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/swap_pager.h>
73 #include <vm/vm_extern.h>
75 #include <sys/thread2.h>
76 #include <vm/vm_page2.h>
78 static void vnode_pager_dealloc (vm_object_t
);
79 static int vnode_pager_getpage (vm_object_t
, vm_page_t
*, int);
80 static void vnode_pager_putpages (vm_object_t
, vm_page_t
*, int, boolean_t
, int *);
81 static boolean_t
vnode_pager_haspage (vm_object_t
, vm_pindex_t
);
83 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 vsetflags(vp
, VOWANT
);
126 tsleep(vp
, 0, "vnpobj", 0);
128 vsetflags(vp
, 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
,
147 OFF_TO_IDX(round_page64(size
)));
149 object
->handle
= handle
;
150 vp
->v_object
= object
;
151 vp
->v_filesize
= size
;
154 if (vp
->v_filesize
!= size
) {
155 kprintf("vnode_pager_alloc: Warning, filesize "
156 "mismatch %lld/%lld\n",
157 (long long)vp
->v_filesize
,
163 vclrflags(vp
, VOLOCK
);
164 if (vp
->v_flag
& VOWANT
) {
165 vclrflags(vp
, VOWANT
);
172 * Add a ref to a vnode's existing VM object, return the object or
173 * NULL if the vnode did not have one. This does not create the
174 * object (we can't since we don't know what the proper blocksize/boff
175 * is to match the VFS's use of the buffer cache).
178 vnode_pager_reference(struct vnode
*vp
)
183 * Prevent race condition when allocating the object. This
184 * can happen with NFS vnodes since the nfsnode isn't locked.
186 while (vp
->v_flag
& VOLOCK
) {
187 vsetflags(vp
, VOWANT
);
188 tsleep(vp
, 0, "vnpobj", 0);
190 vsetflags(vp
, VOLOCK
);
193 * Prevent race conditions against deallocation of the VM
196 while (((object
= vp
->v_object
) != NULL
) &&
197 (object
->flags
& OBJ_DEAD
)) {
198 vm_object_dead_sleep(object
, "vadead");
202 * The object is expected to exist, the caller will handle
203 * NULL returns if it does not.
210 vclrflags(vp
, VOLOCK
);
211 if (vp
->v_flag
& VOWANT
) {
212 vclrflags(vp
, VOWANT
);
219 vnode_pager_dealloc(vm_object_t object
)
221 struct vnode
*vp
= object
->handle
;
224 panic("vnode_pager_dealloc: pager already dealloced");
226 vm_object_pip_wait(object
, "vnpdea");
228 object
->handle
= NULL
;
229 object
->type
= OBJT_DEAD
;
231 vp
->v_filesize
= NOOFFSET
;
232 vclrflags(vp
, VTEXT
| VOBJBUF
);
233 swap_pager_freespace_all(object
);
237 * Return whether the vnode pager has the requested page. Return the
238 * number of disk-contiguous pages before and after the requested page,
239 * not including the requested page.
242 vnode_pager_haspage(vm_object_t object
, vm_pindex_t pindex
)
244 struct vnode
*vp
= object
->handle
;
252 * If no vp or vp is doomed or marked transparent to VM, we do not
255 if ((vp
== NULL
) || (vp
->v_flag
& VRECLAIMED
))
259 * If filesystem no longer mounted or offset beyond end of file we do
262 loffset
= IDX_TO_OFF(pindex
);
264 if (vp
->v_mount
== NULL
|| loffset
>= vp
->v_filesize
)
267 bsize
= vp
->v_mount
->mnt_stat
.f_iosize
;
268 voff
= loffset
% bsize
;
273 * BMAP returns byte counts before and after, where after
274 * is inclusive of the base page. haspage must return page
275 * counts before and after where after does not include the
278 * BMAP is allowed to return a *after of 0 for backwards
279 * compatibility. The base page is still considered valid if
280 * no error is returned.
282 error
= VOP_BMAP(vp
, loffset
- voff
, &doffset
, NULL
, NULL
, 0);
285 if (doffset
== NOOFFSET
)
291 * Lets the VM system know about a change in size for a file.
292 * We adjust our own internal size and flush any cached pages in
293 * the associated object that are affected by the size change.
295 * NOTE: This routine may be invoked as a result of a pager put
296 * operation (possibly at object termination time), so we must be careful.
298 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
299 * we do not blow up on the case. nsize will always be >= 0, however.
302 vnode_pager_setsize(struct vnode
*vp
, vm_ooffset_t nsize
)
304 vm_pindex_t nobjsize
;
305 vm_pindex_t oobjsize
;
306 vm_object_t object
= vp
->v_object
;
312 * Hasn't changed size
314 if (nsize
== vp
->v_filesize
)
318 * Has changed size. Adjust the VM object's size and v_filesize
319 * before we start scanning pages to prevent new pages from being
320 * allocated during the scan.
322 nobjsize
= OFF_TO_IDX(nsize
+ PAGE_MASK
);
323 oobjsize
= object
->size
;
324 object
->size
= nobjsize
;
327 * File has shrunk. Toss any cached pages beyond the new EOF.
329 if (nsize
< vp
->v_filesize
) {
330 vp
->v_filesize
= nsize
;
331 if (nobjsize
< oobjsize
) {
332 vm_object_page_remove(object
, nobjsize
, oobjsize
,
336 * This gets rid of garbage at the end of a page that is now
337 * only partially backed by the vnode. Since we are setting
338 * the entire page valid & clean after we are done we have
339 * to be sure that the portion of the page within the file
340 * bounds is already valid. If it isn't then making it
341 * valid would create a corrupt block.
343 if (nsize
& PAGE_MASK
) {
348 m
= vm_page_lookup(object
, OFF_TO_IDX(nsize
));
349 } while (m
&& vm_page_sleep_busy(m
, TRUE
, "vsetsz"));
352 int base
= (int)nsize
& PAGE_MASK
;
353 int size
= PAGE_SIZE
- base
;
357 * Clear out partial-page garbage in case
358 * the page has been mapped.
360 * This is byte aligned.
363 sf
= sf_buf_alloc(m
, SFB_CPUPRIVATE
);
364 kva
= sf_buf_kva(sf
);
365 bzero((caddr_t
)kva
+ base
, size
);
369 * XXX work around SMP data integrity race
370 * by unmapping the page from user processes.
371 * The garbage we just cleared may be mapped
372 * to a user process running on another cpu
373 * and this code is not running through normal
374 * I/O channels which handle SMP issues for
375 * us, so unmap page to synchronize all cpus.
377 * XXX should vm_pager_unmap_page() have
380 vm_page_protect(m
, VM_PROT_NONE
);
383 * Clear out partial-page dirty bits. This
384 * has the side effect of setting the valid
385 * bits, but that is ok. There are a bunch
386 * of places in the VM system where we expected
387 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
388 * case is one of them. If the page is still
389 * partially dirty, make it fully dirty.
391 * NOTE: We do not clear out the valid
392 * bits. This would prevent bogus_page
393 * replacement from working properly.
395 * NOTE: We do not want to clear the dirty
396 * bit for a partial DEV_BSIZE'd truncation!
397 * This is DEV_BSIZE aligned!
399 vm_page_clear_dirty_beg_nonincl(m
, base
, size
);
401 m
->dirty
= VM_PAGE_BITS_ALL
;
406 vp
->v_filesize
= nsize
;
411 * Release a page busied for a getpages operation. The page may have become
412 * wired (typically due to being used by the buffer cache) or otherwise been
413 * soft-busied and cannot be freed in that case. A held page can still be
417 vnode_pager_freepage(vm_page_t m
)
419 if (m
->busy
|| m
->wire_count
) {
428 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
429 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
430 * vnode_pager_generic_getpages() to implement the previous behaviour.
432 * All other FS's should use the bypass to get to the local media
433 * backing vp's VOP_GETPAGES.
436 vnode_pager_getpage(vm_object_t object
, vm_page_t
*mpp
, int seqaccess
)
442 rtval
= VOP_GETPAGES(vp
, mpp
, PAGE_SIZE
, 0, 0, seqaccess
);
443 if (rtval
== EOPNOTSUPP
)
444 panic("vnode_pager: vfs's must implement vop_getpages\n");
449 * This is now called from local media FS's to operate against their
450 * own vnodes if they fail to implement VOP_GETPAGES.
452 * With all the caching local media devices do these days there is really
453 * very little point to attempting to restrict the I/O size to contiguous
454 * blocks on-disk, especially if our caller thinks we need all the specified
455 * pages. Just construct and issue a READ.
458 vnode_pager_generic_getpages(struct vnode
*vp
, vm_page_t
*mpp
, int bytecount
,
459 int reqpage
, int seqaccess
)
470 * Do not do anything if the vnode is bad.
472 if (vp
->v_mount
== NULL
)
476 * Calculate the number of pages. Since we are paging in whole
477 * pages, adjust bytecount to be an integral multiple of the page
478 * size. It will be clipped to the file EOF later on.
480 bytecount
= round_page(bytecount
);
481 count
= bytecount
/ PAGE_SIZE
;
484 * We could check m[reqpage]->valid here and shortcut the operation,
485 * but doing so breaks read-ahead. Instead assume that the VM
486 * system has already done at least the check, don't worry about
487 * any races, and issue the VOP_READ to allow read-ahead to function.
489 * This keeps the pipeline full for I/O bound sequentially scanned
495 * Discard pages past the file EOF. If the requested page is past
496 * the file EOF we just leave its valid bits set to 0, the caller
497 * expects to maintain ownership of the requested page. If the
498 * entire range is past file EOF discard everything and generate
501 foff
= IDX_TO_OFF(mpp
[0]->pindex
);
502 if (foff
>= vp
->v_filesize
) {
503 for (i
= 0; i
< count
; i
++) {
505 vnode_pager_freepage(mpp
[i
]);
507 return VM_PAGER_ERROR
;
510 if (foff
+ bytecount
> vp
->v_filesize
) {
511 bytecount
= vp
->v_filesize
- foff
;
512 i
= round_page(bytecount
) / PAGE_SIZE
;
515 if (count
!= reqpage
)
516 vnode_pager_freepage(mpp
[count
]);
521 * The size of the transfer is bytecount. bytecount will be an
522 * integral multiple of the page size unless it has been clipped
523 * to the file EOF. The transfer cannot exceed the file EOF.
525 * When dealing with real devices we must round-up to the device
528 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
) {
529 int secmask
= vp
->v_rdev
->si_bsize_phys
- 1;
530 KASSERT(secmask
< PAGE_SIZE
, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask
+ 1));
531 bytecount
= (bytecount
+ secmask
) & ~secmask
;
535 * Severe hack to avoid deadlocks with the buffer cache
537 for (i
= 0; i
< count
; ++i
) {
538 vm_page_t mt
= mpp
[i
];
540 vm_page_io_start(mt
);
545 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
549 ioflags
|= IO_SEQMAX
<< IO_SEQSHIFT
;
551 aiov
.iov_base
= NULL
;
552 aiov
.iov_len
= bytecount
;
553 auio
.uio_iov
= &aiov
;
555 auio
.uio_offset
= foff
;
556 auio
.uio_segflg
= UIO_NOCOPY
;
557 auio
.uio_rw
= UIO_READ
;
558 auio
.uio_resid
= bytecount
;
560 mycpu
->gd_cnt
.v_vnodein
++;
561 mycpu
->gd_cnt
.v_vnodepgsin
+= count
;
563 error
= VOP_READ(vp
, &auio
, ioflags
, proc0
.p_ucred
);
566 * Severe hack to avoid deadlocks with the buffer cache
568 for (i
= 0; i
< count
; ++i
) {
569 vm_page_t mt
= mpp
[i
];
571 while (vm_page_sleep_busy(mt
, FALSE
, "getpgs"))
574 vm_page_io_finish(mt
);
578 * Calculate the actual number of bytes read and clean up the
581 bytecount
-= auio
.uio_resid
;
583 for (i
= 0; i
< count
; ++i
) {
584 vm_page_t mt
= mpp
[i
];
587 if (error
== 0 && mt
->valid
) {
588 if (mt
->flags
& PG_WANTED
)
589 vm_page_activate(mt
);
591 vm_page_deactivate(mt
);
594 vnode_pager_freepage(mt
);
596 } else if (mt
->valid
== 0) {
598 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt
, mt
->object
, (int) mt
->pindex
);
599 /* whoops, something happened */
602 } else if (mt
->valid
!= VM_PAGE_BITS_ALL
) {
604 * Zero-extend the requested page if necessary (if
605 * the filesystem is using a small block size).
607 vm_page_zero_invalid(mt
, TRUE
);
611 kprintf("vnode_pager_getpage: I/O read error\n");
613 return (error
? VM_PAGER_ERROR
: VM_PAGER_OK
);
617 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
618 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
619 * vnode_pager_generic_putpages() to implement the previous behaviour.
621 * Caller has already cleared the pmap modified bits, if any.
623 * All other FS's should use the bypass to get to the local media
624 * backing vp's VOP_PUTPAGES.
627 vnode_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
628 boolean_t sync
, int *rtvals
)
632 int bytes
= count
* PAGE_SIZE
;
635 * Force synchronous operation if we are extremely low on memory
636 * to prevent a low-memory deadlock. VOP operations often need to
637 * allocate more memory to initiate the I/O ( i.e. do a BMAP
638 * operation ). The swapper handles the case by limiting the amount
639 * of asynchronous I/O, but that sort of solution doesn't scale well
640 * for the vnode pager without a lot of work.
642 * Also, the backing vnode's iodone routine may not wake the pageout
643 * daemon up. This should be probably be addressed XXX.
646 if ((vmstats
.v_free_count
+ vmstats
.v_cache_count
) < vmstats
.v_pageout_free_min
)
650 * Call device-specific putpages function
653 rtval
= VOP_PUTPAGES(vp
, m
, bytes
, sync
, rtvals
, 0);
654 if (rtval
== EOPNOTSUPP
) {
655 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
656 rtval
= vnode_pager_generic_putpages( vp
, m
, bytes
, sync
, rtvals
);
662 * This is now called from local media FS's to operate against their
663 * own vnodes if they fail to implement VOP_PUTPAGES.
665 * This is typically called indirectly via the pageout daemon and
666 * clustering has already typically occured, so in general we ask the
667 * underlying filesystem to write the data out asynchronously rather
671 vnode_pager_generic_putpages(struct vnode
*vp
, vm_page_t
*m
, int bytecount
,
672 int flags
, int *rtvals
)
676 int maxsize
, ncount
, count
;
677 vm_ooffset_t poffset
;
683 object
= vp
->v_object
;
684 count
= bytecount
/ PAGE_SIZE
;
686 for (i
= 0; i
< count
; i
++)
687 rtvals
[i
] = VM_PAGER_AGAIN
;
689 if ((int) m
[0]->pindex
< 0) {
690 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
691 (long)m
[0]->pindex
, m
[0]->dirty
);
692 rtvals
[0] = VM_PAGER_BAD
;
696 maxsize
= count
* PAGE_SIZE
;
699 poffset
= IDX_TO_OFF(m
[0]->pindex
);
702 * If the page-aligned write is larger then the actual file we
703 * have to invalidate pages occuring beyond the file EOF.
705 * If the file EOF resides in the middle of a page we still clear
706 * all of that page's dirty bits later on. If we didn't it would
707 * endlessly re-write.
709 * We do not under any circumstances truncate the valid bits, as
710 * this will screw up bogus page replacement.
712 * The caller has already read-protected the pages. The VFS must
713 * use the buffer cache to wrap the pages. The pages might not
714 * be immediately flushed by the buffer cache but once under its
715 * control the pages themselves can wind up being marked clean
716 * and their covering buffer cache buffer can be marked dirty.
718 if (poffset
+ maxsize
> vp
->v_filesize
) {
719 if (poffset
< vp
->v_filesize
) {
720 maxsize
= vp
->v_filesize
- poffset
;
721 ncount
= btoc(maxsize
);
726 if (ncount
< count
) {
727 for (i
= ncount
; i
< count
; i
++) {
728 rtvals
[i
] = VM_PAGER_BAD
;
734 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
735 * rather then a bdwrite() to prevent paging I/O from saturating
736 * the buffer cache. Dummy-up the sequential heuristic to cause
737 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
738 * the system decides how to cluster.
741 if (flags
& (VM_PAGER_PUT_SYNC
| VM_PAGER_PUT_INVAL
))
743 else if ((flags
& VM_PAGER_CLUSTER_OK
) == 0)
745 ioflags
|= (flags
& VM_PAGER_PUT_INVAL
) ? IO_INVAL
: 0;
746 ioflags
|= IO_SEQMAX
<< IO_SEQSHIFT
;
748 aiov
.iov_base
= (caddr_t
) 0;
749 aiov
.iov_len
= maxsize
;
750 auio
.uio_iov
= &aiov
;
752 auio
.uio_offset
= poffset
;
753 auio
.uio_segflg
= UIO_NOCOPY
;
754 auio
.uio_rw
= UIO_WRITE
;
755 auio
.uio_resid
= maxsize
;
757 error
= VOP_WRITE(vp
, &auio
, ioflags
, proc0
.p_ucred
);
758 mycpu
->gd_cnt
.v_vnodeout
++;
759 mycpu
->gd_cnt
.v_vnodepgsout
+= ncount
;
762 krateprintf(&vbadrate
,
763 "vnode_pager_putpages: I/O error %d\n", error
);
765 if (auio
.uio_resid
) {
766 krateprintf(&vresrate
,
767 "vnode_pager_putpages: residual I/O %zd at %lu\n",
768 auio
.uio_resid
, (u_long
)m
[0]->pindex
);
771 for (i
= 0; i
< ncount
; i
++) {
772 rtvals
[i
] = VM_PAGER_OK
;
773 vm_page_undirty(m
[i
]);
780 vnode_pager_lock(vm_object_t object
)
782 struct thread
*td
= curthread
; /* XXX */
785 for (; object
!= NULL
; object
= object
->backing_object
) {
786 if (object
->type
!= OBJT_VNODE
)
788 if (object
->flags
& OBJ_DEAD
)
792 struct vnode
*vp
= object
->handle
;
793 error
= vget(vp
, LK_SHARED
| LK_RETRY
| LK_CANRECURSE
);
795 if (object
->handle
!= vp
) {
801 if ((object
->flags
& OBJ_DEAD
) ||
802 (object
->type
!= OBJT_VNODE
)) {
805 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp
, error
, lockstatus(&vp
->v_lock
, td
));
806 tsleep(object
->handle
, 0, "vnpgrl", hz
);