2 * Copyright (c) 1990 University of Utah.
3 * Copyright (c) 1991 The Regents of the University of California.
5 * Copyright (c) 1993, 1994 John S. Dyson
6 * Copyright (c) 1995, David Greenman
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91
41 * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
42 * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $
46 * Page to/from files (vnodes).
51 * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
52 * greatly re-simplify the vnode_pager.
55 #include <sys/param.h>
56 #include <sys/systm.h>
57 #include <sys/kernel.h>
59 #include <sys/vnode.h>
60 #include <sys/mount.h>
62 #include <sys/vmmeter.h>
64 #include <sys/sfbuf.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_map.h>
71 #include <vm/vnode_pager.h>
72 #include <vm/vm_extern.h>
74 #include <sys/thread2.h>
75 #include <vm/vm_page2.h>
77 static void vnode_pager_dealloc (vm_object_t
);
78 static int vnode_pager_getpages (vm_object_t
, vm_page_t
*, int, int);
79 static void vnode_pager_putpages (vm_object_t
, vm_page_t
*, int, boolean_t
, int *);
80 static boolean_t
vnode_pager_haspage (vm_object_t
, vm_pindex_t
, int *, int *);
82 struct pagerops vnodepagerops
= {
92 static struct krate vbadrate
= { 1 };
93 static struct krate vresrate
= { 1 };
95 int vnode_pbuf_freecnt
= -1; /* start out unlimited */
98 * Allocate (or lookup) pager for a vnode.
99 * Handle is a vnode pointer.
102 vnode_pager_alloc(void *handle
, off_t size
, vm_prot_t prot
, off_t offset
)
108 * Pageout to vnode, no can do yet.
114 * XXX hack - This initialization should be put somewhere else.
116 if (vnode_pbuf_freecnt
< 0) {
117 vnode_pbuf_freecnt
= nswbuf
/ 2 + 1;
120 vp
= (struct vnode
*) handle
;
123 * Prevent race condition when allocating the object. This
124 * can happen with NFS vnodes since the nfsnode isn't locked.
126 while (vp
->v_flag
& VOLOCK
) {
127 vp
->v_flag
|= VOWANT
;
128 tsleep(vp
, 0, "vnpobj", 0);
130 vp
->v_flag
|= VOLOCK
;
133 * If the object is being terminated, wait for it to
136 while (((object
= vp
->v_object
) != NULL
) &&
137 (object
->flags
& OBJ_DEAD
)) {
138 vm_object_dead_sleep(object
, "vadead");
141 if (vp
->v_sysref
.refcnt
<= 0)
142 panic("vnode_pager_alloc: no vnode reference");
144 if (object
== NULL
) {
146 * And an object of the appropriate size
148 object
= vm_object_allocate(OBJT_VNODE
, OFF_TO_IDX(round_page(size
)));
150 object
->handle
= handle
;
151 vp
->v_object
= object
;
152 vp
->v_filesize
= size
;
155 if (vp
->v_filesize
!= size
) {
156 kprintf("vnode_pager_alloc: Warning, filesize "
157 "mismatch %lld/%lld\n",
158 (long long)vp
->v_filesize
,
164 vp
->v_flag
&= ~VOLOCK
;
165 if (vp
->v_flag
& VOWANT
) {
166 vp
->v_flag
&= ~VOWANT
;
173 vnode_pager_dealloc(vm_object_t object
)
175 struct vnode
*vp
= object
->handle
;
178 panic("vnode_pager_dealloc: pager already dealloced");
180 vm_object_pip_wait(object
, "vnpdea");
182 object
->handle
= NULL
;
183 object
->type
= OBJT_DEAD
;
185 vp
->v_filesize
= NOOFFSET
;
186 vp
->v_flag
&= ~(VTEXT
| VOBJBUF
);
190 * Return whether the vnode pager has the requested page. Return the
191 * number of disk-contiguous pages before and after the requested page,
192 * not including the requested page.
195 vnode_pager_haspage(vm_object_t object
, vm_pindex_t pindex
, int *before
,
198 struct vnode
*vp
= object
->handle
;
206 * If no vp or vp is doomed or marked transparent to VM, we do not
209 if ((vp
== NULL
) || (vp
->v_flag
& VRECLAIMED
))
213 * If filesystem no longer mounted or offset beyond end of file we do
216 loffset
= IDX_TO_OFF(pindex
);
218 if (vp
->v_mount
== NULL
|| loffset
>= vp
->v_filesize
)
221 bsize
= vp
->v_mount
->mnt_stat
.f_iosize
;
222 voff
= loffset
% bsize
;
225 * BMAP returns byte counts before and after, where after
226 * is inclusive of the base page. haspage must return page
227 * counts before and after where after does not include the
230 * BMAP is allowed to return a *after of 0 for backwards
231 * compatibility. The base page is still considered valid if
232 * no error is returned.
234 error
= VOP_BMAP(vp
, loffset
- voff
, &doffset
, after
, before
, 0);
242 if (doffset
== NOOFFSET
)
246 *before
= (*before
+ voff
) >> PAGE_SHIFT
;
250 if (loffset
+ *after
> vp
->v_filesize
)
251 *after
= vp
->v_filesize
- loffset
;
252 *after
>>= PAGE_SHIFT
;
260 * Lets the VM system know about a change in size for a file.
261 * We adjust our own internal size and flush any cached pages in
262 * the associated object that are affected by the size change.
264 * NOTE: This routine may be invoked as a result of a pager put
265 * operation (possibly at object termination time), so we must be careful.
267 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
268 * we do not blow up on the case. nsize will always be >= 0, however.
271 vnode_pager_setsize(struct vnode
*vp
, vm_ooffset_t nsize
)
273 vm_pindex_t nobjsize
;
274 vm_pindex_t oobjsize
;
275 vm_object_t object
= vp
->v_object
;
281 * Hasn't changed size
283 if (nsize
== vp
->v_filesize
)
287 * Has changed size. Adjust the VM object's size and v_filesize
288 * before we start scanning pages to prevent new pages from being
289 * allocated during the scan.
291 nobjsize
= OFF_TO_IDX(nsize
+ PAGE_MASK
);
292 oobjsize
= object
->size
;
293 object
->size
= nobjsize
;
296 * File has shrunk. Toss any cached pages beyond the new EOF.
298 if (nsize
< vp
->v_filesize
) {
299 vp
->v_filesize
= nsize
;
300 if (nobjsize
< oobjsize
) {
301 vm_object_page_remove(object
, nobjsize
, oobjsize
,
305 * This gets rid of garbage at the end of a page that is now
306 * only partially backed by the vnode. Since we are setting
307 * the entire page valid & clean after we are done we have
308 * to be sure that the portion of the page within the file
309 * bounds is already valid. If it isn't then making it
310 * valid would create a corrupt block.
312 if (nsize
& PAGE_MASK
) {
317 m
= vm_page_lookup(object
, OFF_TO_IDX(nsize
));
318 } while (m
&& vm_page_sleep_busy(m
, TRUE
, "vsetsz"));
321 int base
= (int)nsize
& PAGE_MASK
;
322 int size
= PAGE_SIZE
- base
;
326 * Clear out partial-page garbage in case
327 * the page has been mapped.
329 * This is byte aligned.
332 sf
= sf_buf_alloc(m
, SFB_CPUPRIVATE
);
333 kva
= sf_buf_kva(sf
);
334 bzero((caddr_t
)kva
+ base
, size
);
338 * XXX work around SMP data integrity race
339 * by unmapping the page from user processes.
340 * The garbage we just cleared may be mapped
341 * to a user process running on another cpu
342 * and this code is not running through normal
343 * I/O channels which handle SMP issues for
344 * us, so unmap page to synchronize all cpus.
346 * XXX should vm_pager_unmap_page() have
349 vm_page_protect(m
, VM_PROT_NONE
);
352 * Clear out partial-page dirty bits. This
353 * has the side effect of setting the valid
354 * bits, but that is ok. There are a bunch
355 * of places in the VM system where we expected
356 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
357 * case is one of them. If the page is still
358 * partially dirty, make it fully dirty.
360 * NOTE: We do not clear out the valid
361 * bits. This would prevent bogus_page
362 * replacement from working properly.
364 * NOTE: We do not want to clear the dirty
365 * bit for a partial DEV_BSIZE'd truncation!
366 * This is DEV_BSIZE aligned!
368 vm_page_clear_dirty_beg_nonincl(m
, base
, size
);
370 m
->dirty
= VM_PAGE_BITS_ALL
;
375 vp
->v_filesize
= nsize
;
380 * Release a page busied for a getpages operation. The page may have become
381 * wired (typically due to being used by the buffer cache) or otherwise been
382 * soft-busied and cannot be freed in that case. A held page can still be
386 vnode_pager_freepage(vm_page_t m
)
388 if (m
->busy
|| m
->wire_count
) {
397 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
398 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
399 * vnode_pager_generic_getpages() to implement the previous behaviour.
401 * All other FS's should use the bypass to get to the local media
402 * backing vp's VOP_GETPAGES.
405 vnode_pager_getpages(vm_object_t object
, vm_page_t
*m
, int count
, int reqpage
)
409 int bytes
= count
* PAGE_SIZE
;
412 rtval
= VOP_GETPAGES(vp
, m
, bytes
, reqpage
, 0);
413 if (rtval
== EOPNOTSUPP
)
414 panic("vnode_pager: vfs's must implement vop_getpages\n");
419 * This is now called from local media FS's to operate against their
420 * own vnodes if they fail to implement VOP_GETPAGES.
422 * With all the caching local media devices do these days there is really
423 * very little point to attempting to restrict the I/O size to contiguous
424 * blocks on-disk, especially if our caller thinks we need all the specified
425 * pages. Just construct and issue a READ.
428 vnode_pager_generic_getpages(struct vnode
*vp
, vm_page_t
*m
, int bytecount
,
440 * Do not do anything if the vnode is bad.
442 if (vp
->v_mount
== NULL
)
446 * Calculate the number of pages. Since we are paging in whole
447 * pages, adjust bytecount to be an integral multiple of the page
448 * size. It will be clipped to the file EOF later on.
450 bytecount
= round_page(bytecount
);
451 count
= bytecount
/ PAGE_SIZE
;
454 * If we have a completely valid page available to us, we can
455 * clean up and return. Otherwise we have to re-read the
458 * Note that this does not work with NFS, so NFS has its own
459 * getpages routine. The problem is that NFS can have partially
460 * valid pages associated with the buffer cache due to the piecemeal
461 * write support. If we were to fall through and re-read the media
462 * as we do here, dirty data could be lost.
464 if (m
[reqpage
]->valid
== VM_PAGE_BITS_ALL
) {
465 for (i
= 0; i
< count
; i
++) {
467 vnode_pager_freepage(m
[i
]);
473 * Discard pages past the file EOF. If the requested page is past
474 * the file EOF we just leave its valid bits set to 0, the caller
475 * expects to maintain ownership of the requested page. If the
476 * entire range is past file EOF discard everything and generate
479 foff
= IDX_TO_OFF(m
[0]->pindex
);
480 if (foff
>= vp
->v_filesize
) {
481 for (i
= 0; i
< count
; i
++) {
483 vnode_pager_freepage(m
[i
]);
485 return VM_PAGER_ERROR
;
488 if (foff
+ bytecount
> vp
->v_filesize
) {
489 bytecount
= vp
->v_filesize
- foff
;
490 i
= round_page(bytecount
) / PAGE_SIZE
;
493 if (count
!= reqpage
)
494 vnode_pager_freepage(m
[count
]);
499 * The size of the transfer is bytecount. bytecount will be an
500 * integral multiple of the page size unless it has been clipped
501 * to the file EOF. The transfer cannot exceed the file EOF.
503 * When dealing with real devices we must round-up to the device
506 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
) {
507 int secmask
= vp
->v_rdev
->si_bsize_phys
- 1;
508 KASSERT(secmask
< PAGE_SIZE
, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask
+ 1));
509 bytecount
= (bytecount
+ secmask
) & ~secmask
;
513 * Severe hack to avoid deadlocks with the buffer cache
515 for (i
= 0; i
< count
; ++i
) {
518 vm_page_io_start(mt
);
523 * Issue the I/O without any read-ahead
526 /*ioflags |= IO_SEQMAX << IO_SEQSHIFT;*/
528 aiov
.iov_base
= (caddr_t
) 0;
529 aiov
.iov_len
= bytecount
;
530 auio
.uio_iov
= &aiov
;
532 auio
.uio_offset
= foff
;
533 auio
.uio_segflg
= UIO_NOCOPY
;
534 auio
.uio_rw
= UIO_READ
;
535 auio
.uio_resid
= bytecount
;
537 mycpu
->gd_cnt
.v_vnodein
++;
538 mycpu
->gd_cnt
.v_vnodepgsin
+= count
;
540 error
= VOP_READ(vp
, &auio
, ioflags
, proc0
.p_ucred
);
543 * Severe hack to avoid deadlocks with the buffer cache
545 for (i
= 0; i
< count
; ++i
) {
548 while (vm_page_sleep_busy(mt
, FALSE
, "getpgs"))
551 vm_page_io_finish(mt
);
555 * Calculate the actual number of bytes read and clean up the
558 bytecount
-= auio
.uio_resid
;
560 for (i
= 0; i
< count
; ++i
) {
564 if (error
== 0 && mt
->valid
) {
565 if (mt
->flags
& PG_WANTED
)
566 vm_page_activate(mt
);
568 vm_page_deactivate(mt
);
571 vnode_pager_freepage(mt
);
573 } else if (mt
->valid
== 0) {
575 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt
, mt
->object
, (int) mt
->pindex
);
576 /* whoops, something happened */
579 } else if (mt
->valid
!= VM_PAGE_BITS_ALL
) {
581 * Zero-extend the requested page if necessary (if
582 * the filesystem is using a small block size).
584 vm_page_zero_invalid(mt
, TRUE
);
588 kprintf("vnode_pager_getpages: I/O read error\n");
590 return (error
? VM_PAGER_ERROR
: VM_PAGER_OK
);
594 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
595 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
596 * vnode_pager_generic_putpages() to implement the previous behaviour.
598 * Caller has already cleared the pmap modified bits, if any.
600 * All other FS's should use the bypass to get to the local media
601 * backing vp's VOP_PUTPAGES.
604 vnode_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
605 boolean_t sync
, int *rtvals
)
609 int bytes
= count
* PAGE_SIZE
;
612 * Force synchronous operation if we are extremely low on memory
613 * to prevent a low-memory deadlock. VOP operations often need to
614 * allocate more memory to initiate the I/O ( i.e. do a BMAP
615 * operation ). The swapper handles the case by limiting the amount
616 * of asynchronous I/O, but that sort of solution doesn't scale well
617 * for the vnode pager without a lot of work.
619 * Also, the backing vnode's iodone routine may not wake the pageout
620 * daemon up. This should be probably be addressed XXX.
623 if ((vmstats
.v_free_count
+ vmstats
.v_cache_count
) < vmstats
.v_pageout_free_min
)
627 * Call device-specific putpages function
630 rtval
= VOP_PUTPAGES(vp
, m
, bytes
, sync
, rtvals
, 0);
631 if (rtval
== EOPNOTSUPP
) {
632 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
633 rtval
= vnode_pager_generic_putpages( vp
, m
, bytes
, sync
, rtvals
);
639 * This is now called from local media FS's to operate against their
640 * own vnodes if they fail to implement VOP_PUTPAGES.
642 * This is typically called indirectly via the pageout daemon and
643 * clustering has already typically occured, so in general we ask the
644 * underlying filesystem to write the data out asynchronously rather
648 vnode_pager_generic_putpages(struct vnode
*vp
, vm_page_t
*m
, int bytecount
,
649 int flags
, int *rtvals
)
653 int maxsize
, ncount
, count
;
654 vm_ooffset_t poffset
;
660 object
= vp
->v_object
;
661 count
= bytecount
/ PAGE_SIZE
;
663 for (i
= 0; i
< count
; i
++)
664 rtvals
[i
] = VM_PAGER_AGAIN
;
666 if ((int) m
[0]->pindex
< 0) {
667 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
668 (long)m
[0]->pindex
, m
[0]->dirty
);
669 rtvals
[0] = VM_PAGER_BAD
;
673 maxsize
= count
* PAGE_SIZE
;
676 poffset
= IDX_TO_OFF(m
[0]->pindex
);
679 * If the page-aligned write is larger then the actual file we
680 * have to invalidate pages occuring beyond the file EOF.
682 * If the file EOF resides in the middle of a page we still clear
683 * all of that page's dirty bits later on. If we didn't it would
684 * endlessly re-write.
686 * We do not under any circumstances truncate the valid bits, as
687 * this will screw up bogus page replacement.
689 * The caller has already read-protected the pages. The VFS must
690 * use the buffer cache to wrap the pages. The pages might not
691 * be immediately flushed by the buffer cache but once under its
692 * control the pages themselves can wind up being marked clean
693 * and their covering buffer cache buffer can be marked dirty.
695 if (poffset
+ maxsize
> vp
->v_filesize
) {
696 if (poffset
< vp
->v_filesize
) {
697 maxsize
= vp
->v_filesize
- poffset
;
698 ncount
= btoc(maxsize
);
703 if (ncount
< count
) {
704 for (i
= ncount
; i
< count
; i
++) {
705 rtvals
[i
] = VM_PAGER_BAD
;
711 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
712 * rather then a bdwrite() to prevent paging I/O from saturating
713 * the buffer cache. Dummy-up the sequential heuristic to cause
714 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
715 * the system decides how to cluster.
718 if (flags
& (VM_PAGER_PUT_SYNC
| VM_PAGER_PUT_INVAL
))
720 else if ((flags
& VM_PAGER_CLUSTER_OK
) == 0)
722 ioflags
|= (flags
& VM_PAGER_PUT_INVAL
) ? IO_INVAL
: 0;
723 ioflags
|= IO_SEQMAX
<< IO_SEQSHIFT
;
725 aiov
.iov_base
= (caddr_t
) 0;
726 aiov
.iov_len
= maxsize
;
727 auio
.uio_iov
= &aiov
;
729 auio
.uio_offset
= poffset
;
730 auio
.uio_segflg
= UIO_NOCOPY
;
731 auio
.uio_rw
= UIO_WRITE
;
732 auio
.uio_resid
= maxsize
;
734 error
= VOP_WRITE(vp
, &auio
, ioflags
, proc0
.p_ucred
);
735 mycpu
->gd_cnt
.v_vnodeout
++;
736 mycpu
->gd_cnt
.v_vnodepgsout
+= ncount
;
739 krateprintf(&vbadrate
,
740 "vnode_pager_putpages: I/O error %d\n", error
);
742 if (auio
.uio_resid
) {
743 krateprintf(&vresrate
,
744 "vnode_pager_putpages: residual I/O %zd at %lu\n",
745 auio
.uio_resid
, (u_long
)m
[0]->pindex
);
748 for (i
= 0; i
< ncount
; i
++) {
749 rtvals
[i
] = VM_PAGER_OK
;
750 vm_page_undirty(m
[i
]);
757 vnode_pager_lock(vm_object_t object
)
759 struct thread
*td
= curthread
; /* XXX */
762 for (; object
!= NULL
; object
= object
->backing_object
) {
763 if (object
->type
!= OBJT_VNODE
)
765 if (object
->flags
& OBJ_DEAD
)
769 struct vnode
*vp
= object
->handle
;
770 error
= vget(vp
, LK_SHARED
| LK_RETRY
| LK_CANRECURSE
);
772 if (object
->handle
!= vp
) {
778 if ((object
->flags
& OBJ_DEAD
) ||
779 (object
->type
!= OBJT_VNODE
)) {
782 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp
, error
, lockstatus(&vp
->v_lock
, td
));
783 tsleep(object
->handle
, 0, "vnpgrl", hz
);