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>
65 #include <sys/thread2.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_pager.h>
71 #include <vm/vm_map.h>
72 #include <vm/vnode_pager.h>
73 #include <vm/vm_extern.h>
75 static void vnode_pager_dealloc (vm_object_t
);
76 static int vnode_pager_getpages (vm_object_t
, vm_page_t
*, int, int);
77 static void vnode_pager_putpages (vm_object_t
, vm_page_t
*, int, boolean_t
, int *);
78 static boolean_t
vnode_pager_haspage (vm_object_t
, vm_pindex_t
, int *, int *);
80 struct pagerops vnodepagerops
= {
90 static struct krate vbadrate
= { 1 };
91 static struct krate vresrate
= { 1 };
93 int vnode_pbuf_freecnt
= -1; /* start out unlimited */
96 * Allocate (or lookup) pager for a vnode.
97 * Handle is a vnode pointer.
100 vnode_pager_alloc(void *handle
, off_t size
, vm_prot_t prot
, off_t offset
)
106 * Pageout to vnode, no can do yet.
112 * XXX hack - This initialization should be put somewhere else.
114 if (vnode_pbuf_freecnt
< 0) {
115 vnode_pbuf_freecnt
= nswbuf
/ 2 + 1;
118 vp
= (struct vnode
*) handle
;
121 * Prevent race condition when allocating the object. This
122 * can happen with NFS vnodes since the nfsnode isn't locked.
124 while (vp
->v_flag
& VOLOCK
) {
125 vp
->v_flag
|= VOWANT
;
126 tsleep(vp
, 0, "vnpobj", 0);
128 vp
->v_flag
|= VOLOCK
;
131 * If the object is being terminated, wait for it to
134 while (((object
= vp
->v_object
) != NULL
) &&
135 (object
->flags
& OBJ_DEAD
)) {
136 vm_object_dead_sleep(object
, "vadead");
139 if (vp
->v_sysref
.refcnt
<= 0)
140 panic("vnode_pager_alloc: no vnode reference");
142 if (object
== NULL
) {
144 * And an object of the appropriate size
146 object
= vm_object_allocate(OBJT_VNODE
, OFF_TO_IDX(round_page(size
)));
148 object
->handle
= handle
;
149 vp
->v_object
= object
;
150 vp
->v_filesize
= size
;
153 if (vp
->v_filesize
!= size
) {
154 kprintf("vnode_pager_alloc: Warning, filesize "
155 "mismatch %lld/%lld\n",
156 (long long)vp
->v_filesize
,
162 vp
->v_flag
&= ~VOLOCK
;
163 if (vp
->v_flag
& VOWANT
) {
164 vp
->v_flag
&= ~VOWANT
;
171 vnode_pager_dealloc(vm_object_t object
)
173 struct vnode
*vp
= object
->handle
;
176 panic("vnode_pager_dealloc: pager already dealloced");
178 vm_object_pip_wait(object
, "vnpdea");
180 object
->handle
= NULL
;
181 object
->type
= OBJT_DEAD
;
183 vp
->v_filesize
= NOOFFSET
;
184 vp
->v_flag
&= ~(VTEXT
| VOBJBUF
);
188 * Return whether the vnode pager has the requested page. Return the
189 * number of disk-contiguous pages before and after the requested page,
190 * not including the requested page.
193 vnode_pager_haspage(vm_object_t object
, vm_pindex_t pindex
, int *before
,
196 struct vnode
*vp
= object
->handle
;
204 * If no vp or vp is doomed or marked transparent to VM, we do not
207 if ((vp
== NULL
) || (vp
->v_flag
& VRECLAIMED
))
211 * If filesystem no longer mounted or offset beyond end of file we do
214 loffset
= IDX_TO_OFF(pindex
);
216 if (vp
->v_mount
== NULL
|| loffset
>= vp
->v_filesize
)
219 bsize
= vp
->v_mount
->mnt_stat
.f_iosize
;
220 voff
= loffset
% bsize
;
223 * BMAP returns byte counts before and after, where after
224 * is inclusive of the base page. haspage must return page
225 * counts before and after where after does not include the
228 * BMAP is allowed to return a *after of 0 for backwards
229 * compatibility. The base page is still considered valid if
230 * no error is returned.
232 error
= VOP_BMAP(vp
, loffset
- voff
, &doffset
, after
, before
, 0);
240 if (doffset
== NOOFFSET
)
244 *before
= (*before
+ voff
) >> PAGE_SHIFT
;
248 if (loffset
+ *after
> vp
->v_filesize
)
249 *after
= vp
->v_filesize
- loffset
;
250 *after
>>= PAGE_SHIFT
;
258 * Lets the VM system know about a change in size for a file.
259 * We adjust our own internal size and flush any cached pages in
260 * the associated object that are affected by the size change.
262 * NOTE: This routine may be invoked as a result of a pager put
263 * operation (possibly at object termination time), so we must be careful.
265 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
266 * we do not blow up on the case. nsize will always be >= 0, however.
269 vnode_pager_setsize(struct vnode
*vp
, vm_ooffset_t nsize
)
271 vm_pindex_t nobjsize
;
272 vm_pindex_t oobjsize
;
273 vm_object_t object
= vp
->v_object
;
279 * Hasn't changed size
281 if (nsize
== vp
->v_filesize
)
285 * Has changed size. Adjust the VM object's size and v_filesize
286 * before we start scanning pages to prevent new pages from being
287 * allocated during the scan.
289 nobjsize
= OFF_TO_IDX(nsize
+ PAGE_MASK
);
290 oobjsize
= object
->size
;
291 object
->size
= nobjsize
;
294 * File has shrunk. Toss any cached pages beyond the new EOF.
296 if (nsize
< vp
->v_filesize
) {
297 vp
->v_filesize
= nsize
;
298 if (nobjsize
< oobjsize
) {
299 vm_object_page_remove(object
, nobjsize
, oobjsize
,
303 * This gets rid of garbage at the end of a page that is now
304 * only partially backed by the vnode. Since we are setting
305 * the entire page valid & clean after we are done we have
306 * to be sure that the portion of the page within the file
307 * bounds is already valid. If it isn't then making it
308 * valid would create a corrupt block.
310 if (nsize
& PAGE_MASK
) {
315 m
= vm_page_lookup(object
, OFF_TO_IDX(nsize
));
316 } while (m
&& vm_page_sleep_busy(m
, TRUE
, "vsetsz"));
319 int base
= (int)nsize
& PAGE_MASK
;
320 int size
= PAGE_SIZE
- base
;
324 * Clear out partial-page garbage in case
325 * the page has been mapped.
328 sf
= sf_buf_alloc(m
, SFB_CPUPRIVATE
);
329 kva
= sf_buf_kva(sf
);
330 bzero((caddr_t
)kva
+ base
, size
);
334 * XXX work around SMP data integrity race
335 * by unmapping the page from user processes.
336 * The garbage we just cleared may be mapped
337 * to a user process running on another cpu
338 * and this code is not running through normal
339 * I/O channels which handle SMP issues for
340 * us, so unmap page to synchronize all cpus.
342 * XXX should vm_pager_unmap_page() have
345 vm_page_protect(m
, VM_PROT_NONE
);
348 * Clear out partial-page dirty bits. This
349 * has the side effect of setting the valid
350 * bits, but that is ok. There are a bunch
351 * of places in the VM system where we expected
352 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
353 * case is one of them. If the page is still
354 * partially dirty, make it fully dirty.
356 * note that we do not clear out the valid
357 * bits. This would prevent bogus_page
358 * replacement from working properly.
360 vm_page_set_validclean(m
, base
, size
);
362 m
->dirty
= VM_PAGE_BITS_ALL
;
367 vp
->v_filesize
= nsize
;
372 * Release a page busied for a getpages operation. The page may have become
373 * wired (typically due to being used by the buffer cache) or otherwise been
374 * soft-busied and cannot be freed in that case. A held page can still be
378 vnode_pager_freepage(vm_page_t m
)
380 if (m
->busy
|| m
->wire_count
) {
389 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
390 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
391 * vnode_pager_generic_getpages() to implement the previous behaviour.
393 * All other FS's should use the bypass to get to the local media
394 * backing vp's VOP_GETPAGES.
397 vnode_pager_getpages(vm_object_t object
, vm_page_t
*m
, int count
, int reqpage
)
401 int bytes
= count
* PAGE_SIZE
;
404 rtval
= VOP_GETPAGES(vp
, m
, bytes
, reqpage
, 0);
405 if (rtval
== EOPNOTSUPP
)
406 panic("vnode_pager: vfs's must implement vop_getpages\n");
411 * This is now called from local media FS's to operate against their
412 * own vnodes if they fail to implement VOP_GETPAGES.
414 * With all the caching local media devices do these days there is really
415 * very little point to attempting to restrict the I/O size to contiguous
416 * blocks on-disk, especially if our caller thinks we need all the specified
417 * pages. Just construct and issue a READ.
420 vnode_pager_generic_getpages(struct vnode
*vp
, vm_page_t
*m
, int bytecount
,
432 * Do not do anything if the vnode is bad.
434 if (vp
->v_mount
== NULL
)
438 * Calculate the number of pages. Since we are paging in whole
439 * pages, adjust bytecount to be an integral multiple of the page
440 * size. It will be clipped to the file EOF later on.
442 bytecount
= round_page(bytecount
);
443 count
= bytecount
/ PAGE_SIZE
;
446 * If we have a completely valid page available to us, we can
447 * clean up and return. Otherwise we have to re-read the
450 * Note that this does not work with NFS, so NFS has its own
451 * getpages routine. The problem is that NFS can have partially
452 * valid pages associated with the buffer cache due to the piecemeal
453 * write support. If we were to fall through and re-read the media
454 * as we do here, dirty data could be lost.
456 if (m
[reqpage
]->valid
== VM_PAGE_BITS_ALL
) {
457 for (i
= 0; i
< count
; i
++) {
459 vnode_pager_freepage(m
[i
]);
465 * Discard pages past the file EOF. If the requested page is past
466 * the file EOF we just leave its valid bits set to 0, the caller
467 * expects to maintain ownership of the requested page. If the
468 * entire range is past file EOF discard everything and generate
471 foff
= IDX_TO_OFF(m
[0]->pindex
);
472 if (foff
>= vp
->v_filesize
) {
473 for (i
= 0; i
< count
; i
++) {
475 vnode_pager_freepage(m
[i
]);
477 return VM_PAGER_ERROR
;
480 if (foff
+ bytecount
> vp
->v_filesize
) {
481 bytecount
= vp
->v_filesize
- foff
;
482 i
= round_page(bytecount
) / PAGE_SIZE
;
485 if (count
!= reqpage
)
486 vnode_pager_freepage(m
[count
]);
491 * The size of the transfer is bytecount. bytecount will be an
492 * integral multiple of the page size unless it has been clipped
493 * to the file EOF. The transfer cannot exceed the file EOF.
495 * When dealing with real devices we must round-up to the device
498 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
) {
499 int secmask
= vp
->v_rdev
->si_bsize_phys
- 1;
500 KASSERT(secmask
< PAGE_SIZE
, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask
+ 1));
501 bytecount
= (bytecount
+ secmask
) & ~secmask
;
505 * Severe hack to avoid deadlocks with the buffer cache
507 for (i
= 0; i
< count
; ++i
) {
510 vm_page_io_start(mt
);
515 * Issue the I/O without any read-ahead
518 /*ioflags |= IO_SEQMAX << IO_SEQSHIFT;*/
520 aiov
.iov_base
= (caddr_t
) 0;
521 aiov
.iov_len
= bytecount
;
522 auio
.uio_iov
= &aiov
;
524 auio
.uio_offset
= foff
;
525 auio
.uio_segflg
= UIO_NOCOPY
;
526 auio
.uio_rw
= UIO_READ
;
527 auio
.uio_resid
= bytecount
;
529 mycpu
->gd_cnt
.v_vnodein
++;
530 mycpu
->gd_cnt
.v_vnodepgsin
+= count
;
532 error
= VOP_READ(vp
, &auio
, ioflags
, proc0
.p_ucred
);
535 * Severe hack to avoid deadlocks with the buffer cache
537 for (i
= 0; i
< count
; ++i
) {
540 while (vm_page_sleep_busy(mt
, FALSE
, "getpgs"))
543 vm_page_io_finish(mt
);
547 * Calculate the actual number of bytes read and clean up the
550 bytecount
-= auio
.uio_resid
;
552 for (i
= 0; i
< count
; ++i
) {
556 if (error
== 0 && mt
->valid
) {
557 if (mt
->flags
& PG_WANTED
)
558 vm_page_activate(mt
);
560 vm_page_deactivate(mt
);
563 vnode_pager_freepage(mt
);
565 } else if (mt
->valid
== 0) {
567 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt
, mt
->object
, (int) mt
->pindex
);
568 /* whoops, something happened */
571 } else if (mt
->valid
!= VM_PAGE_BITS_ALL
) {
573 * Zero-extend the requested page if necessary (if
574 * the filesystem is using a small block size).
576 vm_page_zero_invalid(mt
, TRUE
);
580 kprintf("vnode_pager_getpages: I/O read error\n");
582 return (error
? VM_PAGER_ERROR
: VM_PAGER_OK
);
586 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
587 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
588 * vnode_pager_generic_putpages() to implement the previous behaviour.
590 * All other FS's should use the bypass to get to the local media
591 * backing vp's VOP_PUTPAGES.
594 vnode_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
595 boolean_t sync
, int *rtvals
)
599 int bytes
= count
* PAGE_SIZE
;
602 * Force synchronous operation if we are extremely low on memory
603 * to prevent a low-memory deadlock. VOP operations often need to
604 * allocate more memory to initiate the I/O ( i.e. do a BMAP
605 * operation ). The swapper handles the case by limiting the amount
606 * of asynchronous I/O, but that sort of solution doesn't scale well
607 * for the vnode pager without a lot of work.
609 * Also, the backing vnode's iodone routine may not wake the pageout
610 * daemon up. This should be probably be addressed XXX.
613 if ((vmstats
.v_free_count
+ vmstats
.v_cache_count
) < vmstats
.v_pageout_free_min
)
617 * Call device-specific putpages function
621 rtval
= VOP_PUTPAGES(vp
, m
, bytes
, sync
, rtvals
, 0);
622 if (rtval
== EOPNOTSUPP
) {
623 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
624 rtval
= vnode_pager_generic_putpages( vp
, m
, bytes
, sync
, rtvals
);
630 * This is now called from local media FS's to operate against their
631 * own vnodes if they fail to implement VOP_PUTPAGES.
633 * This is typically called indirectly via the pageout daemon and
634 * clustering has already typically occured, so in general we ask the
635 * underlying filesystem to write the data out asynchronously rather
639 vnode_pager_generic_putpages(struct vnode
*vp
, vm_page_t
*m
, int bytecount
,
640 int flags
, int *rtvals
)
647 vm_ooffset_t poffset
;
653 object
= vp
->v_object
;
654 count
= bytecount
/ PAGE_SIZE
;
656 for (i
= 0; i
< count
; i
++)
657 rtvals
[i
] = VM_PAGER_AGAIN
;
659 if ((int) m
[0]->pindex
< 0) {
660 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
661 (long)m
[0]->pindex
, m
[0]->dirty
);
662 rtvals
[0] = VM_PAGER_BAD
;
666 maxsize
= count
* PAGE_SIZE
;
669 poffset
= IDX_TO_OFF(m
[0]->pindex
);
672 * If the page-aligned write is larger then the actual file we
673 * have to invalidate pages occuring beyond the file EOF. However,
674 * there is an edge case where a file may not be page-aligned where
675 * the last page is partially invalid. In this case the filesystem
676 * may not properly clear the dirty bits for the entire page (which
677 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
678 * With the page locked we are free to fix-up the dirty bits here.
680 * We do not under any circumstances truncate the valid bits, as
681 * this will screw up bogus page replacement.
683 * The caller has already read-protected the pages. The VFS must
684 * use the buffer cache to wrap the pages. The pages might not
685 * be immediately flushed by the buffer cache but once under its
686 * control the pages themselves can wind up being marked clean
687 * and their covering buffer cache buffer can be marked dirty.
689 if (maxsize
+ poffset
> vp
->v_filesize
) {
690 if (vp
->v_filesize
> poffset
) {
693 maxsize
= vp
->v_filesize
- poffset
;
694 ncount
= btoc(maxsize
);
695 if ((pgoff
= (int)maxsize
& PAGE_MASK
) != 0) {
696 vm_page_clear_dirty(m
[ncount
- 1], pgoff
,
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 %d at %lu\n",
745 auio
.uio_resid
, (u_long
)m
[0]->pindex
);
747 for (i
= 0; i
< ncount
; i
++)
748 rtvals
[i
] = VM_PAGER_OK
;
753 vnode_pager_lock(vm_object_t object
)
755 struct thread
*td
= curthread
; /* XXX */
758 for (; object
!= NULL
; object
= object
->backing_object
) {
759 if (object
->type
!= OBJT_VNODE
)
761 if (object
->flags
& OBJ_DEAD
)
765 struct vnode
*vp
= object
->handle
;
766 error
= vget(vp
, LK_SHARED
| LK_RETRY
| LK_CANRECURSE
);
768 if (object
->handle
!= vp
) {
774 if ((object
->flags
& OBJ_DEAD
) ||
775 (object
->type
!= OBJT_VNODE
)) {
778 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp
, error
, lockstatus(&vp
->v_lock
, td
));
779 tsleep(object
->handle
, 0, "vnpgrl", hz
);