Date: Sat, 17 May 1997 16:52:46 +0100 (BST) From: Doug Rabson <dfr@nlsystems.com> To: Christoph Kukulies <kuku@gilberto.physik.rwth-aachen.de> Cc: dyson@freebsd.org, current@freebsd.org Subject: Re: can anyone confirm my NFS/VM problem? Message-ID: <Pine.BSF.3.95q.970517164819.5082B-100000@herring.nlsystems.com> In-Reply-To: <199705161047.MAA00774@gil.physik.rwth-aachen.de>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, 16 May 1997, Christoph Kukulies wrote:
>
> I have created a test case to repeat that weird NFS/VM problem I
> have here with a physics production program - the exposed bug is
> really a showstopper for our production and I'm not sure if it is
> not also in 2.2.1R.
> (Don't know if Doug is already going to deal with this group
> of NFS/VM/MMAP problems). Anyway, to assure it is really a
> repeatable problem independent of some local peculiarity here
> at my site, I'm seeking for someone with the following
> configuration:
Well I think I have a fix. It was bloody painful too. There were some
extremely ugly problems associated with NFS' use of b_validoff and
b_validend. I attempted to fix them but it still feels pretty fragile
although it does pass Ghristoph's test case.
John, could you have a look at these changes and tell me whether they make
sense?
Index: kern/vfs_bio.c
===================================================================
RCS file: /home/ncvs/src/sys/kern/vfs_bio.c,v
retrieving revision 1.115
diff -u -r1.115 vfs_bio.c
--- vfs_bio.c 1997/05/10 09:09:42 1.115
+++ vfs_bio.c 1997/05/17 15:37:52
@@ -1215,6 +1215,10 @@
splx(s);
allocbuf(bp, size);
+ if ((bp->b_flags & (B_VMIO|B_CACHE)) == (B_VMIO|B_CACHE)) {
+ bp->b_validoff = 0;
+ bp->b_validend = bp->b_bufsize;
+ }
#ifdef PC98
/*
* 1024byte/sector support
@@ -1562,7 +1566,7 @@
}
if (bp->b_flags & B_VMIO) {
int i, resid;
- vm_ooffset_t foff;
+ vm_ooffset_t foff, bfoff;
vm_page_t m;
vm_object_t obj;
int iosize;
@@ -1572,6 +1576,7 @@
foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno;
else
foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
+ bfoff = foff;
obj = vp->v_object;
if (!obj) {
panic("biodone: no object");
@@ -1613,8 +1618,18 @@
* here in the read case.
*/
if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) {
- vm_page_set_validclean(m,
- (vm_offset_t) (foff & PAGE_MASK), resid);
+ vm_ooffset_t soff, eoff;
+ if (vp->v_tag == VT_NFS) {
+ soff = max((bfoff + bp->b_validoff + DEV_BSIZE - 1) & -DEV_BSIZE,
+ foff);
+ eoff = min((bfoff + bp->b_validend) & -DEV_BSIZE,
+ foff + resid);
+ vm_page_set_invalid(m, 0, PAGE_SIZE);
+ if (eoff > soff)
+ vm_page_set_validclean(m, soff, eoff - soff);
+ } else
+ vm_page_set_validclean(m,
+ (vm_offset_t) (foff & PAGE_MASK), resid);
}
/*
@@ -1774,14 +1789,16 @@
int i;
if (bp->b_flags & B_VMIO) {
- vm_object_t obj = bp->b_vp->v_object;
- vm_ooffset_t foff;
+ struct vnode *vp = bp->b_vp;
+ vm_object_t obj = vp->v_object;
+ vm_ooffset_t foff, bfoff;
int iocount = bp->b_bufsize;
if (bp->b_vp->v_type == VBLK)
foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno;
else
- foff = (vm_ooffset_t) bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
+ foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
+ bfoff = foff;
vfs_setdirty(bp);
for (i = 0; i < bp->b_npages; i++) {
vm_page_t m = bp->b_pages[i];
@@ -1795,8 +1812,18 @@
}
vm_page_protect(m, VM_PROT_NONE);
if (clear_modify) {
- vm_page_set_validclean(m,
- (vm_offset_t) (foff & PAGE_MASK), resid);
+ vm_ooffset_t soff, eoff;
+ if (vp->v_tag == VT_NFS) {
+ soff = max((bfoff + bp->b_validoff + DEV_BSIZE - 1) & -DEV_BSIZE,
+ foff);
+ eoff = min((bfoff + bp->b_validend) & -DEV_BSIZE,
+ foff + resid);
+ vm_page_set_invalid(m, 0, PAGE_SIZE);
+ if (eoff > soff)
+ vm_page_set_validclean(m, soff, eoff - soff);
+ } else
+ vm_page_set_validclean(m,
+ (vm_offset_t) (foff & PAGE_MASK), resid);
} else if (bp->b_bcount >= PAGE_SIZE) {
if (m->valid && (bp->b_flags & B_CACHE) == 0) {
bp->b_pages[i] = bogus_page;
@@ -1820,13 +1847,15 @@
int i;
if (bp->b_flags & B_VMIO) {
- vm_ooffset_t foff;
+ vm_ooffset_t foff, bfoff;
+ struct vnode *vp = bp->b_vp;
int iocount = bp->b_bufsize;
if (bp->b_vp->v_type == VBLK)
foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno;
else
- foff = (vm_ooffset_t) bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
+ foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno;
+ bfoff = foff;
for (i = 0; i < bp->b_npages; i++) {
vm_page_t m = bp->b_pages[i];
@@ -1835,8 +1864,18 @@
if (resid > iocount)
resid = iocount;
if (resid > 0) {
- vm_page_set_validclean(m,
- ((vm_offset_t) foff & PAGE_MASK), resid);
+ vm_ooffset_t soff, eoff;
+ if (vp->v_tag == VT_NFS) {
+ soff = max((bfoff + bp->b_validoff + DEV_BSIZE - 1) & -DEV_BSIZE,
+ foff);
+ eoff = min((bfoff + bp->b_validend) & -DEV_BSIZE,
+ foff + resid);
+ vm_page_set_invalid(m, 0, PAGE_SIZE);
+ if (eoff > soff)
+ vm_page_set_validclean(m, soff, eoff - soff);
+ } else
+ vm_page_set_validclean(m,
+ (vm_offset_t) (foff & PAGE_MASK), resid);
}
foff += resid;
iocount -= resid;
Index: vm/vm_fault.c
===================================================================
RCS file: /home/ncvs/src/sys/vm/vm_fault.c,v
retrieving revision 1.68
diff -u -r1.68 vm_fault.c
--- vm_fault.c 1997/04/06 16:16:11 1.68
+++ vm_fault.c 1997/05/17 15:12:15
@@ -326,7 +326,7 @@
m->flags |= PG_BUSY;
- if (m->valid &&
+ if (/*m->valid && */
((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
m->object != kernel_object && m->object != kmem_object) {
goto readrest;
Index: vm/vnode_pager.c
===================================================================
RCS file: /home/ncvs/src/sys/vm/vnode_pager.c,v
retrieving revision 1.70
diff -u -r1.70 vnode_pager.c
--- vnode_pager.c 1997/03/08 04:33:47 1.70
+++ vnode_pager.c 1997/05/17 15:28:25
@@ -692,7 +692,12 @@
/*
* if ANY DEV_BSIZE blocks are valid on a large filesystem block
* then, the entire page is valid --
+ * XXX no it isn't
*/
+
+ if (m[reqpage]->valid != VM_PAGE_BITS_ALL)
+ m[reqpage]->valid = 0;
+
if (m[reqpage]->valid) {
m[reqpage]->valid = VM_PAGE_BITS_ALL;
for (i = 0; i < count; i++) {
--
Doug Rabson Mail: dfr@nlsystems.com
Nonlinear Systems Ltd. Phone: +44 181 951 1891
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.BSF.3.95q.970517164819.5082B-100000>