b_flags);
1070 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1071 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1072 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1073 BUF_ASSERT_HELD(bp);
1074 bp->b_flags &= ~(B_RELBUF);
1075 bp->b_iocmd = BIO_WRITE;
1077 if ((bp->b_flags & B_DELWRI) == 0) {
1078 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI;
1080 atomic_add_int(&numdirtybuffers, 1);
1081 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
1088 * Clear B_DELWRI for buffer.
1090 * Since the buffer is not on a queue, we do not update the numfreebuffers
1093 * The buffer must be on QUEUE_NONE.
1097 bundirty(struct buf *bp)
1100 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1101 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1102 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE,
1103 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
1104 BUF_ASSERT_HELD(bp);
1106 if (bp->b_flags & B_DELWRI) {
1107 bp->b_flags &= ~B_DELWRI;
1109 atomic_subtract_int(&numdirtybuffers, 1);
1110 numdirtywakeup(lodirtybuffers);
1113 * Since it is now being written, we can clear its deferred write flag.
1115 bp->b_flags &= ~B_DEFERRED;
1121 * Asynchronous write. Start output on a buffer, but do not wait for
1122 * it to complete. The buffer is released when the output completes.
1124 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1125 * B_INVAL buffers. Not us.
1128 bawrite(struct buf *bp)
1131 bp->b_flags |= B_ASYNC;
1138 * Called prior to the locking of any vnodes when we are expecting to
1139 * write. We do not want to starve the buffer cache with too many
1140 * dirty buffers so we block here. By blocking prior to the locking
1141 * of any vnodes we attempt to avoid the situation where a locked vnode
1142 * prevents the various system daemons from flushing related buffers.
1149 if (numdirtybuffers >= hidirtybuffers) {
1151 while (numdirtybuffers >= hidirtybuffers) {
1153 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1154 msleep(&needsbuffer, &nblock,
1155 (PRIBIO + 4), "flswai", 0);
1157 mtx_unlock(&nblock);
1162 * Return true if we have too many dirty buffers.
1165 buf_dirty_count_severe(void)
1168 return(numdirtybuffers >= hidirtybuffers);
1171 static __noinline int
1172 buf_vm_page_count_severe(void)
1175 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1);
1177 return vm_page_count_severe();
1183 * Release a busy buffer and, if requested, free its resources. The
1184 * buffer will be stashed in the appropriate bufqueue[] allowing it
1185 * to be accessed later as a cache entity or reused for other purposes.
1188 brelse(struct buf *bp)
1190 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X",
1191 bp, bp->b_vp, bp->b_flags);
1192 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1193 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1195 if (bp->b_flags & B_MANAGED) {
1200 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) &&
1201 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) {
1203 * Failed write, redirty. Must clear BIO_ERROR to prevent
1204 * pages from being scrapped. If the error is anything
1205 * other than an I/O error (EIO), assume that retryingi
1208 bp->b_ioflags &= ~BIO_ERROR;
1210 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) ||
1211 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) {
1213 * Either a failed I/O or we were asked to free or not
1216 bp->b_flags |= B_INVAL;
1217 if (!LIST_EMPTY(&bp->b_dep))
1219 if (bp->b_flags & B_DELWRI) {
1220 atomic_subtract_int(&numdirtybuffers, 1);
1221 numdirtywakeup(lodirtybuffers);
1223 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1224 if ((bp->b_flags & B_VMIO) == 0) {
1233 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1234 * is called with B_DELWRI set, the underlying pages may wind up
1235 * getting freed causing a previous write (bdwrite()) to get 'lost'
1236 * because pages associated with a B_DELWRI bp are marked clean.
1238 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1239 * if B_DELWRI is set.
1241 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1242 * on pages to return pages to the VM page queues.
1244 if (bp->b_flags & B_DELWRI)
1245 bp->b_flags &= ~B_RELBUF;
1246 else if (buf_vm_page_count_severe()) {
1248 * The locking of the BO_LOCK is not necessary since
1249 * BKGRDINPROG cannot be set while we hold the buf
1250 * lock, it can only be cleared if it is already
1254 if (!(bp->b_vflags & BV_BKGRDINPROG))
1255 bp->b_flags |= B_RELBUF;
1257 bp->b_flags |= B_RELBUF;
1261 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1262 * constituted, not even NFS buffers now. Two flags effect this. If
1263 * B_INVAL, the struct buf is invalidated but the VM object is kept
1264 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1266 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be
1267 * invalidated. BIO_ERROR cannot be set for a failed write unless the
1268 * buffer is also B_INVAL because it hits the re-dirtying code above.
1270 * Normally we can do this whether a buffer is B_DELWRI or not. If
1271 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1272 * the commit state and we cannot afford to lose the buffer. If the
1273 * buffer has a background write in progress, we need to keep it
1274 * around to prevent it from being reconstituted and starting a second
1277 if ((bp->b_flags & B_VMIO)
1278 && !(bp->b_vp->v_mount != NULL &&
1279 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1280 !vn_isdisk(bp->b_vp, NULL) &&
1281 (bp->b_flags & B_DELWRI))
1290 obj = bp->b_bufobj->bo_object;
1293 * Get the base offset and length of the buffer. Note that
1294 * in the VMIO case if the buffer block size is not
1295 * page-aligned then b_data pointer may not be page-aligned.
1296 * But our b_pages[] array *IS* page aligned.
1298 * block sizes less then DEV_BSIZE (usually 512) are not
1299 * supported due to the page granularity bits (m->valid,
1300 * m->dirty, etc...).
1302 * See man buf(9) for more information
1304 resid = bp->b_bufsize;
1305 foff = bp->b_offset;
1306 VM_OBJECT_LOCK(obj);
1307 for (i = 0; i < bp->b_npages; i++) {
1313 * If we hit a bogus page, fixup *all* the bogus pages
1316 if (m == bogus_page) {
1317 poff = OFF_TO_IDX(bp->b_offset);
1320 for (j = i; j < bp->b_npages; j++) {
1322 mtmp = bp->b_pages[j];
1323 if (mtmp == bogus_page) {
1324 mtmp = vm_page_lookup(obj, poff + j);
1326 panic("brelse: page missing\n");
1328 bp->b_pages[j] = mtmp;
1332 if ((bp->b_flags & B_INVAL) == 0) {
1334 trunc_page((vm_offset_t)bp->b_data),
1335 bp->b_pages, bp->b_npages);
1339 if ((bp->b_flags & B_NOCACHE) ||
1340 (bp->b_ioflags & BIO_ERROR)) {
1341 int poffset = foff & PAGE_MASK;
1342 int presid = resid > (PAGE_SIZE - poffset) ?
1343 (PAGE_SIZE - poffset) : resid;
1345 KASSERT(presid >= 0, ("brelse: extra page"));
1346 vm_page_lock_queues();
1347 vm_page_set_invalid(m, poffset, presid);
1348 vm_page_unlock_queues();
1350 printf("avoided corruption bug in bogus_page/brelse code\n");
1352 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1353 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1355 VM_OBJECT_UNLOCK(obj);
1356 if (bp->b_flags & (B_INVAL | B_RELBUF))
1357 vfs_vmio_release(bp);
1359 } else if (bp->b_flags & B_VMIO) {
1361 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1362 vfs_vmio_release(bp);
1365 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) {
1366 if (bp->b_bufsize != 0)
1368 if (bp->b_vp != NULL)
1372 if (BUF_LOCKRECURSED(bp)) {
1373 /* do not release to free list */
1380 /* Handle delayed bremfree() processing. */
1381 if (bp->b_flags & B_REMFREE)
1383 if (bp->b_qindex != QUEUE_NONE)
1384 panic("brelse: free buffer onto another queue???");
1387 * If the buffer has junk contents signal it and eventually
1388 * clean up B_DELWRI and diassociate the vnode so that gbincore()
1391 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 ||
1392 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0)
1393 bp->b_flags |= B_INVAL;
1394 if (bp->b_flags & B_INVAL) {
1395 if (bp->b_flags & B_DELWRI)
1401 /* buffers with no memory */
1402 if (bp->b_bufsize == 0) {
1403 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1404 if (bp->b_vflags & BV_BKGRDINPROG)
1405 panic("losing buffer 1");
1406 if (bp->b_kvasize) {
1407 bp->b_qindex = QUEUE_EMPTYKVA;
1409 bp->b_qindex = QUEUE_EMPTY;
1411 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1412 /* buffers with junk contents */
1413 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) ||
1414 (bp->b_ioflags & BIO_ERROR)) {
1415 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA);
1416 if (bp->b_vflags & BV_BKGRDINPROG)
1417 panic("losing buffer 2");
1418 bp->b_qindex = QUEUE_CLEAN;
1419 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1420 /* remaining buffers */
1422 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) ==
1423 (B_DELWRI|B_NEEDSGIANT))
1424 bp->b_qindex = QUEUE_DIRTY_GIANT;
1425 else if (bp->b_flags & B_DELWRI)
1426 bp->b_qindex = QUEUE_DIRTY;
1428 bp->b_qindex = QUEUE_CLEAN;
1429 if (bp->b_flags & B_AGE)
1430 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1432 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1434 mtx_unlock(&bqlock);
1437 * Fixup numfreebuffers count. The bp is on an appropriate queue
1438 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1439 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1440 * if B_INVAL is set ).
1443 if (!(bp->b_flags & B_DELWRI))
1447 * Something we can maybe free or reuse
1449 if (bp->b_bufsize || bp->b_kvasize)
1452 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT);
1453 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1454 panic("brelse: not dirty");
1460 * Release a buffer back to the appropriate queue but do not try to free
1461 * it. The buffer is expected to be used again soon.
1463 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1464 * biodone() to requeue an async I/O on completion. It is also used when
1465 * known good buffers need to be requeued but we think we may need the data
1468 * XXX we should be able to leave the B_RELBUF hint set on completion.
1471 bqrelse(struct buf *bp)
1473 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1474 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1475 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1477 if (BUF_LOCKRECURSED(bp)) {
1478 /* do not release to free list */
1483 if (bp->b_flags & B_MANAGED) {
1484 if (bp->b_flags & B_REMFREE) {
1487 mtx_unlock(&bqlock);
1489 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1495 /* Handle delayed bremfree() processing. */
1496 if (bp->b_flags & B_REMFREE)
1498 if (bp->b_qindex != QUEUE_NONE)
1499 panic("bqrelse: free buffer onto another queue???");
1500 /* buffers with stale but valid contents */
1501 if (bp->b_flags & B_DELWRI) {
1502 if (bp->b_flags & B_NEEDSGIANT)
1503 bp->b_qindex = QUEUE_DIRTY_GIANT;
1505 bp->b_qindex = QUEUE_DIRTY;
1506 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1509 * The locking of the BO_LOCK for checking of the
1510 * BV_BKGRDINPROG is not necessary since the
1511 * BV_BKGRDINPROG cannot be set while we hold the buf
1512 * lock, it can only be cleared if it is already
1515 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) {
1516 bp->b_qindex = QUEUE_CLEAN;
1517 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp,
1521 * We are too low on memory, we have to try to free
1522 * the buffer (most importantly: the wired pages
1523 * making up its backing store) *now*.
1525 mtx_unlock(&bqlock);
1530 mtx_unlock(&bqlock);
1532 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))
1536 * Something we can maybe free or reuse.
1538 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1541 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1542 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY))
1543 panic("bqrelse: not dirty");
1548 /* Give pages used by the bp back to the VM system (where possible) */
1550 vfs_vmio_release(struct buf *bp)
1555 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
1556 vm_page_lock_queues();
1557 for (i = 0; i < bp->b_npages; i++) {
1559 bp->b_pages[i] = NULL;
1561 * In order to keep page LRU ordering consistent, put
1562 * everything on the inactive queue.
1564 vm_page_unwire(m, 0);
1566 * We don't mess with busy pages, it is
1567 * the responsibility of the process that
1568 * busied the pages to deal with them.
1570 if ((m->oflags & VPO_BUSY) || (m->busy != 0))
1573 if (m->wire_count == 0) {
1575 * Might as well free the page if we can and it has
1576 * no valid data. We also free the page if the
1577 * buffer was used for direct I/O
1579 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1580 m->hold_count == 0) {
1582 } else if (bp->b_flags & B_DIRECT) {
1583 vm_page_try_to_free(m);
1584 } else if (buf_vm_page_count_severe()) {
1585 vm_page_try_to_cache(m);
1589 vm_page_unlock_queues();
1590 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
1591 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1593 if (bp->b_bufsize) {
1598 bp->b_flags &= ~B_VMIO;
1604 * Check to see if a block at a particular lbn is available for a clustered
1608 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno)
1615 /* If the buf isn't in core skip it */
1616 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL)
1619 /* If the buf is busy we don't want to wait for it */
1620 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1623 /* Only cluster with valid clusterable delayed write buffers */
1624 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) !=
1625 (B_DELWRI | B_CLUSTEROK))
1628 if (bpa->b_bufsize != size)
1632 * Check to see if it is in the expected place on disk and that the
1633 * block has been mapped.
1635 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno))
1645 * Implement clustered async writes for clearing out B_DELWRI buffers.
1646 * This is much better then the old way of writing only one buffer at
1647 * a time. Note that we may not be presented with the buffers in the
1648 * correct order, so we search for the cluster in both directions.
1651 vfs_bio_awrite(struct buf *bp)
1656 daddr_t lblkno = bp->b_lblkno;
1657 struct vnode *vp = bp->b_vp;
1665 * right now we support clustered writing only to regular files. If
1666 * we find a clusterable block we could be in the middle of a cluster
1667 * rather then at the beginning.
1669 if ((vp->v_type == VREG) &&
1670 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1671 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1673 size = vp->v_mount->mnt_stat.f_iosize;
1674 maxcl = MAXPHYS / size;
1677 for (i = 1; i < maxcl; i++)
1678 if (vfs_bio_clcheck(vp, size, lblkno + i,
1679 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0)
1682 for (j = 1; i + j <= maxcl && j <= lblkno; j++)
1683 if (vfs_bio_clcheck(vp, size, lblkno - j,
1684 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0)
1690 * this is a possible cluster write
1694 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1699 bp->b_flags |= B_ASYNC;
1701 * default (old) behavior, writing out only one block
1703 * XXX returns b_bufsize instead of b_bcount for nwritten?
1705 nwritten = bp->b_bufsize;
1714 * Find and initialize a new buffer header, freeing up existing buffers
1715 * in the bufqueues as necessary. The new buffer is returned locked.
1717 * Important: B_INVAL is not set. If the caller wishes to throw the
1718 * buffer away, the caller must set B_INVAL prior to calling brelse().
1721 * We have insufficient buffer headers
1722 * We have insufficient buffer space
1723 * buffer_map is too fragmented ( space reservation fails )
1724 * If we have to flush dirty buffers ( but we try to avoid this )
1726 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1727 * Instead we ask the buf daemon to do it for us. We attempt to
1728 * avoid piecemeal wakeups of the pageout daemon.
1732 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize,
1740 static int flushingbufs;
1744 * We can't afford to block since we might be holding a vnode lock,
1745 * which may prevent system daemons from running. We deal with
1746 * low-memory situations by proactively returning memory and running
1747 * async I/O rather then sync I/O.
1749 atomic_add_int(&getnewbufcalls, 1);
1750 atomic_subtract_int(&getnewbufrestarts, 1);
1752 atomic_add_int(&getnewbufrestarts, 1);
1755 * Setup for scan. If we do not have enough free buffers,
1756 * we setup a degenerate case that immediately fails. Note
1757 * that if we are specially marked process, we are allowed to
1758 * dip into our reserves.
1760 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1762 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1763 * However, there are a number of cases (defragging, reusing, ...)
1764 * where we cannot backup.
1767 nqindex = QUEUE_EMPTYKVA;
1768 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1772 * If no EMPTYKVA buffers and we are either
1773 * defragging or reusing, locate a CLEAN buffer
1774 * to free or reuse. If bufspace useage is low
1775 * skip this step so we can allocate a new buffer.
1777 if (defrag || bufspace >= lobufspace) {
1778 nqindex = QUEUE_CLEAN;
1779 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1783 * If we could not find or were not allowed to reuse a
1784 * CLEAN buffer, check to see if it is ok to use an EMPTY
1785 * buffer. We can only use an EMPTY buffer if allocating
1786 * its KVA would not otherwise run us out of buffer space.
1788 if (nbp == NULL && defrag == 0 &&
1789 bufspace + maxsize < hibufspace) {
1790 nqindex = QUEUE_EMPTY;
1791 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1796 * Run scan, possibly freeing data and/or kva mappings on the fly
1800 while ((bp = nbp) != NULL) {
1801 int qindex = nqindex;
1804 * Calculate next bp ( we can only use it if we do not block
1805 * or do other fancy things ).
1807 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1810 nqindex = QUEUE_EMPTYKVA;
1811 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1814 case QUEUE_EMPTYKVA:
1815 nqindex = QUEUE_CLEAN;
1816 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1827 * If we are defragging then we need a buffer with
1828 * b_kvasize != 0. XXX this situation should no longer
1829 * occur, if defrag is non-zero the buffer's b_kvasize
1830 * should also be non-zero at this point. XXX
1832 if (defrag && bp->b_kvasize == 0) {
1833 printf("Warning: defrag empty buffer %p\n", bp);
1838 * Start freeing the bp. This is somewhat involved. nbp
1839 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1841 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
1844 BO_LOCK(bp->b_bufobj);
1845 if (bp->b_vflags & BV_BKGRDINPROG) {
1846 BO_UNLOCK(bp->b_bufobj);
1850 BO_UNLOCK(bp->b_bufobj);
1853 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d "
1854 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags,
1855 bp->b_kvasize, bp->b_bufsize, qindex);
1860 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1863 * Note: we no longer distinguish between VMIO and non-VMIO
1867 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1870 mtx_unlock(&bqlock);
1872 if (qindex == QUEUE_CLEAN) {
1873 if (bp->b_flags & B_VMIO) {
1874 bp->b_flags &= ~B_ASYNC;
1875 vfs_vmio_release(bp);
1882 * NOTE: nbp is now entirely invalid. We can only restart
1883 * the scan from this point on.
1885 * Get the rest of the buffer freed up. b_kva* is still
1886 * valid after this operation.
1889 if (bp->b_rcred != NOCRED) {
1890 crfree(bp->b_rcred);
1891 bp->b_rcred = NOCRED;
1893 if (bp->b_wcred != NOCRED) {
1894 crfree(bp->b_wcred);
1895 bp->b_wcred = NOCRED;
1897 if (!LIST_EMPTY(&bp->b_dep))
1899 if (bp->b_vflags & BV_BKGRDINPROG)
1900 panic("losing buffer 3");
1901 KASSERT(bp->b_vp == NULL,
1902 ("bp: %p still has vnode %p. qindex: %d",
1903 bp, bp->b_vp, qindex));
1904 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0,
1905 ("bp: %p still on a buffer list. xflags %X",
1916 bp->b_blkno = bp->b_lblkno = 0;
1917 bp->b_offset = NOOFFSET;
1923 bp->b_dirtyoff = bp->b_dirtyend = 0;
1924 bp->b_bufobj = NULL;
1925 bp->b_pin_count = 0;
1926 bp->b_fsprivate1 = NULL;
1927 bp->b_fsprivate2 = NULL;
1928 bp->b_fsprivate3 = NULL;
1930 LIST_INIT(&bp->b_dep);
1933 * If we are defragging then free the buffer.
1936 bp->b_flags |= B_INVAL;
1944 * Notify any waiters for the buffer lock about
1945 * identity change by freeing the buffer.
1947 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) {
1948 bp->b_flags |= B_INVAL;
1955 * If we are overcomitted then recover the buffer and its
1956 * KVM space. This occurs in rare situations when multiple
1957 * processes are blocked in getnewbuf() or allocbuf().
1959 if (bufspace >= hibufspace)
1961 if (flushingbufs && bp->b_kvasize != 0) {
1962 bp->b_flags |= B_INVAL;
1967 if (bufspace < lobufspace)
1973 * If we exhausted our list, sleep as appropriate. We may have to
1974 * wakeup various daemons and write out some dirty buffers.
1976 * Generally we are sleeping due to insufficient buffer space.
1980 int flags, norunbuf;
1985 flags = VFS_BIO_NEED_BUFSPACE;
1987 } else if (bufspace >= hibufspace) {
1989 flags = VFS_BIO_NEED_BUFSPACE;
1992 flags = VFS_BIO_NEED_ANY;
1995 needsbuffer |= flags;
1996 mtx_unlock(&nblock);
1997 mtx_unlock(&bqlock);
1999 bd_speedup(); /* heeeelp */
2000 if (gbflags & GB_NOWAIT_BD)
2004 while (needsbuffer & flags) {
2005 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) {
2006 mtx_unlock(&nblock);
2008 * getblk() is called with a vnode
2009 * locked, and some majority of the
2010 * dirty buffers may as well belong to
2011 * the vnode. Flushing the buffers
2012 * there would make a progress that
2013 * cannot be achieved by the
2014 * buf_daemon, that cannot lock the
2017 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) |
2018 (td->td_pflags & TDP_NORUNNINGBUF);
2019 /* play bufdaemon */
2020 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF;
2021 fl = buf_do_flush(vp);
2022 td->td_pflags &= norunbuf;
2026 if ((needsbuffer & flags) == 0)
2029 if (msleep(&needsbuffer, &nblock,
2030 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
2031 mtx_unlock(&nblock);
2035 mtx_unlock(&nblock);
2038 * We finally have a valid bp. We aren't quite out of the
2039 * woods, we still have to reserve kva space. In order
2040 * to keep fragmentation sane we only allocate kva in
2043 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2045 if (maxsize != bp->b_kvasize) {
2046 vm_offset_t addr = 0;
2050 vm_map_lock(buffer_map);
2051 if (vm_map_findspace(buffer_map,
2052 vm_map_min(buffer_map), maxsize, &addr)) {
2054 * Uh oh. Buffer map is to fragmented. We
2055 * must defragment the map.
2057 atomic_add_int(&bufdefragcnt, 1);
2058 vm_map_unlock(buffer_map);
2060 bp->b_flags |= B_INVAL;
2065 vm_map_insert(buffer_map, NULL, 0,
2066 addr, addr + maxsize,
2067 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
2069 bp->b_kvabase = (caddr_t) addr;
2070 bp->b_kvasize = maxsize;
2071 atomic_add_long(&bufspace, bp->b_kvasize);
2072 atomic_add_int(&bufreusecnt, 1);
2074 vm_map_unlock(buffer_map);
2076 bp->b_saveaddr = bp->b_kvabase;
2077 bp->b_data = bp->b_saveaddr;
2085 * buffer flushing daemon. Buffers are normally flushed by the
2086 * update daemon but if it cannot keep up this process starts to
2087 * take the load in an attempt to prevent getnewbuf() from blocking.
2090 static struct kproc_desc buf_kp = {
2095 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp);
2098 buf_do_flush(struct vnode *vp)
2102 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0);
2103 /* The list empty check here is slightly racy */
2104 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) {
2106 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0);
2111 * Could not find any buffers without rollback
2112 * dependencies, so just write the first one
2113 * in the hopes of eventually making progress.
2115 flushbufqueues(vp, QUEUE_DIRTY, 1);
2117 &bufqueues[QUEUE_DIRTY_GIANT])) {
2119 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1);
2131 * This process needs to be suspended prior to shutdown sync.
2133 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc,
2137 * This process is allowed to take the buffer cache to the limit
2139 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED;
2143 mtx_unlock(&bdlock);
2145 kproc_suspend_check(bufdaemonproc);
2148 * Do the flush. Limit the amount of in-transit I/O we
2149 * allow to build up, otherwise we would completely saturate
2150 * the I/O system. Wakeup any waiting processes before we
2151 * normally would so they can run in parallel with our drain.
2153 while (numdirtybuffers > lodirtybuffers) {
2154 if (buf_do_flush(NULL) == 0)
2160 * Only clear bd_request if we have reached our low water
2161 * mark. The buf_daemon normally waits 1 second and
2162 * then incrementally flushes any dirty buffers that have
2163 * built up, within reason.
2165 * If we were unable to hit our low water mark and couldn't
2166 * find any flushable buffers, we sleep half a second.
2167 * Otherwise we loop immediately.
2170 if (numdirtybuffers <= lodirtybuffers) {
2172 * We reached our low water mark, reset the
2173 * request and sleep until we are needed again.
2174 * The sleep is just so the suspend code works.
2177 msleep(&bd_request, &bdlock, PVM, "psleep", hz);
2180 * We couldn't find any flushable dirty buffers but
2181 * still have too many dirty buffers, we
2182 * have to sleep and try again. (rare)
2184 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10);
2192 * Try to flush a buffer in the dirty queue. We must be careful to
2193 * free up B_INVAL buffers instead of write them, which NFS is
2194 * particularly sensitive to.
2196 static int flushwithdeps = 0;
2197 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps,
2198 0, "Number of buffers flushed with dependecies that require rollbacks");
2201 flushbufqueues(struct vnode *lvp, int queue, int flushdeps)
2203 struct buf *sentinel;
2212 target = numdirtybuffers - lodirtybuffers;
2213 if (flushdeps && target > 2)
2216 target = flushbufqtarget;
2219 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO);
2220 sentinel->b_qindex = QUEUE_SENTINEL;
2222 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist);
2223 while (flushed != target) {
2224 bp = TAILQ_NEXT(sentinel, b_freelist);
2226 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2227 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel,
2232 * Skip sentinels inserted by other invocations of the
2233 * flushbufqueues(), taking care to not reorder them.
2235 if (bp->b_qindex == QUEUE_SENTINEL)
2238 * Only flush the buffers that belong to the
2239 * vnode locked by the curthread.
2241 if (lvp != NULL && bp->b_vp != lvp)
2243 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)
2245 if (bp->b_pin_count > 0) {
2249 BO_LOCK(bp->b_bufobj);
2250 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 ||
2251 (bp->b_flags & B_DELWRI) == 0) {
2252 BO_UNLOCK(bp->b_bufobj);
2256 BO_UNLOCK(bp->b_bufobj);
2257 if (bp->b_flags & B_INVAL) {
2259 mtx_unlock(&bqlock);
2262 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2267 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) {
2268 if (flushdeps == 0) {
2276 * We must hold the lock on a vnode before writing
2277 * one of its buffers. Otherwise we may confuse, or
2278 * in the case of a snapshot vnode, deadlock the
2281 * The lock order here is the reverse of the normal
2282 * of vnode followed by buf lock. This is ok because
2283 * the NOWAIT will prevent deadlock.
2286 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2290 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) {
2291 mtx_unlock(&bqlock);
2292 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X",
2293 bp, bp->b_vp, bp->b_flags);
2294 if (curproc == bufdaemonproc)
2301 vn_finished_write(mp);
2303 flushwithdeps += hasdeps;
2307 * Sleeping on runningbufspace while holding
2308 * vnode lock leads to deadlock.
2310 if (curproc == bufdaemonproc)
2311 waitrunningbufspace();
2312 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
2316 vn_finished_write(mp);
2319 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist);
2320 mtx_unlock(&bqlock);
2321 free(sentinel, M_TEMP);
2326 * Check to see if a block is currently memory resident.
2329 incore(struct bufobj *bo, daddr_t blkno)
2334 bp = gbincore(bo, blkno);
2340 * Returns true if no I/O is needed to access the
2341 * associated VM object. This is like incore except
2342 * it also hunts around in the VM system for the data.
2346 inmem(struct vnode * vp, daddr_t blkno)
2349 vm_offset_t toff, tinc, size;
2353 ASSERT_VOP_LOCKED(vp, "inmem");
2355 if (incore(&vp->v_bufobj, blkno))
2357 if (vp->v_mount == NULL)
2364 if (size > vp->v_mount->mnt_stat.f_iosize)
2365 size = vp->v_mount->mnt_stat.f_iosize;
2366 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
2368 VM_OBJECT_LOCK(obj);
2369 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2370 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
2374 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
2375 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
2376 if (vm_page_is_valid(m,
2377 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
2380 VM_OBJECT_UNLOCK(obj);
2384 VM_OBJECT_UNLOCK(obj);
2391 * Sets the dirty range for a buffer based on the status of the dirty
2392 * bits in the pages comprising the buffer.
2394 * The range is limited to the size of the buffer.
2396 * This routine is primarily used by NFS, but is generalized for the
2400 vfs_setdirty(struct buf *bp)
2404 * Degenerate case - empty buffer
2407 if (bp->b_bufsize == 0)
2411 * We qualify the scan for modified pages on whether the
2412 * object has been flushed yet.
2415 if ((bp->b_flags & B_VMIO) == 0)
2418 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2419 vfs_setdirty_locked_object(bp);
2420 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2424 vfs_setdirty_locked_object(struct buf *bp)
2429 object = bp->b_bufobj->bo_object;
2430 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2431 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2432 vm_offset_t boffset;
2433 vm_offset_t eoffset;
2435 vm_page_lock_queues();
2437 * test the pages to see if they have been modified directly
2438 * by users through the VM system.
2440 for (i = 0; i < bp->b_npages; i++)
2441 vm_page_test_dirty(bp->b_pages[i]);
2444 * Calculate the encompassing dirty range, boffset and eoffset,
2445 * (eoffset - boffset) bytes.
2448 for (i = 0; i < bp->b_npages; i++) {
2449 if (bp->b_pages[i]->dirty)
2452 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2454 for (i = bp->b_npages - 1; i >= 0; --i) {
2455 if (bp->b_pages[i]->dirty) {
2459 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2461 vm_page_unlock_queues();
2463 * Fit it to the buffer.
2466 if (eoffset > bp->b_bcount)
2467 eoffset = bp->b_bcount;
2470 * If we have a good dirty range, merge with the existing
2474 if (boffset < eoffset) {
2475 if (bp->b_dirtyoff > boffset)
2476 bp->b_dirtyoff = boffset;
2477 if (bp->b_dirtyend < eoffset)
2478 bp->b_dirtyend = eoffset;
2486 * Get a block given a specified block and offset into a file/device.
2487 * The buffers B_DONE bit will be cleared on return, making it almost
2488 * ready for an I/O initiation. B_INVAL may or may not be set on
2489 * return. The caller should clear B_INVAL prior to initiating a
2492 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2493 * an existing buffer.
2495 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2496 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2497 * and then cleared based on the backing VM. If the previous buffer is
2498 * non-0-sized but invalid, B_CACHE will be cleared.
2500 * If getblk() must create a new buffer, the new buffer is returned with
2501 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2502 * case it is returned with B_INVAL clear and B_CACHE set based on the
2505 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2506 * B_CACHE bit is clear.
2508 * What this means, basically, is that the caller should use B_CACHE to
2509 * determine whether the buffer is fully valid or not and should clear
2510 * B_INVAL prior to issuing a read. If the caller intends to validate
2511 * the buffer by loading its data area with something, the caller needs
2512 * to clear B_INVAL. If the caller does this without issuing an I/O,
2513 * the caller should set B_CACHE ( as an optimization ), else the caller
2514 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2515 * a write attempt or if it was a successfull read. If the caller
2516 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR
2517 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2520 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo,
2527 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size);
2528 ASSERT_VOP_LOCKED(vp, "getblk");
2529 if (size > MAXBSIZE)
2530 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2535 * Block if we are low on buffers. Certain processes are allowed
2536 * to completely exhaust the buffer cache.
2538 * If this check ever becomes a bottleneck it may be better to
2539 * move it into the else, when gbincore() fails. At the moment
2540 * it isn't a problem.
2542 * XXX remove if 0 sections (clean this up after its proven)
2544 if (numfreebuffers == 0) {
2545 if (TD_IS_IDLETHREAD(curthread))
2548 needsbuffer |= VFS_BIO_NEED_ANY;
2549 mtx_unlock(&nblock);
2553 bp = gbincore(bo, blkno);
2557 * Buffer is in-core. If the buffer is not busy, it must
2560 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK;
2562 if (flags & GB_LOCK_NOWAIT)
2563 lockflags |= LK_NOWAIT;
2565 error = BUF_TIMELOCK(bp, lockflags,
2566 BO_MTX(bo), "getblk", slpflag, slptimeo);
2569 * If we slept and got the lock we have to restart in case
2570 * the buffer changed identities.
2572 if (error == ENOLCK)
2574 /* We timed out or were interrupted. */
2579 * The buffer is locked. B_CACHE is cleared if the buffer is
2580 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set
2581 * and for a VMIO buffer B_CACHE is adjusted according to the
2584 if (bp->b_flags & B_INVAL)
2585 bp->b_flags &= ~B_CACHE;
2586 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2587 bp->b_flags |= B_CACHE;
2591 * check for size inconsistancies for non-VMIO case.
2594 if (bp->b_bcount != size) {
2595 if ((bp->b_flags & B_VMIO) == 0 ||
2596 (size > bp->b_kvasize)) {
2597 if (bp->b_flags & B_DELWRI) {
2599 * If buffer is pinned and caller does
2600 * not want sleep waiting for it to be
2601 * unpinned, bail out
2603 if (bp->b_pin_count > 0) {
2604 if (flags & GB_LOCK_NOWAIT) {
2611 bp->b_flags |= B_NOCACHE;
2614 if (LIST_EMPTY(&bp->b_dep)) {
2615 bp->b_flags |= B_RELBUF;
2618 bp->b_flags |= B_NOCACHE;
2627 * If the size is inconsistant in the VMIO case, we can resize
2628 * the buffer. This might lead to B_CACHE getting set or
2629 * cleared. If the size has not changed, B_CACHE remains
2630 * unchanged from its previous state.
2633 if (bp->b_bcount != size)
2636 KASSERT(bp->b_offset != NOOFFSET,
2637 ("getblk: no buffer offset"));
2640 * A buffer with B_DELWRI set and B_CACHE clear must
2641 * be committed before we can return the buffer in
2642 * order to prevent the caller from issuing a read
2643 * ( due to B_CACHE not being set ) and overwriting
2646 * Most callers, including NFS and FFS, need this to
2647 * operate properly either because they assume they
2648 * can issue a read if B_CACHE is not set, or because
2649 * ( for example ) an uncached B_DELWRI might loop due
2650 * to softupdates re-dirtying the buffer. In the latter
2651 * case, B_CACHE is set after the first write completes,
2652 * preventing further loops.
2653 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2654 * above while extending the buffer, we cannot allow the
2655 * buffer to remain with B_CACHE set after the write
2656 * completes or it will represent a corrupt state. To
2657 * deal with this we set B_NOCACHE to scrap the buffer
2660 * We might be able to do something fancy, like setting
2661 * B_CACHE in bwrite() except if B_DELWRI is already set,
2662 * so the below call doesn't set B_CACHE, but that gets real
2663 * confusing. This is much easier.
2666 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2667 bp->b_flags |= B_NOCACHE;
2671 bp->b_flags &= ~B_DONE;
2673 int bsize, maxsize, vmio;
2677 * Buffer is not in-core, create new buffer. The buffer
2678 * returned by getnewbuf() is locked. Note that the returned
2679 * buffer is also considered valid (not marked B_INVAL).
2683 * If the user does not want us to create the buffer, bail out
2686 if (flags & GB_NOCREAT)
2688 bsize = bo->bo_bsize;
2689 offset = blkno * bsize;
2690 vmio = vp->v_object != NULL;
2691 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2692 maxsize = imax(maxsize, bsize);
2694 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags);
2696 if (slpflag || slptimeo)
2702 * This code is used to make sure that a buffer is not
2703 * created while the getnewbuf routine is blocked.
2704 * This can be a problem whether the vnode is locked or not.
2705 * If the buffer is created out from under us, we have to
2706 * throw away the one we just created.
2708 * Note: this must occur before we associate the buffer
2709 * with the vp especially considering limitations in
2710 * the splay tree implementation when dealing with duplicate
2714 if (gbincore(bo, blkno)) {
2716 bp->b_flags |= B_INVAL;
2722 * Insert the buffer into the hash, so that it can
2723 * be found by incore.
2725 bp->b_blkno = bp->b_lblkno = blkno;
2726 bp->b_offset = offset;
2731 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2732 * buffer size starts out as 0, B_CACHE will be set by
2733 * allocbuf() for the VMIO case prior to it testing the
2734 * backing store for validity.
2738 bp->b_flags |= B_VMIO;
2739 #if defined(VFS_BIO_DEBUG)
2740 if (vn_canvmio(vp) != TRUE)
2741 printf("getblk: VMIO on vnode type %d\n",
2744 KASSERT(vp->v_object == bp->b_bufobj->bo_object,
2745 ("ARGH! different b_bufobj->bo_object %p %p %p\n",
2746 bp, vp->v_object, bp->b_bufobj->bo_object));
2748 bp->b_flags &= ~B_VMIO;
2749 KASSERT(bp->b_bufobj->bo_object == NULL,
2750 ("ARGH! has b_bufobj->bo_object %p %p\n",
2751 bp, bp->b_bufobj->bo_object));
2755 bp->b_flags &= ~B_DONE;
2757 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp);
2758 BUF_ASSERT_HELD(bp);
2759 KASSERT(bp->b_bufobj == bo,
2760 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2765 * Get an empty, disassociated buffer of given size. The buffer is initially
2769 geteblk(int size, int flags)
2774 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2775 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) {
2776 if ((flags & GB_NOWAIT_BD) &&
2777 (curthread->td_pflags & TDP_BUFNEED) != 0)
2781 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2782 BUF_ASSERT_HELD(bp);
2788 * This code constitutes the buffer memory from either anonymous system
2789 * memory (in the case of non-VMIO operations) or from an associated
2790 * VM object (in the case of VMIO operations). This code is able to
2791 * resize a buffer up or down.
2793 * Note that this code is tricky, and has many complications to resolve
2794 * deadlock or inconsistant data situations. Tread lightly!!!
2795 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2796 * the caller. Calling this code willy nilly can result in the loss of data.
2798 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2799 * B_CACHE for the non-VMIO case.
2803 allocbuf(struct buf *bp, int size)
2805 int newbsize, mbsize;
2808 BUF_ASSERT_HELD(bp);
2810 if (bp->b_kvasize < size)
2811 panic("allocbuf: buffer too small");
2813 if ((bp->b_flags & B_VMIO) == 0) {
2817 * Just get anonymous memory from the kernel. Don't
2818 * mess with B_CACHE.
2820 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2821 if (bp->b_flags & B_MALLOC)
2824 newbsize = round_page(size);
2826 if (newbsize < bp->b_bufsize) {
2828 * malloced buffers are not shrunk
2830 if (bp->b_flags & B_MALLOC) {
2832 bp->b_bcount = size;
2834 free(bp->b_data, M_BIOBUF);
2835 if (bp->b_bufsize) {
2836 atomic_subtract_long(
2842 bp->b_saveaddr = bp->b_kvabase;
2843 bp->b_data = bp->b_saveaddr;
2845 bp->b_flags &= ~B_MALLOC;
2851 (vm_offset_t) bp->b_data + newbsize,
2852 (vm_offset_t) bp->b_data + bp->b_bufsize);
2853 } else if (newbsize > bp->b_bufsize) {
2855 * We only use malloced memory on the first allocation.
2856 * and revert to page-allocated memory when the buffer
2860 * There is a potential smp race here that could lead
2861 * to bufmallocspace slightly passing the max. It
2862 * is probably extremely rare and not worth worrying
2865 if ( (bufmallocspace < maxbufmallocspace) &&
2866 (bp->b_bufsize == 0) &&
2867 (mbsize <= PAGE_SIZE/2)) {
2869 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2870 bp->b_bufsize = mbsize;
2871 bp->b_bcount = size;
2872 bp->b_flags |= B_MALLOC;
2873 atomic_add_long(&bufmallocspace, mbsize);
2879 * If the buffer is growing on its other-than-first allocation,
2880 * then we revert to the page-allocation scheme.
2882 if (bp->b_flags & B_MALLOC) {
2883 origbuf = bp->b_data;
2884 origbufsize = bp->b_bufsize;
2885 bp->b_data = bp->b_kvabase;
2886 if (bp->b_bufsize) {
2887 atomic_subtract_long(&bufmallocspace,
2892 bp->b_flags &= ~B_MALLOC;
2893 newbsize = round_page(newbsize);
2897 (vm_offset_t) bp->b_data + bp->b_bufsize,
2898 (vm_offset_t) bp->b_data + newbsize);
2900 bcopy(origbuf, bp->b_data, origbufsize);
2901 free(origbuf, M_BIOBUF);
2907 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2908 desiredpages = (size == 0) ? 0 :
2909 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2911 if (bp->b_flags & B_MALLOC)
2912 panic("allocbuf: VMIO buffer can't be malloced");
2914 * Set B_CACHE initially if buffer is 0 length or will become
2917 if (size == 0 || bp->b_bufsize == 0)
2918 bp->b_flags |= B_CACHE;
2920 if (newbsize < bp->b_bufsize) {
2922 * DEV_BSIZE aligned new buffer size is less then the
2923 * DEV_BSIZE aligned existing buffer size. Figure out
2924 * if we have to remove any pages.
2926 if (desiredpages < bp->b_npages) {
2929 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
2930 vm_page_lock_queues();
2931 for (i = desiredpages; i < bp->b_npages; i++) {
2933 * the page is not freed here -- it
2934 * is the responsibility of
2935 * vnode_pager_setsize
2938 KASSERT(m != bogus_page,
2939 ("allocbuf: bogus page found"));
2940 while (vm_page_sleep_if_busy(m, TRUE, "biodep"))
2941 vm_page_lock_queues();
2943 bp->b_pages[i] = NULL;
2944 vm_page_unwire(m, 0);
2946 vm_page_unlock_queues();
2947 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
2948 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2949 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2950 bp->b_npages = desiredpages;
2952 } else if (size > bp->b_bcount) {
2954 * We are growing the buffer, possibly in a
2955 * byte-granular fashion.
2963 * Step 1, bring in the VM pages from the object,
2964 * allocating them if necessary. We must clear
2965 * B_CACHE if these pages are not valid for the
2966 * range covered by the buffer.
2970 obj = bp->b_bufobj->bo_object;
2972 VM_OBJECT_LOCK(obj);
2973 while (bp->b_npages < desiredpages) {
2977 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2978 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2980 * note: must allocate system pages
2981 * since blocking here could intefere
2982 * with paging I/O, no matter which
2985 m = vm_page_alloc(obj, pi,
2986 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM |
2989 atomic_add_int(&vm_pageout_deficit,
2990 desiredpages - bp->b_npages);
2991 VM_OBJECT_UNLOCK(obj);
2993 VM_OBJECT_LOCK(obj);
2996 bp->b_flags &= ~B_CACHE;
2997 bp->b_pages[bp->b_npages] = m;
3004 * We found a page. If we have to sleep on it,
3005 * retry because it might have gotten freed out
3008 * We can only test VPO_BUSY here. Blocking on
3009 * m->busy might lead to a deadlock:
3011 * vm_fault->getpages->cluster_read->allocbuf
3014 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk"))
3018 * We have a good page.
3020 vm_page_lock_queues();
3022 vm_page_unlock_queues();
3023 bp->b_pages[bp->b_npages] = m;
3028 * Step 2. We've loaded the pages into the buffer,
3029 * we have to figure out if we can still have B_CACHE
3030 * set. Note that B_CACHE is set according to the
3031 * byte-granular range ( bcount and size ), new the
3032 * aligned range ( newbsize ).
3034 * The VM test is against m->valid, which is DEV_BSIZE
3035 * aligned. Needless to say, the validity of the data
3036 * needs to also be DEV_BSIZE aligned. Note that this
3037 * fails with NFS if the server or some other client
3038 * extends the file's EOF. If our buffer is resized,
3039 * B_CACHE may remain set! XXX
3042 toff = bp->b_bcount;
3043 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
3045 while ((bp->b_flags & B_CACHE) && toff < size) {
3048 if (tinc > (size - toff))
3051 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
3064 VM_OBJECT_UNLOCK(obj);
3067 * Step 3, fixup the KVM pmap. Remember that
3068 * bp->b_data is relative to bp->b_offset, but
3069 * bp->b_offset may be offset into the first page.
3072 bp->b_data = (caddr_t)
3073 trunc_page((vm_offset_t)bp->b_data);
3075 (vm_offset_t)bp->b_data,
3080 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3081 (vm_offset_t)(bp->b_offset & PAGE_MASK));
3084 if (newbsize < bp->b_bufsize)
3086 bp->b_bufsize = newbsize; /* actual buffer allocation */
3087 bp->b_bcount = size; /* requested buffer size */
3092 biodone(struct bio *bp)
3095 void (*done)(struct bio *);
3097 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3099 bp->bio_flags |= BIO_DONE;
3100 done = bp->bio_done;
3109 * Wait for a BIO to finish.
3111 * XXX: resort to a timeout for now. The optimal locking (if any) for this
3112 * case is not yet clear.
3115 biowait(struct bio *bp, const char *wchan)
3119 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3121 while ((bp->bio_flags & BIO_DONE) == 0)
3122 msleep(bp, mtxp, PRIBIO, wchan, hz / 10);
3124 if (bp->bio_error != 0)
3125 return (bp->bio_error);
3126 if (!(bp->bio_flags & BIO_ERROR))
3132 biofinish(struct bio *bp, struct devstat *stat, int error)
3136 bp->bio_error = error;
3137 bp->bio_flags |= BIO_ERROR;
3140 devstat_end_transaction_bio(stat, bp);
3147 * Wait for buffer I/O completion, returning error status. The buffer
3148 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR
3149 * error and cleared.
3152 bufwait(struct buf *bp)
3154 if (bp->b_iocmd == BIO_READ)
3155 bwait(bp, PRIBIO, "biord");
3157 bwait(bp, PRIBIO, "biowr");
3158 if (bp->b_flags & B_EINTR) {
3159 bp->b_flags &= ~B_EINTR;
3162 if (bp->b_ioflags & BIO_ERROR) {
3163 return (bp->b_error ? bp->b_error : EIO);
3170 * Call back function from struct bio back up to struct buf.
3173 bufdonebio(struct bio *bip)
3177 bp = bip->bio_caller2;
3178 bp->b_resid = bp->b_bcount - bip->bio_completed;
3179 bp->b_resid = bip->bio_resid; /* XXX: remove */
3180 bp->b_ioflags = bip->bio_flags;
3181 bp->b_error = bip->bio_error;
3183 bp->b_ioflags |= BIO_ERROR;
3189 dev_strategy(struct cdev *dev, struct buf *bp)
3194 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1)))
3195 panic("b_iocmd botch");
3200 /* Try again later */
3201 tsleep(&bp, PRIBIO, "dev_strat", hz/10);
3203 bip->bio_cmd = bp->b_iocmd;
3204 bip->bio_offset = bp->b_iooffset;
3205 bip->bio_length = bp->b_bcount;
3206 bip->bio_bcount = bp->b_bcount; /* XXX: remove */
3207 bip->bio_data = bp->b_data;
3208 bip->bio_done = bufdonebio;
3209 bip->bio_caller2 = bp;
3211 KASSERT(dev->si_refcount > 0,
3212 ("dev_strategy on un-referenced struct cdev *(%s)",
3214 csw = dev_refthread(dev);
3217 bp->b_error = ENXIO;
3218 bp->b_ioflags = BIO_ERROR;
3222 (*csw->d_strategy)(bip);
3229 * Finish I/O on a buffer, optionally calling a completion function.
3230 * This is usually called from an interrupt so process blocking is
3233 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3234 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3235 * assuming B_INVAL is clear.
3237 * For the VMIO case, we set B_CACHE if the op was a read and no
3238 * read error occured, or if the op was a write. B_CACHE is never
3239 * set if the buffer is invalid or otherwise uncacheable.
3241 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3242 * initiator to leave B_INVAL set to brelse the buffer out of existance
3243 * in the biodone routine.
3246 bufdone(struct buf *bp)
3248 struct bufobj *dropobj;
3249 void (*biodone)(struct buf *);
3251 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
3254 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
3255 BUF_ASSERT_HELD(bp);
3257 runningbufwakeup(bp);
3258 if (bp->b_iocmd == BIO_WRITE)
3259 dropobj = bp->b_bufobj;
3260 /* call optional completion function if requested */
3261 if (bp->b_iodone != NULL) {
3262 biodone = bp->b_iodone;
3263 bp->b_iodone = NULL;
3266 bufobj_wdrop(dropobj);
3273 bufobj_wdrop(dropobj);
3277 bufdone_finish(struct buf *bp)
3279 BUF_ASSERT_HELD(bp);
3281 if (!LIST_EMPTY(&bp->b_dep))
3284 if (bp->b_flags & B_VMIO) {
3290 struct vnode *vp = bp->b_vp;
3292 obj = bp->b_bufobj->bo_object;
3294 #if defined(VFS_BIO_DEBUG)
3295 mp_fixme("usecount and vflag accessed without locks.");
3296 if (vp->v_usecount == 0) {
3297 panic("biodone: zero vnode ref count");
3300 KASSERT(vp->v_object != NULL,
3301 ("biodone: vnode %p has no vm_object", vp));
3304 foff = bp->b_offset;
3305 KASSERT(bp->b_offset != NOOFFSET,
3306 ("biodone: no buffer offset"));
3308 VM_OBJECT_LOCK(obj);
3309 #if defined(VFS_BIO_DEBUG)
3310 if (obj->paging_in_progress < bp->b_npages) {
3311 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
3312 obj->paging_in_progress, bp->b_npages);
3317 * Set B_CACHE if the op was a normal read and no error
3318 * occured. B_CACHE is set for writes in the b*write()
3321 iosize = bp->b_bcount - bp->b_resid;
3322 if (bp->b_iocmd == BIO_READ &&
3323 !(bp->b_flags & (B_INVAL|B_NOCACHE)) &&
3324 !(bp->b_ioflags & BIO_ERROR)) {
3325 bp->b_flags |= B_CACHE;
3327 for (i = 0; i < bp->b_npages; i++) {
3331 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3336 * cleanup bogus pages, restoring the originals
3339 if (m == bogus_page) {
3341 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3343 panic("biodone: page disappeared!");
3345 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3346 bp->b_pages, bp->b_npages);
3348 #if defined(VFS_BIO_DEBUG)
3349 if (OFF_TO_IDX(foff) != m->pindex) {
3351 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n",
3352 (intmax_t)foff, (uintmax_t)m->pindex);
3357 * In the write case, the valid and clean bits are
3358 * already changed correctly ( see bdwrite() ), so we
3359 * only need to do this here in the read case.
3361 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) {
3362 KASSERT((m->dirty & vm_page_bits(foff &
3363 PAGE_MASK, resid)) == 0, ("bufdone_finish:"
3364 " page %p has unexpected dirty bits", m));
3365 vfs_page_set_valid(bp, foff, m);
3369 * when debugging new filesystems or buffer I/O methods, this
3370 * is the most common error that pops up. if you see this, you
3371 * have not set the page busy flag correctly!!!
3374 printf("biodone: page busy < 0, "
3375 "pindex: %d, foff: 0x(%x,%x), "
3376 "resid: %d, index: %d\n",
3377 (int) m->pindex, (int)(foff >> 32),
3378 (int) foff & 0xffffffff, resid, i);
3379 if (!vn_isdisk(vp, NULL))
3380 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n",
3381 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize,
3382 (intmax_t) bp->b_lblkno,
3383 bp->b_flags, bp->b_npages);
3385 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n",
3386 (intmax_t) bp->b_lblkno,
3387 bp->b_flags, bp->b_npages);
3388 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n",
3389 (u_long)m->valid, (u_long)m->dirty,
3391 panic("biodone: page busy < 0\n");
3393 vm_page_io_finish(m);
3394 vm_object_pip_subtract(obj, 1);
3395 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3398 vm_object_pip_wakeupn(obj, 0);
3399 VM_OBJECT_UNLOCK(obj);
3403 * For asynchronous completions, release the buffer now. The brelse
3404 * will do a wakeup there if necessary - so no need to do a wakeup
3405 * here in the async case. The sync case always needs to do a wakeup.
3408 if (bp->b_flags & B_ASYNC) {
3409 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR))
3418 * This routine is called in lieu of iodone in the case of
3419 * incomplete I/O. This keeps the busy status for pages
3423 vfs_unbusy_pages(struct buf *bp)
3429 runningbufwakeup(bp);
3430 if (!(bp->b_flags & B_VMIO))
3433 obj = bp->b_bufobj->bo_object;
3434 VM_OBJECT_LOCK(obj);
3435 for (i = 0; i < bp->b_npages; i++) {
3437 if (m == bogus_page) {
3438 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
3440 panic("vfs_unbusy_pages: page missing\n");
3442 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3443 bp->b_pages, bp->b_npages);
3445 vm_object_pip_subtract(obj, 1);
3446 vm_page_io_finish(m);
3448 vm_object_pip_wakeupn(obj, 0);
3449 VM_OBJECT_UNLOCK(obj);
3453 * vfs_page_set_valid:
3455 * Set the valid bits in a page based on the supplied offset. The
3456 * range is restricted to the buffer's size.
3458 * This routine is typically called after a read completes.
3461 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3466 * Compute the end offset, eoff, such that [off, eoff) does not span a
3467 * page boundary and eoff is not greater than the end of the buffer.
3468 * The end of the buffer, in this case, is our file EOF, not the
3469 * allocation size of the buffer.
3471 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK;
3472 if (eoff > bp->b_offset + bp->b_bcount)
3473 eoff = bp->b_offset + bp->b_bcount;
3476 * Set valid range. This is typically the entire buffer and thus the
3480 vm_page_set_valid(m, off & PAGE_MASK, eoff - off);
3484 * vfs_page_set_validclean:
3486 * Set the valid bits and clear the dirty bits in a page based on the
3487 * supplied offset. The range is restricted to the buffer's size.
3490 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m)
3492 vm_ooffset_t soff, eoff;
3494 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
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 * This routine is called before a device strategy routine.
3521 * It is used to tell the VM system that paging I/O is in
3522 * progress, and treat the pages associated with the buffer
3523 * almost as being VPO_BUSY. Also the object paging_in_progress
3524 * flag is handled to make sure that the object doesn't become
3527 * Since I/O has not been initiated yet, certain buffer flags
3528 * such as BIO_ERROR or B_INVAL may be in an inconsistant state
3529 * and should be ignored.
3532 vfs_busy_pages(struct buf *bp, int clear_modify)
3539 if (!(bp->b_flags & B_VMIO))
3542 obj = bp->b_bufobj->bo_object;
3543 foff = bp->b_offset;
3544 KASSERT(bp->b_offset != NOOFFSET,
3545 ("vfs_busy_pages: no buffer offset"));
3546 VM_OBJECT_LOCK(obj);
3547 if (bp->b_bufsize != 0)
3548 vfs_setdirty_locked_object(bp);
3550 for (i = 0; i < bp->b_npages; i++) {
3553 if (vm_page_sleep_if_busy(m, FALSE, "vbpage"))
3558 vm_page_lock_queues();
3559 for (i = 0; i < bp->b_npages; i++) {
3562 if ((bp->b_flags & B_CLUSTER) == 0) {
3563 vm_object_pip_add(obj, 1);
3564 vm_page_io_start(m);
3567 * When readying a buffer for a read ( i.e
3568 * clear_modify == 0 ), it is important to do
3569 * bogus_page replacement for valid pages in
3570 * partially instantiated buffers. Partially
3571 * instantiated buffers can, in turn, occur when
3572 * reconstituting a buffer from its VM backing store
3573 * base. We only have to do this if B_CACHE is
3574 * clear ( which causes the I/O to occur in the
3575 * first place ). The replacement prevents the read
3576 * I/O from overwriting potentially dirty VM-backed
3577 * pages. XXX bogus page replacement is, uh, bogus.
3578 * It may not work properly with small-block devices.
3579 * We need to find a better way.
3582 pmap_remove_write(m);
3583 vfs_page_set_validclean(bp, foff, m);
3584 } else if (m->valid == VM_PAGE_BITS_ALL &&
3585 (bp->b_flags & B_CACHE) == 0) {
3586 bp->b_pages[i] = bogus_page;
3589 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3592 vm_page_unlock_queues();
3593 VM_OBJECT_UNLOCK(obj);
3595 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3596 bp->b_pages, bp->b_npages);
3600 * Tell the VM system that the pages associated with this buffer
3601 * are clean. This is used for delayed writes where the data is
3602 * going to go to disk eventually without additional VM intevention.
3604 * Note that while we only really need to clean through to b_bcount, we
3605 * just go ahead and clean through to b_bufsize.
3608 vfs_clean_pages(struct buf *bp)
3611 vm_ooffset_t foff, noff, eoff;
3614 if (!(bp->b_flags & B_VMIO))
3617 foff = bp->b_offset;
3618 KASSERT(bp->b_offset != NOOFFSET,
3619 ("vfs_clean_pages: no buffer offset"));
3620 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3621 vm_page_lock_queues();
3622 for (i = 0; i < bp->b_npages; i++) {
3624 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3627 if (eoff > bp->b_offset + bp->b_bufsize)
3628 eoff = bp->b_offset + bp->b_bufsize;
3629 vfs_page_set_validclean(bp, foff, m);
3630 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3633 vm_page_unlock_queues();
3634 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3638 * vfs_bio_set_valid:
3640 * Set the range within the buffer to valid. The range is
3641 * relative to the beginning of the buffer, b_offset. Note that
3642 * b_offset itself may be offset from the beginning of the first
3646 vfs_bio_set_valid(struct buf *bp, int base, int size)
3651 if (!(bp->b_flags & B_VMIO))
3655 * Fixup base to be relative to beginning of first page.
3656 * Set initial n to be the maximum number of bytes in the
3657 * first page that can be validated.
3659 base += (bp->b_offset & PAGE_MASK);
3660 n = PAGE_SIZE - (base & PAGE_MASK);
3662 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3663 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3667 vm_page_set_valid(m, base & PAGE_MASK, n);
3672 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3676 * vfs_bio_set_validclean:
3678 * Set the range within the buffer to valid and clean. The range is
3679 * relative to the beginning of the buffer, b_offset. Note that b_offset
3680 * itself may be offset from the beginning of the first page.
3685 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3690 if (!(bp->b_flags & B_VMIO))
3693 * Fixup base to be relative to beginning of first page.
3694 * Set initial n to be the maximum number of bytes in the
3695 * first page that can be validated.
3698 base += (bp->b_offset & PAGE_MASK);
3699 n = PAGE_SIZE - (base & PAGE_MASK);
3701 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3702 vm_page_lock_queues();
3703 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3707 vm_page_set_validclean(m, base & PAGE_MASK, n);
3712 vm_page_unlock_queues();
3713 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3719 * If the specified buffer is a non-VMIO buffer, clear the entire
3720 * buffer. If the specified buffer is a VMIO buffer, clear and
3721 * validate only the previously invalid portions of the buffer.
3722 * This routine essentially fakes an I/O, so we need to clear
3723 * BIO_ERROR and B_INVAL.
3725 * Note that while we only theoretically need to clear through b_bcount,
3726 * we go ahead and clear through b_bufsize.
3729 vfs_bio_clrbuf(struct buf *bp)
3734 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) {
3738 bp->b_flags &= ~B_INVAL;
3739 bp->b_ioflags &= ~BIO_ERROR;
3740 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
3741 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3742 (bp->b_offset & PAGE_MASK) == 0) {
3743 if (bp->b_pages[0] == bogus_page)
3745 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3746 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED);
3747 if ((bp->b_pages[0]->valid & mask) == mask)
3749 if ((bp->b_pages[0]->valid & mask) == 0) {
3750 bzero(bp->b_data, bp->b_bufsize);
3751 bp->b_pages[0]->valid |= mask;
3755 ea = sa = bp->b_data;
3756 for(i = 0; i < bp->b_npages; i++, sa = ea) {
3757 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3758 ea = (caddr_t)(vm_offset_t)ulmin(
3759 (u_long)(vm_offset_t)ea,
3760 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3761 if (bp->b_pages[i] == bogus_page)
3763 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3764 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3765 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED);
3766 if ((bp->b_pages[i]->valid & mask) == mask)
3768 if ((bp->b_pages[i]->valid & mask) == 0)
3771 for (; sa < ea; sa += DEV_BSIZE, j++) {
3772 if ((bp->b_pages[i]->valid & (1 << j)) == 0)
3773 bzero(sa, DEV_BSIZE);
3776 bp->b_pages[i]->valid |= mask;
3779 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
3784 * vm_hold_load_pages and vm_hold_free_pages get pages into
3785 * a buffers address space. The pages are anonymous and are
3786 * not associated with a file object.
3789 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3795 to = round_page(to);
3796 from = round_page(from);
3797 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3799 VM_OBJECT_LOCK(kernel_object);
3800 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3803 * note: must allocate system pages since blocking here
3804 * could intefere with paging I/O, no matter which
3807 p = vm_page_alloc(kernel_object,
3808 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3809 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
3811 atomic_add_int(&vm_pageout_deficit,
3812 (to - pg) >> PAGE_SHIFT);
3813 VM_OBJECT_UNLOCK(kernel_object);
3815 VM_OBJECT_LOCK(kernel_object);
3818 p->valid = VM_PAGE_BITS_ALL;
3819 pmap_qenter(pg, &p, 1);
3820 bp->b_pages[index] = p;
3822 VM_OBJECT_UNLOCK(kernel_object);
3823 bp->b_npages = index;
3826 /* Return pages associated with this buf to the vm system */
3828 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3832 int index, newnpages;
3834 from = round_page(from);
3835 to = round_page(to);
3836 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3838 VM_OBJECT_LOCK(kernel_object);
3839 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3840 p = bp->b_pages[index];
3841 if (p && (index < bp->b_npages)) {
3844 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n",
3845 (intmax_t)bp->b_blkno,
3846 (intmax_t)bp->b_lblkno);
3848 bp->b_pages[index] = NULL;
3849 pmap_qremove(pg, 1);
3850 vm_page_lock_queues();
3851 vm_page_unwire(p, 0);
3853 vm_page_unlock_queues();
3856 VM_OBJECT_UNLOCK(kernel_object);
3857 bp->b_npages = newnpages;
3861 * Map an IO request into kernel virtual address space.
3863 * All requests are (re)mapped into kernel VA space.
3864 * Notice that we use b_bufsize for the size of the buffer
3865 * to be mapped. b_bcount might be modified by the driver.
3867 * Note that even if the caller determines that the address space should
3868 * be valid, a race or a smaller-file mapped into a larger space may
3869 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST
3870 * check the return value.
3873 vmapbuf(struct buf *bp)
3879 struct pmap *pmap = &curproc->p_vmspace->vm_pmap;
3881 if (bp->b_bufsize < 0)
3883 prot = VM_PROT_READ;
3884 if (bp->b_iocmd == BIO_READ)
3885 prot |= VM_PROT_WRITE; /* Less backwards than it looks */
3886 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0;
3887 addr < bp->b_data + bp->b_bufsize;
3888 addr += PAGE_SIZE, pidx++) {
3890 * Do the vm_fault if needed; do the copy-on-write thing
3891 * when reading stuff off device into memory.
3893 * NOTE! Must use pmap_extract() because addr may be in
3894 * the userland address space, and kextract is only guarenteed
3895 * to work for the kernland address space (see: sparc64 port).
3898 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data,
3900 vm_page_lock_queues();
3901 for (i = 0; i < pidx; ++i) {
3902 vm_page_unhold(bp->b_pages[i]);
3903 bp->b_pages[i] = NULL;
3905 vm_page_unlock_queues();
3908 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot);
3911 bp->b_pages[pidx] = m;
3913 if (pidx > btoc(MAXPHYS))
3914 panic("vmapbuf: mapped more than MAXPHYS");
3915 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3917 kva = bp->b_saveaddr;
3918 bp->b_npages = pidx;
3919 bp->b_saveaddr = bp->b_data;
3920 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3925 * Free the io map PTEs associated with this IO operation.
3926 * We also invalidate the TLB entries and restore the original b_addr.
3929 vunmapbuf(struct buf *bp)
3934 npages = bp->b_npages;
3935 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3936 vm_page_lock_queues();
3937 for (pidx = 0; pidx < npages; pidx++)
3938 vm_page_unhold(bp->b_pages[pidx]);
3939 vm_page_unlock_queues();
3941 bp->b_data = bp->b_saveaddr;
3945 bdone(struct buf *bp)
3949 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3951 bp->b_flags |= B_DONE;
3957 bwait(struct buf *bp, u_char pri, const char *wchan)
3961 mtxp = mtx_pool_find(mtxpool_sleep, bp);
3963 while ((bp->b_flags & B_DONE) == 0)
3964 msleep(bp, mtxp, pri, wchan, 0);
3969 bufsync(struct bufobj *bo, int waitfor)
3972 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread));
3976 bufstrategy(struct bufobj *bo, struct buf *bp)
3982 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy"));
3983 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
3984 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp));
3985 i = VOP_STRATEGY(vp, bp);
3986 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp));
3990 bufobj_wrefl(struct bufobj *bo)
3993 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
3994 ASSERT_BO_LOCKED(bo);
3999 bufobj_wref(struct bufobj *bo)
4002 KASSERT(bo != NULL, ("NULL bo in bufobj_wref"));
4009 bufobj_wdrop(struct bufobj *bo)
4012 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop"));
4014 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count"));
4015 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) {
4016 bo->bo_flag &= ~BO_WWAIT;
4017 wakeup(&bo->bo_numoutput);
4023 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo)
4027 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait"));
4028 ASSERT_BO_LOCKED(bo);
4030 while (bo->bo_numoutput) {
4031 bo->bo_flag |= BO_WWAIT;
4032 error = msleep(&bo->bo_numoutput, BO_MTX(bo),
4033 slpflag | (PRIBIO + 1), "bo_wwait", timeo);
4041 bpin(struct buf *bp)
4045 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4052 bunpin(struct buf *bp)
4056 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4058 if (--bp->b_pin_count == 0)
4064 bunpin_wait(struct buf *bp)
4068 mtxp = mtx_pool_find(mtxpool_sleep, bp);
4070 while (bp->b_pin_count > 0)
4071 msleep(bp, mtxp, PRIBIO, "bwunpin", 0);
4075 #include "opt_ddb.h"
4077 #include <ddb/ddb.h>
4079 /* DDB command to show buffer data */
4080 DB_SHOW_COMMAND(buffer, db_show_buffer)
4083 struct buf *bp = (struct buf *)addr;
4086 db_printf("usage: show buffer <addr>\n");
4090 db_printf("buf at %p\n", bp);
4091 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4093 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n"
4094 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n",
4095 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4096 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno,
4097 bp->b_dep.lh_first);
4100 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
4101 for (i = 0; i < bp->b_npages; i++) {
4104 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4105 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4106 if ((i + 1) < bp->b_npages)
4112 lockmgr_printinfo(&bp->b_lock);
4115 DB_SHOW_COMMAND(lockedbufs, lockedbufs)
4120 for (i = 0; i < nbuf; i++) {
4122 if (BUF_ISLOCKED(bp)) {
4123 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4129 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs)
4135 db_printf("usage: show vnodebufs <addr>\n");
4138 vp = (struct vnode *)addr;
4139 db_printf("Clean buffers:\n");
4140 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) {
4141 db_show_buffer((uintptr_t)bp, 1, 0, NULL);
4144 db_printf("Dirty buffers:\n");
4145 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
4146 db_show_buffer((uintptr_t)bp, 1, 0, NULL);