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
= {
91 static struct krate vbadrate
= { 1 };
92 static struct krate vresrate
= { 1 };
94 int vnode_pbuf_freecnt
= -1; /* start out unlimited */
97 * Allocate (or lookup) pager for a vnode.
98 * Handle is a vnode pointer.
101 vnode_pager_alloc(void *handle
, off_t size
, vm_prot_t prot
, off_t offset
)
107 * Pageout to vnode, no can do yet.
113 * XXX hack - This initialization should be put somewhere else.
115 if (vnode_pbuf_freecnt
< 0) {
116 vnode_pbuf_freecnt
= nswbuf
/ 2 + 1;
119 vp
= (struct vnode
*) handle
;
122 * Prevent race condition when allocating the object. This
123 * can happen with NFS vnodes since the nfsnode isn't locked.
125 while (vp
->v_flag
& VOLOCK
) {
126 vsetflags(vp
, VOWANT
);
127 tsleep(vp
, 0, "vnpobj", 0);
129 vsetflags(vp
, VOLOCK
);
132 * If the object is being terminated, wait for it to
135 while (((object
= vp
->v_object
) != NULL
) &&
136 (object
->flags
& OBJ_DEAD
)) {
137 vm_object_dead_sleep(object
, "vadead");
140 if (vp
->v_sysref
.refcnt
<= 0)
141 panic("vnode_pager_alloc: no vnode reference");
143 if (object
== NULL
) {
145 * And an object of the appropriate size
147 object
= vm_object_allocate(OBJT_VNODE
,
148 OFF_TO_IDX(round_page64(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 vclrflags(vp
, VOLOCK
);
165 if (vp
->v_flag
& VOWANT
) {
166 vclrflags(vp
, 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 vclrflags(vp
, VTEXT
| VOBJBUF
);
187 swap_pager_freespace_all(object
);
191 * Return whether the vnode pager has the requested page. Return the
192 * number of disk-contiguous pages before and after the requested page,
193 * not including the requested page.
196 vnode_pager_haspage(vm_object_t object
, vm_pindex_t pindex
)
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
;
227 * BMAP returns byte counts before and after, where after
228 * is inclusive of the base page. haspage must return page
229 * counts before and after where after does not include the
232 * BMAP is allowed to return a *after of 0 for backwards
233 * compatibility. The base page is still considered valid if
234 * no error is returned.
236 error
= VOP_BMAP(vp
, loffset
- voff
, &doffset
, NULL
, NULL
, 0);
239 if (doffset
== NOOFFSET
)
245 * Lets the VM system know about a change in size for a file.
246 * We adjust our own internal size and flush any cached pages in
247 * the associated object that are affected by the size change.
249 * NOTE: This routine may be invoked as a result of a pager put
250 * operation (possibly at object termination time), so we must be careful.
252 * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
253 * we do not blow up on the case. nsize will always be >= 0, however.
256 vnode_pager_setsize(struct vnode
*vp
, vm_ooffset_t nsize
)
258 vm_pindex_t nobjsize
;
259 vm_pindex_t oobjsize
;
260 vm_object_t object
= vp
->v_object
;
266 * Hasn't changed size
268 if (nsize
== vp
->v_filesize
)
272 * Has changed size. Adjust the VM object's size and v_filesize
273 * before we start scanning pages to prevent new pages from being
274 * allocated during the scan.
276 nobjsize
= OFF_TO_IDX(nsize
+ PAGE_MASK
);
277 oobjsize
= object
->size
;
278 object
->size
= nobjsize
;
281 * File has shrunk. Toss any cached pages beyond the new EOF.
283 if (nsize
< vp
->v_filesize
) {
284 vp
->v_filesize
= nsize
;
285 if (nobjsize
< oobjsize
) {
286 vm_object_page_remove(object
, nobjsize
, oobjsize
,
290 * This gets rid of garbage at the end of a page that is now
291 * only partially backed by the vnode. Since we are setting
292 * the entire page valid & clean after we are done we have
293 * to be sure that the portion of the page within the file
294 * bounds is already valid. If it isn't then making it
295 * valid would create a corrupt block.
297 if (nsize
& PAGE_MASK
) {
302 m
= vm_page_lookup(object
, OFF_TO_IDX(nsize
));
303 } while (m
&& vm_page_sleep_busy(m
, TRUE
, "vsetsz"));
306 int base
= (int)nsize
& PAGE_MASK
;
307 int size
= PAGE_SIZE
- base
;
311 * Clear out partial-page garbage in case
312 * the page has been mapped.
314 * This is byte aligned.
317 sf
= sf_buf_alloc(m
, SFB_CPUPRIVATE
);
318 kva
= sf_buf_kva(sf
);
319 bzero((caddr_t
)kva
+ base
, size
);
323 * XXX work around SMP data integrity race
324 * by unmapping the page from user processes.
325 * The garbage we just cleared may be mapped
326 * to a user process running on another cpu
327 * and this code is not running through normal
328 * I/O channels which handle SMP issues for
329 * us, so unmap page to synchronize all cpus.
331 * XXX should vm_pager_unmap_page() have
334 vm_page_protect(m
, VM_PROT_NONE
);
337 * Clear out partial-page dirty bits. This
338 * has the side effect of setting the valid
339 * bits, but that is ok. There are a bunch
340 * of places in the VM system where we expected
341 * m->dirty == VM_PAGE_BITS_ALL. The file EOF
342 * case is one of them. If the page is still
343 * partially dirty, make it fully dirty.
345 * NOTE: We do not clear out the valid
346 * bits. This would prevent bogus_page
347 * replacement from working properly.
349 * NOTE: We do not want to clear the dirty
350 * bit for a partial DEV_BSIZE'd truncation!
351 * This is DEV_BSIZE aligned!
353 vm_page_clear_dirty_beg_nonincl(m
, base
, size
);
355 m
->dirty
= VM_PAGE_BITS_ALL
;
360 vp
->v_filesize
= nsize
;
365 * Release a page busied for a getpages operation. The page may have become
366 * wired (typically due to being used by the buffer cache) or otherwise been
367 * soft-busied and cannot be freed in that case. A held page can still be
371 vnode_pager_freepage(vm_page_t m
)
373 if (m
->busy
|| m
->wire_count
) {
382 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
383 * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
384 * vnode_pager_generic_getpages() to implement the previous behaviour.
386 * All other FS's should use the bypass to get to the local media
387 * backing vp's VOP_GETPAGES.
390 vnode_pager_getpage(vm_object_t object
, vm_page_t
*mpp
, int seqaccess
)
396 rtval
= VOP_GETPAGES(vp
, mpp
, PAGE_SIZE
, 0, 0, seqaccess
);
397 if (rtval
== EOPNOTSUPP
)
398 panic("vnode_pager: vfs's must implement vop_getpages\n");
403 * This is now called from local media FS's to operate against their
404 * own vnodes if they fail to implement VOP_GETPAGES.
406 * With all the caching local media devices do these days there is really
407 * very little point to attempting to restrict the I/O size to contiguous
408 * blocks on-disk, especially if our caller thinks we need all the specified
409 * pages. Just construct and issue a READ.
412 vnode_pager_generic_getpages(struct vnode
*vp
, vm_page_t
*mpp
, int bytecount
,
413 int reqpage
, int seqaccess
)
424 * Do not do anything if the vnode is bad.
426 if (vp
->v_mount
== NULL
)
430 * Calculate the number of pages. Since we are paging in whole
431 * pages, adjust bytecount to be an integral multiple of the page
432 * size. It will be clipped to the file EOF later on.
434 bytecount
= round_page(bytecount
);
435 count
= bytecount
/ PAGE_SIZE
;
438 * We could check m[reqpage]->valid here and shortcut the operation,
439 * but doing so breaks read-ahead. Instead assume that the VM
440 * system has already done at least the check, don't worry about
441 * any races, and issue the VOP_READ to allow read-ahead to function.
443 * This keeps the pipeline full for I/O bound sequentially scanned
449 * Discard pages past the file EOF. If the requested page is past
450 * the file EOF we just leave its valid bits set to 0, the caller
451 * expects to maintain ownership of the requested page. If the
452 * entire range is past file EOF discard everything and generate
455 foff
= IDX_TO_OFF(mpp
[0]->pindex
);
456 if (foff
>= vp
->v_filesize
) {
457 for (i
= 0; i
< count
; i
++) {
459 vnode_pager_freepage(mpp
[i
]);
461 return VM_PAGER_ERROR
;
464 if (foff
+ bytecount
> vp
->v_filesize
) {
465 bytecount
= vp
->v_filesize
- foff
;
466 i
= round_page(bytecount
) / PAGE_SIZE
;
469 if (count
!= reqpage
)
470 vnode_pager_freepage(mpp
[count
]);
475 * The size of the transfer is bytecount. bytecount will be an
476 * integral multiple of the page size unless it has been clipped
477 * to the file EOF. The transfer cannot exceed the file EOF.
479 * When dealing with real devices we must round-up to the device
482 if (vp
->v_type
== VBLK
|| vp
->v_type
== VCHR
) {
483 int secmask
= vp
->v_rdev
->si_bsize_phys
- 1;
484 KASSERT(secmask
< PAGE_SIZE
, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask
+ 1));
485 bytecount
= (bytecount
+ secmask
) & ~secmask
;
489 * Severe hack to avoid deadlocks with the buffer cache
491 for (i
= 0; i
< count
; ++i
) {
492 vm_page_t mt
= mpp
[i
];
494 vm_page_io_start(mt
);
499 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
503 ioflags
|= IO_SEQMAX
<< IO_SEQSHIFT
;
505 aiov
.iov_base
= NULL
;
506 aiov
.iov_len
= bytecount
;
507 auio
.uio_iov
= &aiov
;
509 auio
.uio_offset
= foff
;
510 auio
.uio_segflg
= UIO_NOCOPY
;
511 auio
.uio_rw
= UIO_READ
;
512 auio
.uio_resid
= bytecount
;
514 mycpu
->gd_cnt
.v_vnodein
++;
515 mycpu
->gd_cnt
.v_vnodepgsin
+= count
;
517 error
= VOP_READ(vp
, &auio
, ioflags
, proc0
.p_ucred
);
520 * Severe hack to avoid deadlocks with the buffer cache
522 for (i
= 0; i
< count
; ++i
) {
523 vm_page_t mt
= mpp
[i
];
525 while (vm_page_sleep_busy(mt
, FALSE
, "getpgs"))
528 vm_page_io_finish(mt
);
532 * Calculate the actual number of bytes read and clean up the
535 bytecount
-= auio
.uio_resid
;
537 for (i
= 0; i
< count
; ++i
) {
538 vm_page_t mt
= mpp
[i
];
541 if (error
== 0 && mt
->valid
) {
542 if (mt
->flags
& PG_WANTED
)
543 vm_page_activate(mt
);
545 vm_page_deactivate(mt
);
548 vnode_pager_freepage(mt
);
550 } else if (mt
->valid
== 0) {
552 kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt
, mt
->object
, (int) mt
->pindex
);
553 /* whoops, something happened */
556 } else if (mt
->valid
!= VM_PAGE_BITS_ALL
) {
558 * Zero-extend the requested page if necessary (if
559 * the filesystem is using a small block size).
561 vm_page_zero_invalid(mt
, TRUE
);
565 kprintf("vnode_pager_getpage: I/O read error\n");
567 return (error
? VM_PAGER_ERROR
: VM_PAGER_OK
);
571 * EOPNOTSUPP is no longer legal. For local media VFS's that do not
572 * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
573 * vnode_pager_generic_putpages() to implement the previous behaviour.
575 * Caller has already cleared the pmap modified bits, if any.
577 * All other FS's should use the bypass to get to the local media
578 * backing vp's VOP_PUTPAGES.
581 vnode_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
582 boolean_t sync
, int *rtvals
)
586 int bytes
= count
* PAGE_SIZE
;
589 * Force synchronous operation if we are extremely low on memory
590 * to prevent a low-memory deadlock. VOP operations often need to
591 * allocate more memory to initiate the I/O ( i.e. do a BMAP
592 * operation ). The swapper handles the case by limiting the amount
593 * of asynchronous I/O, but that sort of solution doesn't scale well
594 * for the vnode pager without a lot of work.
596 * Also, the backing vnode's iodone routine may not wake the pageout
597 * daemon up. This should be probably be addressed XXX.
600 if ((vmstats
.v_free_count
+ vmstats
.v_cache_count
) < vmstats
.v_pageout_free_min
)
604 * Call device-specific putpages function
607 rtval
= VOP_PUTPAGES(vp
, m
, bytes
, sync
, rtvals
, 0);
608 if (rtval
== EOPNOTSUPP
) {
609 kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
610 rtval
= vnode_pager_generic_putpages( vp
, m
, bytes
, sync
, rtvals
);
616 * This is now called from local media FS's to operate against their
617 * own vnodes if they fail to implement VOP_PUTPAGES.
619 * This is typically called indirectly via the pageout daemon and
620 * clustering has already typically occured, so in general we ask the
621 * underlying filesystem to write the data out asynchronously rather
625 vnode_pager_generic_putpages(struct vnode
*vp
, vm_page_t
*m
, int bytecount
,
626 int flags
, int *rtvals
)
630 int maxsize
, ncount
, count
;
631 vm_ooffset_t poffset
;
637 object
= vp
->v_object
;
638 count
= bytecount
/ PAGE_SIZE
;
640 for (i
= 0; i
< count
; i
++)
641 rtvals
[i
] = VM_PAGER_AGAIN
;
643 if ((int) m
[0]->pindex
< 0) {
644 kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
645 (long)m
[0]->pindex
, m
[0]->dirty
);
646 rtvals
[0] = VM_PAGER_BAD
;
650 maxsize
= count
* PAGE_SIZE
;
653 poffset
= IDX_TO_OFF(m
[0]->pindex
);
656 * If the page-aligned write is larger then the actual file we
657 * have to invalidate pages occuring beyond the file EOF.
659 * If the file EOF resides in the middle of a page we still clear
660 * all of that page's dirty bits later on. If we didn't it would
661 * endlessly re-write.
663 * We do not under any circumstances truncate the valid bits, as
664 * this will screw up bogus page replacement.
666 * The caller has already read-protected the pages. The VFS must
667 * use the buffer cache to wrap the pages. The pages might not
668 * be immediately flushed by the buffer cache but once under its
669 * control the pages themselves can wind up being marked clean
670 * and their covering buffer cache buffer can be marked dirty.
672 if (poffset
+ maxsize
> vp
->v_filesize
) {
673 if (poffset
< vp
->v_filesize
) {
674 maxsize
= vp
->v_filesize
- poffset
;
675 ncount
= btoc(maxsize
);
680 if (ncount
< count
) {
681 for (i
= ncount
; i
< count
; i
++) {
682 rtvals
[i
] = VM_PAGER_BAD
;
688 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
689 * rather then a bdwrite() to prevent paging I/O from saturating
690 * the buffer cache. Dummy-up the sequential heuristic to cause
691 * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set,
692 * the system decides how to cluster.
695 if (flags
& (VM_PAGER_PUT_SYNC
| VM_PAGER_PUT_INVAL
))
697 else if ((flags
& VM_PAGER_CLUSTER_OK
) == 0)
699 ioflags
|= (flags
& VM_PAGER_PUT_INVAL
) ? IO_INVAL
: 0;
700 ioflags
|= IO_SEQMAX
<< IO_SEQSHIFT
;
702 aiov
.iov_base
= (caddr_t
) 0;
703 aiov
.iov_len
= maxsize
;
704 auio
.uio_iov
= &aiov
;
706 auio
.uio_offset
= poffset
;
707 auio
.uio_segflg
= UIO_NOCOPY
;
708 auio
.uio_rw
= UIO_WRITE
;
709 auio
.uio_resid
= maxsize
;
711 error
= VOP_WRITE(vp
, &auio
, ioflags
, proc0
.p_ucred
);
712 mycpu
->gd_cnt
.v_vnodeout
++;
713 mycpu
->gd_cnt
.v_vnodepgsout
+= ncount
;
716 krateprintf(&vbadrate
,
717 "vnode_pager_putpages: I/O error %d\n", error
);
719 if (auio
.uio_resid
) {
720 krateprintf(&vresrate
,
721 "vnode_pager_putpages: residual I/O %zd at %lu\n",
722 auio
.uio_resid
, (u_long
)m
[0]->pindex
);
725 for (i
= 0; i
< ncount
; i
++) {
726 rtvals
[i
] = VM_PAGER_OK
;
727 vm_page_undirty(m
[i
]);
734 vnode_pager_lock(vm_object_t object
)
736 struct thread
*td
= curthread
; /* XXX */
739 for (; object
!= NULL
; object
= object
->backing_object
) {
740 if (object
->type
!= OBJT_VNODE
)
742 if (object
->flags
& OBJ_DEAD
)
746 struct vnode
*vp
= object
->handle
;
747 error
= vget(vp
, LK_SHARED
| LK_RETRY
| LK_CANRECURSE
);
749 if (object
->handle
!= vp
) {
755 if ((object
->flags
& OBJ_DEAD
) ||
756 (object
->type
!= OBJT_VNODE
)) {
759 kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp
, error
, lockstatus(&vp
->v_lock
, td
));
760 tsleep(object
->handle
, 0, "vnpgrl", hz
);